cartografía predictiva de la distribución de aves

Cartografía predictiva de la distribución deaves terrestres: un estudio piloto en

Andalucía occidental

Javier Seoane Pinilla

Tesis DoctoralUniversidad Autónoma de Madrid – 2002

d3agr

s(d3

agr,

3)

0 5000 10000 15000

-10

12

3

Pagr.350

I((P

agr.

350

- 0.

7) *

(P

agr.

350

> 0.

7))

0.0 0.2 0.4 0.6 0.8 1.0

-2-1

01

2

DEPARTAMENTO DE ECOLOGÍA

FACULTAD DE CIENCIAS

UNIVERSIDAD AUTÓNOMADE MADRID

DEPARTAMENTO DE BIOLOGÍAAPLICADA

ESTACIÓN BIOLÓGICA DE DOÑANA

CONSEJO SUPERIOR DEINVESTIGACIONES CIENTÍFICAS

CARTOGRAFÍA PREDICTIVA DE LA DISTRIBUCIÓN DE AVESTERRESTRES: UN ESTUDIO PILOTO EN ANDALUCÍA OCCIDENTAL

Memoria presentada por el LicenciadoJavier Seoane Pinilla para optar algrado de Doctor en Biología por laUniversidad Autónoma de Madrid

Sevilla, Noviembre de 2002

Carlos Montes del Olmo, Director del Departamento de Ecología de la Universidad

Autónoma de Madrid

CERTIFICA

que la Tesis Doctoral que lleva por título Cartografía predictiva de la

distribución de aves terrestres: un estudio piloto en Andalucía occidental, presentada

por el Licenciado Javier Seoane Pinilla para optar al grado de Doctor en Biología,

reúne los requisitos necesarios para su presentación y defensa pública, si procede, de

acuerdo con la normativa vigente.

Y para que conste a los efectos oportunos, firmo la presente en Madrid, a 1 de

octubre de 2002.

Fdo.: Carlos Montes del Olmo

Javier María Bustamante Díaz, investigador de la Estación Biológica de Doñana,

centro perteneciente al Consejo Superior de Investigaciones Científicas (CSIC)

CERTIFICA

que la Tesis Doctoral que lleva por título Cartografía predictiva de la

distribución de aves terrestres: un estudio piloto en Andalucía occidental, presentada

por el Licenciado Javier Seoane Pinilla, ha sido realizada bajo mi dirección y cuenta

con mi aprobación para su presentación y defensa pública, si procede, de acuerdo con la

normativa vigente. Esta tesis supone, como aportación original al campo de la ecología,

el primer estudio en detalle que aborda la elaboración y aplicación de modelos

estadísticos de distribución de un conjunto numeroso de especies, a escalas local y

regional, en España.

Y para que conste a los efectos oportunos, firmo la presente en Sevilla, a 25 de

septiembre de 2002.

Fdo.: Javier M. Bustamante Díaz

i

Índice

Introducción y objetivos ..................................................................................... 1-15

SECCIÓN PRIMERAReflexiones preliminares

Capítulo I: ....................................................................................................... 16-33Modelos predictivos de la distribución de especies:una revisión de sus limitaciones

SECCIÓN SEGUNDAAspectos metodológicos: Técnicas y estrategias

del modelado de la distribución de especies

Capítulo II: ....................................................................................................... 34-48El muestreo de la presencia/ausencia para construir modelos predictivos:una aproximación de optimalidad usando el teorema del valor marginal

Chapter II:Sampling bird presence/absence to build predictive models:an optimality approach using the marginal value theorem

Capítulo III ....................................................................................................... 49-63¿Incrementa la opinión de experto la habilidad predictiva delos modelos de la distribución de aves?

Chapter IIIDoes expert opinion increase the predictive ability of environmentalmodels of bird distribution?

Capítulo IV ....................................................................................................... 64-78¿Son adecuados los mapas de vegetación existentes para predecirla distribución de las aves?

Chapter IVAre existing vegetation maps adequate to predict bird distributions?

Capítulo V ....................................................................................................... 79-93La elección de la mejor resolución espacial en los modelos predictivosde la distribución de avesApéndice ....................................................................................................... 94-96

ii

Chapter VChoosing the best spatial resolution for predictive modelsof bird distributionAppendix

Capítulo VI ....................................................................................................... 97-113

Una comparación de diferentes variables predictoras para losmodelos de la distribución de aves: el paisaje, la cubierta vegetal,la topografía y el clima

Chapter VI: A comparison of different explanatory variables forpredictive models of breeding bird distribution:competing roles for landscape, land-cover, topography and climate

SECCIÓN TERCERAPuesta en práctica: aplicaciones de la cartografía de especies

Capítulo VII ................................................................................................ 114-135Modelos aditivos generalizados y SIG para predecir la adecuacióndel hábitat de rapaces forestales en el sur de España

Chapter VIIUsing Generalised Additive Models and GIS to predict habitatsuitability for forest raptors in Southern Spain

Capítulo VIII............................................................................................... 136-160El uso de modelos regionales para identificar factores limitantesy áreas con problemas de conservación:la distribución y abundancia del milano real en la península IbéricaApéndice...................................................................................................... 161-163

Chapter VIIIUse of regional models to identify limiting factors andareas with conservation problems: the distribution and abundanceof the Red kite in the Iberian peninsulaAppendix

Capítulo IX .................................................................................................. 164-183Una evaluación con modelos estadísticos de la cartografíade especies generada mediante criterio de expertosApéndice....................................................................................................... 184-204

iii

Chapter IXUsing statistical models to evaluate species cartographyderived from expert opinionAppendix

SECCIÓN CUARTAEsperanzas y desesperanzas de los modelos

Capítulo X:Conclusiones............................................................................................... 205-207

Agradecimientos ........................................................................................ 208-209

Introducción y Objetivos

1

INTRODUCCIÓN Y OBJETIVOS

No sería exagerado afirmar que elhombre desde sus orígenes haobservado con interés los organismosque lo rodeaban y, muy probablemente,ha tratado de predecir dónde y en quéabundancia aparecerían aquellasespecies que utilizaba como alimento,materias primas y medicina (Harris1983). Este interés fundamentalmenteutilitario se ha compaginado desde muypronto en la historia de la humanidadcon otro interés científico, motivado porla simple, aunque siempre provechosa,curiosidad de entender el mundo quenos rodea; así, por ejemplo, Aristótelesdejó escritas hace 2500 años algunasobras (como su investigación sobre losanimales) en las que describía larelación entre diversas especies, muchasde las cuales no tenían una inmediatautilidad para el hombre, y el tipo dehábitats que ocupaban. Talesdescripciones se hicieron muyfrecuentes en los catálogos de especiesque hicieron los naturalistas del sigloXIX (por ejemplo, Wilson y Audubonpara la ornitología norteamericana,Block & Brennan 1993) y secorrespondieron con un esfuerzo porsintetizar cualitativamente como lospatrones de distribución espacial que seobservaban podían deberse a procesosfísicos (el Essai sur la géographie desplantes de Humboldt y Bonpland en1807) o, más tarde, evolutivos (Grinell1904 en Block & Brennan 1993,quienes resaltan con aparente asombroque la evaluación que Grinell hizo sobrela relación entre Parus rufescens y loshábitats en que se encontraba era capazde explicar la expansión de esa especieen EEUU en la década de 1980). Porúltimo, a mediados del siglo XX, lostrabajos de Hutchinson (1978) y MacArthur (1958) impulsaron la estrategia

moderna de análisis cuantitativo de lasrelaciones entre las especies y sushábitats (ver Block & Brennan 1993).

Actualmente, el estudio de ladistribución espacial de las especiestiene una gran importancia en ecología(Lawton 1996; Gaston & Blackburn1999); tanto que para algunos autores lameta principal de esta ciencia esanalizar las causas de que las especiesaparezcan donde lo hacen con lasabundancias en que lo hacen (Begon,Harper & Townsend 1995). Por otrolado, y desde un punto de vistaaplicado, las sociedades industrializadasestán demandando herramientas deplanificación territorial que incluyan untipo de valoración objetiva y repetiblede los recursos naturales, entre los quese encuentra, en un capítulo destacado,la biodiversidad (Colwell & Coddington1994; Díaz, Illera & Hedo 2001). Poreste motivo, se han desarrollado en laúltima década algunos programasregionales que primero muestreanextensivamente distintos gruposanimales y vegetales y después generanuna cartografía de su distribución (comodetallaré más abajo), cuya principalutilidad, desde una vertienteconservacionista, es la de servir comouna guía de la adecuación del territoriopara las distintas especies. Así, losdistintos proyectos denominados “Gap”que se han realizado desde 1993 enEE.UU. (Scott et al. 1993; y verBojórquez-Tapia et al. 1995 para suaplicación en suramérica) tienen comoobjeto evaluar las necesidades deprotección de fauna y flora mediante elexamen de la cobertura de la red deespacios protegidos sobre los hábitatsconsiderados adecuados para lasdistintas especies; los programas que se

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –INTRODUCCIÓN

2

han llevado a cabo en Australia desde1994 (Pearce, Ferrier & Scotts 2001), yen Suiza desde 2000 (Guisan & Harrell2000, ver tambiénhttp://www.wsl.ch/land/products/biomod/ewelcome.html) también pretendíanidentificar las áreas apropiadas paracada especie a una resolución espacialdetallada (típicamente 40 hectáreas enlos proyectos Gap, 4 ha en la cartografíaaustraliana y 1 ha en el proyectoLandspot suizo). En España no tenemosun programa similar, aunque el interésen disponer de una cartografía generalde la diversidad ha conducido aldesarrollo de los atlas regionales devertebrados referidos a cuadrículas de100 km2 (el de peces ya está finalizado,Doadrio 2001, y el resto estabaneditándose mientras se elaboraba estetexto).

La relevancia de la cartografía deespecies en biología de la conservaciónse manifiesta al enumerar susnumerosas aplicaciones, quecomprenden desde unasconceptualmente sencillas (peronecesarias) como, por ejemplo, eldiseño de redes de espacios protegidos(mediante la localización de áreassusceptibles de protegerse: Kiester et al.1996, y el ordenamiento en función desu interés conservacionista: Margules &Austin 1994; Bojórquez-Tapia et al.,1995), y la generación eficaz de mapasde distribución en grandes territorios(e.g., Verlinden & Masogo 1997;Wright, Fielding & Wheater 2000;Osborne, Alonso & Bryant 2001,quienes aprovechan las herramientas dela teledetección), hasta otras máscomplejas como el manejo de especiesamenazadas (e.g., Palma, Beja &Rodrigues 1999; Sánchez-Zapata &Calvo 1999, donde se identificanpatrones de paisaje limitantes para ladistribución de algunas especies y seprevén efectos de los cambios en losusos del suelo), la gestión de

ecosistemas (He et al. 1998, quienesgeneran una proyección espacial y deestructura de la población de bosquesmultiespecíficos), la reintroducción yrecolonización de especies (e.g.,Mladenoff et al. 1997, 1999, donde sepredicen las áreas de expansión y lostamaños poblacionales que podríaalcanzar Canis lupus expansión por eloeste norteamericano; y Yáñez &Floater 2000, donde se describen lasáreas adecuadas para la reintroducciónde la tarántula Brachypelma klaasi), lacomprobación de hipótesisbiogeográficas (e.g., Mourell & Ezcurra1996; Leathwick 1998; Manel, Buckton& Ormerod 2000), o los análisispoblacionales (e.g., Akçakaya,McCarthy & Pearce 1995; Akçakaya &Atwood 1997; donde se da unadimensión espacial a los análisis deviabilidad poblacional, y Dunning et al.1995, quienes introducen los modelospoblacionales espacialmente explícitos).Además, otras áreas de biologíaaplicada se benefician del análisis de ladistribución de especies y de ladelimitación de áreas de distribuciónespacial, por ejemplo, los estudios deespecies invasoras, plagas y vectores deenfermedades (e.g., Venier et al. 1998;Buchan & Padilla 2000, que estudianlas áreas susceptibles de ser invadidas oafectadas por una plaga), otros enrelación con indicadores biológicos(e.g., Utzinger, Roth & Peter 1998,donde el análisis de la distribuciónespacial de especies permite reconocerefectos de la contaminación sobre el pezCottus gobio), o, por último, algunoscon una perspectiva proxima alordenamiento urbano (Le Lay, Clergeau& Hubert-Moi 2001, quienes tratansobre la gestión de especies en unentorno antrópico). Otras posibilidadesde la cartografía de especies ynumerosas referencias se puedenencontrar en Guisan y Zimmerman(2000), Manel et al. (2001) y, conespecial referencia a las limitaciones

http://www.wsl.ch/land/products/biomo

Introducción y objetivos

3

metodológicas, en el capítulo I de estatesis doctoral.

Los modelos de adecuación delhábitat y los Sistemas de InformaciónGeográfica

Se podría suponer que la adquisiciónexperimental del conocimiento precisode los factores que influyen en laaparición y la abundancia de lasespecies permitiría hacer prediccionesexactas de dónde (y cuándo) apareceríacada especie en concreto, es decir,permitiría cartografiar susdistribuciones con exactitud. Sinembargo, el conocimiento de ladistribución de una especie sólo puedeser de tipo probabilístico pues sonvarios los elementos estocásticos queinfluyen en que una especie estépresente en un área independientementede lo adecuada que le resulte (Tyre,Possingham & Lindenmayer 2001).Además, los estudios experimentalesresultan muy costosos y son imposiblesde realizar para ámbitos amplios. Portanto, son necesarias otrasaproximaciones (un caso habitual ylegítimo para algunos ecólogos, e.g.Lawton 1996, pero ver tambiénHairston 1989, cap.1 para una defensaapasionada de la experimentación). Laalternativa es el uso de modelos, bien detipo numérico (o simulación), en los quese resumen los factores más importantesde un proceso y sus efectos posibles; obien de tipo estadístico, en los que seemplean variables descriptoras fácilesde medir y que se confía en que secorrelacionen con los factores causalessubyacentes. Aunque existe unaprevisible tendencia a la unión deambos tipos de modelos en el mismoanálisis de distribución de especies(Akçakaya, McCarthy & Pearce 1995;Akçakaya & Atwood 1997; Hirzel2001), los primeros (modelos basadosen el individuo y de autómatas

celulares) generan prediccionesgenerales y son más adecuados para lacomprobación de hipótesis, mientrasque los segundos tienen un ámbito deaplicación particular y son másapropiados para la cartografía deespecies (Morrison, Marcot & Mannan1998, cap.10). Son estos últimos los quese emplearán a lo largo de esta tesisdoctoral con el nombre de modelos dedistribución de especies (del ingléspredictive distribution modelling) o deadecuación del hábitat (habitatsuitability modelling).

La mayor parte de los ejemplos deaplicaciones que se han enumeradoanteriormente utilizan mapas dedistribución potencial que se elaboranen una secuencia de dos pasos. Primero,se construyen modelos estadísticosmultivariantes que definen la respuestade una especie a un conjunto devariables explicativas que resumen losaspectos físicos y biológicos a los queestá expuesta la especie (es decir, sedefine la adecuación del hábitat).Segundo, se interpola ese resultado alconjunto del área de estudio medianteun Sistema de Información Geográfica(es decir, se genera un mapa de hábitatpotencial, Guisan & Zimmermann2000). Tal secuencia se emplearátambién en los diferentes capítulos deesta tesis doctoral, donde se introducenbrevemente la estrategia de modeladoasí como los métodos estadísticos y lospropios de un Sistema de InformaciónGeográfica (en adelante SIG) que seutilizan. No obstante, es convenienteexponer de manera sucinta algunosfundamentos de los métodosestadísticos y de SIG que se emplearán,para ofrecer una visión de conjunto quecentre al lector. Una exposicióndetallada de los métodos estadísticospuede encontrarse en McCullagh yNelder (1989, capítulos 1,2 y 4), Hastiey Tibshirani (1990, capítulos 2 a 6), yen Chambers y Hastie (1993, capítulos


4

7 y 8), mientras que las referenciaidóneas para los principales temas deGIS y teledetección tratados aquí sonLillesand y Kiefer (1994, capítulos 1 y 5a 7), Burrough y McDonnell (1998,capítulos 1 a 4), y, Gutiérrez-Puebla yGould (1994) como introducción encastellano.

Algunos rudimentos estadísticos

Los primeros análisis de ladistribución de especies se basaban entécnicas de ajuste por mínimoscuadrados, generalmente regresionesmúltiples lineares o análisisdiscriminante (p.e., los estudios quebuscaban identificar las relaciones entrelas especies y sus hábitats, wildlife-habitat relationships, ver referencias enMorrison, Marcot & Mannan 1998).Estos análisis relacionaban una variablerespuesta que informaba sobre lapresencia de una especie en un área,bien en términos de abundancia o biensegún una variable dicotómicaindicativa de la presencia o ausencia, ydistintas variables explicativas (opredictoras) que describían el entornoambiental de los puntos de muestreo. Enestos casos el modelo que se usa es deltipo:

i

p

j ijji XY εβα ++= ∑ =1

donde, siguiendo la terminologíatradicional, Y es la variable respuestaque debe ser contínua y cuya relacióncon los predictores se asume lineal, Bj

los coeficientes que multiplican a cadavariable predictora X; y ei los errores,que se suponen que siguen unadistribución normal y se cancelan unosa otros. Tanto la variable respuestacomo las predictoras habían de sertransformadas frecuentemente paraacercarlas a una distribución normal, loque podría no resultar fácil (nirazonable), especialmente en el casohabitual de analizar una respuestabinaria indicativa de la presencia o

ausencia de una especie en un punto. Enlos últimos 20 años estos modeloslineares se han incorporado a un marcomás amplio de análisis, el de losmodelos lineares generalizados (oGLM, del inglés Generalized LinearModels, Nelder & Wedderburn 1972;McCullagh & Nelder 1989) donde lavariable respuesta puede seguircualquiera de las distribuciones de lafamilia exponencial (normal, Poisson,binomial, gamma o normal inversa), ylos ajustes del modelo ya no se estimanmediante mínimos cuadrados, sinomediante estadísticos de máximaverosimilitud. En contraste con elfuncionamiento de las técnicas demínimos cuadrados, donde los datos seajustan a un modelo determinado (lasdesviaciones se solucionantransformando los datos), el paradigmade la estadística de máximaverosimilitud es el ajuste del modelo alos datos mediante la búsqueda de losvalores de los parámetros del modelo(Bj) que hacen más probable el conjuntode datos observado (McCullagh &Nelder 1989; capítulo 2, y Harrell2001, para una introducción a laestadística de máxima verosimilitud).En los GLM la variable respuesta no semodela directamente, sino a través deuna transformación denominada funciónvínculo (g(Y)) y una distribución de loserrores adecuada a la naturaleza de talrespuesta:

i

p

j ijji XfYg εα ++= ∑ =1)()(

Así, en el modelo adecuado para unavariable respuesta binaria (elequivalente a la regresión logística) nose estima directamente la probabilidadde que la variable adquiera uno de losdos estados posibles, sino que se usa lafunción logit:

εβ

β

+∑−

∑==

X

X

e

eqpEYE

1)/())(logit(

donde p es la probabilidad asociada auno de los estados de la variable


5

dicotómica (en nuestro caso la presenciade una especie en un punto demuestreo) y q la probabilidadcomplementaria (en nuestro caso laausencia de tal especie en el mismopunto); los errores siguen aquí unadistribución binomial. La linearidad delmodelo se mantiene en el denominadopredictor lineal (η=ΣβX) que ya noinforma sobre la variable respuestadirectamente, como en la regresióngaussiana tradicional, sino sobre lafunción que se haya usado (el logit(Y)en el modelo equivalente a la regresiónlogística), por lo que ha de sertransformado para que resulte másinterpretable.

Una limitación de los GLM es quelas relaciones que se modelan sonlineales, es decir, los predictores Xinfluyen sobre la variable respuesta Yde una manera constante determinadapor sus coeficientes β; por ejemplo, unincremento de n unidades en unpredictor X influye en β(X1-Xn) unidadesen la respuesta Y. Aunque estalimitación puede modificarse en algunamedida mediante el uso detransformaciones polinómicas de lasvariables (por ejemplo usando X+X2 enlugar de X), se han desarrolladorecientemente modelos aún másgenerales de los que los GLM puedenconsiderarse un caso particular. Se tratade los modelos aditivos generalizados(GAM, del inglés Generalized AdditiveModels, Hastie & Tibshirani 1990), quedifieren fundamentalmente de los GLMen que la relación entre la respuesta ylos predictores se estima mediante unafunción de suavizado gráfico como lasregresiones locales o los “splines”(término para el que no conozco sutraducción al castellano):

i

p

j ijji XfYg εα ++= ∑ =1)()(

donde los valores de Yi se estiman conun procedimiento doblemente iterativo

considerando los valores de lospredictores X en un entorno próximo alpunto i. Tal mecanismo incluye laestima iterativa de mínimos cuadradosbaremados, iterated reweighted leastsquares o IRLS, y del modeladoiterativo de los residuos parciales,conocido como backfitting (este no es ellugar para entrar en más detallesestadísticos, pero el lector interesado enellos puede dirigirse a los breves eilustrativos capítulos 3 y 6 de laestupenda monografía de Fox 2000).

El uso de GLM está muy establecidoen ecología (Crawley 1993; Crawley2002) y se han usado con frecuenciapara modelar la distribución de especiesy la selección de hábitat (p.e.: Austin etal. 1996; Bustamante 1997;Bustamante et al. 1997). Su cálculo esrelativamente rápido y, al poderseexpresar de forma analítica, son fácilesde transportar a un entorno de SIG(Guisan, Theurillat & Kienast 1998).Por el contario, los GAM son aún rarosen ecología en general y en los estudiosde ditribución de especies en particular(Franklin 1998; Elith 2000; Fewster etal. 2000; Forney 2000). Su mayorflexibilidad tiene como contrapartidauna mayor lentitud (debido al procesodoblemente iterativo que se requierepara su cálculo) y la carencia de unafórmula analítica para resolverlos quedificulta implementarlos en un entornoSIG. Por estos motivos los GAM seusan con frecuencia de maneraexploratoria, para detectar quétransformaciones de los predictorespuede ser adecuada (como se sugiere enHastie & Tibshirani 1990; Brown 1994;y se aplica en una situación práctica enFranklin 1998).

El siguiente esquema resume lascaracterísticas de las técnicasestadísticas que se han introducido enlos párrafos precedentes, y muestra unaexplicación que ayudará a interpretarlas


6

en el contexto de la modelización delhábitat potencial que se empleará másadelante:

- en los modelos lineares (lm):E(Y) = f(y) = ΣβX = β1X1 +...+ βpXp,

es decir, la probabilidad de que unaespecie esté presente en un punto demuestreo depende de una combinaciónlineal de las variables predictoras

- en los modelos linearesgeneralizados (glm):

E(Y) = g[f(y)] = β1X1 +...+ βpXp,

es decir, el cociente entre laprobabilidad de presencia y laprobabilidad de ausencia de una especieen un punto de muestreo (el “logit” deY) depende de una combinación linealde las variables predictoras

- en los modelos aditivosgeneralizados (gam):

E(Y) = g[f(y)] = Σf(X) = f1(X1) +...+fp(Xp),

es decir, el cociente entre laprobabilidad de presencia y laprobabilidad de ausencia de una especieen un punto de muestreo (el “logit” deY) depende de una combinación nonecesariamente lineal de las variablespredictoras

La evaluación de este tipo demodelos se enfrenta con dos problemasfundamentales que pueden hacer que sesobreestime su éxito: la autocorrelaciónespacial de los datos (detallado en elcapítulo I) y la consideración demúltiples modelos alternativos. Enprimer lugar, estos modelos suponenque los errores son independientes entresí, es decir, asumen que un punto demuestreo no ofrece informaciónrespecto sus vecinos. Sin embargo losprocesos ecológicos, y en particular ladistribución y abundancia de las

especies, muestran con gran frecuenciaautocorrelación espacial por la que lospuntos geográficamente próximostienden a parecerse (Legendre 1993;Augustin, Mugglestone & Buckland1996). En esta tesis doctoral se haoptado por obviar el análisis de ladependencia espacial cuando se hacíancomparaciones relativas entre modelosgenerados con los mismos datos(capítulos II a VI donde lasconclusiones de los análisis soninmunes a los posibles sesgos deconstrución de lo modelos por basarseen comparaciones relativas), y se hanusado regresiones de las coordenadasgeográficas y autocovariables cuando sepretendía ofrecer una valoraciónabsoluta de los modelos (ver capítulosVII y VIII). En segundo lugar (y con undesarrollo más detallado porque no sehace mención a este asunto en elcapítulo I), un comportamiento típicoentre quienes practican la estadística oentre los propios estadísticosprofesionales es realizar inferencias apartir de un modelo como si éste sehubiera especificado a priori (Chatfield1995), es decir, ignorando el hecho deque fue escogido entre un conjunto demodelos alternativos (lo que se haconsiderado un escándalo oculto, aquiet scandal, en la literatura estadística(Breiman, 1992 en Chatfield 1995). Setrata de un problema general quecomprende, por ejemplo, los problemasbien conocidos derivados de realizarpredicciones con modelos generadosmediante regresión por pasos, que se hademostrado que pueden incorporarvariables espúreas y sobreestimarse suscapacidades predictivas (Flack & Chang1987; Buckland, Burnham & Augustin1997; en el campo de la estadística yMac Nally 2000, en ecología). Además,recientemente están apareciendo críticasa tal paradigma que abogan por tener encuenta la incertidumbre en laespecificación de un modelo, bienmediante la consideración simultánea de


7

varios modelos, cuyas predicciones sepromediarían de forma que se daría máspeso a aquellas que se dedujeran demodelos más creíbles (Burnham &Anderson 1998; ver la propuesta basadaen la teoría de la información:Anderson, Burnham & Thompson 2000;y otra que se fundamenta en un métodode partición jerárquica: Mac Nally2000); o bien mediante lapreespecificación de la complejidad delos modelos y la incorporación de talincertidumbre de selección de unmodelo a los coeficientes de regresión(lo que Harrell 2001, denominashrinking y podría traducirse como“encogido de los coeficientes”). A pesarde estas críticas (que han aparecidofuera de la literatura estadísticaespecializada con posterioridad alcomienzo de esta tesis doctoral), en loscapítulos siguientes se seguirá elprocedimiento común de construcciónde modelos mediante una selección devariables predictoras paso a paso. Existeun motivo triple para no seguir lasnuevas propuestas: (i) se carece de lainformación básica necesaria paradiseñar un modelo (es decir, paraespecificar sus variables predictoras apriori) para la mayor parte de lasespecies que se tratarán aquí, (ii) lospredictores usados son groseros yfueron diseñados con objetivosdiferentes a la investigación de lasrelaciones entre las especies y sushábitats, y (iii) el número de especiesque se considera en los análisis es altopor lo que se precisan métodosautomáticos para analizarlaseficazmente. La desventaja principal delmétodo de selección por pasos es, en elcaso que ocupa a este trabajo, lasobreestimación de la capacidadpredictiva de los modelos resultantesmediante los estadísticos habituales (elporcentaje de absorción de varianza odevianza) y, en menor medida, laincorporación de variables espúreas alos modelos. Sin embargo, el primer

problema es irrelevante en lascomparaciones relativas entre modelos(como las que se usan en la mayoría delos capítulos de esta tesis doctoral) yambos problemas pueden paliarsemediante técnicas de remuestreo (p.e.,bootstraping y jackknife) y devalidación cruzada (Verbyla & Litvaitis1989), así como por el uso pragmáticode medidas empíricas de capacidadpredictiva (Kapa y AUC, que estiman elporcentaje de aciertos independientesdel azar, ver Pearce & Ferrier 2000;Manel, Williams & Ormerod 2001).Todas estas técnicas se usanextensivamente en los capítulos quesiguen.

Algunos rudimentos de los Sistemasde Información Geográfica

Los Sistemas de InformaciónGeográfica o SIG son bases de datosrelacionados espacialmente cuyo diseñotiene como objetivos: (i) almacenar ymantener datos espacialmenteexplícitos, (ii) mostrarlos y analizarlos,(iii) realizar operaciones espacialescomplejas con ellos y (iv) comunicareficazmente los resultados a los gestoresy al público en general (ver Hirzel 2001,capítulo 1). Para ello, los datos sealmacenan en las llamadas capas deinformación según dos tipos deestructura: vectorial y en rejilla (o“ráster”). En el primer tipo los datos seadjuntan a objetos cuyas coordenadasespaciales se definen con precisión.Estos objetos pueden ser puntos, o tenerforma de líneas o polígonos según lascaracterísticas de la información queincorporen. Así, los puntos suelen servirpara almacenar datos con dimensionesespaciales muy reducidas o sin ellas(p.e., el lugar donde se hizo unmuestreo), las líneas son adecuadas paraformas monodimensionales comocauces o carreteras, y los polígonos paraestructuras bidimensionales comoparcelas de cultivo o de un tipo de


8

vegetación. Los datos que incorporenpuede ser de cualquier tipo: numéricoscontínuos como la temperatura,numéricos discretos como el número deaves detectadas o bien categóricos,como el tipo de vegetación. Laestructura vectorial es muy adecuadapara almacenar y manejar datoscualitativos como son las coberturas deusos del suelo que sirven de base paralas variables predictoras que se usarán alo largo de esta tesis. En la segundaclase de estructura de almacenamiento,de tipo rejilla, los datos se adjuntan a lasceldas (o “píxeles”) de tamañohomogéneo en que se divide el área deestudio. Este tipo de almacenamiento esmás adecuado para datos cuantitativosque muestren una variación espacialgradual, como puede ser la temperatura,humedad, altitud, etc (cada uno de loscuales es una capa de informacióndistinta plasmada en una imagen omapa). Además, el sistema de rejillaincorpora fácilmente la informacióngenerada mediante teledetección porsatélite (que se emplea ampliamente enlos capítulos que siguen), pues estossensores almacenan los registros segúnuna misma estructura en celdasisométricas. En esta tesis doctoral lamayor parte de los datos se almacenarony gestionaron en forma de rejilla. Losprogramas que utilizamos fueronIDRISI (Eastman 1997; Eastman 1999)y MIRAMON (Pons 2000) debido a susrelativos bajos costos y facilidad de uso.

Así pues, el esquema defuncionamiento general de un GIS esconceptualmente sencillo y abordaproblemas para cuya resoluciónsatisfactoria es necesario considerarsimultáneamente distintascaracterísticas del territorio que tienenuna expresión espacial. Cada una detales características (frecuentementeusos y coberturas del suelo, y factoresclimatológicos y topográficos) serepresenta en imágenes (o mapas) que

pueden superponerse para obtenerinformación sobre un punto localizadodel territorio (correspondiente ennuestro caso a los puntos de censo) osobre un entorno arbitrario determinadopor el analista (ver figura 2 del capítuloV). Este modo de proceder sólo difierede los trabajos que, hasta hace poco,debían realizarse a mano (como, porejemplo, los que se aprenden en lasprácticas de la asignatura de ecología enla UAM) en su mayor rapidez y en laposibilidad de llevar a cabo operacionesmuy complejas.

Una de las principales fuentes deinformación para un SIG es lateledetección, que puede definirse comola adquisición de información remota,es decir, alejada del sensor que larecibe. Nuestros ojos, por ejemplo, sonun magnífico sensor de la luz reflejadapor los objetos que nos rodean. Lossatélites que orbitan la tierra recibeninformación de la radiaciónelectromagnética reflejada por lacubierta terrestre siguiendo el mismoprincipio (un emisor de energía –generalmente el sol–, un cuerpo que larefleja y un sensor que la recibe), perode manera no limitada a las radiacionesde longitudes de onda visibles por el ojohumano. Así, el hombre no puedepercibir la radiación infrarroja que,recogida por sensores electrónicos,permite distinguir fácilmente entrecuerpos con distinta temperatura ogrado de humedad, lo que resulta deenorme utilidad para la elaboración depronósticos meteorológicos, elseguimiento del estado fitosanitario demasas vegetales o la generación decartografía, entre otras aplicaciones. Lossatélites comerciales cuya informaciónmás se usa en los SIG transportansensores, denominados multiespectrales,que son capaces de medir la energía endistintas partes del espectroelectromagnético. La cantidad y calidadde los datos que aportan (y la utilidad


9

que se les pueda dar) dependenfundamentalmente de su resoluciónespacial, es decir, del tamaño del áreasobre el terreno para el que el sensorobtiene un valor, y de su resoluciónespectral, es decir, de la magnitud delespectro electromagnético a la que elsatélite es sensible. En este trabajo se hautilizado información procedente de lossensores TM (“Thematic Mapper”) delos satélites Landsat, del sensor LISS-IIIdel satélite IRS, y del sensor AVHRR(“Advanced Very High ResolutionRadiometer”) de los satélitesgestionados por el NOAA (“U.S.National Oceanic and AtmosphericAdministration”). Los sensores TM sonsensibles a siete longitudes de onda quecubren el espectro electromagnético enel segmento correspondiente a la luzvisible, y en partes del infrarrojocercano, medio y térmico. Cada imagencubre 185 km y tiene una resoluciónespacial de 30 metros (900 m2) que esmuy próxima al detalle de la cartografíatemática de que dispusimos (50 metrospara los mapas del SinambA), lo quefacilitó usar tales datos para modificarla(especialmente para distinguir pequeñasformaciones de ribera que no aparecíanreflejadas en los mapas temáticos). ElIRS tiene características similares (625m2 de resolución espacial) y loutilizamos con el mismo objetivo. Encontraste, el sensor AVHRR tiene unasmenores resolución espectral y espacialpues sólo recoge información en cincobandas del espectro electromagnético enunidades de 1.1 km (ca. 1 km2) de área.Por tanto, resulta más adecuado paraevaluaciones en superficies extensas:nosotros lo utilizamos para calcular uníndice de superficie riparia enAndalucía (ver capítulo VII), y uníndice de vegetación de la penínsulaIbérica (ver capítulo VIII).

Objetivos, sujetos de estudio yestructura de la tesis

El objetivo inmediato de este trabajoes explorar las posibilidades delmodelado de la distribución de especiesen un entorno antropizado yheterogéneo, con la intención de sugerirpautas generales para el desarrollo deuna estrategia eficaz de cartografía deespecies. Por este motivo se hanprobado distintas técnicas y estrategiasa cada una de las cuales se dedica unode los capítulos que forman esta tesisdoctoral, como se detalla en los últimospárrafos de este apartado. Enconsecuencia, gran parte de este trabajotiene una notable, pero ineludible,componente técnica que se hapretendido compensar con la puesta enpráctica de los modelos en la SecciónTercera.

El objeto de estudio han sido lasaves, y dentro de ellas las ligadas amedios terrestres. Tal selección se hizoatendiendo a que, por un lado, estegrupo de organismos comprende unconjunto numeroso y variado deespecies, lo que permite ensayarmodelos bajo distintas características deabundancia y selección de hábitat, y,por otro, las aves se muestrean demanera relativamente sencilla, lo quefacilita la adquisición de datos decampo para construir los modelos. Lasaves ligadas a medios acuáticos comomarismas, lagunas y, en general, áreasde aguas libres, no se consideraron eneste trabajo porque requieren unametodología de muestreo de campo y deanálisis en un entorno de SIG muydiferente al resto (sí se analizaron, sinembargo, las especies propias de lossotos de ribera). En un trabajo decomparación de estrategias de modelado(Pearce & Ferrier 2000) mostraron quelos distintos grupos de organismospueden diferir en cuanto a lasestrategias de modelado que les resultan


10

óptimas. Sin embargo, creemos que losresultados generales de esta tesisdoctoral podrían ser extensibles a otrosgrupos de organismos distintos de lasaves terrestres porque la principaldisimilitud que encontraron Pearce yFerrier (op.cit) se debió a lacomplejidad de las relaciones que semodelaban (p.e., los modelos de losreptiles incorporaron los predictorescomo polinomios de un grado superioral del resto de organismos), y en loscapítulos que siguen los modelos que seprueban son muy flexibles (modelosGAM y polinomios que pueden ser dealto grado).

Esta tesis doctoral está dividida encuatro secciones de distinta extensión.La mayoría de los capítulos que lacomponen se han escrito en inglés conel estilo y formato de un artículocientífico y muchos de ellos ya estánenviados a distintas revistas de difusióninternacional para publicarse. Estasituación se aleja del procedimientohabitual en las tesis tradicionales en lasque un manuscrito totalmente inédito sedefendía ante un tribunal y después seextraían los artículos que hubieranlugar, pero se acerca a una opción másmoderna que evita grandes dilacionesentre la realización de un trabajo y supublicación (un ejemplo de esta prácticaen biología de la conservación son estasdos excelentes tesis doctoralesdefendidas recientemente en Europa :Guyonne 2001; Hirzel 2001). La ventajamás relevante para el lector es que lalectura de los capítulos se agiliza, puesaquéllos son breves y están centrados enun problema particular; en contra debeadmitirse que se hallará una ciertarepetición en los apartadosmetodológicos, alguna heterogeneidaden los análisis que se siguen (fruto delaprendizaje durante la elaboración de latesis), y, quizás, la ausencia de unamayor descripción de detalles muytécnicos de interés principalmente para

quien se embarcara en una empresasimilar.

La Sección Primera (Reflexionespreliminares: utilidad y limitaciones delos modelos de distribución de especies)está compuesta por sólo un capítulo(Capítulo I) donde se detalla la utilidady limitaciones de los modelos dedistribución de especies en ecología. ElCapítulo I complementa estaintroducción general justificando elinterés de la cartografía de especies ydetallando sus limitaciones técnicas yconceptuales.

La Sección Segunda (Aspectosmetodológicos: Técnicas y estrategiasdel modelado de la distribución deespecies), que forma la mayor parte dela tesis y tiene un importantecomponente metodológico, comprendelos capítulos II a VI, que explorandistintas técnicas y estrategias demodelado analizadas para valorar suutilidad en la definición general de unprotocolo de modelización. Así, elCapítulo II describe una técnica deoptimización del tiempo dedicado a losmuestreos de aves que se fundamenta enel teorema del valor marginal: lasdistintas especies tendrán un tiempoóptimo de muestreo diferente ypredecible según sus características detamaño, abundancia y tipo de hábitatpreferido.En el Capítulo III se comparanmodelos generados mediante unprocedimiento estadístico automáticocon otros construidos mediante unprotocolo supervisado paso a paso, y seconcluye que los modelos automáticostienen una capacidad predictiva similara la de los modelos supervisados. Elcapítulo IV analiza las fuentes de datosde donde se extraen los predictores quese prueban en los modelos y muestraque la cartografía temática digitalexistente (que se ha elaborado conpropósitos diferentes a las necesidadesde la cartografía de especies) permite


11

crear modelos a gran resolución de altacapacidad predictiva, igual o mayor quela que se alcanza con información desatélite (los resultados sugieren,además, que existe un límite máximo ala capacidad predictiva que se puedealcanzar con estos modelos). Losmodelos que se aplicarán más adelantey, en particular, en el capítulo VII de laSección Tercera, se desarrollaránmediante un procedimiento automáticode selección de predictores derivados ensu mayor parte de la cartografíatemática digital preexistente. Elcapítulo V aborda la selección del gradode detalle espacial de los predictorespara cada especie que haga mayor lacapacidad predictiva de sus modelos, yconcluye que las variables ambientalesque se usan como predictores debenmedirse en un radio muy amplio entorno al punto de muestreo, lo quemuestra de forma indirecta un efecto dela configuración del paisaje sobre laprobabilidad de encontrar a una especieen concreto en un área. Por fín, elcapítulo VI explora qué conjunto devariables explicativas (topo-climático ydescriptivas de la vegetación y elpaisaje) genera modelos máspredictivos, y determina que los mejoresmodelos se construyen con un conjuntomixto de variables e identifica a losdescriptores del paisaje como las másimportantes.

En la Sección Tercera (Puesta enpráctica: aplicaciones de la cartografíade especies) se exploran dosaplicaciones típicas del modelado de ladistribución de especies. En primerlugar (capítulos VII y VIII) se utilizandatos referidos a una malla de 10x10kilómetros, que es la forma quetradicionalmente se ha usado en losesfuerzos de cartografiado regional deespecies. En estos ejemplos se usa lamodelización para detectar algunasáreas geográficas adecuadas paradistintas especies de rapaces e

identificar zonas con problemas deconservación. En segundo lugar(Capítulo IX), se estudia hasta quépunto difieren los modelos empíricosestadísticos que se desarrollan en estatesis doctoral con los que podríancrearse basándose en el criterio deexpertos aplicado a la cartografía deespecies existente (i.e., datos de atlas ymapas de distribución).

La Sección Cuarta (Conclusiones:esperanzas y desesperanzas de losmodelos) comprende sólo un capítulo(Capítulo X) en el que se ofrecen unasconclusiones generales y una valoraciónde los resultados obtenidos.

BIBLIOGRAFÍA

Akçakaya, H. R. & Atwood, J. L.(1997). A habitat-basedmetapopulation model of theCalifornia Gnatcatcher. ConservationBiology, 11(2): 422-434.

Akçakaya, H. R., McCarthy, M. A. &Pearce, J. L. (1995). Linkinglandscape data with populationviability analysis: managementoptions for the helmeted honeyeaterLichenostomus melanops cassidix.Biological Conservation, 73: 169-176.

Anderson, D. R., Burnham, K. P. &Thompson, W. L. (2000). Nullhypothesis testing: problems,prevalence, and an alternative.Journal of Wildlife Management,64(4): 912-923.

Augustin, N. H., Mugglestone, M. A. &Buckland, S. T. (1996). Anautologistic model for the spatialdistribution of wildlife. Journal ofApplied Ecology, 33: 339-347.

Austin, G. E., Thomas, C. J., Houston,D. C. & Thompson, D. B. A. (1996).Predicting the spatial distribution ofbuzzard Buteo buteo nesting areasusing a Geographical Information


12

System and remote sensing. Journalof Applied Ecology, 33: 1541-1550.

Begon, M., Harper, J. L. & Townsend,C. R. (1995). Ecología: individuos,poblaciones y comunidades. Omega,Barcelona

Block, W. M. & Brennan, L. A. (1993).The habitat concept in ornithology:theory and applications. CurrentOrnithology (eds D. M. Power), pp.35-91.

Bojórquez-Tapia, L. A., Azuara, I.,Ezcurra, E. & Flores-Villela, O.(1995). Identifying conservationpriorities in Mexico throughgeographic information systems andmodeling. Ecological Applications,5(1): 213-231.

Brown, D. G. (1994). Predictingvegetation types at treeline usingtopography and biophysicaldisturbance variables. Journal ofVegetation Science, 5: 641-656.

Buchan, L. A. J. & Padilla, D. K.(2000). Predicting the likelihood ofthe water milfoil presence in lakes, amacrophyte monitoring tool.Ecological Applications, 10: 1442-1455.

Buckland, S. T., Burnham, K. P. &Augustin, N. H. (1997). Modelselection: an integral part ofinference. Biometrics, 53: 603-618.

Burnham, K. P. & Anderson, D. R.(1998). Model selection andinference: a practical information-theoretic approach. Springer-Verlag,New-York.

Burrough, P. A. & McDonnell, R. A.(1998). Principles of GeographicalInformation Systems. OxfordUniversity Press, Oxford

Bustamante, J. (1997). Predictivemodels for lesser kestrel Falconaumanni distribution, abundanceand extinction in southern Spain.Biological Conservation, 80: 153-160.

Bustamante, J., Donázar, J. A., Hiraldo,F., Ceballos, O. & Travaini, A.

(1997). Differential habitat selectionby immature and adult Grey Eagle-buzzards Geranoaetus melanoleucus.Ibis, 139: 322-330.

Chambers, J. M. & Hastie, T. J. (1993).Statistical models in S. Chapman &Hall, London.

Chatfield, C. (1995). Model uncertainty,data mining and statistical inference.Journal of the Royal StatisticalSociety, Series A, 158: 419-466.

Colwell, R. K. & Coddington, J. A.(1994). Estimating terrestrialbiodiversity trough exploration.Philosophical Transactions of theRoyal Society. London. Series B,Biological Sciences, 345: 101-118.

Crawley, M. J. (1993). GLIM forecologists. Blackwell Science,London.

Crawley, M. J. (2002). Statisticalcomputing. An introduction to dataanalysis using S-Plus. John Wiley &Sons, Chichester.

Díaz, M., Illera, J. C. & Hedo, D.(2001). Strategic environmentalassessment of plans and programs: amethodology for estimating effectson biodiversity. EnvironmentalManagement, 28(2): 267-279.

Doadrio, I., Ed. (2001). Atlas y librorojo de los peces continentales deEspaña/. Ministerio de MedioAmbiente-Consejo Superior deInvestigaciones Científicas, Madrid.

Dunning, J. B. J., Stewart, D. J.,Danielson, B. J., Noon, B. R., Root,T. L., Lamberson, R. H. & Stevens,E. E. (1995). Spatially explicitpopulation models: current forms andfuture uses. Ecological Applications,5(1): 3-11.

Eastman, J. R. (1997). Idrisi forWindows. User's Guide. Clark Labs,Worcester.

Eastman, J. R. (1999). Idrisi32.Reference guide. Clark Labs,Worcester.

Elith, J. (2000). Quantitative methodsfor modeling species habitat:


13

comparative performance and anapplication to Australian plants.Quantitative methods forconservation biology (eds S. Ferson& M. Burgman), pp. 39-58. Springer,New York.

Fewster, R. M., Buckland, S. T.,Siriwardena, G. M., Baillie, S. R. &Wilson, J. D. (2000). Analysis ofpopulation trends for farmland birdsusing generalized additive models.Ecology, 81(7): 1970-1984.

Flack, V. F. & Chang, P. C. (1987).Frequency of selecting noisevariables in subset regressionanalysis: a simulation study. TheAmerican Statistician, 41: 84-86.

Forney, K. A. (2000). Environmentalmodels of cetacean abundance:reducing uncertainty in populationstrends. Conservation Biology, 14(5):1271-1286.

Fox, J. (2000). Multiple andGeneralized NonparametricRegression. Sage Publications,Thousand Oaks, CA.

Franklin, J. (1998). Predicting thedistribution of shrub species insouthern California from climate andterrain-derived variables. Journal ofVegetation Science, 9: 733-748.

Gaston, K. J. & Blackburn, T. M.(1999). A critique for macro-ecology. Oikos, 84: 353-368.

Guisan, A. & Harrell, F. E. (2000).Ordinal response regression modelsin ecology. Journal of VegetationScience, 11: 617-626.

Guisan, A., Theurillat, J.-P. & Kienast,F. (1998). Predicting the potentialdistribution of plant species in analpine environment. Journal ofVegetation Science, 9: 65-74.

Guisan, A. & Zimmermann, N. E.(2000). Predictive habitat distributionmodels in ecology. EcologicalModelling, 135: 147-186.

Gutiérrez-Puebla, J. & Gould, M.(1994). SIG: Sistemas de

Información Geográfica. Síntesis,Madrid

Guyonne, J. (2001). Birds and power: afield of tension. University ofUtrecht.

Hairston, N. G. (1989). Ecologicalexperiments: Purpose, desing, andexecution. Cambridge UniversityPress, Cambridge.

Harrell, F. E. (2001). Regressionmodeling strategies. Springer, NewYork

Harris, M. (1983). Antropologíacultural. Alianza Editorial, Madrid

Hastie, T. J. & Tibshirani, R. J. (1990).Generalized Additive Models.Chapman & Hall, London.

He, H. S., Mladenoff, D. J., Radeloff,V. C. & Crow, T. R. (1998).Integration of GIS data and classifiedsatellite imagery for regional forestassessment. Ecological Applications,8(4): 1072-1083.

Hirzel, A. (2001). When GIS come tolife. Linking landscape- andpopulation ecology for largepopulation management modelling:the case of Ibex (Capra ibex) inSwitzerland. Université de Lausanne,Lausanne.

Hutchinson, G. E. (1978). Anintroduction to population ecology.Yale University Press, New Haven,Conneticut.

Kiester, A. R., Scott, J. M., Csuti, B.,Noss, R. F., Butterfield, B., Sahr, K.& White, D. (1996). ConservationPrioritization Using GAP Data.Conservation Biology, 10(5): 1332-1342.

Lawton, J. (1996). Patterns in ecology.Oikos, 75: 145-147.

Le Lay, G., Clergeau, P. & Hubert-Moi,L. (2001). Computerized map of riskto manage wildlife species in urbanareas. Environmental Management,27: 451-461.

Leathwick, J. R. (1998). Are New-Zealand's Nothofagus species inequilibrium with their environment?


14

Journal of Vegetation Science, 9:719-732.

Legendre, P. (1993). Spatialautocorrelation: trouble or newparadigm? Ecology, 74(6): 1659-1673.

Lillesand, T. M. & Kiefer, R. W.(1994). Remote sensing and imageinterpretation. John Wiley & Sons,New York.

Mac Nally, R. (2000). Regression andmodel-building in conservationbiology, biogeography and ecology:The distinction between -andreconciliation of- 'predictive' and'explanatory' models. Biodiversityand Conservation, 9: 655-671.

MacArthur, R. H. (1958). Populationecology of some warblers ofnortheastern coniferous forest.Ecology, 39(4): 599-619.

Manel, S., Buckton, S. T. & Ormerod,S. J. (2000). Testing large-scalehypotheses using surveys: the effectsof land use on the habitats,invertebrates and birds of Himalayanrivers. Journal of Applied Ecology,37(5): 756-770.

Manel, S., Williams, H. C. & Ormerod,S. J. (2001). Evaluating presence-absence models in ecology: the needto account for prevalence. Journal ofApplied Ecology, 38: 921-931.

Margules, C. R. & Austin, M. P. (1994).Biological models for monitoringspecies decline: the construction anduse of data bases. PhilosophicalTransactions of the Royal Society.London. Series B, BiologicalSciences, 344: 69-75.

McCullagh, P. & Nelder, J. A. (1989).Generalized Linear Models.Chapman & Hall/CRC, London.

Mladenoff, D. J., Haight, R. G., Sickley,T. A. & Wydeven, A. P. (1997).Causes and implications of speciesrestoration in altered ecosystems.BioScience, 47(1): 21-31.

Mladenoff, D. J., Sickley, T. A. &Wydeven, A. P. (1999). Predicting

Gray Wolf landscape recolonization:logistic regression models vs. newfield data. Ecological Applications,9(1): 37-44.

Morrison, M. L., Marcot, B. G. &Mannan, R. W. (1998). Wildlife-habitat relationships. Concepts andapplications. The University ofWisconsin Press, Madison.

Mourell, C. & Ezcurra, E. (1996).Species richness of Argentine cacti:A test of biogeographic hypotheses.Journal of Vegetation Science, 7:667-680.

Nelder, J. A. & Wedderburn, R. W. M.(1972). Generalised linear models.Journal of the Royal StatisticalSociety A, 135: 370-384.

Osborne, P. E., Alonso, J. C. & Bryant,R. G. (2001). Modelling landscape-scale habitat use using GIS andremote sensing: a case study withgreat bustards. Journal of AppliedEcology, 38(2): 458-471.

Palma, L., Beja, P. & Rodrigues, M.(1999). The use of sighting data toanalyze Iberian lynx habitat anddistribution. Journal of AppliedEcology, 36: 812-824.

Pearce, J. & Ferrier, S. (2000).Evaluating the predictiveperformance of habitat modelsdeveloped using logistic regression.Ecological Modelling, 133: 225-245.

Pearce, J. & Ferrier, S. (2000). Anevaluation of alternative algorithmsfor fitting species distribution modelsusing logistic regression. EcologicalModelling, 128: 127-147.

Pearce, J., Ferrier, S. & Scotts, D.(2001). An evaluation of thepredictive performance ofdistributional models for flora andfauna in north-east New SouthWales. Journal of EnvironmentalManagement, 62(2): 171-184.

Pons, X. (2000). MiraMon: GeographicInformation System and RemoteSensing software. UniversidadAutónoma de Barcelona, Barcelona.


15

Sánchez-Zapata, J. A. & Calvo, J. F.(1999). Raptor distribution inrelation to landscape composition insemi-arid Mediterranean habitats.Journal of Applied Ecology, 36: 254-262.

Tyre, A. J., Possingham, H. P. &Lindenmayer, D. B. (2001). Inferringprocess from pattern: can territoryoccupancy provide information aboutlife history parameters? EcologicalApplications, 11(6): 1722-1737.

Utzinger, J., Roth, C. & Peter, A.(1998). Effects of environmentalparameters on the distribution ofbullhead Cottus gobio with particularconsideration of the effects ofobstructions. Journal of AppliedEcology, 35(6): 882-892.

Venier, L. A., Hopkin, A. A.,McKenny, D. W. & Wang, Y.(1998). A spatial climate-determinedrisk rating for Scleroderris disease ofpines in Ontario. Canadian Journalof Forest Research, 28: 1398-1405.

Verbyla, D. L. & Litvaitis, J. A. (1989).Resampling methods for evaluatingclassification accuracy of wildlifehabitat models. EnvironmentalManagement, 13: 783-787.

Verlinden, A. & Masogo, R. (1997).Satellite remote sensing of habitatsuitability for ungulates and ostrichin the Kalahari of Bostwana. Journalof Arid Environments, 35(3): 563-574.

Wright, A., Fielding, A. H. & Wheater,C. P. (2000). Predicting thedistribution of European badgets(Meles meles) setts over an urbanizedlandscape: a GIS approach.Photogrammetric Engineering &Remote Sensing, 66: 423-428.

Yáñez, M. & Floater, G. (2000). Spatialdistribution and habitat preference ofthe endangered tarantulaBrachypelma klaasi (Araneae:Theraphosidae) in Mexico.Biodiversity and Conservation, 9:795-810.


SECCIÓN PRIMERA

Reflexiones preliminares

Models are like politicians: support them, use them, but don’t unquestioningly trust

them.

—M.L. Morrison, B. Marcot y R. William

Mannan, WILDLIFE-HABITAT RELATIONSHIPS.

University of Wisconsin Press. 1998.

Modelos predictivos de la distribucion de especies: una revisión de sus limitaciones

17

CAPÍTULO I: Modelos predictivos de la distribución de especies:

una revisión de sus limitaciones

RESUMEN

En las últimas dos décadas se ha despertado un enorme interés en el modelado de la

relación entre las especies y sus hábitats, que responde tanto a la demanda de

información aplicable a la gestión del territorio y a la conservación como al fundamento

básico de la ecología en estudiar la distribución y abundancia de los organismos. Sin

embargo, los modelos predictivos de distribución de especies descansan en ciertas

presunciones y tienen unas limitaciones que conviene conocer antes de desarrollarlos.

En este trabajo se ofrece primero un breve sumario de los tipos de modelado que pueden

encontrarse en estudios de ecología, centrándose en los modelos monoespecíficos de

distribución, es decir, en aquellos que relacionan las características del hábitat con la

presencia de una especie en particular. Posteriormente, se presenta una síntesis

comentada de las limitaciones de carácter biológico y estadístico de los modelos

predictivos, analizando en detalle las presunciones en que se sostienen y los problemas

metodológicos que dificultan su aplicación. Se concluye que los modelos de

distribución de especies están sujetos a numerosos defectos, pero su desarrollo puede

ofrecer una interesante herramienta complementaria en la gestión del territorio.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO I

18

CHAPTER I: Predictive models of species distribution:

a review on their limitations

SUMMARY

In the last two decades there has been a growing interest in modelling wildlife-habitat

relationships. This is due both to the necessity of basic information for land

management and conservation, and to the fundamental interest of Ecology in studying

the distribution and abundance of organisms. However, wildlife-habitat models relay on

several asumptions, and have some limitations that must be known. This work offers

first a brief summary of the type of models that can be found in ecological studies. The

focus is on monoespecific models of species distribution, that is, in those that relate

habitat characteristics with the presence/absence of a single species, but the discussion

can be extended to other model types, in particular those which deal with several

species at a time. Second, a commented synthesis on both statistical and biological

limitations of the distribution models is given in detail, with an analysis of the

underlying assumptions and methodological problems. In conclusion, distribution

models have numerous shortcomings but their development may provide a worthy tool

for land management.


19

INTRODUCCIÓN

En las últimas dos décadas se hadespertado un enorme interés en elanálisis de la relación entre las especiesy sus hábitats, extendiéndose losestudios de selección de hábitat a larealización de modelos que predicen ladistribución y abundancia de especies.Esta tendencia es un producto del dobleinterés de estos modelos. Por una parte,existe una fuerte demanda deinformación en numerosos problemasde conservación en los que lasrelaciones de las especies con sushábitats son primordiales y, por otra, laecología tiene un interés primario enestudiar la distribución y abundancia delos organismos, lo que algunos autoreshan identificado como su objetivoprincipal (Begon, Harper & Townsend1995, p.124). En consonancia con estasituación, recientemente han aparecidodiversos trabajos realizados en lapenínsula Ibérica que desarrollanmodelos de distribución de especies(González, Bustamante & Hiraldo 1990;González, Bustamante & Hiraldo 1992;Donázar, Hiraldo & Bustamante 1993;Bustamante 1996; Bustamante 1997;Brito, Crespo & Paulo 1999; Sánchez-Zapata & Calvo 1999; Franco, Brito &Almeida 2000; Martínez Palao et al.2000; Suárez, Balbontín & Ferrer2000), y es de esperar que su númerosiga aumentando en un futuro próximo,dadas las perspectivas optimistas de suposible uso en la gestión del medionatural.

La utilidad general de los modelosde distribución de especies radica enque permiten trabajar con muestrasincompletas acerca de la distribución oabundancia de especies, lo que esespecialmente importante en losestudios en áreas remotas o de difícilacceso, donde no resulta práctico llegar

a la totalidad del territorio, o bien entrabajos en que los recursos seaninsuficientes para ello (Osborne & Tigar1992; Skov & Borchsenius 1997;Manel, Dias & Ormerod 1999). Losdatos recogidos en un muestreo seextienden al conjunto del área de interésmediante la generación de mapas decarácter predictivo (ver p.ej.Mladenoffet al. 1995) entre cuyos valoresprincipales se encuentra, en nuestraopinión, el que pueden ser unaherramienta útil para los gestores delterritorio. Finalmente, si las variablespredictoras pueden derivarse desensores remotos (fotografía aérea,imágenes de satélite) la informaciónproporcionada por sensores remotospodría servir para crear mapaspredictivos fácilmente actualizables(Palmeirim 1988; Avery & Haines-Young 1990; Miller & Conroy 1990;Andries, Gulinck & Herremans 1994;Paruelo & Golluscio 1994). Estosmodelos han sido utilizados paraevaluar las necesidades de protección enun territorio (Scott et al. 1993;Bojórquez-Tapia et al. 1995). Lamodelización por separado de un grannúmero de especies (o la modelizaciónde la riqueza) permite identificar áreasde distinto interés conservacionista,como pueden áreas ricas en especies oen táxones amenazados, para tenerlas encuenta en la creación de espaciosprotegidos. El ejemplo paradigmáticode esta aproximación es el análisis GAP(Scott et al. 1993), actualmente muydesarrollado en EEUU aunque no estáexento de críticas (Short & Hestbeck1995; Conroy & Noon 1996). Enparticular, no hay un acuerdo sobre siexiste una coincidencia geográfica de lariqueza, rareza o grado de amenazaentre diferentes táxones (verp.ej.Williams & Gaston 1994; Castro et


20

al. 1996; Prendergast 1997). Porúltimo, otra extendida aplicación de losmodelos de las relaciones entre lasespecies y sus hábitats es la predicciónde impactos, ya sean naturales, como enlos estudios que sugieren posiblesmodificaciones en la distribución deespecies relacionados con el cambioclimático o incendios (Box,Crumpacker & Hardin 1993; He &Mladenoff 1999); ya sean impactosartificiales, como los provocados porinfraestructuras, actividades extractivaso cambios en el uso del territorio(Avery & Haines-Young 1990; Lavers& Haines-Young 1996). En estos casoslos modelos son herramientas quepermiten decidir entre alternativas degestión del territorio (Turner et al.1995).

A pesar de las numerosasaplicaciones de los modelos de lasrelaciones entre las especies y sushábitats, éstos son una representaciónincompleta de la realidad y, por tanto,tienen limitaciones de las que deben serconscientes quienes los desarrollan yquienes los utilizan. En este trabajo seofrece primero un breve sumario de lostipos de modelado que puedenencontrarse en estudios de ecología,centrándose en los modelosmonoespecíficos de distribución, esdecir, en aquellos que relacionan lascaracterísticas del hábitat con lapresencia de una sola especie (aunque elrazonamiento puede extendersefácilmente a otro tipo de modelos, enparticular los multiespecíficos).Posteriormente, se presenta una síntesiscomentada de las limitaciones decarácter biológico y estadístico de losmodelos predictivos, analizando endetalle las presunciones en que sesostienen y los problemasmetodológicos que dificultan suaplicación.

Una versión resumida de estetrabajo fue presentada en el I CongresoIbérico de Ecología, celebrado enSantiago de Compostela (La Coruña)entre el 25 y 28 de septiembre de 2000y organizado por la AsociaciónEspañola de Ecología Terrestre y laSociedade Portuguesa de Ecología.

Tipos de modelos

En lo que sigue se consideraránsólo los modelos empíricos querelacionan la distribución de una solaespecie (es decir, su existencia y/oabundancia en un área, lo quedenominaremos variable respuesta) conun conjunto de variables del medio quedescriben aspectos bióticos, físicos ohumanos a través de una formulaciónmatemática o lógica (lo quedenominaremos variables predictoras).Como modelo empírico se entenderá eneste trabajo aquellos que se basan endatos reales, como es común en lamayor parte de los trabajos de campo,en oposición a los modelos teóricos(sensu Morrison, Marcot & Mannan1998) cuya formulación parte desupuestos de funcionamiento de unhipotético sistema u organismo. Portanto, no se tendrán en cuenta losmodelos multiespecíficos (p.ej. análisisGAP Scott et al. 1993), los que usancomo variables predictoras rasgosvitales de los organismos (Lawton1993), ni los que se basanexclusivamente en técnicas deinterpolación espacial (Cressie 1993;Maurer 1994). La discusión se centraráen modelos de tipo correlativo,definidos aquí como los que se basan encorrelaciones, no necesariamentecausales, entre variables, aunque granparte de lo que se expone puedeaplicarse a la mayoría de los modelos delas relaciones entre las especies y sushábitats (Morrison, Marcot & Mannan1998).


21

Los modelos que se más se hanusado en ecología para predecir ladistribución de especies podríandividirse en tres grandes gruposatendiendo a su funcionamiento: los queestiman el rango de toleranciaecológica, los modelos de tipocorrelativo y ordenación multivariante,y redes neuronales artificiales.

Dentro del primer tipo de modelosse encuentran los llamados de análisisde superposición (Brito, Crespo &Paulo 1999) y los de envuelta climática(Austin, Nicholls & Margules 1990;Box, Crumpacker & Hardin 1993), quepodrían considerarse extensiones deanálisis típicos de los Sistemas deInformación Geográfica. Elfuncionamiento de estos modelos escomo sigue. Primero se identifican loslugares en las que una especie estápresente y se calculan los valoresmínimos y máximos de las variablesambientales que se considera a priorique afectan a su distribución(generalmente variables descriptoras delclima, la altitud, etc.). La extensión delos resultados al conjunto del área deestudio se hace suponiendo que loslugares adecuados para la especie sonaquellos cuyos valores de todas lasvariables predictoras estén dentro de losrangos en los que se la ha observado.Estos modelos dependen de unaselección adecuada de las variablesambientales y normalmente tienden asobreestimar la extensión areal ocupada(pueden subestimarla si se seleccionandemasiadas variables de escasarelevancia para la especie). La utilidadque se les suele reconocer es la deaportar un primer análisis orientativo,que es particularmente valioso en áreasextensas o escasamente prospectadas(Skov & Borchsenius 1997).

Al segundo grupo pertenece unagran variedad de modelos cuyo patrón

común es que tratan de relacionar lapresencia o la abundancia de unaespecie con distintas variablespredictoras a través de una funciónmatemática. Esta función permiteestablecer el tipo de relación que existeentre la variable respuesta y laspredictoras. En general, el uso detécnicas como análisis discriminante(González, Bustamante & Hiraldo1990; González, Bustamante & Hiraldo1992) y regresiones lineares múltiples(Donázar, Ceballos & Fernández 1989;Carrascal, Bautista & Lázaro 1993) hanido dejando paso a otras enmarcadasdentro de los Modelos LinearesGeneralizados ("Generalized LinearModels" o GLM, de las que lasanteriores pueden considerarse casosparticulares) pues permiten una mayorflexibilidad al tratar los datos (Nicholls1989; Austin, Nicholls & Margules1990; Donázar, Hiraldo & Bustamante1993; Bustamante et al. 1997). Unamención especial merece la regresiónlogística, que es la técnica más usadapues utiliza variables binomiales (p.ej.aquellas cuyas respuestas son 1 y 0)fácilmente entendibles en el contexto deanálisis de presencia/ausencia. Encuanto a las técnicas de ordenaciónmultivariante, éstas son utilizadasgeneralmente como paso previo a lamodelización (Carrascal, Bautista &Lázaro 1993) para resumir un conjuntonumeroso de variables en unas pocasvariables sintéticas, pero pueden usarsepor sí mismas para crear mapaspredictivos de la distribución (análisisfactorial del nicho ecológico,Hausser1995; Hirzel, Hausser & Perrin 2000).Por último, existen técnicas de ajuste noparamétrico como regresiones locales omodelos aditivos generalizados(Generalized Additive Models oGAM,Hastie & Tibshirani 1990) que sehan utilizado para aspectos similares alos que aquí se tratan y son el horizontehacia el que probablemente irántendiendo los próximos trabajos


22

(Thomas & Neil 1991; Fewster et al.2000).

Un último tipo de modelos, que sesepara en esta revisión por suspeculiaridades, es el de redes neuronalesartificiales, denominadas así porqueestán basados en un modelo conceptualdel funcionamiento del cerebro. En estecaso los efectos de las distintasvariables predictoras sobre la respuestase transforman y se combinan en gruposdenominados neuronas cuyo número sedetermina subjetivamente. Estascombinaciones y el peso relativo decada neurona en la respuesta final semodifican iterativamente (a través delos denominados "algoritmos deentrenamiento") hasta dar con un ajustea los datos que se considere apropiado(StatSoft 1999). Se argumenta en favorde su utilización que pueden modelarrelaciones no lineares muy complejas(Lek et al. 1996) y en su contra que nopermiten reconocer fácilmente posiblesrelaciones causales entre los predictoresy la respuesta (originan modelos de tipo"caja negra") y que requieren mayortiempo de computación (Manel, Dias &Ormerod 1999). El modelado de ladistribución de especies a través deredes neuronales se ha emprendidorecientemente y sus ejemplos sontodavía escasos (Mastrorillo et al.1997; Manel, Dias & Ormerod 1999).

LIMITACIONES DE LOS MODELOS

Presunciones

El modelado de la abundancia opresencia de especies en función devariables del hábitat reposa sobre dospresunciones básicas: (i) que la variablerespuesta es independiente entrelocalidades y (ii) que todas las variablespredictoras importantes se incluyen enel modelo (Lennon 1999). Puesto quehabitualmente no se sabe a priori cuálesson las importantes, es necesario hacer

una selección a través de testsestadísticos de manera que el modelofinal retiene sólo las variables que seconsideran significativas, de acuerdocon el principio de parsimonia por elque se prefieren modelos sencillos conpocas variables a otros más complejosque expliquen lo mismo. Sin embargo,la primera de las presunciones esprobablemente falsa en la mayoría delos trabajos ya que las condicionesambientales en un punto de estudiotenderán a ser similares en un áreapróxima y, por tanto, las especiesligadas a tales condiciones tenderán apresentarse también en los puntosvecinos. Además, no es raro que lasespecies aparezcan distribuidas deforma agregada puesto que losindividuos establecidos en un áreapueden ejercer un efecto de atracciónhacia nuevos colonizadores ocondicionar la dispersión de losdescendientes, de forma que laprobabilidad de encontrar a una especieen un lugar podría no ser independientede la probabilidad de encontrarla enlugares vecinos (Legendre &Troussellier 1988; Augustin,Mugglestone & Buckland 1996). Estosdos aspectos originan lo que se conocecomo autocorrelación espacial de lavariable respuesta. Cuando esta existe,los tests estadísticos que seleccionan lasvariables predictoras tienden aincorporar en los modelos aquellasvariables que cambien espacialmente deuna forma gradual, lo cual impide haceruna interpretación biológica del modeloy perjudica su capacidad de ser aplicadoa otros lugares, aunque puede resultarconveniente si el objetivo es explicaruna distribución en un área determinada(Augustin, Mugglestone & Buckland1996; Lennon 1999). El método máscomúnmente utilizado para tener encuenta la autocorrelación espacial sebasa en incorporar a los modelos una odistintas variables predictoras queinformen del estado de la variable


23

respuesta en un área vecina cuyaextensión se decide empíricamente. Portanto, en estos modelos, denominadosautologísticos por ser un refinamientode la regresión logística, las relacionesde vecindad se analizan como unavariable predictora más (Preisler 1993;Smith 1994; Augustin, Mugglestone &Buckland 1996; Chou & Soret 1996;Wu & Huffer 1997).

Por otro lado, los modelos de ladistribución de especies asumenimplícitamente que los hábitats estánsaturados, es decir, que todo hábitatadecuado para una especie estaráocupado por ella. Sin embargo, si unorganismo cuya selección de hábitat sepretende modelar muestra una dinámicapoblacional en la que hay efectos defuente-sumidero, un área adecuadapodría estar vacía si aún no hubiera sidocolonizada o si la población existente sehubiera extinguido por causas naturaleso provocadas por el hombre (Días1996). Además, puede haberinteracciones entre especies (p.ej.,predación o competencia) que haganque un hábitat en otro caso adecuado noesté ocupado (Lawton & Woodroffe1991). Esta situación origina los tiposde error por comisión, en los que sepredice erróneamente la presencia deuna especie en un lugar (Fielding &Bell 1997).

Otra presunción fundamental es quela probabilidad de detección de unaespecie será mayor en sus hábitatsóptimos. En los modelos de distribuciónse mide generalmente la abundancia deuna especie en distintas áreascaracterizadas por un conjunto devariables y se equipara la abundanciacon la calidad del hábitat para esaespecie. No obstante, existen ciertosprocesos naturales que hacen que estapresunción pueda ser falsa en algunoscasos (Van Horne 1983). Por un lado, ladistribución actual de una especie

podría reflejar situaciones pasadas, siexisten cambios en la densidad de losindividuos de frecuencia plurianual quesigan variaciones a escala local enfactores que influyan en la demografía,como la intensidad de depredación o lacantidad de alimento. Por otro lado, enpoblaciones animales que desarrollenjerarquía social, los individuosdesfavorecidos (subadultos inexpertos,ejemplares enfermos, etc...) pueden serdesplazados a ambientes subóptimos enlos que podrían adquirir granabundancia (Días 1996). Finalmente, losíndices de selección de hábitat puedenestar afectados por el tamañopoblacional de manera que, porejemplo, un hábitat de elevados recursospodría usarse mucho hasta que lapoblación creciera tanto que la presiónde competencia intraespecíficacondujera a ocupar hábitats subóptimoscon menor competencia (Hobbs &Hanley 1990). En esta línea derazonamiento se ha destacado que unapoblación más numerosa no significaque esté en mejores condiciones; así seha mostrado (Hobbs & Swift 1985) queun área con abundantes recursos de bajacalidad puede mantener a una granpoblación infraalimentada, mientras queotro área de escasos recursos de altacalidad soporta a pocos individuos de,probablemente, mayor eficaciabiológica (“fitness”). Por último, existensimulaciones en las que se recrea unhábitat fragmentado y revelan que eltamaño medio poblacional de losfragmentos está influido principalmentepor la dispersión de individuos entreellos y no por su capacidad de carga(Fahrig & Paloheimo 1988). Por estosmotivos se ha propuesto que laadecuación de un hábitat se midabaremando la abundancia de lasespecies con la eficacia biológica de losindividuos que lo ocupan (Van Horne1983).


24

Si los modelos trabajan sobre unmuestreo de áreas disponibles ((ounidades de recursos disponibles en laterminología deManly, McDonald &Thomas 1993), se asume que éstas hansido escogidas independientemente y alazar, y que todos los individuos tienenlas mismas probabilidades de acceso aellas (Boyce & McDonald 1999).Además, el significado que en cada casose dé al concepto de "disponibilidad" seha identificado también como unproblema importante en estudios deselección de hábitat (Mac Clean et al.1998; Wilson, Shackleton & Campbell1998) que cabe extender al contexto demodelos de ditribución. Así, en unmodelo de regresión logística típico secomparan las variables predictoras deun conjunto de localidades en el que seha observado a una especie con otro enla que se la supone ausente, de maneraque, a efectos analíticos, el áreadisponible es la suma de las áreas de losdos conjuntos de localidades: a mayorárea en que se midieron los predictoresen torno a cada localidad, mayor áreadisponible y, según los trabajosanteriores, mayor probabilidad decometer error de tipo I al incorporarvariables espúreas a los modelos.

Limitaciones

Unas de carácter biológico ...

Existen varias razones para esperarque los modelos no funcionencorrectamente. Las dos primeras que sesubrayan aquí son producidas por elpropio fenómeno que se quiere modelar.

Por un lado cabría esperar que,hasta cierto punto, la distribución actualde una especie estuviera afectada poracontecimientos pasados (Días 1996;Fielding & Bell 1997) lo cual podríaser especialmente relevante enorganismos sésiles de larga vida, comomuchos táxones vegetales. Así porejemplo, una especie podría habitar unárea que colonizó hace tiempo y quehoy en día carece de las condicionesque le son más favorables; tal especiepodría estar sufriendo un lento decliveen esa zona pero los modelos (quegeneralmente se desarrollan en unintervalo de tiempo breve) no lodetectarían.

.

ObservadoPresencia Ausencia

Presencia a bPredichoAusencia c d

Figura I. Matriz de confusión. a: presencias predichascorrectamente; b: falsos positivos; c: falsos negativos; d:ausencias predichas correctamente.

Figure I. Confusion matrix. a: presences correctly predicted; b:false positives; c: false negatives; d: abscences correctlypredicted


25

Por otro lado, clases distintas deindividuos de una población podríanmostrar una selección de hábitatdiferente, dependiendo por ejemplo desus estatus social (pero no por unasimple expulsión de individuossubordinados a hábitats subóptimos sinopor una selección activa de hábitatsdiferentes con distinta oferta derecursos, ver Ardia & Bildstein 1997)Esta situación se ha encontrado en avesrapaces (Cade 1955; Koplin 1973;Smallwood 1987; Bustamante et al.1997) y podría ser típica de organismosanimales que exhiban una jerarquíasocial cuyos distintos grupos fuerandimórficos (p.ej., en las rapaces diurnaslas hembras suelen ser mayores que losmachos y los tamaños de las presas queles suponen un beneficio óptimo sondiferentes).

Por último, los modelos dedistribución se limitan implícitamente apoblaciones en equilibrio cuya relacióncon el hábitat no cambia (Boyce &McDonald 1999), de otra manera seríanecesario hacer un modelo para cadasituación (Arthur et al. 1996).

... y otras de corte metodológico.

Existen además varios problemasmetodológicos en el tipo demodelización que aquí se trata queimpiden que los modelos sean perfectos.Estos se refieren a la comparación demodelos mediante medidas de error dela predicción, a la conversión deprobabilidades dadas por los modelos avalores de presencia o ausencia y a lanaturaleza correlativa de las relacionesque se establecen entre la variablerespuesta y las predictoras.

Las predicciones de los modelos dela distribución que tratan con datos depresencia/ausencia se analizan con unamatriz de confusión (figura I) y puedenestar erradas de dos formas: en laspresencias (falsos positivos) y en lasausencias (falsos negativos). Lasdistintas medidas de error (tabla I)tienen características diferentes y, enparticular, algunas están influidas por laprevalencia. Por ejemplo, suponiendoque se realizó un muestreo de 100lugares y en sólo 10 apareció la especieobjeto del estudio (N=100, a+c=10,b+d=90), un modelo trivial seríasuponer que ninguno de los lugares esapto para ella (a+b=0, c+d=100) lo quedaría una tasa de clasificación correctadel 90% ((a+d)/N=0+90/100, ver tablaI).

Además, los modelos dangeneralmente valores continuos para laspredicciones dentro del intervalo (0,1)pero los valores de la matriz deconfusión que se utiliza paracompararlos son valores discretos 0 ó 1.Esto hace que los valores deprobabilidad de aparición hayan de ser

convertidos, de manera que seadjudique la presencia de la especie atodas las áreas cuya probabilidad deaparición supere un umbral. Elproblema reside en la elección de estepunto umbral, al que son sensibles lasmedidas de error (ver figura I yBrito,Crespo & Paulo 1999; Franco, Brito &Almeida 2000, para ejemplos reales).La elección de 0,5 como umbral en elejemplo de la figura II conduciría a unbajo poder predictivo positivo(aproximadamente la mitad de laspresencias predichas serían reales). Laadopción de un umbral más bajo, porejemplo 0,3, aumentaría el poderpredictivo para las presencias (hasta0,85), mientras que un umbral mayor,como 0,8, aseguraría un elevado poderpredictivo para las ausencias (cerca del0,9). Cada estrategia se adecuaría adistintos escenarios, por ejemplo, laprimera en la selección de espacios quealbergaran a una especie a proteger y lasegunda en la selección de áreasalternativas para la ubicación deactividades humanas de gran impactopara cierta especie.


26

Medida CálculoPrevalencia (a+c)/N

¿en qué fracción de puntos ha aparecido una especie?Tasa de clasificación correcta total (a+d)/N

¿qué fracción de puntos se predijo correctamente?Tasa de clasificación incorrecta (b+c)/N

¿qué fracción de puntos se predijo incorrectamente?Sensibilidad a/(a+c)

¿qué fracción de las presencias se predijeroncorrectamente?

Especificidad d/(b+d)¿qué fracción de ausencias se predijeron correctamente?

Poder predictivo positivo a/(a+b)de las presencias predichas ¿qué fracción es correcta?

Poder predictivo negativo d/(c+d)de las ausencias predichas ¿qué fracción es correcta?

Kappa¿qué fracción de puntos se predijo correctamente

teniendo en cuenta la prevalencia?[(a+d)-(((a+c)(a+b)+(b+d)(c+d))/N] / [N-(((a+c)(a+b)+(b+d)(c+d))/N)]

Tabla I. Medidas de error (modificado de Fielding & Bell 1997) y preguntas a las que responden.

Table I. Error measures (modified fromFielding & Bell 1997) and questions that they address.

Se ha descrito un método decomparación de modelos que evita elproblema de la influencia del punto decorte sobre las medidas de error. Setrata de los diagramas ROC (de"Receiver Operating Characteristic",Zweig & Campbell 1993) en los que serepresenta la sensibilidad de un modelo,en ordenadas, contra su especificidad,en abcisas, para todos los puntos umbralde forma que el modelo que esté porencima en el diagrama tendrá unamayor exactitud relativa. Los diagramasROC no informan de cuál es el puntoumbral óptimo, pero existen métodospor los que pueden ser utilizados paraconseguir esta información (referenciasenFielding & Bell 1997). Sin embargo,existen pocos ejemplos prácticos enecología del uso de los diagramas ROC(Manel et al. 1999).

Como corolario cabe decir quetanto la elección de las medidas de error

con las que se comparen los modeloscomo, si procede, la elección del puntode corte han de ser escogidos conespecial atención a las preguntas másrelevantes en el contexto de lainvestigación que se esté realizando(Fielding & Bell 1997; Morrison,Marcot & Mannan 1998).

Las relaciones entre las variablesrespuesta y explicativas que se modelansuelen tener una naturaleza correlativapor lo que no revelan necesariamentepautas de causa y efecto. Esto hace quelos modelos de distribución puedanfracasar en su aplicación a otras áreas (otiempos). La solución sería utilizarvariables predictoras causales en losmodelos, pero esto excede nuestroconocimiento actual sobre la mayorparte de las especies. Además, eldesarrollo de modelos causales


27

Figura II. Ejemplo de diagrama de la relación entre la tasa de clasificación correcta (cuadrados para laspresencias, triángulos para las ausencias y círculos para el total) y el punto umbral para convertir losvalores de probabilidad en valores de presencia/ausencia.

Figure II. Example of plot of the relation between correct classification rate (squares for presences,triangles for absences and circles for the overall) and the threshold used to convert probability values onpresence/absence.

probablemente exigiría más tiempo delque permiten las necesidades deconservación que conducen a lamodelización (en particular, ¿cómo seidentifican a que variables del mediorealmente responde una especie?; y unavez hecho esto, ¿cómo conocer ladistribución espacial de esas variablespredictoras, pues la cartografíadisponible se ha creado con otropropósito?). Sin embargo, algunosautores han sugerido que la únicasolución posible reside en el uso devariables causales, pues laconcatenación de presunciones podríaconducir necesariamente a un bajo

poder predictivo (Beutel, Beeton &Baxter 1999).

Por otro lado, los modelos suelendesarrollarse en un contextomultivariante, donde las correlacionesentre las variables (colinearidad) sonmuy probables. La multicolinearidadhace que puedan incorporarse a losmodelos variables espúreas, y quequeden fuera otras más próximas a lacausales (Flack & Chang 1987). Denuevo, este problema reduce lacapacidad de extrapolación de losmodelos (es decir, la fiabilidad con quepueden ser aplicados en otras áreas

0,00

0,10

0,20

0,30

0,40

0,50

0,60

0,70

0,80

0,90

1,00

0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0

Valores umbral

Pro

po

rció

n d

e cl

asif

icac

ión

co

rrec

ta

Alto poderpredictivo positivo

Bajo poderpredictivo negativo

Alto poderpredictivo negativo

Bajo poderpredictivo positivo


28

distintas a aquellas en que se generaron)y hace que las distintas técnicas demodelado produzcan resultadosdiferentes (Mac Nally 2000). Unasolución interesante a este problema escalcular todos los modelos posibles conlas variables de que se dispone yescoger entre ellos según criterios quepromedian la información recogida porel modelo y su complejidad (Akaike1978; Schwarz 1978). Se ha propuestoademás que la incertidumbre asociada ala selección de los modelos, es decir a laelección tanto de las variablespredictoras como de la forma en queestas varían con la variable respuesta, setenga en cuenta en las predicciones,ponderando los resultados procedentesde distintos modelos (Buckland,Burnham & Augustin 1997). Sinembargo, estas aproximaciones no sonmuy frecuentes en ecología (Toner &Keddy 1997). Finalmente, un métododiseñado para identificar variablescausales--pero no para generar modelospredictivos--reduciendo los problemasde multicolinearidad es la particiónjerárquica (Chevan & Sutherland 1991;Christensen 1992), mediante la cual secalcula la influencia relativa de unavariable en todos los modelos en los queaparece. Sus resultados puedencompararse con los modelosseleccionados mediante un criterio deinformación para ilustrar su grado decausalidad (Mac Nally 2000).

Cabe hacer un último comentarioacerca de la posibilidad de validar losmodelos de distribución. La verificaciónde modelos numéricos de sistemasnaturales, si se entiende como tal lademostración de su certeza, esimposible porque tales sistemas no soncerrados y los resultados no sonsingulares (es decir, varios modelospueden originar los mismos resultados).Según Oreskes et al. (1994) el términovalidación se emplea con dossignificados erróneos; el primero es el

de que las predicciones son consistentescon las observaciones, y el segundo elde que el modelo refleja con precisiónla realidad. Estos autores afirman quelos modelos sólo se pueden confirmar,entendiendo este término como lacomprobación de que las observacionescoinciden con las predicciones; ysubrayan que la confirmación nodemuestra la hipótesis (el modelo), sóloapoya su probabilidad (Oreskes,Shrader-Frechette & Belitz 1994).

CONCLUSIÓN

Enfrentados a las limitaciones quese han expuesto, y ante un éxito muyvariable al extrapolar modelos entrezonas geográficas distintas, algunosautores han recomendado cautela en laaplicación de los modelos a problemasde conservación, llegando a sugerir quela distribución de las especies podría serimpredecible (Fielding & Haworth1995).

Sin embargo, los modelos de larelación especies-hábitat han mostradosu utilidad en distintas áreas yproporcionan una herramienta de, almenos, interés heurístico. Laincorporación de variables causales y laatención a medidas de eficacia biológicaaumentarían probablemente su valor. Encualquier caso merece la pena citar aquía Morrison (1999, p.313) que da unaperspectiva práctica a la utilización demodelos "Models are like politicians:support them, use them, but don`tunquestioningly trust them.".

AGRADECIMIENTOSEste trabajo es una contribución al

proyecto "Cartografía predictiva de ladistribución de aves terrestres: unestudio piloto en AndalucíaOccidental", financiado por la DirecciónGeneral de Enseñanza Superior eInvestigación Científica (Ministerio deCiencia y Tecnología) y fondos FEDER


29

de la UE (proyecto # 1DF-97-0648).J.S. disfruta de una beca predoctoral delMinisterio de Educación y Cultura. Losautores desean agradecer loscomentarios de dos revisores anónimosque contribuyeron a mejorar el trabajo.

NOTASEste trabajo ha aparecido publicado

a principios de 2002 en Ecología, 15:9-21 (2001), con Javier Bustamante comocoautor.

REFERENCIAS BI BLIOGRÁFICAS

Akaike, H. (1978). A Bayesian analysisof the minimum AIC procedure.Annals Institute StatisticsMathematics, 30, 9-14.

Andries, A. M., Gulinck, H. &Herremans, M. (1994). Spatialmodelling of the barn owl Tyto albahabitat using landscapecharacteristics derived from SPOTdata. Ecography, 17, 278-287.

Ardia, D. R. & Bildstein, K. L. (1997).Sex-related differences in habitatselection in wintering Americankestrels Falco sparverius. AnimalBehaviour, 53, 1305-1311.

Arthur, S. M., Manly, B. F. J.,McDonald, L. & Garner, G. W.(1996). Assessing habitat selectionwhen availability changes. Ecology,77(1), 215-227.

Augustin, N. H., Mugglestone, M. A. &Buckland, S. T. (1996). Anautologistic model for the spatialdistribution of wildlife. Journal ofApplied Ecology, 33, 339-347.

Austin, M. P., Nicholls, A. O. &Margules, C. R. (1990).Measurement of the realizedqualitative niche: environmentalniches of five Eucalyptus species.Environmental Management, 60(2),161-177.

Avery, M. I. & Haines-Young, R. H.(1990). Population estimates for thedunlin Calidris alpina derived from

remotely sensed satellite imagery ofthe Flow Country of northernScotland. Nature, 344, 860-862.

Begon, M., Harper, J. L. & Townsend,C. R. (1995). Ecología: individuos,poblaciones y comunidades.Omega,Barcelona

Beutel, T. S., Beeton, R. J. S. & Baxter,G. S. (1999). Building betterwildlife-habitat models. Ecography,22, 219-223.

Bojórquez-Tapia, L. A., Azuara, I.,Ezcurra, E. & Flores-Villela, O.(1995). Identifying conservationpriorities in Mexico throughgeographic information systems andmodeling. Ecological Applications,5(1), 213-231.

Box, E. O., Crumpacker, D. W. &Hardin, E. D. (1993). A climaticmodel for location of plant species inFlorida, U.S.A. Journal ofBiogeography, 20, 629-644.

Boyce, M. & McDonald, L. L. (1999).Relating populations to habitatsusing resource selection functions.Trends in Ecology and Evolution,14(7), 268-272.

Brito, J. C., Crespo, E. G. & Paulo, O.S. (1999). Modelling wildlifedistributions: logistic multipleregression vs overlap analysis.Ecography, 22, 251-260.

Buckland, S. T., Burnham, K. P. &Augustin, N. H. (1997). Modelselection: an integral part ofinference. Biometrics, 53, 603-618.

Bustamante, J. (1996). Statistical modelof nest-site selection for the BeardedVulture (Gypaetus barbatus) in thePyrenees and evaluation of thehabitat available with a geographicalinformation system. Biología yConservación de las RapacesMediterráneas, 1994. (eds J. y. M.Muntaner, J.), pp. 393-400. SEO, .Madrid .

Bustamante, J. (1997). Predictivemodels for lesser kestrel Falconaumanni distribution, abundance


30

and extinction in southern Spain.Biological Conservation, 80, 153-160.

Bustamante, J., Donázar, J. A., Hiraldo,F., Ceballos, O. & Travaini, A.(1997). Differential habitat selectionby immature and adult Grey Eagle-buzzards Geranoaetus melanoleucus.Ibis, 139, 322-330.

Cade, T. (1955). Experiments on winterterritoriality of the American kestrel,Falco sparverius. The WilsonBulletin, 67, 5-17.

Carrascal, L. M., Bautista, L. M. &Lázaro, E. (1993). Geographicalvariation in the density of the Whitestork (Ciconia ciconia) in Spain:influence of habitat structure andclimate. Biological Conservation, 65,83-87.

Castro, I., Moreno, J. C., Humphries, C.J. & Williams, P. H. (1996).Strengthening the Natural andNational Park system of Iberia toconserve vascular plants. BotanicalJournal of the Linnean Society, 121,189-206.

Chevan, A. & Sutherland, M. (1991).Hierarchical partitioning. TheAmerican Statistician, 45, 90-96.

Chou, Y.-H. & Soret, S. (1996).Neighborhood effects in birddistributions, Navarre, Spain.Environmental Management, 20(5),675-687.

Christensen, R. (1992). Comment onChevan and Sutherland. TheAmerican Statistician, 46, 74.

Conroy, M. J. & Noon, B. R. (1996).Mapping of species richness forconservation of biological diversity:conceptual and methodologicalissues. Ecological Applications, 6(3),763-773.

Cressie, N. A. C. (1993). Statistics forspatial data.John Wiley & Sons,New York

Días, P. C. (1996). Sources and sinks inpopulation biology. Trends inEcology and Evolution, 11, 326-330.

Donázar, J. A., Ceballos, O. &Fernández, C. (1989). Factorsinfluencing the distribution andabundance of seven cliff-nestingraptors: a multivariate study. Raptorsin the modern world. (eds B.-U.Meyburg & R. D. Chancelor), pp.545-552. WWGBP, . Berlin, London& Paris .

Donázar, J. A., Hiraldo, F. &Bustamante, J. (1993). Factorsinfluencing nest site selection,breeding density and breedingsuccess in the bearded vulture(Gypaetus barbatus). Journal ofApplied Ecology, 30, 504-514.

Fahrig, L. & Paloheimo, J. (1988).Determinants of local population sizein patchy habitats. TheoreticalPopulation Biology, 34, 194-213.

Fewster, R. M., Buckland, S. T.,Siriwardena, G. M., Baillie, S. R. &Wilson, J. D. (2000). Analysis ofpopulation trends for farmland birdsusing generalized additive models.Ecology, 81(7), 1970-1984.

Fielding, A. H. & Bell, J. F. (1997). Areview of methods for the assessmentof prediction errors in conservationpresence/absence models.Environmental Conservation, 24, 38-49.

Fielding, A. H. & Haworth, P. F.(1995). Testing the generality ofbird-habitat models. ConservationBiology, 9, 1466-1481.

Flack, V. F. & Chang, P. C. (1987).Frequency of selecting noisevariables in subset regressionanalysis: a simulation study. TheAmerican Statistician, 41, 84-86.

Franco, A. M. A., Brito, J. C. &Almeida, J. (2000). Modellinghabitat selection of Common CranesGrus grus wintering in Portugalusing multiple logistic regression.Ibis, 142, 351-358.

González, L. M., Bustamante, J. &Hiraldo, F. (1990). Factorsinfluencing the present distribution


31

of the Spanish imperial eagle Aquilaadalberti. Biological Conservation,51, 311-319.

González, L. M., Bustamante, J. &Hiraldo, F. (1992). Nesting habitatselection by the Spanish imperialeagle Aquila adalberti. BiologicalConservation, 59, 45-50.

Hastie, T. J. & Tibshirani, R. J. (1990).Generalized AdditiveModels.Chapman & Hall, London

Hausser, J. (1995). Introduction.Mammifères de la Suisse. (eds , pp.1-14. Birkhäuser, . Basel .

He, H. S. & Mladenoff, D. J. (1999).Spatially explicit and stochasticsimulation of forest-landscape firedisturbance and succession. Ecology,80(1), 81-99.

Hirzel, A. H., Hausser, J. & Perrin, N.(2000). Biomapper. Lausanne,Laboratory for ConservationBiology, University of Lausanne.

Hobbs, N. T. & Hanley, T. A. (1990).Habitat evaluation: douse/availability data reflect carryingcapacity? Journal of WildlifeManagement, 54(4), 515-522.

Hobbs, N. T. & Swift, D. M. (1985).Estimates of habitat carryingcapacity incorporating explicitnutritional constraints. Journal ofWildlife Management, 49, 814-822.

Koplin, J. R. (1973). Differential habitatuse by sexes of American kestrelswintering in northern California.Raptor Research, 7, 39-42.

Lavers, C. P. & Haines-Young, R. H.(1996). Using models of birdabundance to predict the impact ofcurrent land-use and conservationpolicies in the Flow Country ofCaithness and Sutherland, northernScotland. Biological Conservation,75, 71-77.

Lawton, J. H. (1993). Range, populationabundance and conservation. Trendsin Ecology and Evolution, 8(11),409-413.

Lawton, J. H. & Woodroffe, G. L.(1991). Habitat and distribution ofwater voles: why are there gaps in aspecies´ range? Journal of AnimalEcology, 60, 79-91.

Legendre, P. & Troussellier, M. (1988).Aquatic heterotrophic bacteria:modelling in the presence of spatialautocorrelation. Limnology andOceanography, 33(5), 1055-1067.

Lek, S., Delacoste, M., Baran, P.,Dimopoulos, I., Lauga, J. &Aulanier, S. (1996). Application ofneural networks to modellingnonlinear relationships in ecology.Ecological Modelling, 90, 39-52.

Lennon, J. J. (1999). Resource selectionfunctions: taking space seriously?Trends in Ecology and Evolution,14(10), 399-400.

Mac Clean, S., Rumble, M. A., King, R.D. & Baker, W. L. (1998).Evaluation of resource selectionmethods with different definitions ofavailability. Journal of WildlifeManagement, 62(2), 793-801.

Mac Nally, R. (2000). Regression andmodel-building in conservationbiology, biogeography and ecology:The distinction between -andreconciliation of- 'predictive' and'explanatory' models. Biodiversityand Conservation, 9, 655-671.

Manel, S., Dias, J. M., Buckton, S. T. &Ormerod, S. J. (1999). Alternativemethods for predicting speciesdistribution: an illustration withHimalayan river birds. Journal ofApplied Ecology, 36, 734-747.

Manel, S., Dias, J. M. & Ormerod, S. J.(1999). Comparing discriminantanalysis, neural networks and logisticregression for predicting speciesdistributions: a case study with aHimalayan river bird. EcologicalModelling, 120, 337-347.

Manly, B., McDonald, L. & Thomas, D.(1993). Resource selection byanimals. Statistical design and


32

analysis for field studies.Chapman &Hall, London

Martínez Palao, M., Giménez, A.,Martínez, J., Esteve, M. A., Anadón,J. D. & Pérez, I. (2000).Construcción y validación de unmodelo de distribución de tortugamora (Testudo graeca graeca) en laregión de Murcia basado en surespuesta a factores climáticos,Conferencia impartida en el ICongreso Ibérico de Ecología.

Mastrorillo, S., Lek, S., Dauba, F. &Belaud, A. (1997). The use ofartificial neural networks to predictthe presence of small-bodied fish in ariver. Freshwater Biology, 38, 237-246.

Maurer, B. A. (1994). Geographicalpopulation analysis: tools for theanalysis of biodiversity.BlackwellScientific Publications, Oxford

Miller, K. V. & Conroy, M. J. (1990).Spot satellite imagery for mappingKirtland's warbler wintering habitatin the Bahamas. Wildlife SocietyBulletin, 18, 252-257.

Mladenoff, D. J., Sickley, T. A., Haight,R. G. & Wydeven, A. P. (1995). Aregional landscape analysis andprediction of favorable Gray Wolfhabitat in the Northern Great Lakesregion. Conservation Biology, 9(2),279-294.

Morrison, M. L. & Hall, L. S. (1999).Standard terminology: toward acommon language to advanceecological understanding andapplications.

Morrison, M. L., Marcot, B. G. &Mannan, R. W. (1998). Wildlife-habitat relationships. Concepts andapplications.The University ofWisconsin Press, Madison

Nicholls, A. O. (1989). How to makebiological surveys go further withgeneralised linear models. BiologicalConservation, 50, 51-75.

Oreskes, N., Shrader-Frechette, K. &Belitz, K. (1994). Verification,

validation, and confirmation ofnumerical models in the earthsciences. Science, 263, 641-646.

Osborne, P. E. & Tigar, B. J. (1992).Interpreting bird atlas data usinglogistic models: an example fromLesotho, Southern Africa. Journal ofApplied Ecology, 29, 55-62.

Palmeirim, J. M. (1988). Automaticmapping of avian species habitatusing satellite imagery. Oikos, 52,59-68.

Paruelo, J. M. & Golluscio, R. A.(1994). Range assessment usingremote sensing in NorthwestPatagonia (Argentina). Journal ofRange Management, 47(6), 498-502.

Preisler, H. K. (1993). Modelling spatialpatterns of trees atacked by barkbeetles. Applied Statistics, 42, 501-514.

Prendergast, J. R. (1997). Speciesrichness covariance in higher taxa:empirical test of the biodiversityindicator concept. Ecography, 20,210-216.

Sánchez-Zapata, J. A. & Calvo, J. F.(1999). Raptor distribution inrelation to landscape composition insemi-arid Mediterranean habitats.Journal of Applied Ecology, 36, 254-262.

Schwarz, G. (1978). Estimating thedimension of a model. Annals ofStatistics, 6, 461-464.

Scott, J. M., Davis, F., Csuti, B., Noss,R., Butterfield, B., Groves, C.,Anderson, H., Caicco, S., D´erchia,F., Edwards, J., T.C., Ulliman, J. &Wright, R.G. (1993). GAP analysis:a geographic approach to protectionof biological diversity. WildlifeMonographs, 123, 1-41.

Short, H. L. & Hestbeck, J. B. (1995).National biotic resource inventoriesand GAP analysis: problems of scaleand unproven assumptions limit anational program. Bioscience, 45,535-539.


33

Skov, F. & Borchsenius, F. (1997).Predicting plant species distributionpatterns using simple climaticparameters: a case study ofEcuadorian palms. Ecography, 20,347-355.

Smallwood, J. A. (1987). Sexualsegregation by habitat in Americankestrels wintering in southcentralFlorida - vegetative structure andresponses to differential preyavailability. The Condor, 89, 842-849.

Smith, P. A. (1994). Autocorrelation inlogistic regression modelling ofspecies` distributions. GlobalEcology and Biogeography Letters,4, 47-61.

StatSoft, I. (1999). Electronic StatisticsTextbook.StatSoft, Tulsa, OK

Suárez, S., Balbontín, J. & Ferrer, M.(2000). Nesting habitat selection bybooted eagles Hieraaetus pennatusand implications for management.Journal of Applied Ecology, 37, 215-223.

Thomas, Y. & Neil, M. (1991).Generalized additive models in plantecology. Journal of VegetationScience, 2( ), 587-602.

Toner, M. & Keddy, P. (1997). Riverhydrology and riparian wetlands: apredictive model for ecological

assembly. Ecological Applications,7, 236-246.

Turner, M. G., Arthaud, G. J.,Engstrom, R. T., Hejl, S. J., Liu, J.,Loeb, S. & McKelvey, K. (1995).Usefulness of spatially explicitpopulation models in landmanagement. EcologicalApplications, 5(1), 12-16.

Van Horne, B. (1983). Density as amisleading indicator of habitatquality. Journal of WildlifeManagement, 47(4), 893-901.

Williams, P. H. & Gaston, K. J. (1994).Measuring more of biodiversity: canhigher-taxon richness predictwholesale species richness?Biological Conservation, 67, 211-217.

Wilson, S. F., Shackleton, D. M. &Campbell, K. L. (1998). Makinghabitat-availability estimatesspatially explicit. Wildlife SocietyBulletin, 26(3), 626-631.

Wu, H. & Huffer, F. W. (1997).Modelling the distribution of plantspecies using the autologisticregression model. Environmental andEcological Statistics, 4, 49-64.

Zweig, M. H. & Campbell, G. (1993).Receiver-operating characteristic(ROC) plots: a fundamentalevaluation tool in clinical medicine.Clinical Chemistry, 39, 561-577.

SECCIÓN SEGUNDA

Aspectos metodológicos: técnicas y estrategias del modelado de la

distribución de especies

Caminante no hay camino, se hace camino al andar.

—Antonio Machado

(en POESÍAS COMPLETAS. Residencia de

Estudiantes, Madrid. 1917)

Optimización de la duración de los muestreos de aves

35

CAPÍTULO II: El muestreo de la presencia/ausencia para construir modelos

predictivos: una aproximación

de optimalidad usando el teorema del valor marginal

RESUMEN

En este estudio damos una solución basada en el teorema del valor marginal al problema

de la asignación óptima de esfuerzo de muestreo (número de estaciones frente a tiempo

empleado en cada estación) para estudios que utilicen estaciones de escucha, teniendo

en cuenta el tiempo perdido por el observador en desplazarse entre estaciones. El trabajo

se centra en obtener datos de presencia/ausencia para una especie de interés que puedan

usarse para construir un modelo predictivo de su distribución. Las especies que son más

grandes, raras o habitan tipos de vegetación estructuralmente más complejos se

benefician de prospecciones proporcionalmente más largas en cada estación de

muestreo. Las especies comunes y pequeñas que habitan áreas abiertas no necesitan que

las prospecciones se prolonguen más de 5 minutos (en este tiempo una especie pratense

de 10 g que tuviera una frecuencia total del 60% se detectaría en un 93% de los puntos),

mientras que una especie más forestal, más grande o más rara precisaría de un tiempo de

conteo más largo (una especie de 100 g cuya frecuencia total fuera del 10% sólo se

detectaría en el 60% de los puntos durante los primeros 5 min). En este trabajo se

proporcionan modelos de la duración óptima de la prospección en cada estación

considerando varios tiempos de desplazamiento entre estaciones que pueden servir de

ayuda en el diseño del muestreo.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO II

36

CHAPTER II: Sampling bird presence/absence to build predictive models:

an optimality approach using the marginal value theorem

SUMMARY

In this study we offer a solution based in the marginal value theorem to the problem of

allocating sampling effort (number of stations versus time employed at each station) in

point survey sampling schemes, taking into account the time wasted travelling between

stations. We focus in obtaining presence/absence data for a bird species of interest that

can be used to build a predictive model of its distribution. Species that are larger, rarer,

or inhabit vegetation types that are structurally more complex benefit from

proportionally longer surveys at each station. Common and small species inhabiting

open areas do not need more than a 5-min survey (in this period a 10-g grassland

species with a total frequency of 60% would have been recorded as present in 93% of

the points in which it was present), while a larger and rarer forest species would benefit

from a longer survey (a 100-g forest species with a total frequency of 10% would only

be recorded as present in 60% of the points during the first 5 min). We provide models

for optimal survey duration for a variety of travelling times to serve as an aid in

sampling design.


37

INTRODUCCIÓN

Data on speciespresence/absence at point samplingstations are frequently used to buildpredictive models of bird distribution(Green, Osborne & Sears 1994; Bolger,Scott & Rotenberry 1997; Beard,Hengartner & Skelly 1999; Pearce &Ferrier 2001). Presence/absence data,although apparently with lessinformation content than point-countstations, in which all bird individualspresent at the census area are counted,have certain advantages for modelling:(1) Errors of presence/absence datafollow a binomial distribution whilebird counts rarely follow a Poissondistribution and need to be transformedfor modelling. (2) Presence/absencedata are not biased by double countingor by birds entering or leaving thecensus area as bird counts are. (3) Thereis less variability among observerswhen using presence/absence than whenbird counts are used (personalobservation). Empirical data show thatmodels developed withpresence/absence data to assess habitatsuitability tend to perform at least asgood as those developed with birddensity data (Pearce & Ferrier 2001).

One may think of recordedpresence/absence at a single pointsample station as an asymmetricallybiased estimate of true presence/absenceof the species. Recorded presencesindicate true presences of the species(apart from identification errors), whilerecorded absences may result becauseof actual absences or due to the speciespassing unnoticed to the observer. Withthis view in mind, one would have toremain a very long time at each samplestation so that recordedpresence/absence tended to coincidewith true presence/absence.This wouldbe specially true for rare and criptic

species. An alternative way is to thinkof recorded presence/absence instatistical terms. The probability that aspecies will be recorded as present at apoint sample station during a certaintime will be proportional to itsabundance in the area times itsdetectability to the observer. Ifdectability is similar or varies randomlyamong sampling stations, andabundance varies among habitats it willbe possible to fit an environmentalmodel predicting the probability ofrecording a presence and assume thatthose predicted probabilities will beproportional to abundances.

When obtaining field data tobuild these predictive models thereexists a trade-off between the number ofpoint sampling stations that can be doneby an observer in a day and the durationof the survey at each point (Gutzwiller1991; Gutzwiller 1993; Drapeau, Leduc& McNeil 1999). In general, theempirical work shows that most of thespecies are detected within the firstsmin of sampling, so when talking aboutpoint-count stations there seems to bean agreement by experts inrecommending short counts (5 to 10min) (Fuller & Langslow 1984; Hutto,Pletschet & Hendricks 1986; Jiménez2000). However, counts of shortduration (5-10 min) have thedisadvantage that most of the time couldbe wasted travelling between countingpoints and, besides this, a highproportion of species really present atthe point may not be detected (Drapeau,Leduc & McNeil 1999). Moreover,some of these studies suggest that ratesof detection are species-specific,therefore the optimal duration of thesurvey at a point may be affected bycharacteristics of the species thatinfluence the probability of detection.

In the context of obtainingpresence/absence data to model


38

particular bird species the longer thetime at the sample station the greater theprobability of detecting a rare or cripticspecies that inhabits the area. On theother hand, short censuses give theopportunity of exploring more places,which may be crucial when studyinglarge or heterogeneous areas, speciallyconsidering the need of a large sampleof point surveys for modelling (Harrell2001). In this context, a plot ofcumulated recorded presences (dividedby the total number of presences) at thesample station versus time would form acurve of cumulated relative frequency.The shape of this curve would show anincrease towards an asymptote with avalue equal to the mean frequency ofthe species in the study area, thusresembling a curve of diminishingreturns. We can apply then the marginalvalue theorem (Charnov 1976) to thetrade-off between number of surveystations and time spent at each station,reckoning the similarity between thecurve of cumulative energy gain in aplot where an animal gradually depletesa resource, and the curve of cumulatedrelative frequency in an area of studywhere an observer graduallyapproximates the mean frequency of aparticular species in the area. Accordingto the marginal value theorem, thequantity to be maximised is the rate ofenergy gain: E(t)/(t+τ), where E(t) is thecumulated energy up to time t, t is thetime spent feeding in a plot, and τ is thetime spent in travelling between plots.Equivalently, the analysis of optimalsurvey duration to samplepresence/absence can be performed bysubstitution of E(t) for RF(t), thecumulated relative frequency at time t,and t being time spent at the samplestation. Given a certain travelling time(τ), that we assume the observer canroughly fix in advance, the aim is tofind the value of t that maximised RF(t).Optimality theory predicts that the timespent in a plot should increase with

increasing average travelling timebetween plots (Mac Nair 1982).

The aim of our study is to findan optimal solution, by applying themarginal value theorem, to the trade-offbetween the number of survey stationsand time spent at each station to obtainpresence/absence data adequate forpredictive modelling of a particular birdspecies. We will take into considerationthe time wasted by the observertravelling between sampling stations,and biological and ecological factorsthat may affect the detectionprobability: species abundance, bodysize and habitat type. Our focus is onobtaining comparable samples so theobserver will have to remain a fixedtime at each station. We think that ourmodels can be usefull when planingpreliminary surveys for a single specieswith time and man-power constrains.

METHODSWe recorded bird species

presence/absence at point stationssurveyed for 15 min. A total of 1118stations were surveyed between Apriland June in 1999 and 2000, in twoareas of 70 x 70 km in WesternAndalusia, Spain (area centers were: 6°21’ W 37° 39’ N, and 5° 28’ W 36°44’ N). Each year, about 75% of thepoint stations were separated by morethan 1000 m, the rest being 250-300 mapart (although then stations were invery different habitats). The observerstarted to record species about three minafter reaching the survey station, soallowing birds to eventually return tonormal behaviour. Two bands wereconsidered at each survey station: aninternal circular band within 50 m of theobserver and an external band from 50m to unlimited distance. First detectionof a bird species in each band wasrecorded to the second. The reason forusing two bands, one with fixed radiusan the other with unlimited radius, isthat fixed census radius has the


39

advantage of sampling an area of knownsize but has the disadvantage that manyspecies, specially the larger ones, arerarely recorded. The unlimited radiushas the advantage or recording morepresences but the disadvantage that theefective survey radius is unknown andvaries between species. Sampling wasperformed throughout the day, avoidingonly the hottest hours (generally, 1300to 1600). Surveying outside of theoptimal period of the day (Drapeau,Leduc & McNeil 1999) may confermore variability to data, but as points indifferent habitats were sampled atrandom times there is no reason toexpect a bias due to this fact.

Vegetation types (Hall,Krausman & Morrison 1997) variedfrom cattle pastures with little or nopresence of short camephytes (mainlyLavandula stoechas and Thymus spp.)and herbaceous dry cultures (mainlybarley, wheat, and sunflower),Mediterranean scrub formations 50-250centimeters tall, to Evergreen OakQuercus ilex subsp. ballota and CorkOak Quercus suber forests and“dehesas”, olive groves, and pine andeucalyptus (Eucalyptus spp.)plantations. These habitats wereclassified, respectively, as herbaceous(299 survey stations), scrub (94), andforest (751), on the grounds ofpresumed differences in richness anddetectability of birds among habitatswith different structure.

STATISTICAL ANALYSESWe divided the total duration of

each point survey in periods of 30seconds and calculated the proportion ofpoints in which each species wasdetected in each time interval (what wecall "relative frequency" hereafter). Theresult is a cumulative curve ofprobability of detection with time fromsurvey start (RF(t)), which is ourapproximation to obtain an equivalentto the curve of cumulated energy gain

(E(t)) of the marginal value theorem.Data were considered separately for a)detections in the 50-m internal band andb) for first detection in any of both theinternal and external bands (that is,considering time of the first detectionwherever it was recorded). The firstcurve is the result of a 50-m fixed-radius point survey while the secondcorresponds to an unlimited-distancepoint survey. We considered onlyspecies that appeared in more than 5%of the points surveys in adequatehabitats, and used only point surveys inrelevant vegetation types for eachspecies (34 species for fixed-radius and43 species for unlimited-distance pointsurveys). Following the optimizationcriterion used in the marginal valuetheorem (Mac Nair 1982), themaximum for RF(t)/(t+τ) wascalculated, RF(t) being the relativefrequency at a time t, and τ being thetravelling time between point stations.We consider that a field biologist is ableto estimate mean travelling time inadvance (at least to a certain degree)during the sampling design, accountingfor accessibility of terrain, total areathat wants to cover and the desiredseparation between point stations. Theoptimal survey duration was estimatedfor travelling times (τ) between 2.5 and30 min at 2.5 min intervals. The upperlimit for survey duration consideredhere is 15 min (the total duration of thesurvey we performed); consequently, ifthe estimate of optimal duration for aparticular species was 15 min for atravelling time X, then optimal surveyduration would be 15 minutes for alltravelling times greater than X.

We analysed the effect of bodysize, regional abundance, and vegetationtype on the shape of the cumulatedcurves of relative frequency and on theoptimal duration of a survey for severaltravelling times between stations. Mostbird species are registered mainlyduring the first min of the survey (thus


40

having curves of cumulated relativefrequency that soon approach anasymptote) while others may need along survey to be detected (do not showa flattening tendency after 15 min). Tosummarise the shape of the curve for aparticular species with a number, weused the ratio between relativefrequency at 5 min and relativefrequency at 15 min: ratios close to 1indicate that in most points where thespecies was present it was recorded inthe first 5 min, while low ratios indicatethat at many points where the specieswas present it was not detected duringthe first 5 min but during the subsequent10 min. We selected 5 min to computethe ratio because most curves of relativefrequency showed an inflection point ataround this duration (Fig.1), and severalauthors have suggested 5 min as anoptimal survey duration (Fuller &Langslow 1984; Hutto, Pletschet &Hendricks 1986; Jiménez 2000). Toexplore which factors may affect theshape of this curve for each species webuilt a Generalized Linear Model of theratio of relative frequencies in 5 and 15min using the following as predictors:body weight (log-transformed),frequency in the total sampled points(pooling all vegetation types; this is asurrogate for regional abundance), andvegetation structure type. P-values forindividual terms were calculated byanalyzing the change in devianceassociated with the deletion of eachterm from the saturated model;similarly, P-values for the saturatedmodel were calculated by comparisonwith the null model. Weights weretaken from the Handbook of the Birds ofEurope, the Middle East and NorthAfrica (Cramp et al. 1977-1994),choosing data for the spring season(both sexes combined), the appropriatesubspecies, and the closest recordingplaces whenever possible. Weightsranged from 5.3 g (Firecrest Regulusignicapillus) to 508 g (Red-legged

Partridge Alectoris rufa). Totalfrequency in the sample ranged from0.02 (Yellow Wagtail Motacilla flava)to 0.53 (Goldfinch Carduelis carduelis)

Finally, we attempted apredictive model of optimal surveyduration for fixed travelling timesknown in advance. To this end weperformed generalised linear modellingof optimal survey duration for travellingtimes of 5, 10, 15 and 20 min and bothcensus methods (fixed radius andunlimited distance). Explanatoryvariables tested were body weight (log-transformed), total frequency (presencesdivided by total number of samplingpoints irrespective of vegetation type),and vegetation structure type.

For illustration purposes weselected the following subset of species:Sardinian Warbler Sylviamelanocephala and Blackbird Turdusmerula in scrub, Crested Lark Galeridacristata and Calandra LarkMelanocorypha calandra in herbaceousvegetation, and Wren Troglodytestroglodytes and Common ChaffinchFringilla coelebs in forest. Thesespecies were selected to cover areasonable variety of breeding habitatsand frequency in the samples(percentage of points with presence of aspecies varies between 21% of forestpoints for the Wren to 81% of scrubpoints for the Sardinian Warbler).

RESULTSAs expected, curves of

cumulated relative frequency for theselected species were higher for theunlimited-distance point surveys.Curves show, for most of the selectedspecies, a very slow increase after 5-7min of survey duration, almost reachingan asymptote in the case of theSardinian Warbler. The exceptions arethe Blackbird, for which the curve isstill increasing steadily at 15 min and, toa much lesser extent, the Wren (Fig.1).


41

Fig. 1. Predicted cumulative frequency (presences/number of sampled points) for an example of forest (a, b), scrub(c, d), and herbaceous vegetation species (e, f), for 50-m fixed radius surveys and unlimited-distance surveys. Barsare one SE.

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Rel

ativ

e fre

quen

cy

c)

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

d)

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Survey duration (minutes)

Rel

ativ

e fre

quen

cy

e)

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Survey duration (minutes)

f)

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Rel

ativ

e fre

quen

cy

a) Fixed radius

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

b) Unlimited distance

Common ChaffinchWren

Sardinian WarblerBlackbird

Crested LarkCalandra Lark


42

In unlimited radius pointsurveys, the shape of the curves ofcumulated relative frequency for eachspecies (as estimated by the ratio ofrelative frequencies at 5 and 15 min) issignificantly affected by species bodyweight, total frequency, and vegetationstructure (Table 1a), that explainaltogether 40% of the deviance. Bodyweight has a negative effect, totalfrequency a positive effect andvegetation type an effect that decreaseswith vegetation structural complexity(Fig. 2). However, for the 50-m fixed

radius surveys only body weight seemsto have a significant negative effect onthe ratio of relative frequency (Table1b).

Optimal survey durations for thewhole set of individual species showed,in general, a pattern of increase towardsan asymptote when plotted againsttravelling time between points (graphsnot shown), indicating that there was amaximum optimal survey duration notaffected by travelling time

.

Fig. 2. Diagnostic plots for the model for the ratio of relative frequencies at 5 and 15 min (unlimited-distance survey): fitted model and partial effects of terms. Vegetation types are classified as herbaceousvegetation (H) , scrub (S) and forest (F), on the basis of presumed differences of detectability; Totalfrequency: frequency in the total sampled points (pooling all vegetation types). Body weight was log-transformed (natural logarithms) before the analysis.

0.1 0.2 0.3 0.4 0.5

Frequency

-0.2

-0.1

0.0

0.1

0.2

Par

tial f

or r

elat

ive

freq

uenc

y

2 3 4 5 6 7

-0.2

-0.1

0.0

0.1

0.5 0.6 0.7 0.8 0.9

Fitted: Frequency +Vegetation type + Ln(body weight)

0.5

0.6

0.7

0.8

0.9

Rat

io

-0.1

0.0

0.1

0.2

Par

tial f

or v

eget

atio

n ty

pe

Vegetation type

F H S

Par

tial f

o r L

n(b o

d y w

eigh

t)

Ln(body weight)


TABLE 1. Analysis of deviance table for the model of relative frequency ratio in 5 and 15 minDeviance analysis tables for the GLM of ratio of species relative frequencies in 5 and 15 min. P-values for individual terms are calculated by deleting each term from thesaturated model; P-values for the saturated model is calculated by comparison with the null model. a) Results for the unlimited-distance point surveys. b) Results for thefixed-radius point surveys. In both a) and b) rounded treatment coefficients for the vegetation types -- modelled as dummy variables -- are: herbaceous vegetation, 0.10, andscrub, 0.04 (forest is taken as the reference, so entering the linear predictor as 0).

TABLE 1A. Unlimited-distance point surveys

Term Coefficient SE Residualdf

Changein df

Residualdeviance

Change indeviance

F P-value

Null 42 0.5214Saturated 4 0.2073 6.27 <0.001

Intercept 0.687 0.050Total frequency 0.335 0.097 1 0.0980 11.86 0.001Vegetation type - - 2 0.0632 3.82 0.03Ln(weight) -0.027 0.013 1 0.0373 4.51 0.04

TABLE 1B. Fixed-radius point surveys


Changein df

Residualdeviance

Change indeviance

F P-value

Null 33 0.4584Saturated 4 0.0958 1.92 0.14

Intercept 0.773 0.082Total frequency 0.117 0.215 1 0.0037 0.30 0.60Vegetation type - - 2 0.0404 1.62 0.22Ln(weight) -0.059 0.026 1 0.0671 5.36 0.03


TABLE 2. Summary of models of optimal survey duration (T, in min) for a series of travelling times (5, 10, 15 and 20 min) in fixed radius andunlimited-distance point surveys. The complete model includes natural logarithm of body weight (LW), total frequency (TF), and vegetation types:herbaceous (H), scrub (S) and forest. The minimum adequate model contains significant terms only.

Censustype

Travellingtime (min)

τ

P-valueComplete

model

Minimum adequatemodel

P-value R2(%) P-valueindividual terms

Fixed radius

5 ns T= 0.813+1.35 *LW 0.029 13 -

10 ns T= 3.79+1.4 *LW 0.053 11 -15 ns T= 9.24 - - -20 ns T= 11.63– 2.53*H – 3.43*S 0.028 21 -

Unlimited distance

5 0.037 T= 3.93 - - -

10 0.001 T= 4.53+1.13*LW – 8.08*TF 0.002 24 LW: P=0.013TF: P=0.014

15 0.009 T= 7.60+0.915*LW – 9.31*TF 0.006 21 LW: P=0.005TF: P=0.017

20 0.010 T= 12.01 –9.91 *TF 0.010 14 -


45

In models for optimal surveyduration and 50-m fixed radius designsthe complete model that included bodyweight, total frequency and habitat wasnon significant in all cases (Table 2).Only body weight affected optimalsurvey duration and the effect wassignificant in two models ( for 5 and 10min of travelling time, Fig. 3). Inunlimited distance point surveys thecomplete model for optimal surveyduration was significant in all cases,indicating a positive relation with bodyweight, a negative relation withfrequency and suggesting a positiverelation with vegetation structuralcomplexity (coefficients for scrub andherbaceous vegetation were negativeand the latter had generally much higherabsolute values). Simplification of themodels indicated very significant effectof body weight and total frequency for10 and 15 min, and 10, 15 and 20 minof travelling time respectively. Theeffect of habitat was not significant inall cases and only for 20 min travellingtime was marginally significant. R-square of the final models were ratherlow, ranging from 13% to 27%.

DISCUSSIONOptimal survey duration for individualspecies is inversely proportional to itsratio of relative frequencies at 5 and 15min. A ratio of 1 for a species indicatesthat all points with a recorded presenceat 15 min had that contact in the first 5min of the survey, whereas a ratio of 0.5indicates that only half of the pointswith a recorded presence after 15 minhad the contact in the first 5 min. TheGLM model built for the ratios shows anegative correlation with speciesweight, a positive correlation with totalfrequency (a surrogate for regionalabundance), and a negative correlationwith structural complexity of vegetation(the vegetation mean effects follow the

order herbaceous>scrub>forest, Table1).

Consequently, if the aim of astudy is to register presence/absence inas many points as possible where aparticular species could be present, ourresults indicate that survey duration at aeach point should increase with speciesbody size, decrease with expectedregional abundance (longer surveys forrare species), and increase with habitatstructural complexity. This pattern isonly apparent if the method used ispoint surveys with unlimited distance,while only body weight needs to beconsidered for surveys with 50-m fixedradius design. Previous studies haveshown different cumulative curves ofpercentage of total individuals countedwith increasing count duration fordifferent species (Scott & Ramsey1981; Jiménez 2000). Fuller andLangslow (1984) suggest thatcumulative curves, and thereforeadequate sampling schemes (Barker,Sauer & Link 1993), might be speciesspecific. Our analysis indicates thatbody size, regional abundance, andvegetation structure explain some of thedifferences in shape of these curves ofrelative frequency (about 40% of thevariance), and consequently influencethe estimate of optimal survey durationfor a particular species (when we areonly interested in presence/absencedata). For example, our model predicts,for the unlimited-distance method, thata small and frequent grassland species(weighing 10 g and appearing in 60% ofthe total points) will have a ratio ofrelative frequency of 0.93, which meansthat in a 5-min survey this specieswould be detected in 93% of the pointsin which it was present after 15 min. Onthe other hand, a larger and rarer forestspecies (100 g and frequency equal to10%) is predicted by the model to havea ratio of relative frequency of 0.60, thatis, only in 60% of the points where thespecies is recorded as present after 15


46

min would it have been detected duringthe first 5 min.

A similar pattern was foundwhen we tried to model the optimalsurvey duration for fixed values oftravelling time between sample stationsusing body weight, total frequency andvegetation type as explanatory variables(Table 2). In general, the effect of bodyweight is positive (the larger the speciesthe larger the optimal survey duration)and the effect of total frequency isnegative (shorter optimal times forcommon species, see Figs. 3 and 4). Wedid not find a significant effect ofvegetation type, although resultssuggest a positive effect of vegetationstructural complexity (longer optimaltimes for forest and shorter for the moreopen vegetation types) and we think thatdifferences among vegetation types may

be masked by the variability of our data.For example, using our models forunlimited radius surveys, the predictedoptimal survey duration for 10 mintravelling time for the Goldfinch, a verycommon species of 13.2 g body weight,is 3 min, while for the Red-leggedPartridge, less common and heavier(508 g), is around 10 min (Fig. 4) .

To summarize the analyses, onlybody weight affects the shape of theprevalence curve and consequently theoptimal survey duration for a specieswhen using 50-m fixed radius pointsurveys, while body weight, regionalabundance, and vegetation structureaffect the prevalence curve of a speciesand its optimal survey duration fordesigns with unlimited radius.

Fig. 3 (left). Models for optimal survey duration in relation to species body weight for 50-m radius pointsurveys for fixed travelling times between sampling stations (5, 10, 15, and 20 min). Non significanttrends are indicated with point lines and open symbols. Bars above the x-axes indicate the body weightsof the species used in the models.

Fig. 4 (right). Isolines of optimal survey duration in relation to species body weight and species totalfrequency in the study area (a proxy for regional abundance) for unlimited radius point surveys and 10min travelling time between sampling points. Triangles indicate the total frequency and body weight ofthe species used in the model.

0 1 2 3 4 5 6Log body weight (g)

0

4

8

12

Op

timal

cou

nt d

urat

ion

(min

)

Travelling time

5 min10 min15 min20 min

1 2 3 4 5 6Log body weight (g)

0.1

0.2

0.3

0.4

0.5

Tot

al f

requ

ency

3.1 min

4.1

5.1 min

6.2

7.2 min

8.3

9.3 min

Red-legged Partridge

Goldfinch


47

Larger species have greatermobility, larger home ranges, loweraverage densities (Peters 1983), andsongs that can be heard from a greaterdistance (Calder III 1990), all of whichincrease the probability that they willenter the effective survey area thelonger the observer stays in a point.Rare species (those with low abundanceor low detectability for the observer,once these are corrected for body size)benefit from longer surveys only in theunlimited-distance method. This isprobably because, when having a largersurveyed area, the observer requiresmore time to cover the whole area, anda longer time benefits more theinconspicuous species, those at lowabundances, or those to which theobserver is less habituated.Comparatively, rare species, if close tothe observer (50-m fixed-radiusmethod), do not require more time to bedetected than common ones.Structurally complex vegetation typesseem also benefit from longer surveydurations the larger the census radius,but our analyses were not totallyconfirmatory in this regard.

Two limitations of this workmust be noted. First, we performedcensus of 15 min and considered thistime as the upper limit for optimalsurvey duration in the statisticalanalysis. That is, if we estimated anoptimal survey duration of 15 minutesfor a particular species at a giventravelling time, all estimates of optimalsurvey duration for longer travellingtimes were necessarily also 15 min, buthad we surveyed for a longer time alonger optimal time might have beenestimated. This is probably the reasonwhy body weight did not enter themodels for 20 min of travelling time,since larger species have longer optimalsurvey duration, and so they reach theupper limit of 15 min for shortertravelling times. The second limitation

is that we were to build simple modelswith few easily measurable variables,and so we did not add to the analysisbehavioral variables that affectsdetectability of species (for example,foraging and singing behavior), and thisis probably the reason why final modelsexplained a rather low variability ofdata.

As a conclusion, the choice of apoint survey duration to recordpresence/absence data for individualspecies should take into account, atleast, the regional abundance and thebody size of the species, and probablyalso the structural complexity of thevegetation in the area where the surveyis planned to be carried. The models weoffer could help in the sampling designto select an appropriate survey durationto record presence/absence data at pointstations.

ACKNOWLEDGEMENTS

This work is a contribution tothe project "Predictive cartography ofland birds: A pilot study in WesternAndalusia", funded by the DirecciónGeneral de Enseñanza Superior eInvestigación Científica (Ministry ofScience and Technology) and FEDERfunds from the EU, project # 1DF-97-0648. J.S. had a predoctoral fellowshipfrom the Ministry of Education andCulture. The extensive field workpresented in this work could not havebeen done without the help and sense ofhumour of Daniel López Huertas, LuisM. Carrascal, and Mario Díaz who alsobrought intellectual insight to theproject

NOTESA version of this chapter is beingprepared for submittion to Ibis (withJ.Bustamante)


48

REFERENCES

Barker, R. J., Sauer, J. R. & Link, W. A.(1993). Optimal allocation of point-count sampling effort. The Auk,110(4): 752-758.

Beard, K. H., Hengartner, N. & Skelly,D. K. (1999). Effectiveness ofpredicting breeding bird distributionsusing probabilistic models.Conservation Biology, 13(5): 1108-116.

Bolger, D. T., Scott, T. A. &Rotenberry, J. T. (1997). Breedingbird abundance in an urbanizinglandscape in coastal southerncalifornia. Conservation Biology,11(2): 406-421.

Calder III, W. A. (1990). The scaling ofsound output and territory size: arethey matched? Ecology, 71(5): 1810-1816.

Charnov, E. L. (1976). Optimalforaging, the marginal valuetheorem. Theoretical PopulationBiology, 9: 129-136.

Drapeau, P., Leduc, A. & McNeil, R.(1999). Refining the use of pointcounts at the scale of individualpoints in studies of bird-habitatrelationships. Journal of AvianBiology, 30: 367-382.

Fuller, R. J. & Langslow, D. F. (1984).Estimating numbers of birds by pointcounts: how long should census last?Bird Study, 31: 195-202.

Green, R. E., Osborne, P. E. & Sears, E.J. (1994). The distribution ofpasserine birds in hedgerows duringthe breeding season in relation tocharacteristics of hedgerows andadjacent farmlands. Journal ofApplied Ecology, 31: 677-692.

Gutzwiller, K. J. (1991). Estimatingwinter species richness withunlimited-distance point counts. TheAuk, 108: 853-862.

Gutzwiller, K. J. (1993). Refining theuse of point counts for winter studiesof individual species. WilsonBulletin, 105: 612-627.

Hall, L. S., Krausman, P. R. &Morrison, M. L. (1997). The habitatconcept and a plea for standardterminology. Wildlife SocietyBulletin, 25(1): 173-182.

Harrell, F. E. (2001). Regressionmodeling strategies.Springer, NewYork

Hutto, R. L., Pletschet, S. M. &Hendricks, P. (1986). A fixed-radiuspoint-count method for nonbreedingand breeding season use. The Auk,103: 593-602.

Jiménez, J. E. (2000). Effect of samplesize, plot size, and counting time onestimates of avian diversity andabundance in a Chilean rainforest.Journal of Field Ornithology, 71(1):66-87.

Mac Nair, J. N. (1982). Optimal giving-up times and the marginal valuetheorem. The American Naturalist,119: 511-529.

Pearce, J. & Ferrier, S. (2001). Thepractical value of modelling relativeabundance of species for regionalconservation planning: a case study.Biological Conservation, 98: 33-43.

Peters, R. H. (1983). The ecologicalimplications of body size.CambridgeUniversity Press, Cambridge

Scott, J. M. & Ramsey, F. L. (1981).Length of count period as a possiblesource of bias in estimating birddensities. Studies in Avian Biology,6: 409-413.

Añadiendo la opinión de experto a los modelos estadísticos

49

CAPÍTULO III: ¿Incrementa la opinión de experto la habilidad

predictiva de los modelos de la distribución de aves?

RESUMEN

El modelado predictivo del hábitat para la conservación y gestión resulta facilitado por

procedimientos automáticos de selección y transformación de variables explicativas. Se

ha argumentado que los modelos empíricos predictivos se beneficiarían si incluyeran

una opinión de experto en las diferentes fases de la modelización, pero esto supone una

elevada inversión de tiempo y es difícil de estandarizar. Los procedimientos

automáticos, que son más rápidos y fáciles de integrar en un Sistema de Información

Geográfica, pueden producir modelos altamente explicativos que ajustan bien los datos

usados en la construcción del modelo, pero no predicen necesariamente mejor en un

conjunto independiente de observaciones. Por el contrario, los modelos supervisados

pueden incluir más frecuentemente relaciones causales y, por tanto, podrían extrapolarse

mejor a otras áreas. En este trabajo generamos modelos predictivos del hábitat para la

presencia/ausencia de 10 especies de aves en dos áreas de Andalucía (SO España), con

el fin de comparar tres procedimientos de selección de predictores, que van desde uno

automático a otro completamente supervisado (tipos de modelos), y comprobamos su

capacidad discriminativa en tres escenarios de evaluación: (1) en el mismo conjunto de

datos usado para construir los modelos, (2) en un conjunto de datos diferente

(remuestreado) y (3) en datos de un área geográfica diferente. Los modelos automáticos

alcanzaron una capacidad discriminativa significativamente mayor, según AUC y Kapa,

sólo cuando se evaluaron con los datos de construcción. El resto de combinaciones entre

tipo de modelos y escenarios de evaluación no mostraron diferencias significativas,

aunque los modelos automáticos tendieron a resultar ligera pero no significativamente

peores que los supervisados cuando se evaluaron con datos de un área geográfica

diferente. Destaca el hecho de que la capacidad predictiva, medida a través de las

estimas de discriminación en datos remuestreados, no difirió entre los distintos tipos de

modelo. En conclusión, la incorporación de opinión de experto en la modelización (al

menos en la forma que empleamos) no genera modelos con mayor capacidad predictiva.

Por tanto, los procedimientos automáticos para construir modelos predictivos del hábitat

parecen un medio eficaz y rentable para crear mapas de adecuación del hábitat en un

contexto regional.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO III

50

CHAPTER III: Does expert opinion increase the predictive ability of

environmental models of bird distribution?

ABSTRACT

Predictive habitat modeling for conservation and planning is facilitated by automatic

procedures for the selection and transformation of variables to be included into the

models. Empirical predictive models have been claimed to potentially benefit from the

inclusion of expert opinion in different stages of the model-building procedure,

although this is a time-consuming task difficult to standardize. Automated procedures,

faster and easier to integrate into a Geographic Information System, may render highly

explanatory models that fit well the data used to build the model but not necessarily

predict so well independent observations. On the contrary, supervised models may

include more frequently causal relationships and, therefore, they may extrapolate better

to other areas.

We built predictive habitat models for the presence/absence of 10 bird species in two

areas of Andalusia (SW Spain) to compare three different kinds of procedures for

predictor selection, ranging from a completely unsupervised to a fully supervised

method (model types), and tested their discrimination ability in three evaluation

scenarios: (1) on the same data used to build the models, (2) on a different (resampled)

evaluation data set and (3) on data from a different geographic area. Unsupervised

models had a significantly greater discrimination ability, in terms of both AUC and

Kappa, only when evaluated with building data. Other model type-evaluation scenario

combinations did not show significant differences, though unsupervised models tended

to perform slightly but not significantly worse than supervised models when evaluated

with data from a different geographic area. Notably, predictive ability, as measured by

discrimination estimates on resampled data sets, did not differed between model types.

To conclude, incorporating expert opinion in the model building, in the way we have

done, does not render better models measured by their predictive ability. Therefore,

unsupervised fitting procedures for building predictive habitat models seems an

adequate cost-effective way to proceed when aiming to generate habitat suitability maps

in a regional context.


51

INTRODUCTION

Predictive habitat models areincreasingly being used to assessspecies distribution in both conservationand regional planning (Guisan &Zimmermann 2000; Pearce et al. 2001).This is mainly because statisticalmodels of distributional data allow tobenefit the most from wildlife surveys(Nicholls 1989), which is particularlyrelevant when distribution data arescarce or when areas are remote(Osborne & Tigar 1992; Bustamante etal. 1997; Manel et al. 1999). Thesemodels can be built with many differentpurposes. When the aim is generatingdistribution or habitat suitability mapsfor a big number of species that can beused for reserve selection orconservation planning, the predictiveaccuracy of the models is the mostrelevant indicator of model success,while standarization and automatedmodel building are frequently desiredbecause of time constrains and the needof implementing easily the models intoa GIS to generate final maps (Guisan,Weiss & Weiss 1999).

Predictive habitat modeling ishabitually tackled with regression-likeapproaches, among which logisticregression outstands due to itssuitability to model a binary variablesuch as presence/absence (see a reviewin Guisan and Zimmermann 2000). Inthese common cases, a responsevariable —say the presence/absence ofthe species in an area— is related to anumber of predictors with somesuspected discrimination ability, andselection is done among them (however,some authors advocate to prespecifymodel complexity so avoiding thissecond part, see Steyerberg et al. 2000;Harrell 2001). In regression modeling,on one hand, the choice of predictors

may be automated by full forward orbackward algorithms designed to satisfystatistical criteria. Automatedprocedures are desirable because ofboth their quickness and easiness tostandardize, but they are argued toincorporate spurious variables to themodel when predictors are not totallyindependent —a very likely situation—(James & McCulloch 1990; Mac Nally2000). On the other hand, supervisedprocedures to select among predictorsmay lead to models that are morecredible, for example, by excluding ormodifying relationships that do notmeet some biological criteria;unfortunately this can be a tedious taskand may result in overoptimisticestimates of model performance(Harrell 2001). Therefore, choosing oneof the two selection procedures outlinedabove raises a possible conflict betweenthe easiness to built a model and itscredibility; a conflict which address isof prime importance in conservationand planning. Currently, the limitedwork on comparing models built withthe two procedures of predictorsselection suggests that pure statisticalmodels, without supervision, can be asgood as those built with expert opinion(Pearce et al. 2001). However, it can beexpected that the relative performanceof both kind of models changes indifferent scenarios of application. Forinstance, if automated models rely to agreater extent on casual correlationsparticular to a certain area (or, in aextreme example, if they rely onspureous correlations with unsoundpredictors), then they should fail whenapplied to independent data, thoughthey might explain a great amount ofthe variation observed in the data thatwere used to build the model (Verbyla& Litvaitis 1989). The reverse may betrue for supervised models (Lezzoni1999). If they rely more on causalrelationships (or intended to be causal)


52

they may be expected to apply in awider arrange of circumstances.

In this work we address thecomparison between three differentkinds of procedures for predictorselection in predictive speciesdistribution modeling, ranging from acompletely unsupervised (automaticstepwise variable selection by statisticalsoftware) to a fully supervised method(in which an expert onithologist decidedwhether statistical significant relationsmade sense in relation to the ecology ofthe species). We test the predictiveability of models in three scenarios: (1)on the same data used to build themodels, (2) on a different evaluationdata set and (3) on data from a differentgeographic area. The main aim is toexplore whether the inclusion of expertopinion in the building of modelsrenders models with a higher predictiveability, or ones that extrapolate better toother areas.

STUDY AREA AND METHODS

The study areas are two 70x70 kmsquares in Western Andalusia, SouthernSpain. We will refer to them as Aracena(center: 6° 21’ W 37° 39’ N) andGrazalema (center: 5° 28’ W 36° 44’ N;Fig. 1). Both areas have roughly thesame proportion of cropland (mainlywheat, sunflower and olive groves),shrubland and forests (mainlyMediterranean shrubland, evergreen andcork oak forests and “dehesas”), andsimilar and numerous humansettlements. The areas differ mainly inthat Grazalema mountains reach higheraltitudes compared to those in Aracena,ranging from 0 to 1622 m.a.s.l. in thefirst area and from 0 to 960 in thesecond, and in the soil type: mostlycalcareous in Grazalema and mostlyacidic in Aracena.

Figure 1. Areas of study

North m

100000.00

Iberian Peninsula

Aracena

Grazalema


53

We performed 1144 unlimiteddistance point surveys during thesprings of 1999 and 2000, 521 inAracena and 623 in Grazalema. Weselected 10 species from all registered(172) according to three criteria: (i) theyare abundant in both study areas, (ii)they have a variable range ofprevalences (defined as the frequency ofpresences in the sample), and (iii) theyare representative of the mainlandcovers present (cropland, shrublandand forest). The selected species werered-legged partridge Alectoris rufa L.(148 presences in Aracena vs 285presences in Grazalema), linnetCarduelis cannabina L. (228 vs 402),short-toed treecreeper Certhiabrachydactyla C. L. Brehm (306 vs318), robin Erithacus rubecula L. (74 vs237), Thekla lark Galerida theklae C. L.Brehm (170 vs 114), calandra larkMelanocorypha calandra L. (64 vs 88),blue tit Parus caeruleus L. (352 vs307), European nuthatch Sitta europaeaL. (226 vs 124), Sardinian warblerSylvia melanocephala Gmelin (368 vs620) and wren Troglodytes troglodytesL. (76 vs 269). Absences outnumberedpresences for every species in bothstudy areas, so to avoid bias due to thisfact (Fielding & Bell 1997; Cumming2000) we randomly selected a numberof absences equal to the number ofpresences for each species in each studyarea. Sample sizes were similar orhigher to those reported in previousworks to give reliable estimates ofaccuracy (Pearce & Ferrier 2000;Stockwell & Peterson 2002).

The predictive variables in themodels were a large set ofenvironmental predictors (Table 1)extracted and amalgamated from a GISof each study area and aimed tosummarize most relevant environmentalgradients and some landscape features.

These predictors included variablesdescriptive of vegetation, landuse,landscape, topography (resolution 50meters) and climate ( resolution 1 km)that were averaged in a circle of 350meters diameter centered in surveypoints. Extraction of variables from theGIS was done using IDRISI 32(Eastman 1999), IDRISI for Windows(Eastman 1997) and MIRAMON (Pons2000).

We built a generalized additivemodel (GAM, Hastie & Tibshirani1990) for the presence/absence of eachspecies in each study area with binomialerrors and logit link using as predictorsthe environmental variables (Table 1).We built for each species an automaticmodel with stepwise selection ofpredictors using exclusively statisticalcriteria (what we call hereafterunsupervised model). First weperformed a forward-backward stepwiseselection from all possible predictors(with the step.gam procedure of S-PLUS 2000, MathSoft 1999). Westarted from a null model and testedeach predictor sequentially as asmoothing spline with 3 degrees offreedom. The predictor that reduced themost the residual deviance was includedin the model and the procedure wasrepeated until no more predictorsimproved the model. Then, we tried tosimplify the resulting model bydecreasing the complexity of each of thepredictors included (by means of asmoothing spline with 2 degrees offreedom and a linear term). The criteriato enter, remove or simplify a term wasthe Akaike’s Information Criterion(AIC Sakamoto, Ishiguro & Kitagawa1986), that takes into account thereduction both in residual deviance andin residual degrees of freedom due to acertain predictor.


54

Variable description SpeciesMean altitude a 1,2,3,4,5,6,7,8,9,10Mean slope a 1,2,3,4,5,6,7,8,9,10Mean annual temperature b 1,2,3,4,5,6,7,8,9,10Mean annual rainfall b 1,2,3,4,5,6,7,8,9,10Mean annual potential solar radiation a 1,2,3,4,5,6,7,8,9,10Percentage of crop land (crops, olive groves, vineyards) c 1,2,5,6,9,10Percentage of herbaceous vegetation (including cereal crops) c 1,6Percentage of olive groves c -Percentage of forest (including “dehesas” and open forest) c 3,4,7,8,9,10Percentage of dense forest c 3,4,7,8,10Percentage of deciduous forest c 3,4,7,8Percentage of coniferous forest c 3,4,7,8Percentage of shrub c 1,2,4,5,9,10Percentage of riparian vegetation c 4,7,10Presence of sparse tree cover (for example, included in a heterogeneous

crop land area) c6

Presence of dense tree cover (for example, included in a heterogeneouscrop land area) c

6

Presence of sparse srhub or sparse shrub-like structures (such asvineyards) c

5,9

Presence of dense shrub c 2,5,9Length of boundaries between forested landcover categories and the rest

of vegetation categories c3,4,6,7,8,9

Length of boundaries between forest and shrubland c 2,3,4,7,8,9Fractal dimension of NDVI values of a satellite image as an index of

heterogeneity in croplands d1,6

Compactness ratio of dense forest areas (an indirect estimate of surface-perimeter ratio) c

3,4,7,8

Distance to the nearest urban area smaller than 2 ha c -Distance to the nearest urban area sized between 2 and 10 ha c -Distance to the nearest urban area sized between 10 and 100 ha c -The same distances to nearest areas sized <2, 2-10 and 10-100 ha of crop

land, herbaceous vegetation, olive groves, forest land, dense forestland, deciduous forest, coniferous forest, shrubland, and riparianvegetation c

The same as forpercentage of eachcategory

Table 1. Predictors tested in all unsupervised models. The number in the second column indicates forwhich species each predictor was selected a priori in semi- and supervised models: 1 Red-leggedPartridge, 2 Linnet, 3 Short-toed treecreeper, 4 Robin, 5 Thekla Lark, 6 Calandra Lark, 7 Blue Tit, 8European Nuthatch, 9 Sardinian Warbler and, 10 Wren. Sources: a Digital Elevation Model of Andalusia at50 m resolution. b raw meteorological data provided by the Instituto Nacional de Meteorología andinterpolated by regression models and kriging at resolution 1 km2 (own data, unpublished), c 1995 land-use/land-cover cover digital map of Andalusia from the SinambA (Consejería de Medio Ambiente, Junta deAndalucía). d IRS satellite image, sensor LISS III (date: 19/07/99 Aracena, date: 16/07/99 Grazalema).Fractal dimension estimated with IDRISI 32 on a NDVI (Normalized Difference Vegetation Index) image.


55

We incorporated expert opinion inthe building of models in two ways. Inthe first one we chose the predictors tobe submitted to the forward stepwisealgorithm for each species according toprevious knowledge of habitat selectionof the species (see Table 1). Forexample, we judged appropriate toinclude the percentage of crop landamong the predictors to be tested withcalandra lark (but not the percentage offorest). Final model for each species ineach area was obtained using aprocedure analogous to the onedescribed previously. Predictors werefirst entered in a forward fashion(ordered by AIC), but then the modelwas checked after each inclusiondropping those predictors that did notreduce significantly the residualdeviance as measured by a Chi-squaretest (α=0.05%). We call these semi-supervised models.

The second way of incorporatingexpert opinion started from the modelresulting from the semi-supervisedprocedure. In each model we testedtransforming one by one the predictorsselected by the previous automaticprocedure into parametric polynomials,or into linear (or piecewise linear) termsthat were sensible, based on ourprevious knowledge of habitat selectionby the species and visual inspection ofpartial residual plots (Brown 1994;Franklin 1998). Statistically significantrelations with environmental predictorsthat made no sense with species habitatselection and others apparently spuriouswere excluded from the final model. Asmiscelaneous examples: curvilinearforms were transformed when possibleto quadratic polynomials, in increasingor decreasing relationships withdistance to a particular habitat we testedthe inclusion of an asymptote (i.e.assuming a constant effect after acertain distance), and a negative

relationship with forest for a forestspecies would have been excluded fromthe model even if statisticallysignificant. These are called supervisedmodels. Therefore for each species andstudy area we obtained three modeltypes with increasing degree of expertopinion invested in their building:unsupervised, semi-supervised andsupervised.

To assess the performance ofmodels we used Cohen’s Kappa statistic(Titus, Mosher & Williams 1984) andthe area under the curve (AUC) of areceiver operating characteristic plots(ROC, Hanley & McNeil 1982;Murtaugh 1996; Cumming 2000) asmeasures of discrimination ability.Kappa estimates the chance-correctedpercentage of agreement betweenpredictions and observations. Tocalculate Kappa it is necessary to definea threshold of predicted probabilityabove which to consider presence(Fielding & Bell 1997), and this wasselected to be the average between themean of probabilities for absences andthe mean for presences (Fielding &Haworth 1995). The AUC is athreshold-independent measure ofdiscrimination ability (Zweig &Campbell 1993), and it is not affectedby prevalence of presences in thesample (Manel, Williams & Ormerod2001); thus, it is considered to be betterthan kappa. However, AUC has onlybeen recently used in ecology (Fielding& Bell 1997; Cumming 2000; Pearce &Ferrier 2000; Bonn & Schröder 2001;Manel, Williams & Ormerod 2001) andit is more complicated to estimate thanKappa, which has been widely used inprevious works. Thus, we used here thetwo indices as a way of comparison.AUC was calculated with AccuROC 2.5(Vida 1993).


56

Kappa and AUC were calculatedfor the three model types (supervised,semi-supervised and unsupervised) inthree scenarios of evaluation accordingto an increasing degree of independencebetween the data used to build themodel and the data used to evaluate it:(i) building scenario: with the same dataused to build the model, (ii)crossvalidation scenario: with a ten-fold crossvalidation data repeated 20times and (iii) extrapolation scenario:with data of the same species but fromthe other study area. The first is aevaluation (Oreskes, Shrader-Frechette& Belitz 1994) that tends to renderoveroptimistic estimations ofdiscrimination ability, but it may beused both as a maximum reference ofexplanatory ability of the model and to

informally assess the amount ofoveroptimism in the estimates bycomparison with the predictive abilityin the crossvalidation scenario(moreover it is the only one presented inmany research papers). This type ofevaluation informs about theexplanatory ability of the model. Thesecond evaluation is an internalvalidation that follows one of thevariety of resampling approaches mostfrequently used (this one according toHarrell 2001): data is split in 10 groupsof equal size, a model is built using thedata of the first nine groups (90% ofobservations) and evaluated using thedata of the tenth group (10% ofobservations, those not used to build themodel); the procedure is repeated ten

Variable Df SS MS F-value P-valueModel type 2 0.089 0.045 6.396 0.002Evaluation scenario 2 1.078 0.539 77.357 0.000Study zone 1 0.030 0.030 4.275 0.040Model type*evaluation scenario 4 0.112 0.028 4.010 0.004Model type*study area 2 0.017 0.008 1.208 0.302Evaluation type*area 2 0.007 0.004 0.512 0.601Model type*evaluation scenario*study area 4 0.020 0.005 0.725 0.576Residuals 162 1.129 0.007

Table 2. Results of ANOVA of the effect of model type (unsupervised, semi-supervised and supervised),evaluation scenario (building, crossvalidation and extrapolation) and study area (Aracena and Grazalema)on model discrimination ability estimated by AUC.Interactions marked by asterisks.

Variable Df SS MS F-value P-valueModel type 2 0.441 0.221 7.989 0.001Evaluation scenario 2 3.680 1.840 66.593 0.000Study area 1 0.152 0.152 5.484 0.020Model type*evaluation scenario 4 0.499 0.125 4.511 0.002Model type*study area 2 0.061 0.031 1.108 0.333Evaluation type*study area 2 0.018 0.009 0.325 0.723Model type*evaluation scenario*study area 4 0.058 0.014 0.523 0.719Residuals 162 4.476 0.028

Table 3. Results of ANOVA of the effect of model type, evaluation scenario and study area on modeldiscrimination ability estimated by Kappa. Interactions marked by asterisks.


57

times using each group as evaluation setand the remaining nine as building set.This ten-fold crossvalidation wasrepeated 20 times resuffling the data ineach group. The predictors included inthe models were the ones resulting fromthe previously described fittingprocedures (model types) and only newcoefficients were adjusted with thebuilding set. The result is an unbiasedestimate of the predictive ability of themodel within its universe of application.Finally, the third is an externalvalidation (Harrell 2001) used to assessa model outside the universe where it isstatistically valid: the model is appliedto a new scenario different from the onewhere it was built; it evaluates thetransferability of the model (Altman &Royston 2000; Bonn & Schröder2001). From an statistical point of viewthere is no particular interest inestimating the predictive ability of amodel in a universe different fromwhich it was built, but, in real practice,predictive habitat models are frequentlybuilt expecting that they will havecertain predictive ability whentransfered in time (in the future) or inspace (to remote areas).

Differences in predictive(discrimination) ability among modelsdue to differences in model type,evaluation scenario, and study areawere analyzed with a factorial ANOVAof AUC and Kappa values of models.Pairwise differences of means betweenfactor levels were assessed with 95%simultaneous confidence intervals andthe Tukey test. We do not discuss othermeasures of predictive performance ofmodels such as calibration andrefinement (Pearce & Ferrier 2000).

RESULTS

Models better than a null modelwere built for every species and for

each combination of model type,evaluation scenario and study area.Unsupervised models included asignificant (F2,57=48.38, P<0.0001)higher number of predictors (mean=9,sd=4.5) than both semi-supervised(mean=3, sd=1.1) and supervised(mean=4, sd=1.3).

The estimates of discriminationability, both AUC and Kappa (Tables 2and 3), were affected significantly bythe model type (AUC: F2,162=6.40,P=0.002; Kappa: F2,162=7.99, P=0.001),the evaluation scenario (AUC:F2,162=77.36, P<0.0001; Kappa:F2,162=66.60, P<0.0001), and theinteraction model type-evaluationscenario (AUC: F4,162=4.01, P=0.0040;Kappa: F4,162=4.51, P=0.002). Therewas also a slightly significant effect ofthe study area (AUC: F1,162=4.28,P=0.040; Kappa: F1,162=5.48, P=0.020).

As expected, AUC and Kappaestimates were always higher in theevaluation with building data than in thecrossvalidation or in the extrapolation(in this order), ranging from amaximum of AUC=0.95 (SE=0.01) andKappa=0.78 (0.03) in unsupervisedmodels evaluated with building data, toa minimum of AUC=0.68 (0.02) andKappa=0.26 (0.04) in unsupervisedmodels evaluated with extrapolationdata. This change in discriminationability of unsupervised models isresponsible for the significantinteraction detected between model typeand evaluation scenario (Fig.2).Unsupervised models had significantlygreater discrimination ability, both interms of AUC and Kappa, whenevaluated with building data. In thecrossvalidation unsupervised modelsperformed slightly better thansupervised ones, but differences wereonly statistically significant for AUCvalues between unsupervised and


58

supervised models. In the extrapolationall model types performed poorly (butdid not differ significantly), though inthis case unsupervised models tended tobe the worse. Finally, models built withdata from the Aracena zone had greaterestimates of both AUC(mean=0.81[SE=0.01]) and Kappa(0.50[0.02]) than models built with datafrom Grazalema (AUC: 0.78[0.01];Kappa: 0.44[0.02]).

DISCUSION

Empirical models are typically usedin three scenarios. First, they can beused to summarize the information thatwas used in their building, that is, toexplain the pattern in data (e.g. MacNally 2000). Second and morecommon, models are used to predict aresponse given a new data set, withvalues of the predictors that must be inthe range observed in the original data(a review in Guisan & Zimmermann2000). Third, models may be used to

extrapolate to spatial areas (or temporalcontexts) different from the ones inwhich they were built (Schröder &Richter 1999/2000; Bonn & Schröder2001). Extrapolating in a diferent spatialor temporal context may be called intoquestion as models may not behave asexpected outside the universe of theirconstruction. Therefore extrapolation(or forecasting, Morrison, Marcot &Mannan 1998) is better suited to causalmodels, while prediction (orhindcasting, Morrison, Marcot &Mannan 1998) is the only result one canexpect with a certain degree ofreliability from correlational models .

Whatever the scenario, empiricalmodels may potentially benefit fromexpert opinion incorporated in some ofthe stages in the process of modelbuilding: pre-modeling, in whichoriginal variables are synthesized ortransformed, model-fitting, when asubset of potentially explanatory

Figure 2. Mean values (and bars for 95% confidence intervals) of estimates ofdiscrimination ability for each combination of evaluation scenario and model type.Superscript letters show comparisons of model type values within each evaluationscenario: same letters for unsignificant differences (95% simultaneous confidenceintervals and Tukey test).

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Buildin

g

Cross

valid

ation

Extrap

olatio

n

Kap

pa

AU

C

UnsupervisedSemi-supervisedSupervised

0.60.650.7

0.750.8

0.850.9

0.951

Buildin

g

Cross

valid

ation

Extrap

olatio

n

Evaluation scenario

b

a

b a

b

a, baaa

ab

a

b

aa

a aa


59

variables is selected from the originalset, and post-modeling stage, in whichpurely statistical relationships betweenresponse and predictors are modified tosatisfy biological criteria (Pearce et al.2001). Completely unsupervised modelswould not receive expert opinion in anyof these stages. Admittedly, ourunsupervised models incorporate, infact, some degree of expert opinionbecause the explanatory variablesderived from the thematic digitalcartography were selected andamalgamated in the pre-modeling stagelooking for the potential relevanceplayed on predicting bird distribution.However, the distinction that we makehere between unsupervised, semi- andsupervised models reflects a commonsituation very relevant to predictivehabitat modeling practitioners: in mostcases a large pool of potential predictorvariables must be reduced orsynthesized before modeling, inparticular when dealing with vegetation.After that compulsory step (that doesnot seem to affect final model accuracy,according to Pearce et al. 2001), whatmust be considered is to automate or notthe process of selecting predictors andmodifying their functional formsaccording to expert opinion.

On one hand, the unsupervisedbuilding of models—by stepwiseselection of predictors according tostatistical criteria in our study—isrelatively fast and may be easilyintegrated in wider protocols thaninvolve further modeling ormanipulation in a GIS (Guisan, Weiss& Weiss 1999). However, thisapproach is criticized because it tends toincorporate to the model spuriouspredictors or variables difficult tointerpret (Mac Nally 2000; Steyerberget al. 2000; Harrell 2001). On the otherhand, supervised procedures may drawrelationships more credible, but they areusually highly time-consuming, difficult

to standardize and very susceptible tooverestimate the statistical significanceof the predictors (Chatfield 1995;Harrell 2001). Supervised models areexpected to have a greater predictiveability than unsupervised modelsbecause of a thoughtful selection ofpredictors that dismiss relationshipsparticular to a certain set of data or,simply, spurious. They should also beable to extrapolate with a lower loss inaccuracy, because it is expected thatsupervised model will include morefrequently causal relationships betweenbirds and their habitats and less spuriouscorrelations.

Indeed, our results suggest that thepredictive performance of unsupervisedmodels is acceptable, being as good orbetter than that of supervised models (atleast in what concerns discriminationability). Unsupervised models have agood explanatory power (as measuredby the estimates of discriminationability on building data), and the samemedium predictive power thansupervised models (as given byestimates of crossvalidation evaluation).A high explanatory power was partiallyexpected because of the well knowneffect of predicting on the own buildingdata. We think that this effect may havebeen exacerbated in our case, since weentered the predictors as complex non-linear smoothed terms that captured theparticularities of each study area.Furthermore, the unsupervised modelsmay have a higher explanatory powerdue to their generally longer list ofpredictors, which is a consequence ofusing the somewhat lenient AIC as thecriterion to enter or keep predictors(Ludden, Beal & Sheiner 1994;Chatfield 1995; Burnham & Anderson1998). A few of the variables includedin unsupervised models had a difficultinterpretation or one opposite to whatwas expected according to previousknowledge of habitat selection (for


60

example, the probability of presence ofRed-legged Partridge increasing withdistance to small croplands); thesevariables could be spurious or justparticular to a set of observations, and,if present, were disregarded in thesupervised models.

The predictive ability of the threemodel types (as measured oncrossvalidation data) was similar.Although, the results showed a trendtowards higher values of discriminationestimates for the unsupervised models;this suggest that some of the variablesdisregarded in the model-fitting stagehad in fact some predictive power, atleast at the resolution of our study.Models would correctly classify about 8in 10 pairs of presence/absenceobservations (Hanley & McNeil 1982),which is a medium result according tothe standards posed by several authors(Monserud & Leemans 1992; Fielding& Bell 1997; Pearce & Ferrier 2000)that suggest to consider the modeldiscrimination ability to be poor whenKappa is below 0.4 (or AUC below0.7), fair when it is between within 0.4and 0.7, and good for values greaterthan 0.7 (AUC above 0.9). However,the values reported in this study aresimilar to what is commonly found inthe wildlife-modeling literature (Manel,Dias & Ormerod 1999; Tobalske &Tobalske 1999; Cumming 2000; Bonn& Schröder 2001), what may reflect alimit in the accuracy of empiricalmodels based on indirect variables(Guisan & Zimmermann 2000) topredict a response at a high spatialresolution (Fielding & Haworth 1995;Manel, Dias & Ormerod 1999; RicoAlcázar et al. 2001).

Finally, the discrimination abilityof the models evaluated with externaldata was rather low (though it wassignificantly better than that of a nullmodel in all cases), what imply that the

transferability to neighbouring areas ofour models, of either type, is deficient.Our study zones are geographicallyclose and similar in topography,climatology and landscape, at least atthe coarse level of variation measuredby the variables taken and amalgamatedfrom the tematic cartography used inthis work, so we had expected a highersuccess in the extrapolation of models.This failure may be due to unmodeledhistorical factors or local processes,such as intra- and interespecificinteractions, that adjust finely thehabitat distribution of organisms.Currently this factors constitutes anunsolved problem in wildlife-habitatrelationship modeling.

To conclude, incorporating expertopinion in the model building in theway we have done is very time-consuming and does not render bettermodels if we consider their predictiveability. Even when extrapolated toneighbouring areas models benefittingfrom expert opinion do not outperformpure unsupervised models. Consideringall this, unsupervised fitting proceduresfor building predictive habitat modelsseems an adequate cost-effective way toproceed if the aim is generating habitatsuitability maps in a regional context.

ACKNOWLEDGEMENTSThis work is a contribution to the

Project “Predictive cartography of landbirds: A pilot study in WesternAndalusia”, funded by the DirecciónGeneral de Enseñanza Superior eInvestigación Científica (Ministry ofScience and Technology) and FEDERfunds from the EU, project # 1DF-97-0648. During the work, J.S. had apredoctoral fellowship from theMinistry of Education and Culture. Theextensive field work presented in thispaper could not have been done withoutthe help and sense of humour of DanielLópez Huertas, Luis M. Carrascal, and


61

Mario Díaz who also broughtintellectual insight to the project.

NOTESA version of this manuscript,

coauthored with J.Bustamante andR.Díaz-Delgado has been submitted toConservation Biology.

LITERATURE CITED

Altman, D. G. & Royston, P. (2000).What do we mean by validating aprognostic model? Statistics inMedicine, 19: 453-473.

Bonn, A. & Schröder, B. (2001). Habitatmodels and their transfer for singleand multi species groups: a case studyof carabids in an alluvial forest.Ecography, 24: 483-496.

Brown, D. G. (1994). Predictingvegetation types at treeline usingtopography and biophysicaldisturbance variables. Journal ofVegetation Science, 5: 641-656.

Burnham, K. P. & Anderson, D. R.(1998). Model selection andinference: a practical information-theoretic approach.Springer-Verlag,New-York

Bustamante, J., Donázar, J. A., Hiraldo,F., Ceballos, O. & Travaini, A.(1997). Differential habitat selectionby immature and adult Grey Eagle-buzzards Geranoaetus melanoleucus.Ibis, 139: 322-330.

Chatfield, C. (1995). Model uncertainty,data mining and statistical inference.Journal of the Royal StatisticalSociety, Series A, 158: 419-466.

Cumming, G. S. (2000). Using between-model comparisons to fine-tune linearmodels of species ranges. Journal ofBiogeography, 27(2): 441-455.

Eastman, J. R. (1997). Idrisi forWindows. User's Guide.Clark Labs,Worcester

Eastman, J. R. (1999). Idrisi32.Reference guide.Clark Labs,Worcester

Fielding, A. H. & Bell, J. F. (1997). Areview of methods for the assessmentof prediction errors in conservationpresence/absence models.Environmental Conservation, 24: 38-49.

Fielding, A. H. & Haworth, P. F. (1995).Testing the generality of bird-habitatmodels. Conservation Biology, 9:1466-1481.

Franklin, J. (1998). Predicting thedistribution of shrub species insouthern California from climate andterrain-derived variables. Journal ofVegetation Science, 9: 733-748.

Guisan, A., Weiss, S. B. & Weiss, A. D.(1999). GLM versus CCA spatialmodelling of plant speciesdistribution. Plant Ecology, 143: 107-122.


Hanley, J. A. & McNeil, B. J. (1982).The meaning and use of the areaunder a Receiver OperatingCharacteristic (ROC) curve.Radiology, 143: 29-36.



James, F. C. & McCulloch, C. E. (1990).Multivariate analysis in ecology andsistematics: panacea or Pandora'sbox? Annual Review of Ecology andSystematics, 21: 129-166.

Lezzoni, L. I. (1999). StatisticallyDerived Predictive Models: CaveatEmptor. Journal of General InternalMedicine, 14(6): 388-389.

Ludden, T. M., Beal, S. L. & Sheiner, L.(1994). Comparison of the AkaikeInformation Criterion, the SchwarzCriterion and the F Test as guides tomodel selection. Journal of


62

Pharmacokinetics andBiopharmaceutics, 2(5): 431-445.

Mac Nally, R. (2000). Regression andmodel-building in conservationbiology, biogeography and ecology:The distinction between -andreconciliation of- 'predictive' and'explanatory' models. Biodiversityand Conservation, 9: 655-671.

Manel, S., Dias, J. M., Buckton, S. T. &Ormerod, S. J. (1999). Alternativemethods for predicting speciesdistribution: an illustration withHimalayan river birds. Journal ofApplied Ecology, 36: 734-747.

Manel, S., Dias, J. M. & Ormerod, S. J.(1999). Comparing discriminantanalysis, neural networks and logisticregression for predicting speciesdistributions: a case study with aHimalayan river bird. EcologicalModelling, 120: 337-347.

Manel, S., Williams, H. C. & Ormerod,S. J. (2001). Evaluating presence-absence models in ecology: the needto account for prevalence. Journal ofApplied Ecology, 38: 921-931.

MathSoft, I. (1999). S-Plus 2000 User´sGuide.Data Analysis ProductsDivision, Seattle

Monserud, R. A. & Leemans, R. (1992).Comparing global vegetation mapswith the Kappa statistic. EcologicalModelling, 62: 275-293.


Murtaugh, P. A. (1996). The statisticalevaluation of ecological indicators.Ecological Applications, 6(1): 132-139.

Nicholls, A. O. (1989). How to makebiological surveys go further withgeneralised linear models. BiologicalConservation, 50: 51-75.

Oreskes, N., Shrader-Frechette, K. &Belitz, K. (1994). Verification,validation, and confirmation of

numerical models in the earthsciences. Science, 263: 641-646.

Osborne, P. E. & Tigar, B. J. (1992).Interpreting bird atlas data usinglogistic models: an example fromLesotho, Southern Africa. Journal ofApplied Ecology, 29: 55-62.

Pearce, J. & Ferrier, S. (2000).Evaluating the predictive performanceof habitat models developed usinglogistic regression. EcologicalModelling, 133: 225-245.

Pearce, J. & Ferrier, S. (2000). Anevaluation of alternative algorithmsfor fitting species distribution modelsusing logistic regression. EcologicalModelling, 128: 127-147.

Pearce, J. L., Cherry, K., Drielsma, M.,Ferrier, S. & Whish, G. (2001).Incorporating expert opinion and fine-scale vegetation mapping intostatistical models of faunaldistribution. Journal of AppliedEcology, 38: 412-424.

Pons, X. (2000). MiraMon: GeographicInformation System and RemoteSensing software. Barcelona,Universidad Autónoma de Barcelona.

Rico Alcázar, L., Martínez, J. A., Morán,S., Navarro, J. R. & Rico, D. (2001).Preferencias de hábitat del Águila-azor Perdicera (Hieraaetus fasciatus)en Alicante (E de España) a dosescalas espaciales. Ardeola, 48(1):55-62.

Sakamoto, Y., Ishiguro, M. & Kitagawa,G. (1986). Akaike InformationCriterion Statistics.KTK ScientificPublishers, Tokyo

Schröder, B. & Richter, O. (1999/2000).Are habitat models transferable inspace and time? Zeitschrift fürÖkologie und Naturschutz, 8: 195-205.

Steyerberg, E. W., Eijkemans, M. J.,Harrell, F. E. & Habbema, J. D.(2000). Prognostic modelling withlogistic regression analysis: acomparison of selection andestimation methods in small data sets.


63

Statistics in Medicine, 19(8): 1059-1079.

Stockwell, D. R. B. & Peterson, A. T.(2002). Effects of sampling size onaccuracy of species distributionmodels. Ecological Modelling, 148:1-13.

Titus, K., Mosher, J. A. & Williams, B.K. (1984). Chance-correctedclassification for use in discriminantanalysis: ecological applications. TheAmerican Midland Naturalist, 111(1):1-7.

Tobalske, C. & Tobalske, B. W. (1999).Using atlas data to model thedistribution of woodpecker species inthe Jura, France. The Condor, 101:472-483.

Verbyla, D. L. & Litvaitis, J. A. (1989).Resampling methods for evaluatingclassification accuracy of wildlifehabitat models. EnvironmentalManagement, 13: 783-787.

Vida, S. (1993). A computer program fornon-parametric receiver operatingcharacteristic analysis. ComputerMethods and Programs inBiomedicine, 40: 95-101.

Zweig, M. H. & Campbell, G. (1993).Receiver-operating characteristic(ROC) plots: a fundamentalevaluation tool in clinical medicine.Clinical Chemistry, 39: 561-577.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO IV

64

CAPÍTULO IV: ¿Son adecuados los mapas de vegetación existentes

para predecir la distribución de las aves?

RESUMEN

Las especies de aves seleccionan los tipos de vegetación en los que se hallan; sin

embargo, los modelos predictivos de la distribución de aves que se basan en variables

derivadas de mapas de usos y cubiertas del suelo tienen un éxito limitado. Se ha

sugerido que la precisión de los mapas disponibles que se usan para derivar

predicciones es responsable en parte de ese éxito limitado de los modelos de

distribución. En este trabajo se compara la capacidad predictiva de modelos de

distribución de aves derivados de dos mapas de usos y cubiertas del suelo, cuyo diseño

se hizo con un propósito general y que difieren en su resolución y precisión: un mapa de

vegetación de Europa poco detallado (mapa de cubiertas del suelo del CORINE) y un

mapa regional minucioso (mapa de 1995 de cubiertas y usos del suelo del SINAMBA,

Consejería de Medio Ambiente, Junta de Andalucía). El área de estudio son dos

cuadrados de 4900 km2 en Andalucía occidental (España). Se comparan los modelos de

distribución de aves que se derivan de estos mapas de propósito general con otros

derivados de dos o más atributos estructurales de la vegetación, que se construyeron

prestando especial atención a las variables que influyen la selección del hábitat en aves.

Uno se construyó con imágenes de satélite para este estudio, mientras que el otro se

obtuvo mejorando la resolución y precisión del mapa del SINAMBA con datos de

satélite. Se muestreó la presencia/ausencia de especies de aves en 857 puntos usando

estaciones de escucha de 15 min. Se construyeron modelos predictivos para 54 especies

de aves como modelos aditivos generalizados (GAM), usando como predictores

potenciales un conjunto de variables paisajísticas y de estructura de la vegetación

medidas en cada mapa. Para cada especie se comparó la capacidad predictiva del mejor

modelo que se derivó de cada mapa. Las medidas estructurales de la vegetación medidas

en los puntos de muestreo se usaron como “verdad-terreno” (es decir, como referencia)

para comparar la precisión de los mapas de vegetación. Los resultados muestran que

sólo el mapa de cubiertas del suelo de CORINE (el más grosero) produjo modelos

significativamente peores, aunque todos los mapas difirieron en su resolución y

precisión. Los modelos derivados de los mapas detallados de la estructura de la

vegetación que se obtuvieron de los datos de satélite no fueron mejores que aquellos

Poniendo a prueba los mapas de vegetación existentes

65

que se derivaron directamente del mapa del SINAMBA. Nuestros resultados sugieren

que algunos mapas de usos y cubiertas del suelo, diseñados para satisfacer un propósito

general, son suficientemente precisos para derivar buenos modelos de la distribución de

aves, y que existe un cierto límite a la posibilidad de mejorar un mapa por encima del

cual no hay efecto sobre el poder que tienen las variables de la vegetación para predecir

la distribución de las aves.


66

CHAPTER IV: Are existing vegetation maps

adequate to predict bird distributions?

ABSTRACT

Bird species are selective on the vegetation types in which they are found but

predictive models of bird distribution based on variables derived from land-use/land

cover maps tend to have limited success. It has been suggested that accuracy of existing

maps used to derive predictors is in part responsible for the limited success of bird

distribution models. In two areas of 4900 km2 of Western Andalusia, Spain, we

compared the predictive ability of bird distribution models derived from two existing

general-purpose land-use/land-cover maps, that differ in their resolution and accuracy.

A coarse scale vegetation map of Europe, the CORINE land-cover map, and a detailed

regional map, the 1995 land-use/land-cover map of Andalusia from the SINAMBA

(Consejería de Medio Ambiente, Junta de Andalucía). We compared the bird

distribution models derived from these general-purpose vegetation maps with models

derived from two more accurate structural vegetation maps built considering directly

variables that influence bird habitat selection. One built from satellite images for this

study and another obtained by improving the resolution and accuracy of the SINAMBA

map with satellite data. We sampled the presence/absence of bird species at 857 points

using 15-min point surveys. Predictive models for 54 bird species were built with

Generalised Additive Models, using as potential predictors a set of landscape and

vegetation structure variables that was measured on each map. We compared for each

bird species the predictive accuracy of the best model derived from each map.

Vegetation structure measured at bird sample points was used as ground-truth for

comparing the accuracy of vegetation maps. The results show that, although maps

differed in their resolution and accuracy only the less accurate map, the CORINE land-

cover map, produced significantly worse bird distribution models. The models derived

from the more accurate vegetation structure maps obtained from satellite data were not

more accurate than those derived directly from the SINAMBA map. Our results suggest

that some general-purpose land-use/land-cover maps are accurate enough to derive good

bird distribution models, and that there is a certain limit to map improvement above

which there is no effect in the power of vegetation variables to predict of bird

distribution.


67

1. INTRODUCTION

Bird species are selective on thevegetation types in which they inhabit(Cody 1985). It is considered that thevegetation holds a great predictivepotential for the distribution of birds, andseveral ongoing projects are usingvegetation types to map potentialdistribution of bird species, for examplethe GAP project in USA (Scott et al.1993), or are using vegetation variables tobuild predictive models of bird distribution(Pearce & Ferrier 2000). Statistical modelsof bird distribution using as predictorsvariables derived from vegetation arerarely able to explain perfectly observeddistributions. There are several possiblereasons for this fact (Fielding & Bell 1997;Beutel, Beeton & Baxter 1999): (a)Statistical reasons (e.g. when using logisticregression to model presence/absence datathe predicted values are in a continuousscale from 0 to 1 while observed data arediscrete presences and absences). (b)Historical reasons (e.g. an species has notoccupied all potential adequate habitatbecause of geographical barriers, orbecause it has been extirpated by man fromotherwise suitable habitat). (c) Unsaturatedhabitats (small populations are not able tooccupy all suitable habitats, but alsodemographic stochasticity and localizeddispersal generate an imperfect correlationbetween habitat suitability and speciesdistribution Tyre, Possingham &Lindenmayer 2001). (d) Poor quality ofthe response variable (e.g. an inadequatecensus method for a species difficult todetect may render a distribution pattern ofobserved presences that does not reflect thereal pattern of distribution or abundance).(e) Poor quality of the predictive variables(e.g. when the predictors we are measuringare not adequate to explain the distributionof the species or they are measured withtoo much error Guisan & Zimmermann2000).

Land-use/land-cover or vegetationmaps can be used as the source of

predictive variables in statistical models ofthe distribution of bird species (Tobalske &Tobalske 1999; Guisan & Zimmermann2000; Pearce & Ferrier 2000). Maps arethemselves models of reality and as suchthey are always a simplification. Availablevegetation maps may not representadequately the vegetation variablesrelevant for the species of bird whosedistribution we want to predict, or mayhave not the adequate spatial resolution(Pearce et al. 2001). We may be able tomeasure directly at bird sampling pointsthose vegetation variables that an expert onthe species would consider more relevant,but the final model obtained will not beuseful to map the potential habitat for thespecies if the maps of these vegetationvariables are not available. Vegetation,land use and land cover maps are currentlybuilt by governmental agencies at differentresolutions and for different purposes.Existing environmental maps are cheappredictors for mapping potential habitat forbirds while the best potential predictors wemight think about may never be mapped.On the other hand, remote sensing is apotential tool for mapping the vegetationvariables that we might consider morerelevant for the distribution of a species ofbird (Palmeirin 1988; Avery & Haines-Young 1990; Franklin & Steadman 1991;Andries, Gulinck & Herremans 1994;Paruelo & Golluscio 1994; Wu & Strahler1994; Roy, Sharma & Jain 1996; Trodd1996; Ormerod & Watkinson 2000).Remote sensing imagery sensu lato(airborne sensors, aerial photography andsatellite images) is the tool most widelyused nowadays to create new land covermaps (CORINE project in Europe,MIOMBO project in Southern Africa, etc.),to improve thematic maps accuracy(Stehman 1996), or final map spatialresolution (Defries & Belward 2000).

In this paper we compare thecapacity to predict the distribution of 54species of birds of variables derived fromtwo existing vegetation maps (a coarsescale vegetation map of Europe and a more


68

detailed regional land use/land cover map),a vegetation map derived from satellitedata, and a vegetation map obtained byimproving the accuracy of the existingregional map with satellite data. It has beensuggested that higher accuracy andresolution of input maps is necessary toimprove predictions of plants and animaldistributions models (Guisan &Zimmermann 2000). As we were moreinterested in the effect of map accuracy inpredictive capacity than in the potential ofdifferent maps to measure differentpredictors we measured the same set ofpredictors in all vegetation maps.

2. STUDY AREA AND METHODS

We performed 1144 unlimited-distance bird point surveys in two 70x70km squares in Western Andalusia,

Southern Spain (centres: 6° 21’ W 37° 39’N, and 5° 28’ W 36° 44’ N), during thesprings of 1999 and 2000 (Fig. 1). Bothareas have a similar proportion of differentland-cover types and have approximately20% of cropland (mainly wheat, sunflowerand olive groves), 70% of scrubland andforests (mainly Mediterranean scrubland,evergreen oak Quercus rotundifolia andcork oak Q. suber forests and “dehesas”open oak forest with pastures). At eachsurvey point the presence/absence ofbreeding bird species was recorded during15 minutes. For subsequent modelling, weselected 857 points in natural andseminatural areas (that means, excludingthose in agricultural and urban areas), andwe selected 54 bird species that appearedin more than 5% of these sampling points.

Figure 1.- Location of study areas.


69

2.1. Vegetation maps

We used seven different vegetationmaps of each study area to derive thepredictive variables to model birddistribution: Three of them wereindependent vegetation maps: 1) CORINEland cover map of Europe from theEuropean Environmental Agency(CORINE_250). Original data in rasterformat at 250 m resolution (v.12/1989)were obtained from the European TopicCenter on Land Cover, Kiruna, Sweden.The CORINE map legend has 44 landcover classes for the whole Europe. Sourcedata correspond nominally to the period1989-91. 2) The SinambA land use/landcover digital map of Andalusia(SINAMBA_50) from the Environmental

Department of the Junta de Andalucía.Original data in vector format wererasterised to 50 m resolution. The maplegend has 112 classes. Source datacorrespond nominally to 1995. 3) Avegetation structure map derived fromsatellite images from 1999 and 2000(SATELLITE_30, see below for details).Original data were in raster format at 30 mresolution and consisted of two maps in acontinuous scale: degree of tree cover anddegree of shrub cover. We generatedanother vegetation map 4) by combininginformation of SINAMBA map andsatellite images (MIXED_30). This map at30 m resolution consisted also in a treecover and a shrub cover map and had asignificantly greater accuracy thanSATELLITE_30.

Figure 2.- Tree cover coverages from the CORINE_250, SINAMBA_50 and SATELLITE_30 maps for one ofthe study areas.


70

To test whether differences inpredictive accuracy were due to differencesin data quality, predictors used, or mapresolution we generated another threemaps: 5) SINAMBA_250 was obtainedresampling the SINAMBA_50 map at 250m resolution. 6) SINAMBA_250R wasalso obtained by resampling theSINAMBA_50 map at 250 m resolutionbut then using only the reduced set ofpredictors that could be measured inCORINE_250. 7) SATELLITE_50 wasgenerated by resampling theSATELLITE_30 map at 50 m, the sameresolution as SINAMBA_50.

In each vegetation map the originallegend categories (or the values in acontinuous scale) were reclassed into three

categories of shrub cover and threecategories of tree cover: 1) no cover, 2)disperse cover, and 3) dense cover (Fig. 2).With the help of a GIS we derived fromeach vegetation map at each bird surveypoint the set of vegetation structure andlandscape predictors indicated in Table 1.These predictors included variablesdescriptive of vegetation structure in acircle of 350 m diameter centred in the birdsurvey point and variables indicatingdistances to landscape features of differentsizes. The same predictors were measuredin each map with a few exceptions. TheCORINE_250 map legend did notdistinguish clearly between disperse anddense tree cover and between disperse and

Table 1. List of variables measured in each map and used in the initial sets of potential predictors of birddistribution models.

Variable descriptionShrub cover fraction +

Forest cover fraction (including dense and disperse forest categories like “dehesas”)Dense forest fraction *Length of boundaries between forested landcover categories and the rest of vegetation categories *Length of boundaries between forest and shrublandCompactness ratio of dense forest areas (an indirect estimate of surface-perimeter ratio) *Presence/absence of disperse tree cover (for example, included in a heterogeneous shrubland area) *Presence/absence of dense tree cover (for example, included in a heterogeneous shrubland area) *Presence/absence of disperse shrubland (for example in areas dominated by forest) *Presence/absence of dense shrubland (for example in areas dominated by forest) *Distance to the nearest patch of shrubland (dense or disperse) ++

Distance to the nearest patch of forest (dense or disperse) ++

Distance to the nearest patch of dense forest * ++

+ Variable values are estimated in a circle of 350 m radius centered in bird point surveys.For example, shrubcover fraction is the fraction of 30 or 50 m pixels in the circle that have dense or disperse shrub cover (In mapsof 250 m pixel resolution the value can only be 0 or 1).

* Variables that could not be measured in the CORINE_250 map and consecuently were also excluded aspotential predictors in the SINAMBA_250R models.

++ Each distance was 4 variables: (1) distance to the nearest patch of any size, (2) distance to the nearest patch 2-10 ha. in size, (3) distance to the nearest patch 10-100 ha. in size, (4) distance to the nearest patch > 100 ha.


71

dense shrub cover in its legend, and somepredictors involving these variables couldnot be measured (see Table 1). For thisreason we generated the SINAMBA_250Rmap that had the same predictors asCORINE_250. Also, a few predictorschanged their range of possible valueswhen measured at 250 m resolution.

Extraction of variables from theGIS was done using IDRISI 32 (Eastman1999), IDRISI for Windows (Eastman1997) and MIRAMON (Pons 2000).

2.2. Vegetation maps derived from satellitedata

At each bird survey point theobserver recorded structural attributes ofvegetation that we had considered a prioriimportant for bird distribution (Table 2).We performed a Principal ComponentsAnalysis (PCA) on these structuralvariables (excluding the survey points inagricultural or urban areas). The two firstcomponents explained 32.9 and 23.2 %respectively of the original variance (a totalof 56.2%). The first component defined agradient of tree cover (high loads forvariables: Mean tree DBH, and cover oftrees > 6m high). The second componentdefined a gradient of shrub cover (highpositive loads for variables like cover ofshrubs < 0.50 m tall, and cover of shrubs0.5 to 2m tall, and high negative load forcover of herbaceous vegetation). Then weused the first two components of the PCAas the response variable in a generalisedadditive model (GAM, Hastie & Tibshirani1990) with normal errors and identity link.We used as predictors reflectance values ofbands 1 to 7, and NDVI of three Landsatscenes (TM and ETM+) for each studyarea corresponding to early spring, midspring and summer of the years 1999 and2000. Images were geometrically correctedwith the aid of a Digital Elevation Model(Palà & Pons 1995) and radiometricallycalibrated according to Pons and Solé-

Sugrañes model (1994). GAM models (J.Bustamante and R.Díaz-Delgado, unpub.data) explained 37-40 % variance of thetree cover gradient (each study arearespectively) and 21-30 % variance of theshrub cover gradient. GAM modelspredicted tree cover and shrub cover valuesin a continuous scale (0-255) for each 30 mpixel in the study area. We selected cut-points in this gradient to recode tree andshrub cover in three classes (no cover,disperse cover, and dense cover), so thatsurface covered by each tree cover andshrub cover class was as close as possibleto that of the SINAMBA_50 map. Theresulting coverages defined theSATELLITE_30 vegetation map. The treecover models and shrub cover modelsimproved significantly if the land-use/land-cover class of the SINAMBA map at thelocation of each sampling point wasincluded as a factor. We refitted the GAMmodels for the tree cover and shrub covergradients of each study area using theSINAMBA class as a factor , satellitereflectance values and NDVI values. Thesenew GAM models explained 55-56 % ofthe variance in tree cover and 26-49% ofthe variance in shrub cover. The gradientswere reclassified to three discrete classesand generated the MIXED_30 map.

2.3. Predictive models for birds

We built a generalised additivemodel (GAM, Hastie & Tibshirani 1990)for the presence/absence of each species ineach study area with binomial errors andlogit link using as predictors the variablesin Table 1. Seven models were generatedfor each bird species with the predictorsderived from each one of the sevenvegetation maps. We selected the variablesto include in the models with a forward-backward stepwise selection from thecomplete set of predictive variablesmeasured from each map (with thestep.gam procedure of S-PLUS 2000,MathSoft 1999). We started from a nullmodel and tested each predictor


72

sequentially as a smoothing spline with 3degrees of freedom. The predictor thatreduced the most the residual deviance wasincluded in the model and the procedurewas repeated until no more predictorsimproved the model. Then, we tried tosimplify the resulting model by decreasingthe complexity of each of the predictorsincluded (by means of a smoothing splinewith 2 degrees of freedom and a linearterm). The criteria to enter, remove orsimplify a term was the Akaike’sInformation Criterion (AIC Sakamoto,Ishiguro & Kitagawa 1986), that takes intoaccount the reduction both in residualdeviance and in residual degrees offreedom due to a certain predictor.Automatic procedures for selection ofpredictors have been criticised becausethey can yield ecologically implausiblemodels (Greenland 1989; James &McCulloch 1990); but it is a method thatallows for rapid development of models(Pearce & Ferrier 2000), and it has beenshown empirically that frequently performbetter than tedious manual selectiontechniques incorporating opinion of experts(Pearce et al. 2001). In our study, therandom inclusion of spurious correlationsin the predictive models could affectequally the models derived from each mapand would not bias the comparitionbetween models.

2.4. Comparition of predictive accuracy ofmaps

The predictive ability of eachmodel was assessed by the Area Under theCurve (AUC) of Receiver OperatingCharacteristics (ROC) plots (Murtaugh1996; Pearce & Ferrier 2000). AUC wascalculated with AccuROC 2.5 (Vida 1993).The interest of the analysis is in thepotential differences in predictiveperformance of the models generated withdifferent data sources, and not in theabsolute values of AUC, therefore we didnot evaluate the data with an independentdata set. Differences among model types

were tested with a repeated measuresfactorial ANOVA (with an error term dueto species to control for the between-species variation, of no interest in thisstudy). Preplanned comparisons(Montgomery 2001) were performed totest differences between particular models.

First we compared the differentvegetation maps to see if they differed intheir accuracy regarding the structuralvegetation classes defined, using asground-truth the vegetation data measuredat the 857 bird survey points. Then wetested if there were differences inpredictive accuracy of bird distributionrelated to the original map source ofpredictors: CORINE_250, SINAMBA_50or SATELLITE_30. Then we tested ifdifferences in predictive ability betweenCORINE and SINAMBA maps were dueto differences in: map quality (comparingCORINE_250 vs. SINAMBA_250R),predictors used (SINAMBA_250R vs.SINAMBA_250), or map spatial resolution(SINAMBA_250 vs. SINAMBA_50).Then we tested if there were anydifferences in predictive accuracy relatedto a difference in spatial resolution of 50 to30 m (SATELLITE_30 vs.SATELLITE_50) or if a more accuratevegetation map derived from two sources(MIXED_30) differed in predictiveaccuracy from the original maps(SINAMBA_50 and SATELLITE_30) .

3. RESULTS

3.1. Accuracy of vegetation maps

Each sampling point was classifiedinto one of nine exclusive categories(Table 3) using the coordinates in the treecover and shrub cover gradients of thePCA and the cut-points selected for thesatellite vegetation maps. These pointswere used as ground-truth for allvegetation maps. A confusion matrix wasgenerated comparing ground-truthclassification with classification from each


73

map. Percentage of agreement and Kappavalues (classification rate corrected forchanceTitus, Mosher & Williams 1984)indicated that greatest map quality (oraccuracy) corresponded to the MIXED_30map. Map quality declined in this orderMIXED_30 > SATELLITE_30 >SATELLITE_50 > SINAMBA_50 >SINAMBA_250 > CORINE_250 (Table 4)

3.2. Bird distribution models

It was possible to build predictivemodels significantly better than a nullmodel for 48 out of 54 species of birdsusing maps at a spatial resolution of 250 m.All bird species gave models better thanthe null model when predictors werederived from maps at spatial resolution of50 or 30 meters. Mean AUC for each mapranged from 0.59(SE=0.05) forCORINE_250 to 0.80(SE=0.06) forSINAMBA_50 (Table 5).

There were significant differencesin bird predictive ability (AUC values)when comparing the models derived fromdifferent data sources (CORINE_250 vs.SINAMBA_50 vs. SATELLITE_30)(Table 6). CORINE_250 gave birddistribution models of significantly lowerpredictive accuracy than SINAMBA_50,while SINAMBA_50 and SATELLITE_30did not differ. CORINE_250 differed fromSINAMBA_50 in map quality, the numberof predictors derived, and the spatialresolution of the source map. To study theeffect of each of these factorsindependently we compared the modelsderived from CORINE_250,SINAMBA_250R, SINAMBA_250 andSINAMBA_50 (Table 7). There weresignificant differences in predictive abilityof models derived from each map. Plannedcomparitions indicated that differences

Table 2. Vegetation variables measured at bird survey points in a circle of 50 m radius.

Variables possible valuesCover of herbaceous vegetation <10%, 10-50%, >50%Cover of shrubs < 0.5 m tall absence, < 25%, > 25%Cover of shrubs 0.5-2 m tall absence, < 25%, > 25%Cover of trees 2-6 m tall absence, < 25%, > 25%Cover of trees > 6 m tall absence, < 25%, > 25%Mean diameter at breast height (DBH) of the 5 biggest trees m (continuous)Number of trees with DBH > 0.2 m in a circle 25 m radius integer

Table 3. Categories used in vegetation maps (structural categories) to compare map quality (accuracy).

Categoriesno tree cover no shrub coverno tree cover disperse shrub coverno tree cover dense shrub coverdisperse tree cover no shrub coverdisperse tree cover disperse shrub coverdisperse tree cover dense shrub coverdense tree cover no shrub coverdense tree cover disperse shrub coverdense tree cover dense shrub cover


74

could be attributed in this order: first tomap source quality (CORINE_250 vs.SINAMBA_250R, F = 937.59, P < 0.001),second to the reduced set of predictors thatcould be derived from the CORINE map(SINAMBA_250 vs. SINAMBA_250R, F= 39.37, P <0.001) and third to mapresolution (SINAMBA_50 vs.SINAMBA_250, F = 32.80, P < 0.001).Improving the quality of the SINAMBAmap by generating a MIXED map, ordegrading the spatial resolution of theSATELLITE map from 30 to 50 m had nosignificant effect on the predictive abilityof bird models derived from each map(Table 8). Planned comparitions indicatedthat there was a statistically significant (but

very small) difference attributed to thesource (SINAMBA_50 vs.SATELLITE_50, F = 4.75, P = 0.031) butthis was not related to map quality.SATELLITE_50 was a slightly better map(Table 4) but rendered predictive models ofbird distribution with slightly lower AUC.There was no significant effects of eithermap resolution (SATELLITE_30 vs.SATELLITE_50, F = 3.80, P = 0.05), ormap quality improvement (SINAMBA_50vs. MIXED_30, F = 0.17, P = 0.7,SATELLITE_30 vs. MIXED_30, F = 0.68,P = 0.4).

Table 4. Percentage of agreement between map categories and ground-truth data (n= 857) for each map, andKappa values that indicate percentage improvement over a random classification.

Map Percentage ofagreement *

Kappa Sk (C.I. 95%) z value (p)

CORINE_250 36.6 0.01 0.06 0.22 (=0.41)SINAMBA_250R 44.4 0.12 0.05 4.39 (<0.001)SINAMBA_50 20.3 0.12 0.03 10.12 (<0.001)SATELLITE_50 23.7 0.14 0.03 11.37 (<0.001)SATELLITE_30 26.5 0.17 0.03 13.56 (<0.001)MIXED_30 29.7 0.22 0.03 19.16 (<0.001)

* Percentage agreement value of CORINE_250 can only be compared with SINAMBA_250R that has a reducedset of 4 classes also: no tree cover no shrub cover, no tree cover disperse or dense shrub cover, disperse ordense tree cover no shrub cover , and disperse or dense tree cover disperse or dense shrub cover.

Table 5. Bird prediction accuracy, mean AUC (and SE)values, for the models generated with the different maps

Model (map source) AUC (SE)CORINE_250 0.595 (0.051)SINAMBA_250R 0.742 (0.069)SINAMBA_250 0.788 (0.070)SINAMBA_50 0.801 (0.059)SATELLITE_50 0.798 (0.065)SATELLITE_30 0.789 (0.071)MIXED_30 0.799 (0.070)


75

Table 6. Results of repeated measures ANOVA testing the effect in bird prediction accuracy (model AUC) oforiginal independent map sources (CORINE_250, SINAMBA_50 and SATELLITE_30), and plannedcomparisons between them.

Variable Df SS MS F P

Error: species

Residuals 53 0.409 0.008 - -

Error: within

Map source 2 1.446 0.723 432.17 <0.0001CORINE_250 vs. SINAMBA_50 1 1.442 1.442 861.86 <0.0001SATELLITE_30 vs. SINAMBA_50 1 0.004 0.004 2.47 0.118

Residuals 106 0.177 0.0017

4. DISCUSSION

Our original vegetation maps show agradient in quality (accuracy) for severalreason. The CORINE map has a coarserresolution (250 m), has a reduced set ofland cover classes (44 for the wholeEurope) it does not reflect well differencesin vegetation structure (for example, mostclasses do not distinguish between denseand disperse tree and shrub cover) and isten year older than our ground-truth data.The SINAMBA map has a finer spatialresolution (50 m), has more land coverclasses (that are easier to reclassify asdisperse or dense tree and shrub cover), butis five year older that ground-truth data.The SATELLITE map has the finer spatialresolution (30 m), is contemporaneous withground truth data (1999-2000) and usesmodels to discriminate directly thestructural variables we were interested in(degree of tree cover and degree of shrubcover). The confusion matrix of mapclassification and ground-truth data forsampling points indicates that mapaccuracy increases in a gradient CORINE< SINAMBA < SATELLITE.Interestingly, measuring the samepredictors on each map --that reflect

characteristics of vegetation structure in a350 m radius around the bird survey points,and distances to landscape features-- wefind that CORINE is a significantly poorerpredictor of bird distribution but thatSINAMBA and SATELLITE maps do notdiffer in their predictive ability. Thedifferences in predictive ability betweenCORINE and SINAMBA can be attributedmainly to the difference in quality(accuracy), and to a lesser extent to boththe difference in the number of predictorsthat can be measured on each map, and thedifference in resolution between the twomaps. SINAMBA and SATELLITEvegetation maps do not differ in theirpredictive accuracy of bird distribution,even though the SATELLITE map wastemporally closer to our bird survey, had afiner spatial resolution, and actually betteragreement with ground-truth data. Evenimproving the SINAMBA map withsatellite data (MIXED_30) does notimprove the predictive performance ofmodels. The MIXED_30 map derives frommodels that are significantly better thanthose from the SATELLITE map, and alsoshows a better agreement with ground-truth data (Table 4).


76

Table 7. Results of repeated measures ANOVA testing the effect in bird prediction accuracy (model AUC) ofmap source type, and planned comparisons between them to test the effect of map quality, spatial resolution andset of predictors derived.

Variable Df SS MS F PError: speciesResiduals 53 0.605 0.011 - -

Error: withinMap source type 3 1.178 0.393 308.25 <0.0001

quality (CORINE_250 vs. SINAMBA_250R) 1 1.345 1.345 937.59 <0.0001resolution (SINAMBA_50 vs. SINAMBA_250) 1 0.047 0.047 32.80 <0.0001predictors (SINAMBA_250 vs. SINAMBA_250R) 1 0.056 0.056 39.37 <0.0001

Residuals 159 0.228 0.0014

Table 8. Results of repeated measures ANOVA testing the effect in bird prediction accuracy (model AUC) ofmap source type, and planned comparitions between them to test the effect of map quality, resolution and mapimprovement (map sources: SINAMBA_50, SATELLITE_50, SATELLITE_30 and MIXED_30).


Error: withinMap source type 3 0.0049 0.0016 2.91 0.036

source (SINAMBA_50 vs SATELLITE_50) 1 0.0027 0.0027 4.75 0.031resolution (SATELLITE_30 vs SATELLITE_50) 1 0.0021 0.0021 3.80 0.053improvement (MIXED_30 vs SINAMBA_50) 1 0.0000 0.0000 0.17 0.678

Residuals 159 0.0889 0.0006

Our results show that there aredifferences in information content ofavailable, general purpose, vegetationmaps that affect the predictive ability ofbird distribution models. Differences inpredictive ability of the two maps wecompared were due to differences in mapquality (accuracy to distinguish vegetationstructural classes) but also the number ofland cover classes initially distinguishedand the difference in spatial resolutioninfluenced the result. Our results also showthat there is a certain limit to mapimprovement. The SINAMBA landuse/land cover map is clearly not perfect,but increasing its resolution, or improvingits accuracy with ancillary data likesatellite images does not improve birddistribution models. Our results are

encouraging because existing vegetationmaps produced with a general purpose canshow a relatively high predictive ability ofbird distribution. It has to be consideredthat demographic stochasticity anddispersal can prevent a perfect adjustmentbetween predictive models and wildlifedistribution as Tyre et al. (2001) haveshown with computer simulation models.Also, it is interesting to note that whenmaps of enough accuracy and resolutionare not available there is the alternative ofderiving vegetation structuralcharacteristics from satellite images

ACKNOWLEDGEMENTS

This work is a contribution to theProject “Predictive cartography of land


77

birds: A pilot study in Western Andalusia”,funded by the Dirección General deEnseñanza Superior e InvestigaciónCientífica (Ministry of Science andTechnology) and FEDER funds from theEuropean Union (EU), project # 1DF-97-0648. During the work, J.S. had apredoctoral fellowship from the Ministryof Education and Culture. The extensivefield work presented in this paper could nothave been done without the help and senseof humour of Daniel López Huertas, LuisM. Carrascal, and Mario Díaz who alsobrought intellectual insight to the project.We are grateful to the EnvironmentalDepartment of the Junta de Andalucía forproviding access to the 1995 SinambA landuse/land cover map of Andalusia and to theEuropean Environmental Agency (EEA)and the European Topic Center on LandCover (ETC/LC) for providing access tothe CORINE Land Cover databases.

NOTESA version of this chapter has beensubmitted to Ecological Modelling, withJ.Bustamante and R.Díaz-Delgado.

REFERENCESAndries, A. M., Gulinck, H. & Herremans,

M. (1994). Spatial modelling of the barnowl Tyto alba habitat using landscapecharacteristics derived from SPOT data.Ecography, 17: 278-287.

Avery, M. I. & Haines-Young, R. H.(1990). Population estimates for thedunlin Calidris alpina derived fromremotely sensed satellite imagery of theFlow Country of northern Scotland.Nature, 344: 860-862.

Beutel, T. S., Beeton, R. J. S. & Baxter, G.S. (1999). Building better wildlife-habitat models. Ecography, 22: 219-223.

Cody, M. L., Ed. (1985). Habitat selectionin birdsAcademic Press, Inc., . SanDiego

Defries, R. S. & Belward, A. S. (2000).Global and regional land cover

characterization from satellite data: anintroduction to the Special Issue.International Journal of RemoteSensing, 21: 1083-1092.

Eastman, J. R. (1997). Idrisi for Windows.User's Guide.Clark Labs, Worcester

Eastman, J. R. (1999). Idrisi32. Referenceguide.Clark Labs, Worcester

Fielding, A. H. & Bell, J. F. (1997). Areview of methods for the assessment ofprediction errors in conservationpresence/absence models.Environmental Conservation, 24: 38-49.

Franklin, J. & Steadman, D. W. (1991).The potential for conservation ofPolynesian birds through habitatmapping and species translocation.Conservation Biology, 5: 506-519.

Greenland, S. (1989). Modelling variableselection in epidemiologic analysis.American Journal of Public Health, 79:719-732.

Guisan, A. & Zimmermann, N. E. (2000).Predictive habitat distribution models inecology. Ecological Modelling, 135:147-186.

Hastie, T. J. & Tibshirani, R. J. (1990).Generalized Additive Models.Chapman& Hall, London

James, F. C. & McCulloch, C. E. (1990).Multivariate analysis in ecology andsistematics: panacea or Pandora's box?Annual Review of Ecology andSystematics, 21: 129-166.

MathSoft, I. (1999). S-Plus 2000 User´sGuide.Data Analysis Products Division,Seattle

Montgomery, D. C. (2001). Design andanalysis of experiments.John Wiley &Sons, New York

Murtaugh, P. A. (1996). The statisticalevaluation of ecological indicators.Ecological Applications, 6: 132-139.

Ormerod, S. J. & Watkinson, A. R. (2000).Editors' Introduction: Birds andAgriculture. Journal of AppliedEcology, 37(5): 699-705.

Palà, V. & Pons, X. (1995). Incorporationof relief into a geometric correctionbased on polynomials.


78

Photogrammetric Engineering andRemote Sensing, 61: 935-944.

Palmeirin, J. M. (1988). Automaticmapping of avian species habitat usingsatellite imagery. Oikos, 52: 59-68.

Paruelo, J. M. & Golluscio, R. A. (1994).Range assessment using remote sensingin Northwest Patagonia (Argentina).Journal of Range Management, 47(6):498-502.

Pearce, J. & Ferrier, S. (2000). Evaluatingthe predictive performance of habitatmodels developed using logisticregression. Ecological Modelling, 133:225-245.

Pearce, J. & Ferrier, S. (2000). Anevaluation of alternative algorithms forfitting species distribution models usinglogistic regression. EcologicalModelling, 128: 127-147.

Pearce, J. L., Cherry, K., Drielsma, M.,Ferrier, S. & Whish, G. (2001).Incorporating expert opinion and fine-scale vegetation mapping into statisticalmodels of faunal distribution. Journal ofApplied Ecology, 38: 412-424.


Pons, X. & Solé-Sugrañes, L. (1994). Asimple radiometric correction model toimprove automatic mapping ofvegetation from multispectral satellitedata. Remote Sensing of Environment,48: 191-204.

Roy, P. S., Sharma, K. P. & Jain, A.(1996). Stratification of density in drydeciduous forest using satellite remotesensing digital data - an approach basedon spectral indices. Journal ofBiosciences, 21: 723-734.

Sakamoto, Y., Ishiguro, M. & Kitagawa,G. (1986). Akaike Information CriterionStatistics.KTK Scientific Publishers,Tokyo

Scott, J. M., Davies, F., Csuti, B., Noss, R.,Butterfield, B., Groves, C., Anderson,H., Caicco, S., D'Erchia, F., EdwardsJr., T. C., Ulliman, J. & Wright, R. G.(1993). GAP Analysis: a geographicapproach to protection of biologicaldiversity. Wildlife Monographs, 123: 1-41.

Stehman, S. V. (1996). Use of AuxiliaryData to Improve the Precision ofEstimators of Thematic Map Accuracy.Remote Sensing of Environment, 58:169-176.

Titus, K., Mosher, J. A. & Williams, B. K.(1984). Chance-corrected classificationfor use in discriminant analysis:ecological applications. The AmericanMidland Naturalist, 111(1): 1-7.

Tobalske, C. & Tobalske, B. W. (1999).Using atlas data to model thedistribution of woodpecker species inthe Jura, France. The Condor, 101: 472-483.

Trodd, N. M. (1996). Analysis andrepresentation of heathland vegetationfrom near-ground level remotely-senseddata. Global Ecology and BiogeographyLetters, 5: 206-216.


Vida, S. (1993). A computer program fornon-parametric receiver operatingcharacteristic analysis. ComputerMethods and Programs in Biomedicine,40: 95-101.

Wu, Y. C. & Strahler, A. H. (1994).Remote estimation of crown size, standdensity, and biomass on the Oregontransect. Ecological Applications, 4:299-312.

Selección de la resolución espacial de los predictores

79

CAPÍTULO V: La elección de la mejor resolución espacial en los modelos

predictivos de la distribución de aves

RESUMEN

¿Cuál es el grano apropiado para generar modelos predictivos de la distribución de especies?.

Se ha sugerido que el grano, o máxima resolución espacial, debería seleccionarse de acuerdo

con el uso que se espera hacer del modelo, o bien con la percepción del ambiente que se

asuma que tenga la especie. Sin embargo, no existe todavía una guía cuantitativa en la que

basarse para seleccionar el grano de medición de predictores dentro del contínuo entre los

extremos de baja y de alta resolución. En este trabajo exploramos el efecto que tiene medir

predictores ambientales a diferentes resoluciones espaciales sobre la capacidad discriminativa

de distintos modelos de la distribución reproductora de aves. Los modelos más

discriminativos fueron en promedio aquellos que se hicieron con predictores medidos en

círculos de gran diámetro (es decir, con baja resolución espacial: 2450 m). El área de estos

círculos (471 ha) fue generalmente mucho mayor que el tamaño del área de campeo para la

mayor parte de las especies (con la excepción de las rapaces), lo que sugiere que la

probabilidad de detectar a una especie en particular depende de las características del hábitat

en un entorno amplio, y no sólo en los aspectos particulares a un punto de muestreo. Se

discute después que este resultado puede deberse a efectos de la dinámica metapoblacional

sobre los patrones de distribución.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO V

80

CHAPTER V: Choosing the best spatial resolution

for predictive models of bird distribution

ABSTRACT

What is the appropriate grain for predictive models of species distribution? The grain, or

maximum spatial resolution, has been suggested to be selected according to either the

expected use of the model, or to the assumed species’ perception of the environment.

However, there is no widely available quantitative guidelines to select the grain of predictors

from a low-high resolution continuum. In this work we explore the effect of measuring

environmental predictors at different spatial resolutions on the discriminative ability of

breeding bird distribution models. The more discriminative models were on average those

made with predictors measured in circles of large diameters (i.e., low grain: 2450 m). The area

of these circles (471 ha) were generally much larger than the area of home range for most

species, with the exception of raptors, what suggest that the probability of detecting a

particular species depend on habitat characteristics measured in an wide neighboring area, and

not only on in-site specific features of the sampling point. We discuss that this result may be

due to metapopulation dynamics on patterns of occurrence.


81

INTRODUCTION

Scale is a main theme in ecology(Levin 1992) because of its wideimplications in theoretical and practicalstudies (Peterson & Parker 1998). Scaleis also such a central issue in predictivehabitat modeling (Guisan &Zimmermann 2000) that it has recentlybeen the focus of a entire symposium onthe subject (Scott et al. in press). And,among the multiple facets of scale,modelers are mainly concerned with theselection of an appropriate grain tomeasure environmental variables thatwill serve as predictors of speciesdistribution (Morrison, Marcot &Mannan 1998, p. 141). The grain, ormaximum spatial resolution (O'Neill &King 1998), has been suggested to beselected according to the expected useof the model (Morrison, Marcot &Mannan 1998). Thus, variablesmeasured at low spatial resolution(typically broad macrohabitatsummaries) suit models of generaldistribution, while variables measuredwith detail at high resolution (generallymicrohabitat descriptors) are adequatefor models of spatial use on particularhabitats or areas. However, there is nowidely available quantitative guidelinesto select the grain of predictors from thelow-high resolution continuum.Besides, an alternative to this criterionis to consider the species’s perception ofits environment. Animal’s habitatselection, according to the hierachicaltheory (Johnson 1980; Holling 1992),proceeds first at the broad level ofgeographical area, then at the homerange level, then within specific siteswithin home ranges, and finally atmicrosites --or specific features ofselected sites (see two examples inHolling 1992, for Egretta alba y ). Thus,the appropriate grain for predictors maybe chosen depending on our interest inone or other level of habitat selection.

Moreover, a practical considerationmust be made to select the appropriategrain for predictors in habitatdistribution modelling. There are severelogistic and budget limits to both thespatial extent and the grain used torecord habitat features during fieldwork;besides, sample units may be difficult tostandardize. For example, in a study ofbirds of the Baja California Xerophyticscrub (R. Rodríguez-Estrella and J.Bustamante, unpublished), one of us(JB) employed between 30 to 60minutes to reach each aleatorily-selected sampling point and more than120 minutes to record habitat variables(vegetation floristic and structure)within an 50x5 m linear transect.Measuring vegetation on a bigger areaseemed desiderable but impractical. Onthe contrary, logistics and budget arenot so limiting if environmentalpredictors are measured within a GISframework, and researchers may chooseamong almost limitless possibilities forthe grain of predictors (though, grain isconditioned to be a multiple of theoriginal spatial resolution of thematiclayers). In the example above, habitatvariables could have been measuredequally almost effortlessly in circles ofdiameter of, say, 100, 200 or 500meters, so which one to use then?

Many studies on predictive modelsof bird distribution for a single speciesmeasure environmental factors andspecies distribution at the sameresolution, approximately that of thespecies home range (i.e.: González,Bustamante & Hiraldo 1992; Donázar,Hiraldo & Bustamante 1993;Bustamante 1997), othersopportunistically measure both speciesdistribution and environmental factorsat the resolution provided by preexistingregional atlas (typically with 1x1 or10x10 km squares). When using point-surveys to record presence/absence orabundance of several bird species


82

simultaneously, two strategies arepossible to record bird data. (1) To use afixed radius detection circle, so that the

census area is known and it is the samefor all species. This has the

Figure 1. Scale has two dimensions: grain and extent (O'Neill & King 1998). In the context of predictivehabitat modelling the extent is the area of study while the grain is the minimum unit in which the area ofstudy is divided to measure environmental predictors. The figure is a idealized representation of threepossible choices of grain for environmental variables in predictive habitat modelling. Hatched areasdelimit home ranges for a single individual of a species, circles are the grain. Inside circles, it is shownthe radius (thick line) and the area in which detections of birds are possible (crosses): a bird is recorded ifthe cross hit the home range area. a) grain approximately equal to home range. b) grain approximatelyequal to several neighboring home ranges. In both a and b the left figure shows that most of the areasampled for predictors is suitable for the species, while in the right figure the species is detected at theborder of home range (or group of neighboring home ranges) and thus a large part of the area sampled forpredictors is unsuitable for the species (we expect that, in the average, center of sampling circles coincidewith center of home ranges (or group of home ranges). c) large grain approximately equal to the extent ofa local subpopulation.

XX

XX

a

b

c

grain

X

extent


83

disadvantage that contacts with birdsoutside the circle are wasted. (2) To usean unlimited detection circle. Thismakes an optimal use of all datarecorded, but has the disadvantage thatcensus area is unknown and varies withspecies body-size, because larger birdsare seen or heard from greater distance(Calder III 1990). Census-area size mayinfluence the grain at whichenvironmental variables have greaterpredictive power of birds-speciesoccurrences. In that case the grain atwhich environmental predictors giveoptimal predictive power should besmaller when using fixed-radius vs.unlimited-radius designs. If speciesbody size, by its effect on species home-range size, affects the grain at whichenvironmental predictors give optimalpredictive power this effect should begreater with unlimited-radius designs inwhich the effect of home-range size andsize of censused area add to each other.

Even though, it is not necessarilyclear if the grain at which bird species ishabitually thought to perceive theenvironment –the size of the homerange—is of the same magnitude as thegrain at which environmental variableshave optimal predictive power of theirdistributions (fig. 1). An individual of abird species might be present at asampling point depending exclusivelyon the environmental characteristics ofthe habitat in a radius that encompasesits home range (that is, in-site specifichabitat features could affect the mostthe probability of occurrence or theabundance of the species). If weconsider the effect of conspecificattraction and dispersal it may be moreimportant for the presence of theindividual at the sampling point(Goodwing & Fahrig 1998) the habitatin a larger circle that encompases theterritories of several conspecifics orperhaps a whole subpopulation. In sucha case, species characteristics that

influence abundance and dispersalmight be more important to determinethe grain at which environmentalpredictors give optimal predictivepower than home-range size or theeffect of body-size in detection distance.

In this work we explore the effectof measuring environmental predictorsat different spatial resolutions (or with adifferent grain) on the discriminativeability of breeding bird distributionmodels. We assume that the grain atwhich models are more discriminativeshows the level at which theenvironment is affecting more theobserved pattern of species distribution.Our aims with this study are to explore:(1) if the grain at which environmentalvariables are measured has an effect onthe predictive power of bird occurrencemodels, (2) if sampling-area size has aneffect on the grain at whichenvironmental predictors have greaterpredictive power (fixed vs unlimited-radius designs), (3) if optimal grain isequivalent or proportional to specieshome-range size, or is affected byfactors operating at larger spatial scales,(4) if species body-size or otherecological factors influence the grain atwhich environmental variables haveoptimal predictive power. Ourcomparison across spatial resolutionsserve also as a guide of how muchpredictive ability one is losing whenselecting a certain grain to measurepredictors.

METHODS

Study area

The study area are two 70 x 70 kmsquares in Western Andalusia, SouthernSpain. Both of them include low, flatareas, mainly devoted to agriculture,surrounded by mountainous areas withmore natural vegetation (altitude rangesfrom 0 to 1600 m a.s.l.). In both areas


84

land-cover is mainly dominated(approximately 70%) by Mediterraneanvegetation (shrubland and evergreenoak Quercus ilex subsp. ballota L. andcork oak Quercus suber L. forests anddehesas), and pine and eucalyptusplantations, and 30% is mainlyagricultural land, mainly non-irrigatedwheat and sunflower crops and olivegroves. Villages and urbanized areas arewidely interspersed.

We analyzed occurrence of 79species detected in 1144 point surveysof 15 min duration made between Apriland June in 1999 and 2000. Wedifferentiated among birds detectedwithin a circle of 100 m diametercentered in the sampling point, andbirds detected outside the circle (wepooled the data for some analyses). Birdspecies were mostly passerines (80%),and their sizes ranged between FirecrestRegulus ignicapillus Temminck (~5,3gr) to Short-toed eagle Circaetusgallicus Gmelin (~1700 gr). Weights (asurrogate for body size) and homeranges for each species were taken fromPerrins (1998). We selected weights andhome range sizes for the samesubespecies and the location closest to,and climatically more similar to, ourstudy area, and averaged values acrosssexes (or areas) if appropriate(Appendix 1). Home range estimateswere of varying quality, and a numberof species lacked reliable measures; insome cases we had to estimate homerange size as the inverse of populationdensity. Moreover, the distinctionbetween territory size (the areasurrounding the nest that is activelydefended) and home range (the totalarea used by a breeding pair to forage)was not clear for all species. For thesereasons we focused the analysis onweights rather than on home ranges,assuming that they are well correlated(Peters 1983).

Environmental predictors used asexplanatory variables in the models(table 1 and see below) were a set ofland-cover variables extracted andamalgamated from the 1995 land-use/land-cover digital map of Andalusia(SinambA), provided by theEnvironmental Department of the Juntade Andalucía (Moreira & Fernández-Palacios 1995). Original data in vectorformat were rasterized to 50 m spatialresolution. To simulate differentavailable spatial resolutions ofexplanatory variables, predictors wereaveraged in circles of increasingdiameter (150, 350, 650, 1250, 2450and 4850 m, fig. 2) centered insampling points, and models were builtseparately for each diameter. For someanalyses we considered diameter as anordered factor with 6 levels.

Statistical analyses

We performed GeneralizedAdditive Models (Hastie & Tibshirani1990) of presence/absence of each birdspecies for each diameter, usingbinomial errors and logit link.Explanatory variables for each modelwere selected from each set of potentialpredictors (table 1) by a forward-backward stepwise procedure (with thestep.gam function of S-PLUS 2000software, MathSoft 1999) that used anapproximation of AIC (Akaike´sInformation Criterion, Burnham &Anderson 1998) as the criterion to enteror to remove variables. Predictors wereallowed to enter the models as linearterms or as smoothing splines with 2 or3 degrees of freedom (to achieve,respectively, a lower or higher degree ofsmoothing). The procedure weimplemented tested first each predictoras a smoothing spline with 3 degrees offreedom, and then tried to simplify themodel by testing the variables enteredpreviously as smoothing splines with 2


85

degrees of freedom and, finally, aslinear terms.

The discriminative ability of thebird distribution models were assessedby the area under the curve (AUC) of aReceiver Operator Characteristic (ROC)plot (Swets 1988; Murtaugh 1996). TheAUC summarizes ROC plots with ameasure of discrimination independentof a threshold (Fielding & Bell 1997).AUC ranges between 0.5 (chanceperformance) to 1 (perfect

discrimination), and can be interpretedas the probability of a model to render ahigher predicted value of presence for aspecies in a site where the species existsthan for a site where the species is notpresent (Zweig & Campbell 1993;Cumming 2000)

We used two methods to analyzethe relationship between grain ofpredictors and discriminative ability ofmodels. Firstly, we select the more

Figure 2. Comparison of the different grains tested. Circles to measure predictors increase from diameter150 m (17.6 ha, but only 12.5 ha with our raster approximation) to diameter 4850 m (185 ha that weapproximated to 183 ha).

4850

2450

1250

650

350

150


86

discriminative diameter level for eachspecies (i.e., the one used to build themodel with highest AUC among the 6models for each particular species). Thespatial resolution of predictors thatachieve maximum discrimination maybe influenced by body size (as asurrogate of home range anddetectability: the bigger the bird, thebigger the diameter to measureenvironmental predictors), by frequencyof occurrence in the sample (becauserarer birds may be responding to verylocal features of landscape andperceiving the environment with a finergrain), and by particular characteristicsof the species (that we resumed in afactor grouping species similar in itshabitat requirements and trophic habits,see \Boone, 1999 #25, and appendix 1).Then we performed an ordinal logisticregression (Guisan & Harrell 2000) ofdiameter level on species’ weight (logtransformed), prevalence and ecologicalgroup (Harestad & Bunnel 1979).Preliminar analysis of ordinality ofdiameter category showed the datasatisfied the assumptions of acontinuation ratio model, that we finallyimplemented in its extended version(Harrell 2001, pages 339-340). Thecontinuation ratio model fits an ordinallogistic regression for each diameterlevel, and its extended version allowsfor different slopes among levels(Harrell, 2001 #12, pages 331-373, andsee Harrell et al. 1998, for background;Guisan & Harrell 2000). Weacknowledge that the spatial resolutionof predictors that yield the morediscriminatory models can be bestconsidered as a continuous measure,rather than having to select one of adiameter levels that were sometimesvery similar in its resulting AUC’s.Therefore, we secondly estimate theoptimum diameter (OD) as the averageddiameter weighted by AUC, OD=ΣriAUCi/ΣAUCi, where ri is thediameter (in meters) for category i, and

AUCi is the estimated AUC for themodel performed with predictorsmeasured in a circle of diameter i. ODwas linearly regressed on body weight(log transformed), frequency ofoccurrence in the sample (namedprevalence) and ecological group. In thetwo analyses we only considered thespecies with at least one fairlydiscriminative model (AUC>0.75, Elith2000).

All analyses were made using S-plus 2000 (MathSoft 1999). Theextended continuation ratio was madewith the Design library (described inHarrell 2001). AUC’s were estimatedwith AccuROC 2.5 (Vida 1993).

RESULTS

For unlimited-radius census (recallthat in these surveys birds recordingswere taken into account only if theywere made within a circle of diameter100 m), the ordinal regression does notshow a significant effect of neither oneof the predictors tested: weight (logtransformed), prevalence or ecologicalgroup, on the spatial resolution at whichmodels are more discriminative (table2). However, the effect of cohort ishighly significant (p<0.0001), what canbe interpreted as different base rates ofdiscrimination among levels ofdiameter. The observed number ofmodels with maximum discriminativevalues (AUC) at each category ofspatial resolution for predictors differssignificantly from what expected atrandom: level 5 (diameter 2450 m) hastwice the expected value for a regulardistribution, while level 1 (150 m) and 2(350 m) have about half the expectedvalue (χ2=14.05, df=5, p=0.015, fig. 3).That is, the “best” diameter isfrequently 2450 m. On the contrary, forfixed-radius census the effect of cohortwas significant (p=0.01), but there werenot differences in discriminative values


87

Table 1. Variables tested as environmental predictors in the six models of breeding bird speciesoccurrence in two areas of Western Andalusia. Values were averaged for circles of increasing diameters(150, 350, 650, 1250, 2450 and 4850 m) centered in sampling points, and a model was built for eachspecies and diameter of circle. Sources: a 1995 land-use/land-cover cover digital map of Andalusia fromthe SinambA (Consejería de Medio Ambiente, Junta de Andalucía), b IRS satellite image, sensor LISS III(dates: 19/07/99 and 16/07/99 for the two study areas). Fractal dimension estimated with IDRISI 32 on aNDVI (Normalized Difference Vegetation Index) image, c Digital Elevation Model of Andalusia at 50 mresolution.

Predictive variable type and name Description1.-Vegetation

variables1.1.- Land-cover

Forest a Proportion of pixels belonging to any forest typecategory (including sparse forested areas such asdehesas but not olive groves)

Dense Forest a Proportion of pixels belonging to dense forestcategory (excluding dehesas)

Coniferous forest a Proportion of pixels belonging to coniferous foresttype category (mainly reafforestations of coniferousbut we included those of eucalyptus)

Broad-leave forest a Proportion of pixels belonging to any broad-leaveforest type category

Riparian vegetation b, c Proportion of pixels belonging to any riparian typecategory (from shrubs to riparian)

Shrub a Proportion of pixels belonging to any shrub typecategory

Agriculture a Proportion of pixels belonging to agricultural landuse categories (irrigated and non-irrigated crops andolive groves)

Herbaceous a Proportion of pixels belonging to any herbaceoustype category (whether natural or cultivated)

Tree cultures a Proportion of pixels belonging to tree cultures(mainly olive groves)

Urbanized a Proportion of pixels in urbanized or industrial areas1.2.- Landscape

Borders (all) a Length of boundaries between forested landcovercategories and the rest of vegetation categories

Borders (forest-shrub) a Length of boundaries between forest and shrublandCropland heterogeneity b Fractal dimension of Normalized Differenced

Vegetation Index values of a satellite image as anindex of heterogeneity in croplands

2.- Topographicvariables

Altitude c Mean altitude (in m)Slope c Mean slope (in degrees)


88

among levels of spatial resolution(χ2=5.92, df=5, p=0.314, fig.3). Usingeither the home range extracted fromthe bibliography (appendix 1), or theallometric predictions from weightestimated with our data (HR=W1.06),instead of body weight made nodifferences in the results of ordinalregressions.

Estimates of OD for unlimited-distance surveys did not differ fromestimates for fixed-distance pointsurveys (wilcoxon test, Z=0.91, p=0.36,n=54). Equally, the “best” diameter forunlimited-distance point surveys did notdiffer from that for fixed-distance pointsurveys (wilcoxon test, Z=1.17, p=0.24,n=54). For the 65 especies with data forboth the unlimited-distance and thefixed-radius surveys, there were notdifferences between survey type, noramong the six levels of spatialresolution (ANOVA: spatial resolution,

F=0.90, p=0.48; survey type, F=1.41,p=0.24; interaction, F=0.03, p=0.99).

There is a small positive andsignificant effect of weight on theoptimum (averaged) diameter forunlimited-distance surveys (OD, table4). However, the range of predicted ODis low (1598-1678 m), and thecorresponding areas of circles withthose diameters vary only between 200and 221 ha. Using recorded home rangeor estimated home range did not changethe results. The inclusion of thecomplete set of species (rather thanselecting those with at least one fairlydiscriminative models) renders similarresults (both for the single morediscriminative diameter and the OD)and is not discussed further. No effectof weight (or home range), prevalenceor ecological group could be found forfixed-radius surveys.

Table 2. Ordinal logistic regression (extended continuation ratio model) for the length of diameter thatrenders the more discriminative models (unlimited-distance census). Predictive factors were log(weight),prevalence (frequency of occurrence in the point surveys taken as a whole) and ecological group. Cohortis a factor that was included for testing if the effect of weight or prevalence on predictive ability variedamong the different diameter lengths.

Wald statistic (response: diameter level)Factor Chi-square d.f. P

Cohort (Factor + Higher order factors) 45.31 12 <0.0001 All interactions 10.05 8 0.2617Log(weight) (Factor + Higher order factors) 9.30 5 0.0977 All interactions 7.83 4 0.0981Prevalence (Factor + Higher order factors) 2.62 5 0.7585 All interactions 2.62 4 0.6236Ecological group 9.67 6 0.1394Cohort X log(weight) (Factor + Higher order factors) 7.83 4 0.0981Cohort X prevalence (Factor + Higher order factors) 2.62 4 0.6236Total interaction 10.05 8 0.2617Total 48.09 20 0.0004


89

The sampling circles correspondingfor both the single “best” and theoptimal diameters had surfaces largerthan home ranges (appendix 1). Forexample, Certhia brachydactyla, asmall passerine with home range of 4ha, had an optimal diameter of 1547 m,which corresponds to a circle of 188 hafor sampling predictors. Thus, largerbirds could have single “best” andoptimal diameters actually larger thanthe ones we analyzed (4850 m) and,consequently, an artifact may arise dueto the limited range of diameters weconsidered. However, excluding the 8bigger species, heavier than 300 g (forwhich the point surveys may not besuch an adequate sampling method) didnot change noticeably the results.

Figure 3. Histograms of predictive ability(AUC) among the six levels of spatial resolutionfor the two survey types.

DISCUSSION

A surprising result of our analysesis that the more discriminative modelswere on average those made withpredictors measured in circles of largediameters (i.e., low grain: 2450 m).Indeed, the area of these circles (471 ha)were generally much larger than thearea of home range for most species,with the exception of raptors. We hadexpected a closer match between sizesof home ranges and “best” grain,because both the quality and originalspatial resolution of our environmentalpredictors (they are broad descriptors ofvegetation measured at spatialresolution 50 m) were more suited todetect habitat selection at the homerange level (for example, we wouldhave needed more detailed vegetationvariables had we been interested inhabitat selection of within-home rangeor nesting site features). What are thereasons for this disagreement? Ourresults mean that the probability ofdetecting a particular species depend onhabitat characteristics measured in anwide neighboring area, and not only onin-site specific features of the samplingpoint. That is, a certain species will bepresent in a survey point in part becauseof what habitat characteristics it mayfind in this point, but also because ofwhat habitat characteristics exists in thesurroundings (Goodwing & Fahrig1998; Saab 1999). It is known thatsuboptimal areas close to sourcenucleus can be occupied morefrequently than expected according toquality of habitat, because of socialinteractions (less experienced orcompetitive individuals may be forcedto settle in the perimeter of a preferredarea Van Horne 1983; Hobbs & Hanley1990). Also, patches of habitat aresubject to the effects of demographicstochasticity and limited dispersal ofindividuals, by which isolated or

05

1015

2025

1 2 3 4 5 6

Unlimited-radius census

AU

C0

24

68

1012

1 2 3 4 5 6

Fixed-radius census

AU

C


90

peripheral fragments are expected to befrequently unoccupied even if they aresuitable for the species (Tyre,Possingham & Lindenmayer 2001).This view of the relevance of theneighborhood on the probability ofoccurrence of a species in a point issupported by neither optimum nor bestdiameter differing for unlimited-distance or fixed-distance point surveys(since unlimited-distance implies alarger sampling area, we expected ODto be greater for these surveys than forfixed-distance points).

For unlimited-distance surveys, wealso found a positive relation betweenbody size and the spatial extent tomeasure predictors that yields the morediscriminative models. This result is inagreement with what may have beenexpected: models for larger birds aremore predictive when built withpredictors measured in larger circles.However, this relationship is not strongand clear predictions of optimal

diameter to measure environmentalvariables can not be made. Indeed, theeffect of weight is so small thatoptimum diameter can be seen as aconstant (around 1430 m or 160 ha) forour set of species. Body size and homerange —a variable difficult to measurereliably in the field— are correlated.The extension of home range (HR) inbirds grows with body size (estimatedby weight: W) according to theallometric expression: HR=W1.14

(Schoener, 1968 #9, and see also Peters1983), that is, the regression of(log)home range on (log)body weighthas a slope of 1.14 using our data(appendix 1). For our data HR=W1.37 (orHR=W1.06 if we exclude the 8 biggerspecies). Recalling that optimumdiameter is nearly a constant, theallometric expressions imply thatpredictors for small birds should besampled in circles much larger thantheir home ranges, while the difference

Table 3. Ordinal logistic regression (extended continuation ratio model) for the length of diameter thatrenders the more discriminative models (fixed-distance census). Predictive factors were log(weight),prevalence (frequency of occurrence in the point surveys taken as a whole) and ecological group. Cohortis a factor that was included for testing if the effect of weight or prevalence on predictive ability variedamong the different diameter lengths.

Wald statistic (response: diameter level)Factor Chi-square d.f. P

Cohort (Factor + Higher order factors) 28.38 12 0.0049 All interactions 10.32 8 0.2432Log(weight) (Factor + Higher order factors) 8.00 5 0.1564 All interactions 7.93 4 0.0942Prevalence (Factor + Higher order factors) 2.39 5 0.7934 All interactions 2.29 4 0.6828Ecological group 4.38 5 0.4963Cohort X log(weight) (Factor + Higher order factors) 7.93 4 0.0942Cohort X prevalence (Factor + Higher order factors) 2.29 4 0.6828Total interaction 10.32 8 0.2432Total 29.80 19 0.0544


91

Table 4. Linear regression of optimum diameter on logarithm of weight forunlimited-distance census (neither prevalence nor ecological group weresignificant).

Terms Coefficient Change indeviance

F value P value

Intercept 1610Log(weight) 4.6 2662 (6%) 4.19 0.045

is not that large for big species. Forexample, a 20-g bird is predicted tohave the most discriminative modelwhen environmental predictors aremeasured in an area of 207 ha (diameterof 1624 m), and this area equals 8.6predicted home ranges (24 ha each),according to the relationship HR=W1.06.In contrast, a 150-g bird is predicted tohave the most discriminative model forpredictors measured in an area of 211ha (diameter of 1637 m), what equalsonly 1 predicted home ranges (203 haeach). Such a difference between smalland big birds may be explainedfollowing the same reasoning aboutmetapopulation dynamics outlinedabove. Small species can be expected tohave a relatively limited dispersalability (compared with large species)from established breeding nucleus tonew, unoccupied, and favorable areas.Also, small species live usually shorterthan large species, and may exhibit ahigher demographic variation.Therefore, the effects of localextinctions (lose of occupied homeranges) and limited dispersal (fail toreoccupy lost or new home ranges)would probably result in more suitableareas being unoccupied for smallspecies than for big species.

The areas that our analyses suggestas optima to sample environmentalpredictors may seem very large but, infact, they are of a magnitude similar tothe spatial resolution utilized in thefirsts Gap projects (100 ha, what is the

area covered by a circle of diameter1128 m, Scott et al. 1993, and see theGap analysis program WWW page:http://www.gap.uidaho.edu/). On theother hand, that magnitude is larger thanspatial resolutions used in othersuccessful regional modelling (4 ha, ordiameter 226 m, in Australian forestsPearce, Ferrier & Scotts 2001), thoughin this case some of the predictivevariables considered in the models weremeasured within 2 km. Therefore, ourfindings, though initially surprising, arein accord with common practice inregional modelling.

To summarize, our empiricalresults suggest that the grain to measureenvironmental predictors in habitatmodelling for birds should be relativelylarge (a general recommendation ofabout 200 ha that correspond to circlesof diameter ~1.5 km), probably becauseof the effect of metapopulationdynamics on patterns of occurrence. Itremains, however, to articulate thisqualitative assertion with a moreappropriate quantitative model. A finalcaveat should be made. Habitats in ourstudy area are quite fragmented, andthis may exacerbate the metapopulationdynamics discussed above, thus makingmore relevant to consider ample areas tomeasure environmental variables inpredictive habitat modelling. It maywell be that that our results are not fullyextrapolable to more homogeneousareas.

http://www.gap.uidaho.edu/


92

REFERENCES

Burnham, K. P. & Anderson, D. R.(1998). Model selection andinference: a practical information-theoretic approach.Springer-Verlag,New-York

Bustamante, J. (1997). Predictivemodels for lesser kestrel Falconaumanni distribution, abundanceand extinction in southern Spain.Biological Conservation, 80: 153-160.

Calder III, W. A. (1990). The scaling ofsound output and territory size: arethey matched? Ecology, 71(5): 1810-1816.

Cumming, G. S. (2000). Usingbetween-model comparisons to fine-tune linear models of species ranges.Journal of Biogeography, 27(2):441-455.

Donázar, J. A., Hiraldo, F. &Bustamante, J. (1993). Factorsinfluencing nest site selection,breeding density and breedingsuccess in the bearded vulture(Gypaetus barbatus). Journal ofApplied Ecology, 30: 504-514.

Elith, J. (2000). Quantitative methodsfor modeling species habitat:comparative performance and anapplication to Australian plants.Quantitative methods forconservation biology. (eds S. Ferson& M. Burgman), pp. 39-58. Springer,. New York .


González, L. M., Bustamante, J. &Hiraldo, F. (1992). Nesting habitatselection by the Spanish imperialeagle Aquila adalberti. BiologicalConservation, 59: 45-50.

Goodwing, B. J. & Fahrig, L. (1998).Spatial scaling and animal population

dynamics. Ecological scale: Theoryand applications. (eds D. L. Peterson& V. T. Parker), pp. 3-15. ColumbiaUniversity Press, . New York .

Guisan, A. & Harrell, F. E. (2000).Ordinal response regression modelsin ecology. Journal of VegetationScience, 11: 617-626.


Harestad, A. L. & Bunnel, F. L. (1979).Home Range and Body Weight--AReevaluation. Ecology, 60(2): 389-402.


Harrell, F. E., Margolis, P. A., Gove, S.,Mason, K. E., Mulholland, E. K.,Lehmann, D., Muhe, L., Gatchalian,S. & Eichenwald, H. F. (1998).Development of a clinical predictionmodel for an ordinal outcome: theWorld Health OrganizationMulticentre Study of Clinical Signsand Etiological agents of Pneumonia,Sepsis and Meningitis in YoungInfants. WHO/ARI Young InfantMulticentre Study Group. Statisticsin Medicine, 17(8): 909-944.


Hobbs, N. T. & Hanley, T. A. (1990).Habitat evaluation: douse/availability data reflect carryingcapacity? Journal of WildlifeManagement, 54(4): 515-522.

Holling, C. S. (1992). Cross-scalemorphology, geometry, anddynamics of ecosystems. EcologicalMonographs, 62: 447-502.

Johnson, D. H. (1980). The comparisonof usage and availabilitymeasurements for evaluatingresource preference. Ecology, 61(1):65-71.


93

Levin, S. A. (1992). The Problem ofPattern and Scale in Ecology: TheRobert H. MacArthur AwardLecture. Ecology, 73(6): 1943-1967.

MathSoft, I. (1999). S-Plus 2000 User´sGuide.Data Analysis ProductsDivision, Seattle

Moreira, J. M. & Fernández-Palacios,A. (1995). Usos y coberturasvegetales del suelo en Andalucía.Seguimiento a través de imágenes desatélite.Junta de Andalucía.Consejería de medio ambiente,Sevilla



O'Neill, R. V. & King, A. W. (1998).Homage to St. Michael; or, why arethere so many books on scale?Ecological scale: Theory andapplications. (eds D. L. Peterson &V. T. Parker), pp. 3-15. ColumbiaUniversity Press, . New York .

Pearce, J., Ferrier, S. & Scotts, D.(2001). An evaluation of thepredictive performance ofdistributional models for flora andfauna in north-east New SouthWales. Journal of EnvironmentalManagement, 62(2): 171-184.

Perrins, C., Ed. (1998). The completebirds of the Western Paleartic onCD-ROMOxford University Press, .Oxford

Peters, R. H. (1983). The ecologicalimplications of body size.CambridgeUniversity Press, Cambridge

Peterson, D. L. & Parker, V. T., Eds.(1998). Ecological scale: Theory and

applicationsColumbia UniversityPress, . New York

Saab, V. (1999). Importance of spatialscale to habitat use by breeding birdsin riparian forests: a hierarchicalanalysis. Ecological Applications,9(1): 135-151.

Scott, J. M., Davis, F., Csuti, B., Noss,R., Butterfield, B., Groves, C.,Anderson, H., Caicco, S., D´erchia,F., Edwards, J., T.C., Ulliman, J. &Wright, R.G. (1993). GAP analysis:a geographic approach to protectionof biological diversity. WildlifeMonographs, 123: 1-41.

Scott, J. M., Heglund, P., Wall, W.,Samson, F. & Haufler, J., Eds. (inpress). Predicting speciesoccurrences: issues of scale andaccuracyIsland Press, . Covello

Swets, J. A. (1988). Measuring theaccuracy of diagnostic systems.Science, 240: 1285-1293.


Van Horne, B. (1983). Density as amisleading indicator of habitatquality. Journal of WildlifeManagement, 47(4): 893-901.

Vida, S. (1993). A computer programfor non-parametric receiver operatingcharacteristic analysis. ComputerMethods and Programs inBiomedicine, 40: 95-101.

Zweig, M. H. & Campbell, G. (1993).Receiver-operating characteristic(ROC) plots: a fundamentalevaluation tool in clinical medicine.Clinical Chemistry, 39: 561-577.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO V: APÉNDICE (CONTINUACIÓN)

Appendix 1. Species considered initially in the analyses, with their estimated body weight, home-range size, prevalence in the sampled points, andecological group at which they belong (see text for details). It is also shown the “best” level of factor diameter (the spatial resolution of predictors thatyield maximum discriminative ability), its corresponding AUC, and the optimum diameter (the averaged diameter weighted by AUC). FOREST, forest-dwelling species, OPEN, species of open areas, GROUND, species linked to mixed habitats that usually feed on the ground, SHRUB, species occupyinga variety of shrub habitats, RIPARIAN, species linked to a variety of riparian habitats, RAPTOR, big species (mainly raptors) that forage in largeextensions of terrain. A perfect circle cannot be made in a raster GIS, so predictors were measured within figures that approximated circles with surfacesof 1.3 ha (for circles of 150 m diameter), 9.3 ha (350 m), 31.3 ha (650 m), 120.3 ha (1250 m), 465.3 ha (2450 m), 1832.3 ha (4850 m).

Species Weight(g)

Ecologicalgroup

Home-rangesize (ha)

Prevalence “Best” diameter (m) AUC OptimumDiameter (m)

Aegithalos caudatus 8 FOREST 16.7 0.05 4850 0.81 1649Alectoris rufa 508 OPEN 20 0.19 2450 0.79 1633Athene noctua 168 GROUND 20 0.04 2450 0.84 1591Burhinus oedicnemus 462 OPEN 100 0.01 2450 0.97 1628Calandrella brachydactyla 22.3 OPEN 2.5 0.01 350 0.95 1619Carduelis cannabina 14.5 SHRUB 16.7 0.28 650 0.74 1629Carduelis chloris 25.9 GROUND 6.7 0.43 2450 0.73 1683Carduelis carduelis 13.2 GROUND 3.2 0.56 4850 0.74 1639Cercotrichas galactotes 23.4 SHRUB 5 0.01 4850 0.46 1732Certhia brachydactyla 8.2 OPEN 4 0.27 2450 0.29 1547Cettia cetti 13.8 RIPARIAN 10.6 0.06 650 0.88 1597Circaetus gallicus 1700 RAPTOR 7875 0.02 2450 0.84 1627Circus pygargus 316 RAPTOR 20106 0.02 2450 0.96 1618Cisticola juncidis 8.7 OPEN 1.44 0.15 350 0.92 1609Coccothraustes coccothraustes 54.7 FOREST 66 0.02 650 0.88 1538Columba palumbus 489 GROUND no data 0.13 2450 0.78 1671Corvus corax 1131 RAPTOR 3475 0.03 2450 0.82 1636Corvus monedula 234 GROUND 20 0.03 1250 0.94 1651Coturnix coturnix 98.8 OPEN 1.5 0.06 2450 0.83 1619Cuculus canorus 103 FOREST 30 0.16 4850 0.71 1686Cyanopica cyana 74.6 FOREST 26.3 0.09 4850 0.90 1642Delichon urbica 19.5 AEREAN 3.33 0.09 2450 0.79 1653Dendrocopos major 74.4 FOREST 32 0.05 2450 0.88 1639Emberiza cia 23.3 SHRUB 10.2 0.08 2450 0.81 1652Emberiza cirlus 25.6 FOREST 0.94 0.04 650 0.83 1623Erithacus rubecula 16.5 RIPARIAN 7.3 0.14 2450 0.92 1627


Species Weight(g)

Ecologicalgroup

Home-rangesize (ha)


Falco tinnunculus 233 RAPTOR no data 0.03 2450 0.86 1716Fringilla coelebs 23.1 GROUND 2.6 0.45 2450 0.89 1631Galerida cristata 41.4 OPEN 4 0.22 2450 0.90 1632Galerida theklae 36.8 SHRUB 1 0.12 1250 0.80 1613Garrulus glandarius 171 FOREST 10 0.07 2450 0.84 1634Hieraaetus pennatus 842 RAPTOR 314 0.03 1250 0.89 1632Hippolais pallida 11.9 RIPARIAN no data 0.01 2450 0.97 1633Hippolais polyglotta 10.5 RIPARIAN 2 0.10 4850 0.73 1643Hirundo daurica 19.1 AEREAN no data 0.08 4850 0.76 1642Hirundo rustica 22.2 AEREAN 10 0.24 1250 0.75 1624Jynx torquilla 38.4 FOREST 78.5 0.07 4850 0.85 1652Lanius excubitor 63.4 GROUND 55 0.03 650 0.52 1604Lanius senator 29 GROUND 8 0.16 1250 0.78 1631Lullula arborea 26.1 GROUND 16.8 0.18 1250 0.85 1621Luscinia megarhynchos 20.5 RIPARIAN 0.67 0.34 2450 0.81 1622Melanocorypha calandra 65 OPEN 3.25 0.07 1250 0.98 1620Merops apiaster 55.2 AEREAN 314 0.33 4850 0.75 1658Miliaria calandra 47.6 OPEN 2 0.50 2450 0.82 1629Monticola solitarius 58 SHRUB 2.5 0.05 150 0.95 1592Motacilla cinerea 17.4 RIPARIAN 150 0.01 650 0.98 1610Motacilla flava 16.2 OPEN 0.9 0.02 650 0.96 1612Muscicapa striata 14.2 GROUND 0.64 0.02 350 0.94 1611Oenanthe hispanica 16.4 OPEN 1.7 0.09 2450 0.83 1653Oenanthe leucura 38 SHRUB 32 0.01 2450 0.93 1604Oriolus oriolus 68.5 FOREST 186 0.07 2450 0.76 1663Parus caeruleus 10.8 FOREST 0.53 0.29 2450 0.82 1626Parus cristatus 10.4 FOREST 7.7 0.07 650 0.91 1641Parus major 18 FOREST 1.54 0.35 650 0.76 1586Passer domesticus 28.5 GROUND 0.5 0.23 1250 0.77 1632Petronia petronia 32 FOREST 0.1 0.08 2450 0.88 1658Phoenicurus ochruros 16.5 SHRUB 0.66 0.03 650 0.95 1611Phoenicurus phoenicurus 15.7 FOREST 1 0.04 4850 0.91 1678Phylloscopus bonelli 7 FOREST 1 0.08 2450 0.87 1627Phylloscopus collybita 7.7 FOREST 9.9 0.05 1250 0.90 1620


Species Weight(g)

Ecologicalgroup

Home-rangesize (ha)


Picus viridis 175 FOREST 467 0.07 4850 0.79 1647Ptyonoprogne rupestris 23 AEREAN 12 0.03 350 0.85 1602Pyrrhocorax pyrrhocorax 322 RAPTOR 941 0.05 4850 0.98 1646Regulus ignicapillus 5.3 FOREST 0.1 0.07 1250 0.89 1622Saxicola torquata 14.5 GROUND 3 0.12 150 0.82 1558Serinus serinus 11.9 GROUND 1 0.45 1250 0.72 1611Sitta europaea 21.5 FOREST 1.5 0.15 1250 0.87 1624Streptopelia turtur 125 FOREST 200 0.13 4850 0.76 1655Sturnus unicolor 86.5 GROUND 201 0.13 1250 0.79 1630Sylvia atricapilla 17.6 RIPARIAN 7.4 0.15 650 0.90 1608Sylvia cantillans 10.9 SHRUB 5 0.03 1250 0.85 1619Sylvia conspicillata 8.8 OPEN 15 0.02 2450 0.93 1659Sylvia hortensis 21.1 FOREST 2.4 0.06 350 0.76 1620Sylvia melanocephala 10.9 SHRUB 0.21 0.44 650 0.78 1607Sylvia undata 9.5 SHRUB 1.6 0.08 2450 0.77 1620Troglodytes troglodytes 9 FOREST 0.71 0.15 1250 0.82 1596Turdus merula 95.9 SHRUB 1 0.55 1250 0.84 1635Turdus viscivorus 117 GROUND 16 0.04 2450 0.85 1633Upupa epops 66.2 FOREST 257 0.18 4850 0.80 1640

¿Qué tipo de variables predictoras es más útil?

97

CHAPTER VI: A comparison of different explanatory variables for

predictive models of breeding bird distribution: competing roles for

landscape, land-cover, topography and climate

ABSTRACT

Predictive habitat models rely on the relationship between a response variable (either

occurrence or abundance of species) and a set of explanatory variables. Vegetation is

habitually preferred as a source of potential predictors because of having a more direct

link with reproductive necessities of species than topography and climate. However

vegetation cartography is costly to produce and update, and most land-cover maps are

usually made with a general purpose, focused on land management. On the contrary,

basic topographic and climatic data are easier to obtain. In this study we compare the

predictive ability (as estimated by the Area Under the Curve, AUC, of Receiver

Operating Characteristic plots) of different predictive bird distribution models generated

for 79 species in Southwestern Spain. The presence of each species were modelled with

Generalized Additive Models with binomial errors and logit link. Within each species,

several models were created that differ in the set of candidate variables (either

vegetation or topography and climate), or in the order in which those were tested. A

similar strategy were used to ascertain the relative relevance of landscape and land-

cover variables within vegetation variables. Vegetation models were significantly more

accurate than topo-climatic models, but the difference was due to the higher number of

potential predictors in the set of vegetation variables. Landscape models were

significantly more accurate than land-cover models, even when controlling the number

of candidate predictors. Mixed models improved slightly the predictive ability. Our

results suggests that the selection of a set of candidate variables (if any) should be done

on the grounds of data availability, though a mixed set is best, and that landscape

reflects an important information not revealed by land-cover measures. Thus regional

modelling programmes would gain predictive ability by including landscape measures.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO VI

98

CAPÍTULO VI: Una comparación de diferentes variables predictoras para los

modelos de la distribución de aves:

el paisaje, la cubierta vegetal, la topografía y el clima

RESUMEN

Los modelos predictivos de la distribución de especies se basan en la relación entre

una variables respuesta (bien la frecuencia de aparición, o bien la abundancia) y un

conjunto de variables explicativas. La vegetación se suele preferir a la topografía y el

clima como fuente de predictores potenciales, ya que está ligada más directamente con

los requerimientos reproductivos de las especies. Sin embargo, la cartografía de la

vegetación es costosa de producir y actualizar, y la mayoría de los mapas de coberturas

se crean con un propósito general, centrado en la gestión del territorio. Por el contrario,

ciertos datos básicos de topografía y clima son fáciles de obtener. En este estudio

comparamos la capacidad predictiva (estimada por el área bajo la curva, AUC, de

gráficos característicos de operador-receptor) de diferentes modelos predictivos de la

distribución de aves, que se generaron para 79 especies en el suroeste de España. La

presencia de cada especie se modeló mediante modelos aditivos generalizados (GAM)

con errores binomiales y vínculo logístico. Dentro de cada especie se generaron varios

modelos que diferían en el conjunto de potenciales variables explicativas (de vegetación

o bien topográfico-climáticas), o en el orden en que se comprobaba su inclusión en los

modelos. Una estrategia similar se usó para investigar la importancia relativa de las

variables de paisaje y de cobertura dentro del conjunto de variables de vegetación.

Los modelos de vegetación alcanzaron una capacidad predictiva significativamente

mayor que los topográfico-climáticos, pero la diferencia se debió al mayor número de

predictores potenciales en el conjunto de variables de vegetación. Los modelos

paisajísticos tuvieron una capacidad predictiva significativamente mayor que los

modelos de cobertura incluso cuando se controló la cantidad de variables potenciales.

Los modelos mixtos mejoraron ligeramente la capacidad predictiva. Nuestros resultados

sugieren que la selección de un conjunto de variables explicativas potenciales debe

hacerse en función de su disponibilidad (aunque un conjunto mixto es la mejor opción),

y que el paisaje tiene una información que no revelan las medidas de cobertura. Por


99

tanto, los programas regionales de modelado ganarían en capacidad predictiva si

incluyeran medidas de paisaje.


100

INTRODUCTION

The study of the relationships betweenspecies and their habitats has beentraditionally a central issue in ecologyand is nowadays of prime importance inconservation and planning (see a reviewin Morrison, Marcot & Mannan 1998).Accurate distribution maps are welcomein, for example, the selection and designof natural parks (Scott et al. 1993), theassessment of human impacts onbiodiversity (Lavers & Haines-Young1996), or the testing of biogeographicalhypotheses (Mourell & Ezcurra 1996;Leathwick 1998). However, even aperfect knowledge of the biology of aspecies cannot guarantee that a static mapwill reflect dynamic properties of speciesdistribution (Tyre, Possingham &Lindenmayer 2001). Moreover, humanand logistic limitations make impracticalto survey extense areas and, inevitably,our knowledge of the spatial distributionof most species will have many gaps.Then, a common solution is to resort topredictive habitat modelling (reviewed inGuisan & Zimmermann 2000; and see afine recent example in Osborne, Alonso& Bryant 2001) and regard the results aspotential habitat, able to be reached andcolonized by a species (Tyre, Possingham& Lindenmayer 2001).

The predictive habitat models relatetypically the occurrence pattern of aspecies (either presence/absence orabundance) with some predictors selectedfrom a set of biologically plausiblecandidate variables. A large number ofpotentially explanatory variables arecurrently easy to obtain thanks to theincreasing development of GIStechniques and digital cartography(thematic maps and satellite imagery).Thus, potential predictors such astopographic and climatic data on the one

hand, and vegetation digital mapas on theother, are widely spread as potentialsources of information for modelling(Goodchild, Parks & Steyaert 1996).Now, the raw data must be preprocessedbefore the analysis and this can be verytime-consuming (Lillesand & Kiefer1994; Goodchild, Parks & Steyaert1996), so given time and budgetconstraints, what kind of data should amodeller prioritize?

It may be argued that vegetationaffects the distribution of animalsproximally, by providing shelter, foodand potential nest-sites, while topographyand climate affect indirectly, bymodifying the relationships of birdspecies with vegetation (e.g., differenthabitat selection may be selected underdifferent climatic conditions, Tellería,Pérez-Tris & Carbonell 2001; Tellería etal. 2001) or, simply, by modifying thevegetation itself. Therefore, vegetation isexpected to generally be a better predictorof animal distribution than topographyand climate at local scales. However,high-resolution digital vegetation mapsare costly to produce and update.Moreover, land-cover cartography ishabitually built by govermental agencieswith a general purpose (often focused onland-use) and the categories of naturalland are not described in enough detail forthe modeler of species distribution (see anexample for Southern Spain in Moreira &Fernández-Palacios 1995). On thecontrary topographic digital bases(Digital Elevation Models) and, to alesser extent, meteorological data, aremore easily available and will be spatiallycorrelated with vegetation in mostinstances (so conveying informationredundant to some degree). It is desirableto know whether one source may renderpredictors that lead to more accuratemodels (such a case would be if both sets


101

of variables had a completely redundantinformation) or both are necessary toachieve high predictive ability.

We assume that, within vegetationvariables, the relative effect of site-specific local descriptors or landscapeconfiguration on the distribution ofanimals is of great interest forconservation management. If landscapehas any relevance in the patterns ofoccurrence for a particular species, thenthe design of reserves or corridors for thisspecies should take landscape intoaccount. This is so because apparentlysuitable areas may be in fact unsuitable,for example, if the areas are too small orclose to a border. A major effect oflandscape would render useless all effortsof habitat description in the field.However the effects of landscape and itsrelative importance compared to site-specific features are not yet wellunderstood. Landscape has been found toseriously affect patterns of abundance anddistribution in some studies (Bolger, Scott& Rotenberry 1997; Vander Haegen,Dobler & Pierce 2000), but the ubiquityand relevance of this effect iscontroversial (Mac Garigal & Mac Comb1995), and it is not explicitly consideredin some successful regional modellingprogrammes like the GAP program inUSA (Scott et al. 1993). Once one has aGIS database, obtaining landscape andsite-specific vegetation variables requiresapproximately the same effort, soselecting either one or the other (or both)type of variables should be done on thegrounds of accuracy of predictions.

In this work we build predictivemodels for several breeding birds inSouthwestern Spain and address thefollowing questions: (i) what set ofvariables has a greater predictive ability,the vegetation variables or the

topographic and climatic variables?; (ii)both sets of variables are expected to becorrelated, but do they have some degreeof independent information?; (iii) is asingle set of variables enough to generatereasonably accurate models, or do weneed mixed sets?; (iv) within vegetationvariables, does landscape contribute toobtain a greater predictive accuracy?;and (v) as a last, methodological, issue,are the results influenced by the particularcharacteristics of the analysis (namely:the number of predictors in each set andthe order in which variables areincorporated to the models)?

STUDY AREA AND METHODS

The study area are two 70 x 70 kmsquares in Western Andalusia, SouthernSpain. Both of them include low, flatareas, mainly devoted to agriculture,surrounded by mountainous areas withmore natural vegetation (altitude rangesfrom 0 to 1600 m a.s.l.). In both areasland-cover is mainly dominated(approximately 70%) by Mediterraneanvegetation (shrubland and evergreen oakQuercus ilex subsp. ballota L. and corkoak Quercus suber L. forests anddehesas), and pine and Eucalyptus spp.plantation, and 30% is mainly cultivatedland, mainly non-irrigated wheat andsunflower crops and olive groves.Villages and urbanized areas are widelyinterspersed.

We analyzed occurrence of 79 speciesdetected in 1144 unlimited-distance pointcounts of 15 min duration made betweenApril and June in 1999 and 2000. Birdspecies were mostly passerines (80%),and their sizes ranged between FirecrestRegulus ignicapillus Temminck (~5,3 gr)to Short-toed eagle Circaetus gallicusGmelin (~1600 gr).


102

Table 1. Predictive variables tested to model breeding bird species occurrences in two areas of WesternAndalusia.

Predictive variable type and name Description1.-Vegetationvariables

1.1.- Land-cover

Agriculture Proportion of pixels belonging to agricultural land use categories(non-irrigated cereal crops and olive groves) in a circle of radius of150 mts centered in sampling point

Herbaceous Proportion of pixels belonging to any herbaceous type category(whether natural or cultivated) in a circle of radius of 150 mtscentered in sampling point

Forest Proportion of pixels belonging to any forest type category (includingsparse forested areas such as dehesas but not olive groves) in a circleof radius of 150 mts centered in sampling point

Shrub Proportion of pixels belonging to any shrub type category in a circleof radius of 150 mts centered in sampling point

Riparian vegetation Proportion of pixels belonging to any riparian type category (fromscrub to forested masses) in a circle of radius of 150 mts centered insampling point

1.2. –LandscapeDistance toagricultural patches

(3 variables) Distance (in m.) to the nearest patch with anagricultural land use (1) larger than 2 ha. (2) larger than 10 ha. (3)larger than 100 ha.

Distance toherbaceousvegetation patches

(3 variables) Distance (in m.) to the nearest patch with herbaceosuvegetation (1) larger than 2 ha. (2) larger than 10 ha. (3) larger than100 ha.

Distance to forestpatches

(3 variables) Distance (in m.) to the nearest forest patch (1) largerthan 2 ha. (2) larger than 10 ha. (3) larger than 100 ha.

Distance to shrubpatches

(3 variables) Distance (in m.) to the nearest shrub patch (1) largerthan 2 ha. (2) larger than 10 ha. (3) larger than 100 ha

Distance to riparianvegetation patches

(3 variables) Distance (in m.) to the nearest riparian vegetation patch(1) larger than 2 ha. (2) larger than 10 ha. (3) larger than 100 ha

Perimeter/area Perimeter/area ratio of the patch were the sampling point wasperformed considering any of the above land-cover variables (exceptriparian vegetation)

2.- Topo-climaticvariable

Altitude Mean altitude in a circle of radius of 150 mts centered in samplingpoint

Slope Mean slope in a circle of radius of 150 mts centered in samplingpoint

Precipitation Mean annual precipitation in mm. (modelled to a resolution of 1km2)

Temperature Mean annual temperature in ºC (modelled to a resolution of 1 km2)Radiation

(insolation)Mean annual potential solar radiation (Kj/m2) in a circle of radius of150 mts centered in sampling point


103

We defined two sets of predictivevariables (table 1). The first set(vegetation) included 24 variables aimedto describe the land-cover (5 variables)and landscape structure (19 variables); thesecond set (topography and climate)included 5 variables descriptive oftopography (3 variables) and climate (2variables). We assume that the two setsconvey partially different (but correlated)information about the environmentaround the point counts. We think thatvegetation has a proximal effect onpatterns of breeding bird occurrence atthe scale of our study, while topographyand climate may affect indirectly or,simply, their combination may reflectchanges of vegetation cover at a finergrain than that of our land-cover mapreflects: topography and climate mayinform about microclimatic conditionsand we therefore decided to join thesetwo types of variables in a single set(topo-climatic).

Landscape structure variablesestimated both the relationship betweenarea and perimeter (AP-ratio) of thepatches belonging to different land-covercategories, and the distance to the nearestpatch of a given size (i.e.: d1X would bedistance to nearest patch of land-cover Xlarger than 2 ha, d2X the same for thenearest patch larger than 10 ha, and d3Xfor the nearest patch larger than 100 ha).We account for size of patches because itaffects occupancy patterns in a species-specific way (for an analysis in ourgeographical area, see: Tellería & Santos1997; Santos & Tellería 1998).Vegetation variables (landscape structureand land-cover variables), except thoserelated to riparian vegetation (see below),were extracted from the 1995 land-use/land-cover digital map of Andalusia(SinambA) provided by the

Environmental Department of the Juntade Andalucía (Moreira & Fernández-Palacios 1995). Original data in vectorformat were rasterized to 50 m resolution.Topographic variables were obtainedfrom a Digital Elevation Model (DEM) ofAndalusia (50 m horizontal resolution).Mean precipitation and temperature datawere obtained from the Instituto Nacionalde Meteorología and interpolated byregression models and kriging at pixelresolution of 1 km2 (own data,unpublished). Finally, a map of meanannual potential solar radiation wasestimated from the DEM (followingNinyerola, Pons & Roure 2000).Published digital cartography of riparianvegetation was unsatisfactory. This typeof vegetation occupies mostly narrow andfragmented patches in the study area, andthus small vegetated watercourses arelikely to be disregarded and lumped withneighboring land-use/land-cover classes(land-use and land-cover patches smallerthan 25 ha tend to be grouped in thevegetation map). Therefore, weelaborated a cartography of riparianvegetation through interpretation of IRSsatellite images (sensor LISS III) aided byan overlay of watercourse informationextracted from the DEM. All variables,except distances, were averaged forcircles of radius 150 m centered insampling points. Admittedly, theresolution at which define local variablesshould ideally be species-specific, but wehave no reliable clues to select optimalresolutions for predictors (indeed, suchclues are habitually inexistent for mostspecies), so we choose to average incircles of radius 150 m to achieve a highresolution that is above expectedgeoreferencing errors. Extraction ofvariables from the GIS was done usingIDRISI 32 (Eastman 1999), IDRISI for


104

Windows (Eastman 1997) andMIRAMON (Pons 2000).

Statistical analysis

We performed Generalized AdditiveModels (Hastie & Tibshirani 1990) ofpresence/absence of each bird speciesusing binomial errors and logit link.Explanatory variables for each modelwere selected from each set of potentialpredictors (table 1) by a forward-backward stepwise procedure (with thestep.gam function of S-PLUS 2000software, MathSoft 1999) that used anapproximation of AIC (Akaike´sInformation Criterion, Sakamoto,Ishiguro & Kitagawa 1986) as thecriterion to enter or to remove variables.Predictors were allowed to enter themodels as linear terms or as smoothingsplines with 2 or 3 degrees of freedom (toachieve, respectively, a lower or higherdegree of smoothing). The procedure weimplemented tested first each predictor asa smoothing spline with 3 degrees offreedom, and then tried to simplify themodel by testing the variables enteredpreviously as smoothing splines with 2degrees of freedom and, finally, as linearterms. Automated methods of variableselection such the one outlined abovehave been criticized (Burnham &Anderson 1998) because they can revealspureous relationships (Flack & Chang1987; Mac Nally 2000) or rendbiologically implausible models (James &McCulloch 1990). However, they arefrequently used in ecological analyses(Scott et al. in press) because theyfacilitate a rapid generation of models;moreover, automated models comparewell with models that include expertopinion (Pearce et al. 2001; Seoane,Bustamante & Díaz-Delgado submitted),so we think our procedure is justified (atleast as a heuristic comparison).

We built several models for each birdspecies that differed in the type ofexplanatory variables allowed to enterwith the aim to answer several questions.First, in order to compare the predictiveability of vegetation versus topographyand climate, we modelled speciesoccurrence using only vegetationvariables as predictors (models V), andthen we tested if these models could beimproved by the inclusion of topo-climatic variables (keeping the vegetationpredictors that entered previously; modelsV-T). Analogously, we built models usingonly topo-climatic variables as predictors(models T), and then tested if vegetationvariables could improve these models(models T-V). We also tried a completemodel in which all of the variables wereallowed to enter simultaneously. Second,to explore the relative relevance of thedifferent vegetation variables, we builtspecies predictive models with onlyvegetation variables reflecting land-coverat the sampling point (models C), andwith only vegetation variables reflectinglandscape structure (L), to finally test ifthe models with landscape variables couldbe improved with land-cover variables (Lvs L-C), and if the models with land-cover variables could be improved withlandscape variables (C vs C-L).

The set of potential vegetationpredictors is considerably larger than theset of potential topographic and climatepredictors (24 vs. 5, table 1), and thus thevegetation models could be better onaverage because they are considering awider set of measures of environmentalconditions (Elith 2000). To take intoaccount this fact on the comparisons, werandomly distributed the set of vegetationvariables in five groups of five variables(to be able to generate five groups of fivepredictors from 24 variables, the last


105

group of four variables was completedwith a 5th predictor –D1-F– selected atrandom from those already included inthe other groups). Then, we built fivevegetation models (VR1 to VR5) witheach of those reduced sets of fivepotential predictors (the same number asin topo-climatic models). The set oflandscape structure variables is also largerthan the set of land-cover variables (19 vs5); thus, as we did previously, we builtfour landscape structure models (LR1 toLR4) each one using a set of 5 randomlyselected landscape variables (n=5, D3-Rwas randomly selected to appear in twogroups).

To assess model accuracy, weestimated the area under the curve ofROC plots (the AUC, Swets 1988;Murtaugh 1996) with AccuROC 2.4 forWindows (Vida 1993). The ROC curve isbuilt by plotting the sensitivity of a model(or true-positive rate) on the ordinateagainst 1-specificity (or false-positiverate) on the abcissa. This was made forevery possible threshold value that can bechosen to convert the predictedprobability of occurrence (a continuousvalue given by the models in the interval(0,1)) to predicted presence or absence (adichotomous variable). The AUCsummarizes ROC plots with a measure ofoverall accuracy independent of athreshold (Fielding & Bell 1997). AUCranges between 0.5 (chance performance)to 1 (perfect discrimination), and can beinterpreted as the probability of a modelto render a higher predicted value ofpresence for a species in a site where thespecies exists than for a site where thespecies is not present (Zweig & Campbell1993; Cumming 2000). AUC for modelsVR1 to VR5 and LR1 to LR4 wereaveraged to obtain a single estimate (VRand LR, respectively). Differences amongmodel types were tested with a repeatedmeasures factorial ANOVA with an error

term due to species to control for thebetween-species variation, of no interestin this study.

RESULTS

It was possible to build predictivemodels better than a null model for everyspecies and set of variables (except forSouthern Grey Shrike Laniusmeridionalis and topo-climatic variables).Mean AUC (Table 2) ranged from0.69[SE=0.099] for models with land-cover predictors only (C) to 0.85[0.100]for models with topo-climatic andvegetation predictors (included in thisorder). Most of the specific models withineach model type reached at least amoderate accuracy (we consider asubjective threshold of AUC>0.7 for themodels to be considered potentiallyuseful), except for models C, LR, and VR(table 2).

Vegetation models (V) weresignificantly more accurate than topo-climatic models (T) showing on theaverage a 6% difference in AUC(mean±SE, V=0.81±0.104,T=0.75±0.113; F=90.4, p<0.0001, table3). Topo-climatic models were improvedby the inclusion of vegetation variables(there was a significant increase in AUC:T vs T-V, F=250.84, p<0.0001, table 3),and vegetation models were improved bythe inclusion of topo-climatic variables(V vs V-T, F=16.22, p<0.0001, table 3).The improvement in accuracy wasrelatively important (10%) whenvegetation variables were allowed to enterin the topo-climatic model, but wasrelatively minor (3 %) when topo-climaticvariables improved the vegetation models(T-V=0.85±0.100, V-T=0.84±0.104).However, mean AUC of vegetationmodels built with a reduced set of


106

potential explanatory variables (VR) didnot differ from mean AUC of topo-

climatic models (VR=0.75±0.081; T vsVR, F=0.040, p=0.84, table 3).

Table 2. Bird prediction accuracy: mean AUC (and SE) values for the models generated with the differentset of predictors. Models considered the following variables: 1) T, topography and climate; V, vegetation;T-V, topography and, afterwards, vegetation; V-T, vegetation and, afterwards, topography; All, allvariables simultaneously; VR, a reduced set of five randomly-selected vegetation variables (AUC is theaverage of five VR1 to VR5 models). 2) C, land-cover; L, landscape structure; L-C, landscape structureand, afterwards, land-cover; C-L, land-cover and, afterwards, landscape structure; LR, a reduced set of fiverandomly-selected landscape variables (AUC is the average of four LR1 to LR4 models). It is also giventhe percentage of models with AUC >0.7 (a subjective threshold to consider models accurate enough to beof any use).

Model AUC (SE) %AUC > 0.7T 0.75 (0.113) 70V 0.81 (0.104) 94

T-V 0.85 (0.100) 99V-T 0.84 (0.104) 96All 0.84 (0.102) 99VR 0.75 (0.081) 67C 0.69 (0.099) 38L 0.80 (0.084) 91

L-C 0.81 (0.082) 96C-L 0.81 (0.104) 94LR 0.78 (0.083) 67

Table 3. Results of repeated measures ANOVA testing the effect in model accuracy (AUC) of usingvegetation or topo-climatic variables as predictors. Planned comparitions test if vegetation differs from topo-climatic variables in predictive accuracy (T vs V), if vegetation improves topo-climatic models (T vs T-V), iftopo-climatic variables improve vegetation models (V vs V-T), and if mean predictive accuracy ofvegetation variables differs from that of topo-climatic variables (T vs VR). Names of models as in table 2.


Error: withinVegetation vs topo-climatic variables 4 0.760 0.190 90.384 <0.0001

T vs V 1 0.199 0.199 94.438 <0.0001T vs T-V 1 0.527 0.527 250.835 <0.0001V vs V-T 1 0.034 0.034 16.221 <0.0001T vs VR 1 0.000 0.000 0.040 0.84

Residuals 312 0.656 0.002


107

Table 4. Results of repeated measures ANOVA testing the effect in model accuracy (AUC) of using land-cover or landscape variables as predictors. Planned comparitions test if landscape models differ in accuracyfrom land-cover models (L vs C), if landscape variables improve land-cover models (C vs C-L), if land-cover variables improve landscape models (L vs L-C) and if mean predictive accuracy of a landscapevariable differs from mean predictive accuracy of a land-cover variable (C vs LR). Names of models as intable 2.


Error: withinVegetation structure vs landscape 4 0.888 0.222 119.330 <0.0001

L vs C 1 0.275 0.275 148.070 <0.0001C vs C-L 1 0.408 0.408 239.398 <0.0001L vs L-C 1 0.008 0.008 4.418 0.036LR vs C 1 0.196 0.196 105.436 <0.0001

Residuals 312 0.580 0.002

Models built using only landscapestructure variables (L) were significantlymore accurate than models built usingonly land-cover variables (C), showing onaverage a 11% difference in AUC(mean±SE, L=0.80±0.084,C=0.69±0.099; F=148.1, p<0.0001, table4). The inclusion of landscape variablesimproved greatly (12% on average) andsignificantly the land-cover models (C vsC-L, F=239.4, p<0.0001, table 4), but theinclusion of land-cover variablesimproved only sligthly the landscapemodels (1% on the average) and thisdifference was only slightly significant(L=0.80±0.084, L-C=0.81±0.082; L vs L-C, F=4.4, p=0.04, table 4). Mean AUC oflandscape models built with a reducedrandom set of potential predictors (LR)was significantly higher (LR vs C,F=105.4, p<0.0001, table 4) than that ofland-cover models (a 9% of averagedifference).

The order in which each general set ofvariables (vegetation and topo-climaticvariables) was added did not clearlyaffect the prediction ability of theresulting models, though there was a

slight significant improvement (1%) formodels in which topography was addedfirst (T-V vs All variables, F=10.3,p=0.002; V-T vs All variables, F=0.07,p=0.791).

DISCUSSION

Our analysis show correlationalrelationships between breeding birddistribution and some coarse explanatoryvariables typically available fromcommon cartographical data. Byproviding breeding substrates andforaging grounds, vegetation is likely tohave an effect on breeding birddistribution at a fine scale closer tocausality than topography and climate.Therefore, vegetation variables may bemore promising to build accuratepredictive models at the scale of thispresent study. Accordingly, ourvegetation models (V) have a greaterpredictive ability than topo-climaticmodels (T), and the mixed models (T-V,V-T, and All variables) only improvedconspicuosly topo-climatic models.However, this relative improvementseems to be due to the number of


108

potential predictors tested to enter themodels, because there were nodifferences between vegetation and topo-climatic models when controlling for thenumber of the potential predictors in theinitial set. A larger set of potentialpredictors may contain more informationabout the environment, and is likely toinclude more variables that correlate withthe presence/absence of a species. This ispossibly the reason why the vegetationmodels here developed were moreaccurate than topo-climatic models.Vegetation, climate and topography areexpected to be correlated (Woodward1987; Brown 1995); for example, in ourstudy area the more xeric, low-altitudeand flat zones are mainly covered bycereal crops. However, our results suggestthat both sets of potential predictors havesome degree of independent informationabout the environment, because mixedmodels reached the higher predictiveabilities and, in particular, the inclusionof vegetation variables notably improvedthe predictive ability of topo-climaticmodels. This pattern is in agreement witha previous study by Beard et al. (1999),who found similar results in a coarser-

scale study that considered other climateand land-cover variables (but notincluding landscape structure) andanalyzed with a different modellingapproach data from the Breeding BirdSurvey at Idaho (area of study ~200000km2). This agreement in the results of twodisparate analysis further support ourfeeling that topographic and climatic dataare a source of potential predictors asadequate for breeding bird distributionmodelling as vegetation data. Mixedmodels (vegetation plus topo-climaticvariables) somewhat improve predictiveability but, what is more important, thesemodels generates reasonably accuratepredictions (AUC>0.7) for almost everyspecies (>95%).

Within vegetation variables, landscapestructure had a greater predictive abilitythan land-cover categories as predictors.This is so even when controlling thenumber of the potential predictors in theinitial set, what means that, in theaverage, a single landscape variable ismore useful for modelling than a singleland-cover variable. Landscape

Table 5. Results of repeated measures ANOVA testing the effect in model accuracy (AUC) of the order ofinclusion of topo-climatic or vegetation variables. Planned comparitions test if including topo-climaticvariables first renders models that differ in accuracy from those in which all variables are testedsimultaneously (T-V vs All), and if including vegetation variables first renders models that differ in accuracyfrom those in which all variables are tested simultaneously (V-T vs All). Names of models as in table 2.


Error: withinOrder of inclusion of variables 2 0.0023 0.0011 5.191 0.007

T-V vs All 1 0.0023 0.0023 10.312 0.002V-T vs All 1 0.0000 0.0000 0.070 0.791

Residuals 156 0.0343 0.0002


109

characteristics and land-cover categoriesare well known to be spatially correlated(Mac Garigal & Mac Comb 1995), but atour study scale this two sets of potentialpredictors are not totally redundant, sincethe inclusion of landscape variablesgreatly improved land-cover models.Therefore, our results suggest thatlandscape may have indeed a profoundeffect on breeding bird distribution, whatcould be of particular concern inheterogenous areas. A caveat to made isthat we did not thoroughly exploreindividual models, so there may be somespecies for which land-cover predictorsare as good as (or even better than)landscape, since species differ in theirsensibility to landscape features (Knick &Rotenberry 1995; Bolger, Scott &Rotenberry 1997; Santos & Tellería1998).

The relative importance of landscapeconfiguration and site-specific vegetationvariables in bird species distribution isopen to discussion. So, some studies havefound a relevant and direct influence ofsimple landscape patterns (e.g., distanceto borders) on bird species distribution(Bolger, Scott & Rotenberry 1997;Sánchez-Zapata & Calvo 1999), whileothers have found more moderate andcomplex effects (Mac Garigal & MacComb 1995). Landscape may be relevantin explaining bird distribution, as ourresults suggests, for two reasons. First,according to a hierarchical view of habitatselection (Johnson 1980), landscapepatterns provide environmental clues thatare used for birds to select their homerange, so that potential resourceful areaswithin the range of a species may remainunoccupied or sub-occupied if they lackthose clues (Rolstad, Loken & Rolstad2000). Second, the adequacy ofapparently homogenous habitats for aparticular species may not be spatially

constant, for example by changing withdistances to the limits between suitableand unsuitable habitats (Bolger, Scott &Rotenberry 1997). Indeed, there is someevidence of negative effects of landscapequality on birds physiology (individualswith feathers growing slower were foundin small fragments of forests, Stratford &Stouffer 2001).

If the relevance of landscape on animaldistribution is prevalent among differentcommunities and taxa, then the regionalprogrammes that (courageously) aim tomodel distributions at a high spatialresolution, such as the GAP in USA(Scott et al. 1993), LANDSPOT inSwitzerland (Guisan et al. 2000), or theNFBS in Australia (see Pearce & Ferrier2000; Pearce & Ferrier 2001) shouldconsider landscape among their potentialpredictor variables (or, at least, in whatconcerns bird distribution). To ourknowledge, only the latter explicitlyincludes explanatory variables related tolandscape configuration (e.g., theprobability of certain vegetation types ina surrounding area), and their modelshave proven useful and accurate for manytaxa (Pearce & Ferrier 2001).

To conclude, our results have twoimplications in the modelling of bird-habitat relationships. First, the selectionamong sources of potential explanatorydata (if any) should be done on thegrounds of data availability, since modelaccuracy is likely to be similar for modelsderived from general land-cover and land-use vegetation maps, and from modelsderived from topographic and climaticinformation. However, we wouldrecommend using the two sources ofinformation. Second, local models that donot take into account landscapeparameters are probably missing arelevant source of variation because the


110

many implications of such variables onbird distribution here mentioned.

ACKNOWLEDGEMENTSThis work is a contribution to the

project “predictive cartography of landbirds: a pilot study in western Andalusia”,funded by the Dirección General deEnseñanza Superior e InvestigaciónCientífica (Ministry of Science andTechnology) and FEDER funds from theEU, project # 1DF-97-0648. during thework, J.S. had a predoctoral fellowshipfrom the Ministry of Education andCulture. The extensive field workpresented in this paper could not havebeen done without the help and sense ofhumour of Daniel López Huertas, Luis M.Carrascal, and Mario Díaz who alsobrought intellectual insight to the project.

NOTEA version of this chapter has been

submitted to Conservation Biology withJ.Bustamante and R.Díaz-Delgado.

LITERATURE CITED

Beard, K. H., Hengartner, N. & Skelly, D.K. (1999). Effectiveness of predictingbreeding bird distributions usingprobabilistic models. ConservationBiology, 13(5): 1108-116.

Bolger, D. T., Scott, T. A. & Rotenberry,J. T. (1997). Breeding bird abundancein an urbanizing landscape in coastalsouthern california. ConservationBiology, 11(2): 406-421.

Brown, J. H. (1995).Macroecology.University of ChicagoPress, Chicago

Burnham, K. P. & Anderson, D. R.(1998). Model selection and inference:a practical information-theoreticapproach.Springer-Verlag, New-York



Eastman, J. R. (1999). IDRISI32: Guideto GIS and Image Processing.ClarkLabs, Clark University, Worcester

Elith, J. (2000). Quantitative methods formodeling species habitat: comparativeperformance and an application toAustralian plants. Quantitativemethods for conservation biology. (edsS. Ferson & M. Burgman), pp. 39-58.Springer, . New York .


Flack, V. F. & Chang, P. C. (1987).Frequency of selecting noise variablesin subset regression analysis: asimulation study. The AmericanStatistician, 41: 84-86.

Goodchild, M. F., Parks, B. O. &Steyaert, L. T., Eds. (1996).Environmental Modeling withGISOxford University Press, . NewYork

Guisan, A., Gonseth, Y., Cherix, D.,Zimmerman, N. E., Kienast, F.,Edwards, T., Walter, T. & Weiss, A.D. (2000). Landscape potential foranimal species survival, colonizationand dispersal: a spatial simulationstudy:http://www.unine.ch/cscf/PROJETS/landspot/index.htm.

Guisan, A. & Zimmermann, N. E. (2000).Predictive habitat distribution modelsin ecology. Ecological Modelling, 135:147-186.

http://www.unine.ch/cscf/PROJETS/la


111

Hastie, T. J. & Tibshirani, R. J. (1990).Generalized Additive Models.Chapman& Hall, London

James, F. C. & McCulloch, C. E. (1990).Multivariate analysis in ecology andsistematics: panacea or Pandora's box?Annual Review of Ecology andSystematics, 21: 129-166.

Johnson, D. H. (1980). The comparisonof usage and availabilitymeasurements for evaluating resourcepreference. Ecology, 61: 65-71.

Knick, S. T. & Rotenberry, J. T. (1995).Landscape Characteristics ofFragmented Shrubsteppe Habitats andBreeding Passerine Birds.Conservation Biology, 9(5): 1059-1071.

Lavers, C. P. & Haines-Young, R. H.(1996). Using models of birdabundance to predict the impact ofcurrent land-use and conservationpolicies in the Flow Country ofCaithness and Sutherland, northernScotland. Biological Conservation, 75:71-77.

Leathwick, J. R. (1998). Are New-Zealand's Nothofagus species inequilibrium with their environment?Journal of Vegetation Science, 9: 719-732.

Lillesand, T. M. & Kiefer, R. W. (1994).Remote sensing and imageinterpretation.John Wiley & Sons,New York

Mac Garigal, K. & Mac Comb, W. C.(1995). Relationships betweenlandscape structure and breeding birdcommunities in the Oregon coastrange. Ecological Monographs, 65:235-260.

Mac Nally, R. (2000). Regression andmodel-building in conservationbiology, biogeography and ecology:The distinction between -andreconciliation of- 'predictive' and

'explanatory' models. Biodiversity andConservation, 9: 655-671.

MathSoft, I. (1999). S-Plus 2000 Guide toStatistics.Data Analysis ProductsDivision, Seattle

Moreira, J. M. & Fernández-Palacios, A.(1995). Usos y coberturas vegetalesdel suelo en Andalucía. Seguimiento através de imágenes de satélite.Junta deAndalucía. Consejería de medioambiente, Sevilla


Mourell, C. & Ezcurra, E. (1996). Speciesrichness of Argentine cacti: A test ofbiogeographic hypotheses. Journal ofVegetation Science, 7: 667-680.


Ninyerola, M., Pons, X. & Roure, J. M.(2000). A methodological approach ofclimatological modelling oftemperature and precipitation throughGIS techniques. International Journalof Climatology, 20: 1823-1841.

Osborne, P. E., Alonso, J. C. & Bryant, R.G. (2001). Modelling landscape-scalehabitat use using GIS and remotesensing: a case study with greatbustards. Journal of Applied Ecology,38(2): 458-471.

Pearce, J. & Ferrier, S. (2000). Anevaluation of alternative algorithms forfitting species distribution modelsusing logistic regression. EcologicalModelling, 128: 127-147.

Pearce, J. & Ferrier, S. (2001). Thepractical value of modelling relativeabundance of species for regionalconservation planning: a case study.Biological Conservation, 98: 33-43.

Pearce, J. L., Cherry, K., Drielsma, M.,Ferrier, S. & Whish, G. (2001).


112

Incorporating expert opinion and fine-scale vegetation mapping intostatistical models of faunaldistribution. Journal of AppliedEcology, 38: 412-424.


Rolstad, J., Loken, B. & Rolstad, E.(2000). Habitat selection ashierarchical spatial process: the greenwoodpecker at the northern edge of itsdistribution range. Oecologia, 124:116-119.


Sánchez-Zapata, J. A. & Calvo, J. F.(1999). Raptor distribution in relationto landscape composition in semi-aridMediterranean habitats. Journal ofApplied Ecology, 36: 254-262.

Santos, T. & Tellería, J. L., Eds. (1998).Efectos de la fragmentación de losbosques sobre los vertebrados de lasmesetas ibéricasOrganismo Autónomo"Parques Nacionales", . Madrid

Scott, J. M., Davis, F., Csuti, B., Noss, R.,Butterfield, B., Groves, C., Anderson,H., Caicco, S., D´erchia, F., Edwards,J., T.C., Ulliman, J. & Wright, R.G.(1993). GAP analysis: a geographicapproach to protection of biologicaldiversity. Wildlife Monographs, 123:1-41.

Scott, J. M., Heglund, P., Wall, W.,Samson, F. & Haufler, J., Eds. (inpress). Predicting species occurrences:issues of scale and accuracyIslandPress, . Covello

Seoane, J., Bustamante, J. & Díaz-Delgado, R. (submitted). Does expertopinion increase the predictive abilityof environmental models of birddistribution? .

Stratford, J. A. & Stouffer, P. C. (2001).Reduced Feather Growth Rates of TwoCommon Birds Inhabiting CentralAmazonian Forest Fragments.Conservation Biology, 15(3): 721-728.

Swets, J. A. (1988). Measuring theaccuracy of diagnostic systems.Science, 240: 1285-1293.

Tellería, J. L., Pérez-Tris, J. & Carbonell,R. (2001). Seasonal changes inabundance and flight-relatedmorphology reveal different migrationpatterns in Iberian forest passerines.Ardeola, 48(1): 27-46.

Tellería, J. L., Pérez-Tris, J., Ramírez, A.,Fernández-Juricic, E. & Carbonell, R.(2001). Distribution of Robins(Erithacus rubecula) in winteringgrounds: effects of conspecific density,migratory status and age. Ardea, 89:363-373.

Tellería, J. L. & Santos, T. (1997).Seasonal and interannual occupation ofa forest archipelago by insectivorouspasserines. Oikos, 78: 239-248.


Vander Haegen, W. M., Dobler, F. C. &Pierce, D. J. (2000). Shrubsteppe BirdResponse to Habitat and LandscapeVariables in Eastern Washington,U.S.A. Conservation Biology, 14(4):1145-1160.

Vida, S. (1993). A computer program fornon-parametric receiver operatingcharacteristic analysis. ComputerMethods and Programs inBiomedicine, 40: 95-101.

Woodward, F. I. (1987). Climate andplant distribution.CambridgeUniversity Press, Cambridge, UK

Zweig, M. H. & Campbell, G. (1993).Receiver-operating characteristic


113

(ROC) plots: a fundamental evaluationtool in clinical medicine. ClinicalChemistry, 39: 561-577.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO V: APÉNDICE

SECCIÓN TERCERA

Puesta en práctica: aplicaciones de la cartografía de especies

There is not one correct way to do ecology. Mathematical models, model ecosystems, field

manipulation experiments and the search for large-scale patterns are all valid approaches,

and all have their strengths and weaknesses.

—John Lawton, OIKOS 75: 145-147, 1996.

Modelos para rapaces forestales

115

CAPÍTULO VII: Modelos aditivos generalizados y SIG para predecir la

adecuación del hábitat de rapaces forestales en el sur de España

RESUMEN

Los gestores de recursos naturales necesitan predecir la distribución y abundancia de las

especies y la adecuación de los hábitats existentes. En este trabajo se comprueba la

efectividad de las variables topográficas y de vegetación, medidas con un Sistema de

Información Geográfica (SIG), para predecir la distribución del Busardo ratonero Buteo

buteo, Culebrera europea Circaetus gallicus, Aguililla calzada Hieraaetus pennatus y

Milano negro Milvus migrans en el sur de España. Se realizaron censos por carretera en

cuadrículas de 10x10 km para muestrear la distribución de esas rapaces y se ajustaron

modelos aditivos generalizados (GAM, del inglés “Generalised Additive Models”) con un

procedimiento de selección de variables automático por pasos. En la mayor parte de las

circunstancias, y usando sólo variables topográficas o de vegetación, fue posible construir

modelos predictivos que mejoraron significativamente una clasificación al azar, pero los

modelos no fueron precisos. Los modelos mejoraron su capacidad predictiva si se incluían

las variables de ambos conjuntos y, además, en tres de las cuatro especies la inclusión de

las coordenadas espaciales mejoró los modelos. Los mejores modelos, en cuanto a su

capacidad predictiva, fueron los de la Culebrera europea y los de la Aguililla calzada; pero

estos modelos incluyeron algunas variables de difícil interpretación ecológica. Los modelos

para el milano negro alcanzaron altas tasas de clasificación correcta pero no fueron

robustos. La distribución del Busardo ratonero resultó la más difícil de modelar

probablemente debido a que esta especie está muy extendida y el hábitat parece no tener

una calidad demasiado heterogénea a la escala en que fue medido. Nuestros resultados

indican que es posible construir satisfactoriamente modelos predictivos para las rapaces

usando predictores derivados de la cartografía temática digital disponible, e integrar tales

modelos en un SIG para producir mapas precisos de la distribución del hábitat adecuado

para cada especie.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO VII

116

CHAPTER VII: Using Generalised Additive Models and GIS to predict habitat

suitability for forest raptors in Southern Spain

ABSTRACT

Resource managers need to be able to predict the distribution and abundance of species and

the suitability of available habitats. We test the effectiveness of topographic and vegetation

variables estimated with a Geographic Information System (GIS) to predict the distribution

of the common buzzard Buteo buteo, short-toed eagle Circaetus gallicus, booted eagle

Hieraaetus pennatus and black kite Milvus migrans in Southern Spain. We used road

census in 10x10 km squares to sample raptor distribution and adjusted Generalised

Additive Models with a stepwise variable selection procedure. In most cases it was possible

to build predictive models that improved significantly a classification by chance with only

topographic or only vegetation variables, but models were not accurate. Models improved

their predictive ability if variables from both sets were included, and, further, in three out

of four species the inclusion of spatial co-ordinates to account for neighbourhood effects

improved these models. The best models considering their predictive ability were those for

the short-toed eagle and for the booted eagle; but they included some variables difficult to

interpret from an ecological point of view. Models for the black kite gave high correct

classification rates but were not robust. The distribution of the buzzard resulted the most

difficult to model probably because the species is very widespread and the habitat seems

not very heterogeneous in quality at the scale we measured it. Our results indicate it is

possible to build accurate predictive models for raptors using predictors derived from the

digital environmental cartography available, and to integrate these models in a GIS to

render accurate distribution maps of habitat suitability for each species.


117

1. INTRODUCTION

Resource managers need to knowhow species are distributed, howabundant they are in the landscape andhow suitable for a certain speciesdifferent habitats are. Distribution mapsin books and field guides have beencompiled traditionally from the records oflocalities were a species is known to bepresent plus a certain degree ofinterpolation and expert knowledge guess,usually in unknown proportions (see e.g.Harrison 1982). Atlas works providedistribution maps that are built in a moresystematic way, but data are costly toobtain and are usually not detailedenough for all applications (see e.g.Hagemaijer & Blair 1997). In most Atlasit is not possible to distinguish betweenreal absences and areas that have not beenwell covered with field work. Also, areaswere the species is abundant tend toappear indicated in the same way as areaswhere the species is rare or accidental(Purroy 1997).

Predictive models provide analternative way to build distribution,abundance and/or habitat suitability mapsfor a species (Austin et al. 1996;Morrison, Marcot & Mannan 1998;Beard, Hengartner & Skelly 1999; Brito,Crespo & Paulo 1999; Osborne, Alonso& Bryant 2001). They are based on theknowledge that species are habitatselective (Cody 1985), and they assume itis possible to find environmentalvariables that are good predictors of theirdistribution or abundance (Nicholls 1989;Buckland & Elston 1993). Consideringthat abundance is in many instances agood indicator of habitat suitability (butsee Van Horne 1983; Vickery, Hunter &Wells 1992) it is possible to build habitatsuitability maps with predictive models.

Geographic Information Systems(GIS) provide tools that allow to measureeasily environmental variables that areavailable in a digital format for any pointwhere the distribution of the species hasbeen surveyed. These variables measuredin a GIS can be tested statistically asprospective predictors of the distributionof the species. The resulting statisticalmodels can generate predictive maps ofthe distribution of the species with thehelp of a GIS, provided that we havedigital maps for the predictors for thestudy area (Pereira & Itami 1991; Guisan,Theurillat & Kienast 1998; He et al.1998; Rico Alcázar et al. 2001). It isdesirable that the causal relation betweenpredictors and the distribution of thespecies is known, as this would allowexperts to know when the extrapolation toother areas is possible. Also, is importantthat a detailed cartography for thepredictors is available and that it can beupdated easily and at low cost. Modelsbased on predictors that are as difficult toupdate as the distribution of the speciesitself are of limited use for a resourcemanager.

It is expected that vegetation willbe a better predictor of bird distributionthan topography, because bird speciestend to be associated with certainvegetation types more than with certaintopographic features (Cody 1985). On theother hand, it is easier and cheaper togenerate a Digital Elevation Model(DEM) for an area than a vegetation mapthat is up to date. Also land-use/land-cover maps are very dependent on thecriteria used to generate them, so that twoland-use/land-cover maps of the samearea can be very different if they havebeen generated for different purposes orby different agencies (Cherrill &


118

McClean 1999). As vegetation and land-use over a territory is not independent oftopography, we expect that topographicvariables measured on a DEM wouldhave a certain predictive ability of birddistribution. We want to know: (1) if theinformation contained in a DEM isenough to predict the distribution of somebird species, (2) if the models derivedfrom a DEM are better or worse thanthose derived from a vegetation map, (3)if models derived from one group ofvariables (topography or vegetation) canbe improved with variables from thesecond group, and (4) if consideringneighbourhood effects we can improvethe predictive ability of habitat models.

In this paper we have tested thepossibility of predicting the distributionof four species of forest raptors: commonbuzzard Buteo buteo Linnaeus, bootedeagle Hieraaetus pennatus Gmelin, short-toed eagle Circaetus gallicus Gmelin andblack kite Milvus migrans Boddaert, inSouthern Spain using as predictorstopographic variables derived from aDEM, vegetation variables derived from aland-use/land-cover digital map andspatial co-ordinates to correct forneighbourhood effects.

Figure 1. Location of study area (black striped polygons) and censused squares (solid squares) in theAutonomous Community of Andalusia (Spain).


119

2. METHODS

2.1. Study area

We restricted our study to theAutonomous Community of Andalusia(Southern Spain) and divided the area in10x10 km squares using the UTM grid.We selected all continuous squares wheredominant vegetation was Mediterraneanforest or Mediterranean scrubland. Thestudy area was divided in 9 zones thatcovered a surface of 37,700 km2 . Weselected a sample of 10x10 km squares ineach zone to be censused for raptors (Fig.1). Census work was initially designed toestimate red kite Milvus milvus breedingpopulation –a species not considered inthis paper–, so the sample was stratifiedbetween zones, and the number of squaressampled within each zone wasproportional to expected breeding densityof red kites. Within each zone the squaresto be censused were selected at random.After the censuses were carried weselected the four species of forest raptorsthat resulted more abundant: commonbuzzard, booted eagle, short-toed eagle,and black kite, and tried to buildpredictive models for them. As we hadexcluded a priori those continuous zoneswere dominant land-use is agriculture wedid not made predictions from our modelsin those areas.

2.2. Raptor census

A total of 88 squares of 10x10 kmwere censused (23% of the study area,Fig. 1). In each square we performedapproximately 40 km of road census witha vehicle, using dirt roads or roads withlow traffic that allowed us to census at aspeed of 20 km/h. Two persons carriedout each census, one driving and the otherrecording all raptors. All squares werecensused in spring 1996, between May

and July. Although observers recordednumber of individuals of each species andthe co-ordinates of each contact with araptor, for our models we only used thepresence/absence of each species in eachsquare. Prevalence (the ratio of positivesquares to total sample) was similar forthe four species: common buzzard (0.45),short-toed eagle (0.43), booted eagle(0.42), and black kite (0.39)

2.3. Predictive variables

Estimates of all environmentalvariables within each square wereperformed in a GIS, using IDRISI forWindows v.2.0 (Eastman 1997) andIDRISI32 v.1.01 (Eastman 1999).Topographic variables (T) were estimatedfrom a DEM of the study area (50 mhorizontal resolution, 20 m verticalresolution), that had been derived frominterpolation of 1:50.000 topographicmaps. Overall DEM accuracy waschecked by comparing a random sampleof point co-ordinates with the altitudemeasured from 1:50.000 topographicmaps. This gave an error of less than 20m (or one contour line). The DEM waschecked for errors at the joints ofdifferent map sheets, as these errors couldaffect our estimates of slope. Pixels thatgave unreliable slopes were excludedfrom the estimates of slope for eachsquare. Topographic variables tested aspredictors in the models are given inTable 1. Land-use/land-cover variables(U) were estimated from the SinambA1995 digital land-use/land-cover map forAndalusia (Consejería de MedioAmbiente, Junta de Andalucía, unpubl.data). The map has 112 land-use/land-cover classes that have been updated withsatellite image and aerial photographs andrecords all land-use/land-cover polygonsthat have more than 25 Ha. We rasterisedthe original Arc-Info coverage at a 50 m


120

resolution, and all our variables wereestimated on this raster image. Theestimated Vegetation Index for eachsquare was derived from theExperimental Calibrated GlobalVegetation Index from NOAA-AVHRR(NOAA 1992). Land-use/land-covervariables tested as predictors in themodels are given in Table 1.

2.4. Stastistical models

We used Generalised Linear Models(GLM) (Nelder & Wedderburn 1972;Dobson 1983; McCullagh & Nelder1989) and Generalised Additive Models(GAM) (Hastie & Tibshirani 1990) with abinomial error and a logistic link to modelthe presence/absence of each raptorspecies in each 10x10 km square. Weused as predictors a set of topographicvariables (T models), a set of vegetation,land-use/land-cover variables (U models),

both sets of variables (TU models), andboth sets of variables plus spatial co-ordinates to correct for possibleneighbourhood effects (TUC models).These final TUC models were alsosimplified with more stringent statisticalcriterion looking for possible causalrelationships between environmentalvariables and species distribution (TUCSmodels)

GLM are mathematical models inwhich a relation is stablished between aresponse variable and a linearcombination of explanatory variablesusing an error distributions and a linkfunction adequate to the nature of theresponse variable. Their use is wellstablished in ecology (Crawley 1993) andhave been previously used to modelraptor distribution and habitat

Table 1. Description of explanatory variables tested in predictive models

Acronym Description and source

Topograhic variablesAltitude Mean altitude a.s.l. estimated from a DEM(1) of the study area.

Slope Mean slope (%) in the 10x10 km square for a 50 m pixel as estimatedfrom the DEM using the SURFACE module of IDRISI32 that uses arook's case procedure (Monmonier 1982; Eastman 1999)

Southern Orientation Fraction of 50 m pixels showing a South-east to South-west orientation inthe 10x10 km square. Orientation calculated from the DEM withSURFACE module of IDRISI32. Flat pixels are considered to have asouthern orientation

Rivers Fraction of 50 m pixels that are crossed by a river or a stream obtainedconverting from vector to raster the 1:50.000 hydrology coverage ofAndalusia (2)

(1) Digital Elevation Model of Andalusia obtained by interpolation of 20 m contours from the 1:50.000topographic maps. Horizontal resolution 50 m. Source: SinambA (1999), Consejería de Medio Ambiente,Junta de Andalucía., unpubl. data.

(2) Hydrology cover of Andalusia digitised from 1:50.000 topographic maps. All permanent rivers andstreams are considered irrespective of their size. Source: SinambA (1999), Consejería de MedioAmbiente, Junta de Andalucía, unpubl. data.


121

Table 1 (Cont.). Description of explanatory variables tested in predictive models

Acronym Description and source

Land-use/land-cover variablesUrban Fraction of 50 m pixels classified in the land-use/land-cover raster map(3)

in the urbanised and infrastructure classes.Agricultural Fraction of 50 m pixels classified in the land-use/land-cover raster map in

the agricultural classes.Natural Fraction of 50 m pixels classified in the land-use/land-cover raster map as

natural vegetation.Dense Forest Fraction of 50 m pixels included in classes with tree coverage >50 % in

the land-use/land-cover raster map.Dispersed Forest Fraction of 50 m pixels included in classes with tree coverage from 5 to

50% in the land-use/land-cover raster map.Dense Scrubland Fraction of 50 m pixels included in classes with a scrub coverage >50 %

in the land-use/land-cover raster map.Disperse Scrubland Fraction of 50 m pixels included in classes with a scrub coverage from 20

to 50% in the land-use/land-cover raster map.Pine Forest Fraction of 50 m pixels included in classes of natural or planted

coniferous forest (tree coverage > 5 %) in the land-use/land-cover rastermap.

Eucalyptus Forest Fraction of 50 m pixels included in classes of planted Eucalyptus spp.Forest (tree coverage > 5 %) in the land-use/land-cover raster map.

Broad-leaved Forest Fraction of 50 m pixels included in classes of natural broad-leaved forest(mostly Quercus spp. forest, tree coverage > 5 %) in the land-use/land-cover raster map.

Olive/fruit Groves Fraction of 50 m pixels included in classes of cultivated trees (mainlyolive, fruit and almond groves) in the land-use/land-cover raster map.

Forest Perimeter Border in meters between forested and non-forested classes divided bynumber of 50 m pixels with land-use information.

MSSVI Mean Spring-Summer Vegetation Index (March to August). First a meanimage for each month for the period April 1985-August 1991 wasobtained from monthly values of the Experimental Calibrated VegetationIndex (version 2, monthly values from the U.S. Geological Survey)(NOAA 1992). Then the Mean Index was computed from mean monthlyvalues. The NDVICOMP module of IDRISI32 was used to compute meanvalues as quadratic means of original values according to the formula x’=√Σxj

2⁄n. The mean index (original data at approximate 15 km resolution)was reprojected to UTM co-ordinates at 1 km resolution, filtered treetimes with a 3x3 mean filter, and mean values were extracted for 10x10km squares

(3) Land-use/land-cover 1995 digital map of Andalusia. Source: SinambA (1999), Consejería de MedioAmbiente, Junta de Andalucía, unpubl. data, but see Moreira and Palacios (1995) for a description ofland-use/land-cover 1991 digital map.


122

selection(Donázar, Hiraldo & Bustamante1993; Austin et al. 1996; Bustamante1996; Bustamante 1997; Sánchez-Zapata& Calvo 1999; Suárez, Balbontín &Ferrer 2000). Generalised AdditiveModels (GAM) are a more general classof models from which GLMs constitute aparticular case (Hastie & Tibshirani1990). In GAMs the relation between theresponse and the explanatory variable isreplaced by a scatterplot smooth functionof the data (s).

y = s(x) eqn 1

As linear functions can be seen asparticular cases of scatterplot smoothingwhen span = 1, GLM models can be seenas a restricted subset of GAM models(Hastie & Tibshirani 1990).

Model fitting wasperformed using S-Plus 2000 (MathSoft,1999). To find the best T, U, TU modelfor each species, we chose an automaticforward-backward stepwise variableselection procedure (procedure step.gam)testing in turns polynomial fits of all theexplanatory variables in a set up to thirddegree. The step.gam procedure uses astepwise search to select the best model interms of Akaike’s Information Criterion(AIC), given a range of models toconsider. The AIC statistic has intoaccount both the information explainedby the model and its complexity,according to the expression:

AIC = Deviance + 2 * Scale * residualdegrees of freedom (eqn 2)

where the scale is the scaled Chi-squared (Sakamoto, Ishiguro & Kitagawa1986). The gam procedure in S-Plusallows to fit all models even those inwhich the scatterplot smoother is a linear

or polynomial function. The particularstepwise procedure that we implementedstarts from a null model and builds allpossible univariate models containing thepredictor as a first degree polynomial. Itcompares all these univariate models andthe null model and keeps the one with alower AIC. In a second step it testsincluding in the model previouslyselected each of the remaining predictorsas first degree polynomials ortransforming to a second degreepolynomial the predictor included in theprevious step. The model with the lowestAIC is selected at each of the steps. In athird step the program tests excluding allthe predictors included, one at a time, orreducing the degree of the polynomialsincluded. The procedure cycles thesesteps until the AIC can not be decreased.

As the procedure does nottest automatically a second orderpolynomial if the first order polynomial isnot significant in the first case, we alsorun the procedure a second timeeliminating all non-significant first-orderterms, and a third time eliminating allnon-significant second-order terms. Usingthis procedure we obtained for eachspecies (i) the best predictive modelderived from the topographic variables set(T model), (ii) the best predictive modelderived from the land-use/land-covervariable set (U model), and the bestpredictive model derived from alltopographic or land-use/land-covervariables (TU model).

The best TU model wascorrected for neighbourhood effects(Legendre 1993) fitting a nonparametricsurface of latitude and longitude (abivariate local regression surface,LOESS, with span equal to 0.5), and thenwe tested by removal (procedure


123

step.gam) if T variables, U variables andthe spatial surface remained significant.The resulting model was the TUC model.We decided to use linear models with amaximum 3 df instead of scatterplotsmoothers for topographic and land-use/land-cover variables both becausepolynomials are easier to implement in aGIS than scatterplot smoothers (Guisan,Theurillat & Kienast 1998), and in orderto keep low the model degrees of freedomconsidering our sample size (see a similarapproach inPreisler, Rappaport & Wood1997). We used a smoother for spatial co-ordinates because we had no a prioriexpectation of spatial trends and wewanted to be strict in checking ifenvironmental variables remainedsignificant after correcting for anypossible neighbourhood effects.

In order to avoidoverparametrization and to obtain asimpler model, we further modified theTUC model by a backward stepwiseprocedure using the χ2 statistic (moreconservative than AIC Ludden, Beal &Sheiner 1994; Burnham & Anderson1998) and produced the TUCS models.We only kept variables (or terms ofvariables) when removing them yielded asignificant increment of residualdeviance. A variable was removed fromthe model if P > 0.01 , but the order of apolynomial was only reduced if P > 0.05 .We believe that the TUCS modelsalthough having lower predictive powerthan the TUC models provide us withmore appropriate cues of actual habitatselection by the species.

Predicted probabilities ofappearance given by the models wereconverted to 1 (presence) or 0 (absence)values by choosing a threshold value foreach species so that the number of

predicted and actual presences wereequal. Success of predictions was thenmeasured by the correct classification rateand by the Kappa statistic whichmeasures the correct classification ratecorrected by chance (Titus, Mosher &Williams 1984; Fielding & Bell 1997).

Stability of models was studiedusing a Leave-One-Out (LOO)resampling technique (each square wasleft out in turns and the model wasrefitted with the remaining 87 squares, theprobability of the square left out wasestimated from the model not containingit), and we compared presence/absencepredictions of the original model withthose of the models generated byresampling.

3. Result

3.1. Common buzzard

The models that contained bothtopographic and land-use variables (TUmodel) predicted significantly better thatthose based on one type of variables (T orU models) (Table 2). The spatial co-ordinates did not improve the TU model,so TU and TUC models are actually thesame model. All models except the Tmodel predicted significantly better thanchance. The value of Kappa was poor forthe U model (0.24) (Landis & Koch1977) that consider poor predictivemodels those with Kappa < 0.4) but goodfor TU and TUCS models (0.45 and 0.50respectively) (Fig. 2). Models were quitestable, as measured by a the percentage ofcoincidences with LOO models, that wereabove 90% in every case (Fig. 3). Thespatial predictions of U, TU, TUC andTUCS models were similar (only theTUC model is represented in Fig. 4)while the T model was the one giving amore different location for suitable areas.


124

Figure 2. Classification improvement over chance (Kappa statistics) for all models. Variable sets used ineach model: Topography (T), land-use/land-cover(U), topography and land-use/land-cover (TU),topography, land-use/land-cover and co-ordinates (TUC), and simplified TUC model (TUCS). StandardError for kappa is 0.1 in all cases. n.s. = not significant. Dashed lines are suggested boundaries for poor,good and excellent agreement according to Landis and Koch (1977)

According to the TUCS model theprobability of presence of the commonbuzzard increases with the variablesDense Forest and Forest Perimeter whiledecreases with Slope and EucalyptusForest. These four variables were presentin all the mixed models (TU and TUCmodel), but Altitude entered instead ofSlope in the T model, and Pine Forestentered instead of Eucalyptus Forest inthe U model. However, within these pairsof variables the relationship with theresponse was very similar (a strong lineardecrease in both cases). Altitude ispositively correlated with Slope andEucalyptus Forests are to a great extentcorrelated with Pine Forests so thesepairs of explanatory variables probablyindicate the same habitat and land-usefeatures.

3.2. Short-toed eagle

The models that contained bothtopographic and land-use variables (TUand TUC models) predicted significantlybetter that those based on one type ofvariables (T or U models) (Table 2). Allmodels predicted significantly better thanchance, having significant Kappasvarying from a poor (T and U model, 0.33and 0.37 respectively) to a goodagreement (TU, TUC and TUCS models,0.51, 0.68 and 0.47 respectively, Fig. 2).The model with highest correctedclassification rate was the TUC model(Fig. 5). Coincidence of predictionbetween each model and those generatedby the LOO procedure were above 95%for the T and U models, and between 84and 92% for the mixed TU, TUC andTUCS models (Fig. 3). The spatialpredictions derived from all modelsshowed a good spatial agreement

0

0.2

0.4

0.6

0.8

1

Commonbuzzard

Short-toedeagle

Booted eagle Black kite

Kap

pa

TU

TUTUC

TUCSn.s


125

Table 2. Best models for each species, based on topography (T), land-use/land-cover(U), topography andland-use/land-cover (TU), topography, land-use/land-cover and co-ordinates (TUC), and simplified TUCmodel (TUCS)

Species Model Variables (%) correctclassification

AIC (deviance,df)

Commonbuzzard

T Altitude + Rivers 58 118.1(112.1, 85)

U (Dense Forest)2 + (Disperse Scrubland)2 + PineForest

63 113.6(101.6, 82)

TU Slope + (Dense Forest)2 + Dense Scrubland +Eucalyptus Forest + Forest Perimeter + MSSVI

73 96.6(80.6, 80)

TUC Slope + (Dense Forest)2 + Dense Scrubland +Eucalyptus Forest + Forest Perimeter + MSSVI

73 96.6(80.6, 80)

TUCS Slope + Dense Forest + Eucalyptus Forest +Forest Perimeter

75 -(90.6, 83)

Short-toedeagle

T (Altitude)2 + (Rivers)2 + Southern Orientation 67 109.8(97.8, 82)

U (Dense Forest)2 + (Pine Forest)2 + (MSSVI)2 69 111.3(97.3, 81)

TU (Rivers)2 + Southern Orientation + (DenseForest)3 + (Pine Forest)2 + (Eucalyptus Forest)2 +(MSSVI)2

76 99.7(73.7, 75)

TUC (Rivers)2 + Southern Orientation + (DenseForest)3 + (Pine Forest)2 + (Eucalyptus Forest)2 +(MSSVI)2 + LOESS(Latitude, Longitude)

84 98.3(56.3, 67)

TUCS (Rivers)2 + MSSVI + LOESS(Latitude,Longitude)

74 -(81.0, 76)

Booted eagle T Slope + Southern Orientation 63 118.8(112.8, 85)

U (Dense Forest)2 64 115.3(109.3, 85)

TU Southern Orientation + (Dense Forest)2 +(Disperse Scrubland)2 + (Broad-leaved Forest)2

75 105.7(89.7, 80)

TUC Southern Orientation + (Dense Forest)2 +(Disperse Scrubland)2 + (Broad-leaved Forest)2 +LOESS(Latitude, Longitude)

82 105.1(72.4, 72)

TUCS (Dense Forest)2 + Disperse Scrubland +Southern Orientation

70 -96.6, 83)

Black kite T (Slope)3 76 98.4(90.4, 84)

U (Agricultural)2 + (Natural)2 + Disperse Scrubland+ Olive/fruit Groves + (Pine Forest)3 +(Eucalyptus Forest)3

82 88.7(62.7, 75)

TU Slope + (Natural)2 + (Disperse Scrubland)2 +(Pine Forest)3

82 87.7(69.8, 79)

TUC (Natural)2 + (Pine Forest)3 + LOESS(Latitude,Longitude)

85 83.4(55.5, 74)

TUCS Natural + (Pine Forest)2 + LOESS(Latitude,Longitude)

88 -(57.1, 76)


126

The TUCS model retained onlytwo of the six topographic and habitatvariables of the TU and TUC model(Table 2): Rivers and MSSVI, thatentered, respectively, as a quadratic and apositive linear function. Both enteredwith a similar form in the single T and Umodels. Co-ordinates entered the TUCmodel without removing variables ormodifying noticeably their form in theTU model. The shape of the spatialsurface in the model indicates acontagious distribution and that theprobability of presence of short-toedeagle increases in the north and centralparts of the study area.

Figure 3. Percentage of agreement betweenpredictions of the original models and thosegenerated by the Leave-One-Out procedure.Diamonds, Common buzzard; squares, Short-toedeagle; Triangles, Booted eagle; Crosses, Blackkite.

3.3. Booted eagle

The models that contained bothtopographic and land-use variables (TUand TUC models) predicted significantlybetter that those based on one type ofvariables (T or U models) (Table 2). Allmodels predicted significantly better than

chance, but Kappas were only good forthe TU and TUC models (0.49 and 0.63respectively) (Fig. 2). Coincidences withpredictions of LOO models variedbetween 84 (TUC) and 95% (TUCS, Fig.3). The spatial predictions derived fromthe different models agree in some areasbut disagree in others. Spatial predictionfrom the best model in terms of predictiveability (TUC model) are given in Fig. 6.

The explanatory variablesSouthern Orientation and Dense Forestthat were included in the simplifiedTUCS model, were the more stable of allthe variables entering models for thisspecies (single and mixed, Table 2). Theprobability of presence of the eagleincreased with Southern Orientationfollowing a positive linear function andalso increased as a quadratic function ofDense Forest . The TUCS model alsoindicated an increase in the probability ofpresence with Disperse Scrubland thathad not entered in U model. The spatialco-ordinates surface of the TUC modelindicated a higher probability of presencein the north of the study area but it was nolonger significant in the TUCS model.The co-ordinates entered the TUC modelwithout removing or changing the form ofthe variables that were previously in theTU model.

3.4. Black kite

Models for the black kite had AICstatistics below 98, what is similar orlower than the best achieved for the restof the species. The models that containedboth topographic and land-use variables(TU and TUC models) performed betterthan those based on one type of variables(T and U models), although the differencebetween the U and the TU model wasonly slight (Table 2). Correct

50

60

70

80

90

100

T TU TUCS

Model

Per

cent

age

of a

gree

men

t

U TUC


127

classification rates were over 75% for allmodels (Table 2). They were significantlybetter than chance and showed a goodagreement with original data (Fig. 2).However, the stability of the models waslow, as coincidences with predictions ofLOO models were only between 55-65 %(Fig. 3). The spatial predictions from theTUC model are given in Fig. 7.

The explanatory variablesNatural and Pine Forest were present inboth the single and the mixed models. Inthe TUCS model the probability ofpresence of the black kite decreasedlinearly with Natural and increased withPine Forest. The spatial co-ordinatessurface improved the TU model andindicated an increase in the probability ofpresence to the north and west of thestudy area.

4. DISCUSSION

4.1. Common buzzard

The models we were able to buildfor the common buzzard had a poorperformance. The best models (TU andTUCS models), although significantlybetter than chance were only 45 to 50%better. This is not too high if we considerthat models under 40% are consideredpoor models (Landis & Koch 1977). Thiscould be due to an incorrect selection ofeither the explanatory variables or thescale to which those and the response areanalysed. Indeed, there are severalreasons to think that scale is precludingan accurate prediction. First, we believethat the selected pool of variables waswide enough to cover, direct or indirectly,most of the ecological needs of buzzard(and the other raptors in general). Thevariables that entered in the simplifiedTUCS model are easily interpreted and

are in agreement with what we expectedbefore building the models: the commonbuzzard seems to favour forested areasinterdigitated with other open habitats,and avoids poor homogeneouslyreforested land (areas dominated byeucalyptus and pine trees). This kind ofhabitat is most common at low andmedium altitude (where Slope tends to bemoderate). This preference with forestedhabitat has also been seen in otherMediterranean areas in Spain (Sánchez-Zapata & Calvo 1999), and a strongrelationship between breeding distributionand border between forest and openhabitats was also apparent in the previousstudy and in another study in Scotland(Austin et al. 1996). Co-ordinates did notenter in TUC model, suggesting that thecommon buzzard is widespread within thestudy area and has no clear spatial trend.Moreover, the low average influence of asingle sample square in the models (thepercentage of agreement with LOOmodels was very high) is in support ofthis view. It could be that within thelimits of the study area chosen and withthe resolution of our 10x10 km samplingsquares there are not big differences insuitability for the buzzard. That wouldexplain why our best model onlyimproves slightly over a null model thatgives equal probabilities to all squares. Inorder to build better models for thisspecies it would be necessary to extendthe study area to increase the range ofenvironmental conditions met by thespecies, or to work at a finer resolutionso that sampling units would be morevariable in relation to habitat suitabilityfor the buzzard. Austin et al.(1996)working at a finer resolution were able tofind adequate predictive models for thebuzzard in an area of Scotland.


128

Figure 4. Spatial predictions of the best model for the common buzzard (TUC) White areas indicate where themodel is not applicable.

Figure 5. Spatial predictions of the best model for the short-toed eagle (TUC) White areas indicate where themodel is not applicable.

NorthGridm

100000.00

Without predictions0%0 - 12.5%12.6 - 25.0%25.1 - 37.5%37.6 - 50.0%50.1 - 62.5%62.6 - 75.0%75.1 - 87.5%87.6 - 99.9%100%

common buzzard

NorthGrid m

100000.00


short-toed eagle


129

Figure 6. Spatial predictions of the best model for the booted eagle (TUC) White areas indicate where themodel is not applicable

Figure 7. Spatial predictions derived from the best model for the black kite (TUC) White areas indicate wheremodels are not applicable

NorthGridm

100000.00


booted eagle

NorthGrid m

100000.00

Without predictions0%0 - 12.5%12.6 - 25.0%25.1 - 37.5%37.6 - 50.0%50.1 - 62.5%62.6 - 75.0%75.1 - 87.5%

black kite


130

4.2. Short-toed eagle

Mixed models for this species aresatisfactory in terms of predictive ability.However, the best models (the TU andTUC models) have numerous variablesand are difficult to interpret. Some ofthese may have entered by chance, sinceour more strict procedure of variableremoval led to a very simplified TUCSmodel. This simple model is still difficultto interpret from the point of view of theecology of the species. It suggests aselection for the more productive areasregarding vegetation and a noticeableneighbourhood effect identified by thesignificance of the spatial co-ordinates,but we have no interpretation for thestrong apparent avoidance of squareswith intermediate levels of Rivers

4.3. Booted eagle

Similarly to what happened withthe short-toed eagle, the mixed models forthe booted eagle were satisfactory interms of predictive ability but difficult tointerpret. They suggest a slightneighbourhood effect, since co-ordinatesentered the TUC model (but were rejectedin the stricter TUCS model). TUCSmodel suggest also an expectedpreference for more forested squares anda positive relation (although slight) withDisperse Scrubland and SouthernOrientation. A similar preference forforested areas and areas of scrubland wasfound by Sánchez-Zapata & Calvo (1999)in South-eastern Spain.

4.4. Black kite

Models for this species, bothsingle and mixed, have the greatestpredictive ability among those generatedin this work. This was to be expected

since the black kite is the species with amore localised distribution in the west ofthe study area (it is abundant mainly inthe Doñana National Park andsurrounding areas ), and thus it is a priorieasy to model: any variable that identifiesthe clumped zone of distribution willhave a high predictive ability. This seemto be the case for the positive relationbetween the presence of the species andPine Forest (TUCS model), since thehabitat the black kite occupies in Doñanais mainly pine-tree forest. On the otherhand, TUCS model shows a negativerelation with Natural that could be closerto the actual habitat selection of thisspecies, known to breed and feed inhumanised areas more frequently than therest of the raptors analysed in this work.

4.5. General conclusions

Topography and vegetation (asderived from a land-use/land-cover map)have a certain predictive ability on thedistribution of forest raptors, but neitherof them alone seem to be able to provideaccurate predictions of the distribution ofthe species we studied. Considering theextreme results, topography was not ableto predict better than chance thedistribution of the buzzard but, on theother hand, topography or vegetationalone gave relative good predictivemodels (> 40% better than chance) for theblack kite. Models derived fromvegetation variables had a slightly higherpredictive ability and they were better (interms of AIC) than topographic models,but we cannot exclude that this could be aconsequence of the former having a morenumerous set of variables from which tochoose. Mixed models were needed toobtain a fair (40-75%) improvement overchance in predictions, what shows thatboth set of variables provide a different


131

information content. Probablytopographic variables complement theinformation on changes inhabitat/vegetation that are not adequatelycovered in the land-use/land-cover map.Our results are in agreement with those ofBeard et al.(1999) who found that modelsbased on vegetation, climate or spatialautocorrelation were better than nullmodels to predict the distribution of birdsin Idaho, but that the best predictivemodels were those combining variablesfrom two sets.

As has been shown in otherstudies (Smith 1994; Augustin,Mugglestone & Buckland 1996; Chou &Soret 1996; Lennon 1999; Merrill et al.1999; Osborne, Alonso & Bryant 2001)there is much to gain in incorporatingterms in the model building that takeaccount of neighbourhood effects. Thesecan arise because of habitat being moresimilar between neighbouring areas orbecause the probability of finding aindividual in a place may not beindependent of the probability of findingindividuals in neighbouring places(Augustin, Mugglestone & Buckland1996). Raptors are very mobile and theirdistribution may be less influenced bylocal habitat features than in other animalgroups(Chou & Soret 1996). If thepresence of raptors in 10x10 km squaresis autocorrelated, the “spaceless” modelsT, U and TU may be formed by the moreautocorrelated variables (Lennon 1999)and then a biologically meaningfulinterpretation of the models could beprecluded. This does not seem to be thegeneral case in our study, since theincorporation of co-ordinates (TUC)improved the prediction ability but didnot remove any variables that had enteredpreviously. The only exception to thisrule were the models for the black kite,whose distribution is clearly contagious,

with peak numbers around the DoñanaNational Park, and in which variables likeSlope and Disperse Scrubland were nolonger significant when co-ordinatesentered the models.

Our study was able to producepredictive maps for all four species thatwhere significantly better than chance(TUC maps Figs 4-7). These maps can beuseful for pointing out the best areas foreach species. Although maps producedare statistically accurate it is difficult toknow how reliable will be the predictionsif the models are extrapolated to otherareas. The models for the short-toed eagleand the booted eagle were relativelygood, although some of the variables thatentered the models probably do not havea direct causal relation with thedistribution of the species and aredifficult to interpret ecologically. Thedistribution of the black kite in the studyarea was relatively easy to predict butmodels proved to be very unstable. Thesemodels probably can provide a verylimited insight to which factors affect thedistribution of the species and we do notexpect that they can make accuratepredictions outside the study area. Thefact that the inclusion of the spatial co-ordinates modifies the previous modelindicates that the significance of somevariables can be attributed, at leastpartially, to a neighbourhood effect. Onthe other hand, the models for the buzzardhad a poor predictive performance at thisscale but the selection of variables wasrelatively stable and coincident withvariables that have shown to affect thedistribution of the species in other areasor at other spatial scales. In a similarstudy in Argyll (Scotland) --in which theextent of the study area (140 km2) wassimilar to our resolution (100 km2)--Austin et al. (1996) were able to buildgood predictive models for the


132

distribution of common buzzard nestingsites in 0.5x0.5 km squares. All thissupports our conclusion that our studyarea is too homogeneous for the commonbuzzard at the scale of 10x10 km squares.Still the map produced by our modelprobably reflects some subtle differencesin suitability across the study area, andcould be useful as a management tool forthe species.

Finally, our results show it ispossible to produce accurate predictivemaps for the distribution of raptors usingavailable environmental cartographyelaborated for other purposes. But cautionshould be taken when using these modelsas good predictive accuracy does notnecessarily imply that they are goodexplanatory models of the habitatselection of each species (Mac Nally2000) .

ACKNOWLEDGEMENTS

Data for this study were collectedwhile conducting a red kite surveyfinanced by the Consejería de MedioAmbiente, Junta de Andalucía anddirected by F. Hiraldo. We thank E.López, F. Gavilán, R. Cañete, J.M. Pérez,E. Urbano, F. Buenestado, S. Buenestado,F. J. Ávila, A. Matos, J. Bustamante, F.Hiraldo, J. Ballester, J. Royo, J.C.Nevado, A. Montalban, J. L. Barroso, M.A. Pineda, J. R. Benítez, L. García, J. A.Oña, M. Carrasco, J. López, M. de laRiva, J. A. Sánchez-Zapata, M. Moral, S.Moreno, J. M. Sayago, P. Moreno, P.Córdoba, R. Pulido, M. López and R.López for conducting or helping with theroad census. Y. Menor de Gaspar and M.de la Riva helped with databasemanagement. The SinambA, Consejeríade Medio Ambiente, Junta de Andalucíaprovided GIS data coverages. J. S.

acknowledges a predoctoral fellowship ofthe Ministerio de Educación y Cultura.The analysis of these data and writing ofthis paper was funded by project #1DF97-0648 to J. B. from the DirecciónGeneral de Enseñanza Superior eInvestigación Científica.

NOTESA version of this manuscript has

been submitted to Ecography withJ.Bustamante.

REFERENCES

Augustin, N. H., Mugglestone, M. A. &Buckland, S. T. (1996). Anautologistic model for the spatialdistribution of wildlife. Journal ofApplied Ecology, 33: 339-347.

Austin, G. E., Thomas, C. J., Houston, D.C. & Thompson, D. B. A. (1996).Predicting the spatial distribution ofbuzzard Buteo buteo nesting areasusing a Geographical InformationSystem and remote sensing. Journal ofApplied Ecology, 33: 1541-1550.

Beard, K. H., Hengartner, N. & Skelly, D.K. (1999). Effectiveness of predictingbreeding bird distributions usingprobabilistic models. ConservationBiology, 13(5): 1108-116.

Brito, J. C., Crespo, E. G. & Paulo, O. S.(1999). Modelling wildlifedistributions: logistic multipleregression vs overlap analysis.Ecography, 22: 251-260.

Buckland, S. T. & Elston, D. A. (1993).Empirical models for the spatialdistribution of wildlife. Journal ofApplied Ecology, 30: 478-495.

Burnham, K. P. & Anderson, D. R.(1998). Model selection and inference:a practical information-theoreticapproach.Springer-Verlag, New-York

Bustamante, J. (1996). Statistical modelof nest-site selection for the bearded


133

vulture (Gypaetus barbatus) in thePyrenees and evaluation of the habitatavailable with a geographicalinformation system. Biología yConservación de las RapacesMediterráneas, 1994. (eds J. Muntaner& J. Mayol), pp. 393-400. SEO, .Madrid .

Bustamante, J. (1997). Predictive modelsfor lesser kestrel Falco naumannidistribution, abundance and extinctionin southern Spain. BiologicalConservation, 80: 153-160.

Cherrill, A. & McClean, C. (1999).Between-observer variation in theapplication of a standard method ofhabitat mapping by environmentalconsultants in the UK. Journal ofApplied Ecology, 36: 989-1008.

Chou, Y.-H. & Soret, S. (1996).Neighborhood effects in birddistributions, Navarre, Spain.Environmental Management, 20(5):675-687.

Cody, M. L., Ed. (1985). Habitatselection in birdsAcademic Press, Inc.,. San Diego

Crawley, M. J. (1993). GLIM forecologists.Blackwell Science, London

Dobson, A. J. (1983). Introduction tostatistical modelling.Chapman andHall, London

Donázar, J. A., Hiraldo, F. & Bustamante,J. (1993). Factors influencing nest siteselection, breeding density andbreeding success in the bearded vulture(Gypaetus barbatus). Journal ofApplied Ecology, 30: 504-514.


Eastman, J. R. (1999). IDRISI32.Worcester, MA, Clark Labs, ClarkUniversity.

Eastman, J. R. (1999). IDRISI32: Guideto GIS and Image Processing.ClarkLabs, Clark University, Worcester,MA


Guisan, A., Theurillat, J.-P. & Kienast, F.(1998). Predicting the potentialdistribution of plant species in analpine environment. Journal ofVegetation Science, 9: 65-74.

Hagemaijer, E. J. M. & Blair, M. J., Eds.(1997). The EBCC Atlas of EuropeanBreeding Birds: Their distribution andAbundanceT & A D Poyser, . London

Harrison, C. (1982). An Atlas of the Birdsof the Western Palearctic.Collins,London

Hastie, T. J. & Tibshirani, R. J. (1990).Generalized Aditive Models.Chapman& Hall, London

He, H. S., Mladenoff, D. J., Radeloff, V.C. & Crow, T. R. (1998). Integrationof GIS data and classified satelliteimagery for regional forest assessment.Ecological Applications, 8(4): 1072-1083.

Landis, J. R. & Koch, G. C. (1977). Themeasurement of observer agreementfor categorical data. Biometrics, 33:159-174.

Legendre, P. (1993). Spatialautocorrelation: trouble or newparadigm? Ecology, 74(6): 1659-1673.

Lennon, J. J. (1999). Resource selectionfunctions: taking space seriously?Trends in Ecology and Evolution,14(10): 399-400.

Ludden, T. M., Beal, S. L. & Sheiner, L.(1994). Comparison of the AkaikeInformation Criterion, the SchwarzCriterion and the F Test as guides tomodel selection. Journal ofPharmacokinetics andBiopharmaceutics, 2(5): 431-445.

Mac Nally, R. (2000). Regression andmodel-building in conservation


134

biology, biogeography and ecology:The distinction between -andreconciliation of- 'predictive' and'explanatory' models. Biodiversity andConservation, 9: 655-671.

McCullagh, P. & Nelder, J. A. (1989).Generalised linearmodelling.Chapman and Hall, London

Merrill, T., Mattson, D. J., Wright, R. G.& Quigley, H. B. (1999). Defininglandscape suitable for restoration ofgrizzly bears Ursus arctos in Idaho.Biological Conservation, 87: 231-248.

Monmonier, M. (1982). Computer-Assisted Cartography: Principles andProspects. Pages 76-80. Prentice-Hall,Inc., Englewood Cliffs, N.J.

Moreira, J. M. & Fernández-Palacios, A.(1995). Usos y coberturas vegetalesdel suelo en Andalucía. Seguimiento através de imágenes de satélite.Junta deAndalucía. Consejería de medioambiente,


Nelder, J. A. & Wedderburn, R. W. M.(1972). Generalised linear models.Journal of the Royal Statistical SocietyA, 135: 370-384.

Nicholls, A. O. (1989). How to makebiological surveys go further withgeneralised linear models. BiologicalConservation, 50: 51-75.

NOAA (1992). Experimental CalibratedGlobal Vegetation Index fromNOAA's Advanced Very HighResolution Radiometer, 1985-1991.(KGRD: 28). Global Change DataBase. (eds , . , . Boulder, Colorado .

Osborne, P. E., Alonso, J. C. & Bryant, R.G. (2001). Modelling landscape-scalehabitat use using GIS and remotesensing: a case study with great

bustards. Journal of Applied Ecology,38(2): 458-471.

Pereira, J. M. C. & Itami, R. C. (1991).GIS-based habitat modelling usinglogistic multiple regression: a study ofthe Mt. Graham red squirrel.Photogrammetric Engineering &Remote Sensing, 57: 1475-1486.

Preisler, H. K., Rappaport, N. G. &Wood, D. L. (1997). Regressionmethods for spatially correlated data:an example using beetle attacks in aseed orchard. Forest Science, 43: 71-77.

Purroy, F. J., Ed. (1997). Atlas de las avesde España (1975-1995)Lynx editions, .Barcelona

Rico Alcázar, L., Martínez, J. A., Morán,S., Navarro, J. R. & Rico, D. (2001).Preferencias de hábitat del Águila-azorPerdicera (Hieraaetus fasciatus) enAlicante (E de España) a dos escalasespaciales. Ardeola, 48(1): 55-62.


Sánchez-Zapata, J. A. & Calvo, J. F.(1999). Raptor distribution in relationto landscape composition in semi-aridMediterranean habitats. Journal ofApplied Ecology, 36: 254-262.

Smith, P. A. (1994). Autocorrelation inlogistic regression modelling ofspecies` distributions. Global Ecologyand Biogeography Letters, 4: 47-61.

Suárez, S., Balbontín, J. & Ferrer, M.(2000). Nesting habitat selection bybooted eagles Hieraaetus pennatus andimplications for management. Journalof Applied Ecology, 37: 215-223.

Titus, K., Mosher, J. A. & Williams, B.K. (1984). Chance-correctedclassification for use in discriminantanalysis: ecological applications. TheAmerican Midland Naturalist, 111(1):1-7.


135

Van Horne, B. (1983). Density as amisleading indicator of habitat quality.Journal of Wildlife Management,47(4): 893-901.

Vickery, P. D., Hunter, J., M.L. & Wells,J. V. (1992). Is density an indicator ofbreeding success? The Auk, 109(4):706-710.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO VIII

136

CAPÍTULO VIII: El uso de modelos regionales para identificar factores

limitantes y áreas con problemas de conservación: la distribución

y abundancia del milano real en la península Ibérica

RESUMEN

Los modelos predictivos de hábitat generan hipótesis acerca de los requerimientos

ecológicos de las especies y de los factores que afectan a su distribución, por lo que pueden

guiar la práctica de la conservación. En este trabajo se presenta un modelo regional de la

distribución y abundancia de milano real Milvus milvus en la península Ibérica. La

distribución y abundancia de milano real en cuadrados UTM de 100 km2, que se estimó

mediante censos en carretera, se modela con variables explicativas de grano grueso

obtenidas de imágenes de satélite, cartografía temática digital, datos meteorológicos y

coordenadas espaciales. El modelo de distribución incorporó principalmente variables

climáticas y alcanzó una gran capacidad discriminatoria en un conjunto de datos

independiente (Kapa=0.48[SE=0.07], AUC=0.92[0.01]). Por el contrario, el modelo de

abundancia incorporó en mayor medida variables de cobertura de vegetación y tuvo un

menor poder explicativo (r2=0.14). Las predicciones subestimaron algo los datos

registrados, lo que está de acuerdo con el declive de la población y del rango areal que se

ha observado para esta especie en el área de estudio. Los modelos son relevantes para la

conservación del milano real por dos razones principales: primero, sugieren los factores

limitantes para el milano real en la península Ibérica, y segundo, generan mapas predictivos

que destacan tanto los lugares donde existen serios problemas de conservación (aquellas en

las que áreas óptimas están desocupadas), como los lugares de los que se carece de datos

pero donde es probable que la especie esté presente.

Distribución del milano real en la península ibérica

137

CHAPTER VIII: Use of regional models to identify limiting factors and areas

with conservation problems: the distribution and abundance

of the Red kite in the Iberian peninsula

ABSTRACT

Predictive habitat modeling render insights into the ecological requirements of the species

and the factors affecting its distribution, and so can guide conservation practice. In this

work we present a regional model for the distribution and abundance of breeding red kite

Milvus milvus in the Iberian peninsula. Red kite occurrence and estimated abundance in 100

km2 UTM squares resulting from road census was modeled with coarse explanatory

variables obtained from satellite imagery, thematic digital cartography, meteorological data

and spatial coordinates. The occurrence model incorporated mainly climatic variables and,

assessed in a independent data set, have good discrimination ability

(Kappa=0.48[SE=0.07], AUC=0.92[0.01]), while the abundance model incorporated

mainly land-use variables and had a lower explanatory power (r2=0.14). The predictions

somewhat underestimate actual outcomes, what agrees with the declining of population size

and range observed for this species in the study area. These models are relevant in the

conservation of the species for two main reasons: first, they suggest the limiting factors for

red kite in the Iberian peninsula, and, second, they generate predictive maps that point out

both areas in which conservation problems may be acute (suitable locations are

unoccupied), and areas where no data is available but red kite is likely to be present.


138

INTRODUCTION

The development of effectiveconservation programs for a given speciesrequires clear understanding of its ecologicalrequirements, and of the factors determiningits distribution and abundance. However,regional-scale studies on the abundance anddistribution of species are difficult to perform,and results from regional-scale census arerarely analyzed in detail for conservationpurposes. The wider availability of digitalcartography and environmental data derivedfrom sensors onboard of satellites, plusrecently developed methods based on the useof GIS (Geographical Information Systems)and statistical modeling techniques such asGLM (Generalized Linear Models) or GAM(Generalized Additive Models), providepowerful tools that can be used to model thedistribution and abundance of species (see a

review in Guisan & Zimmermann 2000)considering relevant variables, such asclimate, topography, habitat type or structure,or human pressure (e.g.: Austin et al. 1996;Lavers & Haines-Young 1996; Corsi, Duprè& Boitani 1999; Sánchez-Zapata & Calvo1999; Osborne, Alonso & Bryant 2001).These models may then be used to evaluate ifthe census has covered adequately allpotential areas for the species and indicatewhere coverage should be improved. Modelscan also detect suitable areas for the speciesbut currently unoccupied, and consequently,may give insight into conservation problemsat regional or local scales, indicating areaswhere conservation actions should beprioritary (e.g.: Lawton & Woodroffe 1991;Donázar, Hiraldo & Bustamante 1993;Sánchez-Zapata & Calvo 1999; Osborne,Alonso & Bryant 2001; Teixeira, Ferrand &Arntzen 2001) .

Figure 1. (a) Study area. Topography of the Iberian peninsula with geographical names as used in the text. Darkershades correspond to lower altitudes. (b) Location of 10x10 UTM squares with red kite presence (black squares),absence (grey squares) or no data available (but presumably absence in most cases, white squares). The main breedingareas are highlighted: 1. southwestern Pyrenees, 2. central Pyrenees, 3. Northern Plateau (nucleus of Salamanca-Zamora), 4. Central Mountains, 5. Extremadura , 6. Eastern Sierra Morena, 7. Western Sierra Morena, 8. Doñanamarshland.

1

3 4

5

2

6

7

8

Pyrenees

Central m

ountains

Extremadura

Iberian mountains

CataloniaNorthernplateau

Sierra morena

Cantabric mountainsGalicia

Andalusia

Southernplateau

Por

tug

al

Spain

France

a b


139

The red kite (Milvus milvus) is theonly species of raptor that may be consideredas almost exclusively european (Cramp &Simmons 1980). With the exception of somesmall relict populations of uncertain status inMediterranean Northern Africa (Viñuela1996), the Iberian peninsula is the southernedge of the distribution of the species. Redkite populations disappeared or were stronglyreduced over all its range during the XIX andfirst half of XX centuries, mainly due tohuman persecution, and thus it wasconsidered a globally endangered species upto the 80s (Collar & Andrew 1988; Evans &Pienkowski 1991). After implementation ofprotection laws of raptors during the 60s-70s,populations of red kites in central Europequickly recovered, and the species evenrecolonized countries where it had becomeextinct long time ago (Evans & Pienkowski1991; Tucker & Heath 1994). As aconsequence, the red kite was removed fromthe list of endangered birds by the early 90s(Tucker & Heath 1994). However, by thattime the species had still a poor conservationstatus in the southern edges of its range(Viñuela 1996), and during the last 10 yearsalarming population declines have beendetected in the three main strongholds of thespecies, Germany, Spain, and France(Viñuela, Martí & Ruiz 1999; Mammen 2000;Mammen & Stubbe 2001; Thiollay 2001).The causes of these declines might be at thebreeding areas (Hille 1995; Mammen 2000;Thiollay 2001), but also at the main winteringarea, Spain, where most German and Frenchred kites spend about half their lives, andwhere serious conservation problems for thespecies have been detected in the last decade(Viñuela, Martí & Ruiz 1999; Viñuela &Contreras 2001; Viñuela & Villafuerte inpress). Spain held the second most importantEuropean breeding population of the speciesin 1994, estimated to be 3300-4100 breedingpairs (Viñuela, Martí & Ruiz 1999), but bythat time most populations for which datawere available were declining, apparently dueto illegal predator control by hunting and

poisoning (Villafuerte, Viñuela & Blanco1998).

The red kite is considered a relativelyeclectic species, with no important habitatrequirements, and able to breed in a widerange of climates (Cramp & Simmons 1980;Carter 2001). The only requirement oftencited for this species is a mixture of forestpatches to breed and open areas to search forfood. It has been suggested that humanlandscapes created in some agricultural areas,with a mixture of forest patches and opencroplands, probably favoured the expansionof the species in the past (Carter 2001).However, only partial and local data about thefactors determining distribution andabundance of the species have been published(reviewed in Viñuela, Martí & Ruiz 1999;Carter 2001), and it has been suggested thathabitat alterations induced by regional-scaleland-use changes may strongly affect itsbreeding success, abundance, and distribution(George 1995; Hille 1995; Mammen 2000;Thiollay 2001). Furthermore, the distributionof red kites in France, Germany and Spain ishighly fragmented, but the reasons for thatirregular distribution are poorly known(Cramp & Simmons 1980; De Juana 1989).It has been suggested that this irregularpattern of distribution may be explained byhuman factors, such as small populationspersisting where persecution levels are lower(De Juana 1989; Villafuerte, Viñuela &Blanco 1998; Carter 2001). However, naturalfactors such as habitat or climate couldcontribute to explain that irregulardistribution. This is why it is relevant toidentify the biological or ecological factorsdetermining red kite distribution, in order toimprove the identification of human-relatedconservation problems (e.g., poisoning orhunting).

In this paper we present regional-scalepredictive models of distribution andabundance of breeding red kites for theIberian peninsula (~600000 km2). The modelsconsider climatic, topographic, and habitat


140

factors, with the aim of improving ourunderstanding of which are the main naturalfactors determining the occurrence andabundance of the species. The Iberianpeninsula may be particularly adequate forthis study, because it is the southern edge ofthe range of red kites, and then presumablythe factors affecting its distribution areprobably more easily identified than at thecore of its range, where some limiting factormay be overlooked. Furthermore, we developa predictive model of the occurrence andabundance of red kites in Spain, with the aimof identifying areas suitable for red kites butcurrently unoccupied, and that thus may givea clue about which are the areas whereconservations problems are acting morestrongly, and where conservation programsshould have priority.

METHODS

Red kite data and variables consideredOur basic data are the results of the

national red kite census performed in 1994 formost of Spain (Viñuela, Martí & Ruiz 1999),plus data of a more detailed census ofAndalusia performed in 1996 (Bustamante,Donázar & Hiraldo 1997). In both censusesan average of 40 kilometers of road transectswere driven at a low speed in each of anumber of 10x10 km UTM squares. For the1994 census, the study area was stratifiedfollowing habitat and topographic criteria,and the census was conducted by > 500volunteer observers and regionalornithologists. Every stratum was sampledcompletely when possible, but if there wereinsufficient observers, a random sample ofsquares was selected trying to cover aminimum of 50 % of the area of each stratum.In a sample of 61 10x10 km UTM squares,red kite populations were surveyedsimultaneously by road transects and nestsearching/detection of territorial pairs bystandardized observation (Viñuela 1997;Viñuela, Martí & Ruiz 1999) . Within thissample we calculated an index of relativedensity (IRD, no of kites /100 km of transect)for each square sampled. A linear regression

of IRDs on estimated populations throughnest detection for those squares explained >85 % of the variance in IRDs, and thus roadtransect proved to be an adequate method tocensus this species in most of the Iberianpeninsula (Viñuela 1997). For the 1996census, the Andalusia AutonomousCommunity was stratified in nine zones thatcovered all historical and potential breedingareas for red kite. The UTM sampled squareswere those known to have had at least onebreeding pair the previous years and a setselected randomly within each zone in anumber proportional to the a priori expectedbreeding density (thus the known core zonesfor breeding had more squares sampled thanthe zones with a lower density of kites).About a 25% of total potential area wasfinally sampled. Some squares where no kiteswere seen in road transects, were additionallysurveyed by nest searching/detection ofterritorial pairs, because road transects maybe not an adequate method to detect thisspecies when breeding densities are low (< 3pairs/100 km2; Viñuela, 1997). Some areas ofSpain were not exhaustively sampled, butenough previous information existed toassume that no kites bred there; these squareswere included in the analysis as squareswhere the species was absent (Fig. 1). Thus,we had information about the occurrence ofthe species in 2990 10x10 km UTM squares(386 with presence of red kite and 2604 withabsence) and an estimate of breeding pairs foreach sampled square.

To build models to predict thedistribution and abundance of the species, wetested as predictors 11 variables: six land-use/land-cover variables derived fromCORINE digital map (CORINE 1991), twotopographic variables (altitude and slope)derived from a digital elevation model (DEM,a computerized representation of altitudecurves), two climatic variables (rainfall andtemperature), and an index of interannualplant productivity derived from satelliteimagery (PPI, Table 1). The raw data for thelatter is a monthly maximum value composite(an image) of a radiometric vegetation index


141

(NDVI) from the sensor AVHRR of theNOAA satellite (Mather 1999; Díaz-Delgado& Pons 2001). We performed a principalcomponent analysis on 166 images datedbetween 1984 and 2000, and selected the firstcomponent (our index PPI), which explained95.8% of variation in data (see for detailsEastman & Fulk 1993; Lillesand & Kiefer1994; Osborne, Alonso & Bryant 2001).Predictors were in origin at differentresolutions (from DEM at 100 m to Corine at250 m and NOAA imagery at 10x10 km) anddata were averaged within 10x10 km UTMsquares (Table 1).

Statistical analysisWe built Generalized Additive Models

(GAM, Hastie & Tibshirani 1990) ofoccurrence and abundance of breeding redkites in 10x10 km UTM squares. For the firsttype of model the response variable was thepresence/absence of red kite in each square,and we used a binomial error and a logisticlink, that is equivalent to logistic regression.

To model the breeding abundance theresponse variable was the estimated numberof breeding pairs per square, and we used aPoisson error with a log link that is equivalentto Poisson regression. In the abundancemodel we included only the squares with oneor more kites (n=386). We started buildingfull models that included all the predictors assmooth terms (a smoothing spline with 3degrees of freedom) and performed abackwards stepwise search of a best subsetmodel. In every step, we tested thesignificance of variables by a likelihood ratiotest of the current full model versus thereduced model without each particularvariable (Crawley 1993). Non-significantvariables (p>0.05) were tested with a simplerform (that is, with less degrees of freedom)and, if the effect was non-significant, theywere excluded from the model. We aimed toobtain a parametric model to facilitate thetransfer of results to a GIS for the generationof maps, and because parametric models may

Table 1. Predictive variables tested in the models for occurrence and abundance of breeding red kite in the Iberianpeninsula. A single mean value was obtained for each 10x10 km UTM square in the analysis.

Predictor Description SourceFOR Percentage of forest Modified from CORINE1

DEH Percentage of dehesas (sparsely forested areas, mainly ofQuercus ilex subsp. ballota and Q. suber.

Modified from CORINE1

PAS Percentage of pastureland Modified from CORINE1

TRE Percentage of tree cultures (mainly olive groves) Modified from CORINE1

IRR Percentage of irrigated cultures Modified from CORINE1

NIC Percentage of non-irrigated cultures Modified from CORINE1

ALT Mean altitude (m) Digital Elevation Model2

SLO Mean slope (degrees) Digital Elevation Model2

TEMP Mean annual temperature (10-1 ºC) Meteorological stations3

RAIN Mean annual precipitation (mm) Meteorological stations3

PPI Plant productivity index Satellite imagery4

1 Variables obtained by pooling the original 54 categories of the CORINE land cover cartography as follows: FOR isForests category; DEH, Agro-forestry areas; PAS, Pastures and Natural grassland; TRE, Permanent crops; IRR,Permanently irrigated land and Rice fields; and NIC, Heterogeneous agricultural and non-irrigated areas (except Agro-forestry areas). Resolution is 250 meters.2 Variables obtained from a Digital Elevation Model of the Iberian peninsula at 100 meters horizontal resolution.3 Raw data provided by the Spanish Instituto Nacional de Meteorología and spatially modeled at resolution 1km2 (owndata, unpublished).4 Raw data provided by the LATUV (Laboratorio de Teledetección de la Universidad de Valladolid) is a monthlymaximum value composite of a radiometric vegetation index (NDVI) from the sensor AVHRR of the NOAA satelliteat resolution 10x10 km (see text for details).


142

be more interpretable than complexnon-parametric curves (Guisan &Zimmermann 2000). So we finallytransformed all the smoothed variables tosuitable parametric terms guided by visualinspection of partial residual plots (see asimilar approach in Brown 1994; Franklin1998). For example, we transformedcurvilinear forms that showed a maximumeither to quadratic polynomials (see ALT inFig. 2b) or to piecewise linear functions witha threshold beyond which the response isconstant (see PAS in Fig. 2d). These modelswill be called hereafter environmentalmodels.

We converted the environmentalmodels into autologistic models (Augustin1996) to take into account the possible spatialautocorrelation of the squares (Legendre1993; Smith 1994), that is, the fact thatneighbouring squares are likely to havesimilar environmental characteristics or thathigh red kite density in a given square mayinfluence density in neighbour, lessfavourable squares. To this aim, we estimatedthe predicted probabilities of the models foreach 10x10 km UTM square in the Iberianpeninsula, then we took the average ofpredicted probability in each group of 9adjacent squares and included this newvariable —an autocovariable— within theenvironmental models (Wu & Huffer 1997;Merrill et al. 1999; Araújo & Williams2000). Finally, we modeled the amount ofunexplained variation in the autologisticmodels with the cartesian coordinates(latitude, longitude and their interaction) inUTM projection to account for regionaltrends in the pattern of distribution (Legendre1993; Preisler, Rappaport & Wood 1997). Weentered coordinates as non linear terms(natural cubic splines with 3 knots in the 0.1,0.5 and 0.9 quantiles, Harrell 2001).

There is not an unanimously acceptedmeasure of performance for logistic modelssuch as the coefficient of determination R2 inlinear regression (but see Ash & Shwartz1999). Therefore, we assessed the

discrimination ability (Pearce & Ferrier 2000)of the occurrence models with three differentmeasures. First, the commonly used correctclassification rate, which is affected by boththe unbalance between presences andabsences and the need to choose a thresholdto convert the estimated probabilities inpresences or absences (Fielding & Bell 1997).Second, the Cohen’s Kappa statistic, whichestimates the correct classification rateadjusting by chance (Titus, Mosher &Williams 1984). In this case we chose thethreshold to be the mid-point between themean estimate for presence and the meanestimate for absence (Fielding & Haworth1995). And third, the Area Under the Curve(AUC) of a Receiver Operating Characteristic(ROC) plot (Swets 1988; Cumming 2000),which is an index of rank correlation betweenpredicted probability of presence and actualobservations (Harrell 2001). From all possiblepairs of squares, one with and the other onewithout breeding kites, the AUC measures theproportion of such pairs in which theoccupied square has a higher probability ofpresence than the unoccupied square (Centor1991; Zweig & Campbell 1993). AUC doesnot require to choose a threshold to convertprobabilities in presences or absences, and itis unaffected by the unbalance between them(Manel, Williams & Ormerod 2001). Finally,we used Spearman correlation (rs) to analyzethe agreement of predicted and actualabundance, but Pearson correlation (r) tomeasure the amount of explained variation inboth the occurrence and abundance models(Mittlböck & Schemper 1996).

To evaluate the models we followed adata-splitting strategy (Verbyla & Litvaitis1989; Picard & Berk 1990), developing themodels with a random selection of 75% of thesquares (the training set) and holding the restof data to evaluate the models (the test set).The correct estimates of discrimination abilityof the models in new scenarios are those forthe test set, that will be reported here. Finalestimates of model coefficients were obtainedwith the complete dataset, and predictionswere produced for all squares in the Iberian


143

peninsula, including Portugal and theunsampled Spanish areas.

Models were developed with S-PLUS2000 (MathSoft 1999) and AUC calculatednon-parametrically with AccuROC 2.5 (Vida1993).

RESULTSOccurrence model

The environmental model was highlysignificant (p<0.0001) and included 8 of theoriginal set of 11 variables (Table 2). Thevariable PPI (plant productivity) generatedthe major change in deviance (about 30% ofthe total), and was included in the model as aquadratic polynomial (Fig. 2a). Red kiteoccurrence showed a maximum at a value of1896 units of our index (range 0-2500). Thetopographic variables accounted for a further

25% of the model change in deviance (Table2). Red kite probability of occurrence had aquadratic relationship with ALT and apiecewise linear relationship with SLO, withmaximum probability of occurrence foraltitudes around 850 m, and decreasingprobability of occurrence with topographicruggedness (Figs. 2b and 2d). The mostimportant land-use/land-cover variable relatedto red kite occurrence was PAS, so that anincreasing coverage of pasturelands wasassociated with increasing probability ofoccurrence (Fig. 2c). Other land cover andclimatic variables had less importance, interms of change in deviance (Table 2). Themodel showed a linear decrease in probabilityof occurrence with increasing TRE, a slightlinear increase with increasing NIC, a smallmaximum for intermedious values of FOR(65%) and, finally, a stairway-like decreasewith TEMP (Figs. 2f-2h).

Table 2. Deviance table of the environmental model for occurrence of red kite in 10x10 km UTM squares. Change indegrees of freedom and change in deviance associated with each variable is estimated by comparison of the reducedmodel without each particular variable against the saturated model. Names of variables as in table 1.


Change indf

Residualdeviance

Change indeviance

P-value

Null 2416 1882Saturated 2403 13 1289 593Intercept -17.406 2.346PPI 0.022 0.002PPI2 -5.8*10-6 -5.9*10-7 -2 -168 <0.0001

SLO* -0.851 0.117 -1 -89 <0.0001PAS* 0.068 0.008 -1 -87 <0.0001ALT 0.007 0.001ALT2 -4.1*10-6 6.2*10-7 -2 -58 <0.0001

TEMP1* -5.634 1.133

TEMP2 -4.868 1.031TEMP3 -7.032 1.252

-3 -44 <0.0001

TRE -0.061 0.019 -1 -8 0.005NIC 0.047 0.004 -1 -8 0.005FOR 0.026 0.012FOR2 -2.0*10-4 1.6*10-4 -2 -7 0.029

*SLO, PAS and TEMP modeled as piecewise linear functions (see Figs. 2c, 2d, and 2h).


144

Figure 2. Partial residual plot for the individual predictors in the occurence model and their corresponding map. Thepartial residual plot is a plot of ri+bkXik versus Xik, where ri is the deviance residual for the i-th observation, Xik is thevalue for the k-th predictor and the i-th observation, and bk is the regression coefficient for the k-th predictor (Hastieand Tibshirani 1990). Lines show the relationship between each predictor and the response variable (probability of redkite presence, vertical axes), keeping constant the rest of predictors. Circles correspond to partial residuals for eachobservation. To have a spatial representation of the effect of each variable, we show maps of predicted values for eachterm (in the predictor scale) categorized in five 20-quantile levels. Darker squares indicate a higher predictedprobability of presence. PPI: plant productivity index, ALT: mean altitude (m), SLO: mean slope (degrees), PAS:percentage of pastureland, TRE: percentage of tree cultures, NIC: percentage of non-irrigated cultures, FOR:percentage of forest, TEMP: mean annual temperature (10-1 ºC).

ALT0 500 1000 1500 2000

-10

-50

5

SLO0 2 4 6 8 10 12 14

-50

510

PAS0 20 40 60 80

-50

5

Percentiles of predicted presence

0.2 0.4 0.6 0.80 1

PPI500 1000 1500 2000 2500

-15

-10

-50

5

a

d

c

b


145

Figure 2 (cont.). Partial residual plot for the individual predictors in the occurence model and their corresponding map.

FOR0 20 40 60 80

-50

5

NIC0 20 40 60 80 100

-50

5

TRE0 20 40 60 80

-50

5

Percentiles of predicted presence

0.2 0.4 0.6 0.80 1

TEMP50 100 150

-50

5

e

h

g

f


146

The autocovariable was alsosignificant (change in deviance = 128.8,change in df = -1, p<0.0001) and improvedthe discrimination ability of the model (Table3). Final models with the addition of thespatial coordinates had a correct classificationrate of some 8 to 9 in each 10 squares, whichmeans a 48% more than what expected bychance (as estimated by Kappa). According toAUC, 9 in each 10 pairs of squares (onesquare occupied and the other oneunoccupied) are correctly rated. Overall, theoccurrence model could explain about a 40%(r=0.63, t=44.6, p<0.0001) of the variability

of the data according to the Pearsoncorrelation between prediction and actualoutcomes (Mittlböck & Schemper 1996),which is a fair amount considering the lowvalues of explained variance (or deviance)obtained typically in logistic regressions (Cox& Wermuth 1992; Ash & Shwartz 1999).Accordingly, a calibration plot shows a highagreement between observations andpredictions (Fig. 3a). It is interesting to notethat predicted probabilities of occurrencecorrelated also significantly with estimatedabundance (rs=0.51, z=27.9, p< 0.0001).

Table 3. Estimates of discrimination ability for the red kite occurrence models in 10x10 km UTM squares. Theenvironmental model is described in table 2, the autologistic model is the same but entering an autocovariate (themean of predicted probabilities in 9 adjacent 10x10 km UTM squares) simultaneously with the rest of predictors, andthe final model included the spatial coordinates to fit the unexplained variation in the autologistic model. Standarderrors are given between parentheses (approximate SE for Kappa following Titus et al. (1984) and asymptotic SE forAUC according to Vida (1993)).

Model type Correct classificationrate

Kappa AUC

Environmental 0.84 0.35(0.07) 0.84(0.02)Autologistic 0.86 0.42(0.07) 0.88(0.02)Final (plus spatial coordinates) 0.87 0.48(0.07) 0.92(0.01)

Table 4. Deviance table of the environmental model for abundance. Change in degrees of freedom and in devianceassociated with each variable is estimated by comparison of the reduced model without a particular variable against thesaturated model. P-values corrected for underdispersion. Names of variables as in table 1.


Change inDf

Residualdeviance

Change indeviance

F P-value

Null 278 12.07Saturated 272 -6 9.54 -2.53 9.01 <0.0001Intercept -4.366 2.108SLO* -0.302 0.319 -1 -0.90 19.33 <0.0001ALT 0.003 0.005ALT2 -1.5*10-6 3.1*10-6 -2 -0.46 4.91 0.008

PAS* 0.017 0.027 -1 -0.39 8.32 0.004IRR -0.034 0.075 -1 -0.28 6.03 0.015TRE -0.053 0.125 -1 -0.24 5.20 0.023

* SLO and PAS modeled as piecewise linear functions (see Figs. 5a and 5c).


147

The final model predicted a highprobability of occurrence in the four mainbreeding areas of the species in the Iberianpeninsula (Fig. 4): southern slopes of thePyrenees, western side of the NorthernPlateau (between Salamanca and Zamoraprovinces), both sides of the CentralMountains, and Extremadura (see also areas2, 3, 4, 5 and 6 in Fig. 1). However, the modelpredicted the occurrence of red kites in anarea larger than the current range of thespecies, where red kites are absent or presentonly sparsely (e.g. Cantabric and Iberianmountains, and Catalonia). On the other hand,the model predicts a low probability ofoccurrence in the Doñana marshlands(population 8, Fig. 1), where a small densepopulation occurs in an atypical area (inDoñana red kites breed in the narrow edge ofmarshland with pine forest, at sea level).

Abundance model

The environmental model was highlysignificant (p<0.0001) and included 5 of theoriginal set of 11 variables (Table 4), amongwhich topographic variables accounted formost of the change in deviance (about 54%).SLO again suggests a lower predictedabundance for more rugged areas (Fig. 5a),while ALT, that entered as a quadraticpolynomial (Fig. 5b), suggests a higherabundance for intermedious altitudes(maximum at 900-1000 m). The rest of thevariables in the model belong to the land-use/land-cover type: IRR and TRE, with anegative sign, and PAS, with a positive sign(Figs. 5c-5e).

Figure 3. (a) Calibration plot for the red kite occurrence model (following Pearce and Ferrier, 2000). Thepredicted probability of occurrence (in ten equi-interval classes) is plotted against the observed proportion ofoccupied squares in the test set. The number of evaluation squares and the 95% confidence interval for theobserved occurrence is shown for each class. (b) Calibration plot for the abundance model. The predictednumber of breeding pairs is plotted against the observed number in the test set (the model did not predict anysquare having 11 to 12 and 13 to 15 pairs, so the line joining the points is broken). The number of evaluationsquares and the 95% confidence interval for the mean abundance is shown. In both graphs the thindiscontinuous line shows a perfect relationship between observations and predictions.

Predicted probability of occurrence

Obs

erve

d oc

curr

ence

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0 .4

0 .6

0 .8

1 .0

333

91

52

36

20

7 11

7

5 11

Predicted number of pairs

Obs

erve

d nu

mbe

r of

pai

rs

0 6 10

24

68

1012

9

12

22

19

15 9

1 3

1

2

1 2

11

2 4 8 12 14 16

a b


148

Figure 4. Final predictions of red kite occurrence (probability of presence) in 10x10 km UTMsquares in the Iberian peninsula.

The autocovariable was alsosignificant (change in deviance = 0.48,change in df = 1, F=10.48, p=0.001), andimproved the predictive ability of the modelfrom rs=0.21 (z=2.2, p= 0.003) to rs=0.29(z=2.9, p= 0.003). The inclusion of thegeographical coordinates further enhanced theaccuracy of the model (rs=0.41, z=4.3, p<0.0001). Overall, the abundance model couldexplain only a small amount of the variabilityof the data (r2=0.14, p=0.0001, Mittlböck &Schemper 1996). The average modelpredictions agreed closely with observationsin squares with low abundance, butoverestimated the number of breeding pairsfor squares with predicted numbers above 9pairs per 100 km 2 ( Fig. 3b).

Final predicted abundances werehighest for Western side of the Northernplateau, Southern Pyrenees, Centralmountains and some areas in Extremadura,where the maximum densities of red kiteactually occurr. Predicted abundance was lowin areas of the Cantabric Mountains and theSouthern Plateau were low abundances havebeen observed. However, predictedabundance was high in other areas where nohigh density of red kites actually occur, suchas some parts of the Iberian Mountains,Catalonia and a single square in Andalusia(Fig. 6).

Background

0 to 0.2

0.2 to 0.4

0.4 to 0.6

0.6 to 0.8

0.8 to 1


149

Figure 5. Partial residual plot for the individual predictors in the abundance model and their corresponding map. Thepartial residual plot is a plot of ri+bkXik versus Xik, where ri is the deviance residual for the i-th observation, Xik is thevalue for the k-th predictor and the i-th observation, and bk is the regression coefficient for the k-th predictor (Hastieand Tibshirani 1990). Lines show the relationship between each predictor and the response variable (abundance ofkites, vertical axes), keeping constant the rest of predictors. Circles correspond to partial residuals for eachobservation.. To have a spatial representation of the effect of each variable, we show maps of predicted values for eachterm (in the predictor scale) categorized in five 20-quantile levels. Darker squares indicate a higher predictedabundance. IRR, percentage of irrigated cultures; rest of variables as in Figure 2.

ALT0 500 1000 1500

-1.5

-0.5

0.5

1.0

SLO0 2 4 6 8 10 12 14

-1.0

0.0

0.5

1.0

1.5

PAS0 20 40 60 80

-1.0

0.0

0.5

1.0

1.5

Percentiles of predicted abundance

0.2 0.4 0.6 0.80 1

a

c

b


150

Figure 5 (Cont.). Partial residual plot for the individual predictors in the abundance model and their correspondingmap.

DISCUSSION

Factors explaining red kite distribution

Although considered an eclecticspecies our models show that the red kitedistribution can be easily predicted fromclimate, topography and vegetation cover.The irregular and fragmented distributioncurrently observed in Spain can be explainedlargely by these factors although it may havebeen also influenced by the intensity ofpresent or past human persecution upon thespecies (Viñuela, Martí & Ruiz 1999). Thevariables PPI, ALT and TEMP represent 45%of the deviance explained by ourenvironmental model. We consider thesethree variables as representing the response of

red kite to climate (although ALT wasderived from the digital elevation model, asSLO, we will consider it a surrogate forclimatic conditions in the discussion).

PPI is an index of plant productivityderived from NOAA AVHRR imagery that,at this spatial scale, represents the response ofvegetation growth to rainfall and temperature.The bell-shape response of red kite ocurrenceto altitude probably also reflects a selection ofparticular climatic conditions in the Iberianpeninsula. The shape of the decliningresponse to mean anual temperature shows anavoidance of the more Mediterranean andsemiarid locations. We do not know theultimate cause of the response of red kites toPPI. The highest values of the index are

TRE0 10 20 30

-2.0

-1.0

0.0

1.0

IRR0 20 40 60

-2.0

-1.0

0.0

Percentiles of predicted abundance

0.2 0.4 0.6 0.80 1

d

e


151

obtained in rainy areas of Atlantic climate inthe north of the Iberian peninsula and thelowest values in the semi-arid Mediterraneansoutheast. PPI values optimum for the red kite(intermediate-high values) occur in areassurrounding the mountain ranges and increasein areal extent following a southeast tonorthwest gradient (Fig. 2a). The distributionof optimum PPI values in the peninsulafollows quite closely the known red kitedistribution with the exception of some smallareas in the southeastern half that currently donot hold breeding populations (compare Fig.1b and 2a). The predictive ability of PPIcould be due to an avoidance by red kites ofrainy climates and at the same time apreference for areas of relatively highbiological productivity. In a similar species,the black kite Milvus migrans, it has beenobserved that rain in spring has a directnegative effect on hatching success (Viñuela& Sunyer 1992), and on nestling’s growthrate (Hiraldo, Veiga & Mañez 1990). Theremnant red kite population in Wales (U.K.),living in a rainy climate, has one of the lowestbreeding rates known for the species, due tolow hatching success and high nestlingmortality, apparently associated to poorfeeding rates (Lovegrove 1990; Newton,Davis & Moss 1994). In contrast, thereintroduced population in east England, inone of the driest areas in the country, ishaving a relatively high breeding success(Carter 2001, pers. comm.). The PPI indexprobably also reflects average biologicalproductivity of the area. The red kite isbasically an opportunistic searcher, able tohunt only small size, handicapped or easilycaught live prey, and relies largely on smallcarrions (see Cramp & Simmons 1980;García, Viñuela & Sunyer 1998; Carter2001). Areas with low PPI values must beareas with relatively low average biologicalproductivity, where the kind of prey orcarrion searched by red kites may be morescarce. This could be especially true duringthe typical hot summer drought periods ofMediterranean latitudes. Furthermore, thisunfavourable summer period covers a criticalstage, such as the period of transition to

independence of fledglings (Bustamante1993). The preference of relatively highground (optimum around 850 m) wheresummer drought is shorter and less intense(Font Tullot 1983) and the negativerelationship between red kite occurrence andmean annual temperature gives aditionalsupport to the idea that hot summers may be alimiting factor on the distribution of thespecies.

To test the idea that biologicalproductivity may be setting a lower limit tothe distribution of the red kite and that rainfallhas a negative effect once productivity iscontrolled for, we built an alternativeenvironmental model of occurrence in whichPPI was modeled as a piecewise linearfunction with a constant effect after themaximum of 1896 units (that is, we assumethat productivity has a positive linear effecton red kite distribution below this value but aconstant effect above) and RAIN wasintroduced as a linear term. This alternativemodel indicated that there was a significantnegative relationship between red kiteoccurrence and rainfall once the effect of PPIwas controlled for. This model had apredictive ability similar to the environmentalmodel (Kappa= 0.45[SE=0.08],AUC=0.83[0.03]) (compare with Table 3),but being less parsimonious it was notselected by our statistical analysis procedure.

Topography was the next mostimportant factor affecting red kitedistribution. The variable SLO wasresponsable for a 15% of the reduction indeviance of the environmental model.According to this variable, red kites show aclear preference for rather flat areas (at ourcoarse resolution: 10x10 km) or an avoidanceof the more rugged areas, what corresponds toa selection of mid-mountain locations asindicated by the relationship with altitude.This effect of topography in red kitedistribution had been previously suggested byseveral authors (Meyburg 1973; Elósegui1985). The absence of the species from themore rugged terrain in mountain areas could


152

look surprising at first sight, because these areusually the best areas for raptors in Spain,since they provide adequate breeding sitesand usually are exposed to lower illegalpredator control (González, Bustamante &Hiraldo 1990; Sánchez-Zapata & Calvo1999). The reason could be the scarcity of theextensive open lands prefered for hunting oran avoidance of species of birds of prey oflarger size such as imperial eagles Aquilaadalberti, golden eagles Aquila chrysaetos oreagle owls Bubo bubo that can predate on redkites (Cramp & Simmons 1980; Ferrer 1993;Serrano 1998; Serrano 2000), and are moreabundant in these areas. The selection of mid-mountain areas could also be related to the

foraging technique of the species. Red kitesare superb gliders with one of the lowest wingloading among raptors, and show a markeduse of slope air-currents (Lovegrove 1990).Hilly terrain typical of mid-mountain areascan be specially favourable for this kind offlight, providing cover from strong winds,and lifting on days with less wind, and thusprobably allowing an optimal energeticbalance for the long searching flights typicalof its foraging method. This relationshipbetween relief and red kite distribution is alsoknown for the Welsh population, where this isprobably one of the main factors affectingselection of nesting sites (Lovegrove, 1996;Carter, 2001; Doody, pers. comm.).

Figure 6. Final predictions of red kite abundance (in pairs/100km2) in 10x10 km UTM squares in theIberian peninsula.We represent only the squares for which presence of breeding red kite waspreviously predicted using a threshold of probability P>0.30, which is the average between thepredicted probabilities for squares with recorded presence of red kite and the probabilities for squareswith absences, (see Fielding and Haworth, 1995; Guisan et al. 1998).

no breeding

1 pair

2 pairs

3-4 pairs

5-6 pairs

7-9 pairs

> 9 pairs


153

Interestingly, this pattern ofdistribution affected by climate andtopography probably could also be applied tothe red kite population in France, where theyare absent from the major part of the Atlanticwest of the country, are very scarce inMediterranean areas in the southeast, andreach highest densities in mid-mountain areasand mountain valleys of Pyrenees, MassifCentral, the Vosgues and Alsacie (Voisin1994).

The variables representing land-use/land-cover constitute the third group ofpredictors of red kite distribution. Theyrepresent aproximately 20% of the devianceexplained by the model, but a single variable,the surface occupied by pasturelands, is themost important variable explaining by itself15% of the deviance.The other threevariables: extent of forest, non-irrigatedcultures and tree cultures represent a minimalinfluence (approximately 1 % each) in themodel. The importance of pasturelandsconfirms at a larger scale the importance ofthis kind of habitat reported locally inGermany (Hille 1995). In fact, recent declinesin some German populations have beenclaimed to be caused by a reduction in surfaceor change of management of pasturelands(Stubbe, Mammen & Gedeon 1995).Pasturelands may constitute the optimalforaging habitat for red kites, because theseare extensive open areas, often with relativelylow vegetation cover due to grazing ormowing, and where they may find plenty ofeasy prey such as Microtines andinvertebrates (Hille 1995). In Spain, largeconcentrations of red kites during breedingseason have been observed in recently mowedpasturelands of Pyrenees, and this is a habitatcommonly used for food searching by redkites in other areas of Spain too (Viñuela,Martí & Ruiz 1999). This association withpasturelands could aditionally explainselection of mid-mountain areas, because inmost of Spain the presence of mowing fieldsis restricted to mountain areas (e.g. typically

between 900 and 1300 m. in the PyreneesLasanta 1989).

Although playing a minor role indistribution, probability of occurrence of redkites was maximal at intermediatepercentages of forest cover and was positivelyaffected by the percentage of surface coveredby non-irrigated cultures, supporting the ideathat red kites select areas with some forestproviding breeding sites, but avoiding denselyforested areas. Cereal fields are also a habitatoften used for foraging in Spain, especially inwinter and during harvesting (García et al.1998; pers.obs.). Furthermore, voles Microtusarvalis and Arvicola terrestris, the main preysof red kite (Cramp & Simmons 1980; García,Viñuela & Sunyer 1998), have invadedagricultural habitats of Northern Spain duringlast 30 years, and now there are plagues asthose found at more northern latitudes (Bonal& Viñuela 1998). Preference for open areaswith herbaceous vegetation (pastures, cerealfields) could also explain the negative effectof cover of tree cultures on probability ofoccurence.

Somehow surprisingly, the variableDEH, indicating surface covered by dehesas(extensive pasturelands devoted to cattle,sheep or pig raising, with scattered oaksQuercus spp. or ashes Fraxinus spp.), did notenter the distribution or abundance models. Ithas been argued that these open dehesasconstitute the most important breeding habitatfor red kites in Spain (Viñuela, Martí & Ruiz1999), having the optimal landscapestructure: extensive open lands for huntingand large scattered trees for breeding.However, dehesas does not seem to be a goodpredictor by themselves, as there areextensive areas of dehesas in Extremaduraand Andalusia without breeding red kites, andthose where the species breeds are alreadyidentified by the combination of climatic andtopographic variables. Also, the class agro-forestry areas of the CORINE land-use/land-cover map that we used to assess the extent ofdehesas, may be quite heterogeneous in


154

regards of tree and shrub cover, since thatclass includes a range of different landscapes,from dehesas with too closed canopies or toomuch shrub cover to others highly managedin which trees are small or subject to severepruning, and thus not suitable for breeding oflarge raptors (Viñuela, Martí & Ruiz 1999).

Factors explaining red kite abundance

Red kite abundance resulted moredificult to model than distribution. Two of theclimatic factors with a large effect in thedistribution model, PPI and TEMP, were notsignificant in the abundance model. Onlyaltitude indicated a greater red kite abundancein mid-mountain areas and was responsablefor aproximately 18% of the reduction indeviance of the model. Topography (meanslope) was responsable for another 35%reduction and land-use/land-cover variablesfor another 36%. The abundance modelindicated in general a pattern that agrees withthe distribution model: greater abundances inmid-mountain areas with relatively plainrelief and more than 30% pasture lands, and arejection of the more intensively managedagricultural landscapes (irrigated cultures andtree crops). Contrary to the occurrence model,in the abundance model land-use factors hadmore weight than topography and climate (ifwe consider altitude as a surrogate forclimatic conditions).

The calibration plot of the abundancemodel (Fig. 3b) indicated a very goodadjustment of the model to test data for lowdensities but a clear overestimation forpredicted values > 9 breeding pairs per 100km 2. This result is congruent with an upperdensity limit in most habitats caused by redkite territoriality; but could also be the resultof the species decline by illegal predatorcontrol that left habitats insaturated. Thissecond explanation seems more realistic if weconsider that there exist a few areas with veryhigh red kite breeding densities (20 pairs per100 km 2 ) above those predicted by ourmodel, and the fact that the calibration plotfor the distribution model (Fig 3a) also

indicates a slight overestimation (8 out of 10values are below the equal probability line).

Observed vs. predicted distribution andabundance

Overall, the occurrence modelexplained a fair amount of variance and itspredictive power was high, according to thestandards suggested by several authors(Monserud & Leemans 1992; Fielding &Bell 1997; Pearce & Ferrier 2000) andcomparing our results with those reported inprevious habitat models (Austin et al. 1996;Manel, Dias & Ormerod 1999; Tobalske &Tobalske 1999; Cumming 2000; Bonn &Schröder 2001; Osborne, Alonso & Bryant2001; Rico Alcázar et al. 2001). Moreover,the correlation between predictedprobabilities and estimated abundance (higherprobabilities of presence for squares withmore kites) gives further confidence in themodel. However, the abundance model had arelatively low predictive power.

The models predicted occurrence ofred kites in areas where the species is veryscarce or does not currently breed. The moststriking case are the southern slopes of theCantabric mountains (provinces of León andPalencia), where the habitat seems to be goodfor the species, but where only isolated pairsor small populations in restricted areas of themountains have been found. The highprobability of presence predicted in this area(Fig. 4) contrast with the low values predictedby the abundance model (Fig. 6) . Perhapsthis discrepancy could be explained by factorsnot considered in the models, and related withhuman activities such as an intense use ofrodenticides (to which red kite is particularlyvulnerable Carter 2001; Thiollay 2001), andthe recent increase of irrigated cultures in thatzone. The fact that land-use variables havemore weight in the abundance modelcompared to climatic variables in thedistribution model support this view.

Our models also predict the presenceof red kites in some areas of southern Spain in


155

the western border of the Southern Plateauwhere only small relict populations or evenisolated pairs remain. The reasons behind thisdiscrepancy are likely human factors, mainlyillegal predator control (Villafuerte, Viñuela& Blanco 1998). Game hunting has a higheconomic importance in this part of Spain,and illegal activities to control predators, suchas poisoning, are still culturally rooted in therural societies of these areas. Furthermore,population crash in the two main small gamespecies in Spain (rabbits Oryctolaguscunniculus and red-legged partridgesAlectoris rufa), along with increasing use ofintensive game management (e.g. massivereleases of partridges), have induced anstrong illegal persecution of predators duringlast 10 years (Villafuerte, Viñuela & Blanco1998; Viñuela & Villafuerte in press). InExtremadura, where the problem of illegalpredator control seems to be lower(Villafuerte, Viñuela & Blanco 1998;Viñuela, Martí & Ruiz 1999; Mañosa 2002),there is still a healthy population of red kites.

A remarkable case is the population inDoñana National Park (population 8 in Fig.1b) situated in an area where our modelpredicts a low probability of occurence. Thisdense population is situated at the highlyproductive narrow edge of a large marshlandin an atypical breeding habitat from the pointof view of our models: a flat area of scrublandwith scattered trees at sea level, far from themountains, that has low average rainfall andhigh mean annual temperature. Red kites mayexist there because the exceptionalproductivity of the marshes apparently notwell reflected by our variable PPI . Thepredicted probabilities of occurrence are verylow for this area, which agrees well with lowbreeding success of red kites and theimpression that Doñana population –that hasbeen declining for the last decade– issupported by immigration from otherpopulations (Bustamante, Donázar & Hiraldo1997; Viñuela, Martí & Ruiz 1999, andpersonal observation).

The red kite had a much largerdistribution in the Iberian peninsula in the

past than the one observed during 1994 and1996 (Cramp & Simmons 1980; Purroy1997). Our occurrence model predicts an areamuch smaller than the distribution providedby Cramp & Simmons (1980). Comparingboth, the red kite seems to have disappearedfrom the most arid areas in the southeasternborder where our model indicates the habitatwas suboptimal for the especies. The speciesdecline seems to continue and evenpopulations in optimal habitat of northwesternSpain are suffering a marked decrease(estimated reduction of 50 % in NorthernPlateau between 1994 and 2001 Viñuela &Contreras 2001).

Conclusions derived from models

The occurrence model built for the redkite was discriminative according to thevalidation with independent data. It predictsoccurrence in areas that we can consider asthe optimal for breeding according to climate,topography and vegetation cover. The mapselaborated for each of the predictorsindividually (Fig. 2 and 5) suggest which arethe limiting factors for the species in differentparts of the Iberian peninsula. Mountainlocations in eastern and southern Spain,where the species is absent, tend to havesmall areas with optimum values of PPI andthese tend to be obtained at higher altitudesand in more rugged landscapes than what redkite apparently prefer. So there is not a spatialcoincidence of the main limiting factors inthese locations.

Another point indicated by our modelsis that some areas of optimum habitat werenot thoroughly covered by the 1994-1996census. We suggest that these areas should beprospected more thoroughly in the future,since the red kite may still breed but haveremained unnoticed. A real absence of thespecies may indicate some conservationproblem that needs to be identified. Evenareas where species is known to be absent byrecent detailed atlas work (e.g. Catalonia)could be colonized by the species in thefuture, and should receive adequate attention.


156

On the other hand, they can haveconservation problems that our models arenot able to detect. We think that the spatialpredictions from our models can help tooptimize the allocation of the always limitedresources to estimate the species populationsize and trend in the future.

Our model also remarks that the redkite population in Doñana National Parkoccupies a very atypical habitat. Although,most of the studies on the ecology of thespecies in Spain have taken place in thispopulation (Blanco, Hiraldo & Heredia 1990;Veiga & Hiraldo 1990; Heredia, Alonso &Hiraldo 1991; Hiraldo, Blanco & Bustamante1991; Blanco, González & Hiraldo 1992;Viñuela & Bustamante 1992; Bustamante1993; Hiraldo, Heredia & Alonso 1993;Bustamante, Donázar & Hiraldo 1997; VanKleef & Bustamante 1999), but their resultsmight not be extrapolable to otherpopulations. Considering our models, wesuggest that future research effort should bedirected to more typical habitats whereconclusions can be more widely applicablefor the management of the species.

Summarizing, it is very possible thatthe general pattern of occurrence of red kitesin the Iberian peninsula is mainly determinedby natural factors such as climate, topographyand vegetation cover, while their currentpopulation abundance is modulated byhuman-related factors (e.g. González,Bustamante & Hiraldo 1990; Donázar,Hiraldo & Bustamante 1993). We have notincluded anthropogenic factors in the models,that are very difficult to quantify, and thiscould explain the relatively low predictivepower of the abundance model. Taking intoaccount just natural factors such as climate,topography or habitat, red kites should bemore extended in Spain than they currentlyare.

ACKNOWLEDGEMENTS

The national census of Red kites wasfunded by the Royal Society for theProtection of Birds, and coordinated bySEO/Birdlife. More than 500 people (listed inthe appendix), participated in the census. Redkite census in Andalusia was funded by theConsejería de Medio Ambiente (Junta deAndalucía) and coordinated by F. Hiraldo.LATUV (Laboratorio de Teledetección de laUniversidad de Valladolid) kindly providedraw data for our index of interannual plantproductivity. Part of the data analysis andwriting of this paper was funded by project#1DF-97-0648 from the Ministerio de Cienciay Tecnología and FEDER program from theEuropean Union. JS acknowledges apredoctoral fellowship (Ministerio deEducación y Ciencia) during part of thewriting of the paper.

NOTESA version of this manuscript, with J.Viñuela,R.Díaz-Delgado y J.Bustamante, is beingreviewed in Biological Conservation

LITERATURE CITED

Araújo, M. B. & Williams, P. H. (2000).Selecting areas for species persistenceusing occurrence data. BiologicalConservation, 96(3): 331-345.

Ash, A. & Shwartz, M. (1999). R2: a usefulmeasure of model performance whenpredicting a dichotomous outcome.Statistics in Medicine, 18: 375-384.

Augustin, N. H. (1996). Regression methodswith spatially referenced data. Aspects ofApplied Biology, 46: 67-74.

Austin, G. E., Thomas, C. J., Houston, D. C.& Thompson, D. B. A. (1996). Predictingthe spatial distribution of buzzard Buteobuteo nesting areas using a GeographicalInformation System and remote sensing.Journal of Applied Ecology, 33: 1541-1550.

Blanco, J. C., González, J. L. & Hiraldo, F.(1992). Trophic and spatial relationshipsbetween wintering red kites (Milvusmilvus) and marsh harriers (Circus


157

aeruginosus) in the Guadalquivir marshes.Micelania Zoologica, 14: 161-166.

Blanco, J. C., Hiraldo, F. & Heredia, B.(1990). Variations in the diet and foragingbehaviour of a wintering red kite (Milvusmilvus) population in response to changesin food availability. Ardeola, 37: 267-278.

Bonal, R. & Viñuela, J. (1998). Las plagas detopillo en España: enigmas, folcklore yproblemas de conservación. Quercus, 146:35-39.

Bonn, A. & Schröder, B. (2001). Habitatmodels and their transfer for single andmulti species groups: a case study ofcarabids in an alluvial forest. Ecography,24: 483-496.

Brown, D. G. (1994). Predicting vegetationtypes at treeline using topography andbiophysical disturbance variables. Journalof Vegetation Science, 5: 641-656.

Bustamante, J. (1993). Post-fledgingdependence period and development offlight and hunting behaviour in the red kiteMilvus milvus. Bird Study, 40: 181-188.

Bustamante, J., Donázar, J. A. & Hiraldo, F.(1997). Factores que condicionan ladistribución reproductora del milano real(Milvus milvus) en Andalucía: elaboraciónde un plan de conservación. Sevilla, AMA-CSIC.

Carter, I. (2001). The Red Kite.ArlequinPress, Essex, UK

Centor, R. M. (1991). The use of ROC curvesand their analyses. Medical DecisionMaking, 11: 102-106.

Collar, N. J. & Andrew, A. (1988). Birds towatch: the ICBP world checklist ofthreatened birds.ICBP, Cambridge, UK

CORINE (1991). CORINE Land Cover Map,European Comission.

Corsi, F., Duprè, E. & Boitani, L. (1999). ALarge-Scale Model of Wolf Distribution inItaly for Conservation Planning.Conservation Biology, 13(1): 150-159.

Cox, D. R. & Wermuth, N. (1992). Acomment on the coefficient ofdetermination for binary responses.American Statistician, 46: 1-4.

Cramp, S. & Simmons, K. E. (1980). Thebirds of the Western Palearctic, Vol.II.Oxford University Press, Oxford, UK

Crawley, M. J. (1993). GLIM forecologists.Blackwell Science, London


De Juana, F. (1989). Situación actual de lasrapaces diurnas (O. Falconiformes) enEspaña. Ecología, 3: 237-292.

Díaz-Delgado, R. & Pons, X. (2001). Spatialpatterns of forest fires in Catalonia (NE ofSpain) along the period 1975-1995.Analysis of vegetation recovery after fire.Forest Ecology and Management, 147(1):67-74.

Donázar, J. A., Hiraldo, F. & Bustamante, J.(1993). Factors influencing nest siteselection, breeding density and breedingsuccess in the bearded vulture (Gypaetusbarbatus). Journal of Applied Ecology, 30:504-514.

Eastman, J. R. & Fulk, M. (1993). LongSequence Time Series Evaluation UsingStandardized Principal Components.Photogrammetric Engineering and RemoteSensing, 59(8): 1307-1312.

Elósegui, J. (1985). Navarra: Atlas de avesnidificantes.Caja de Ahorros de Navarra,Pamplona, Spain

Evans, I. M. & Pienkowski, M. W. (1991).World status of the red kites. Abackground to the experimentalreintroduction to England and Scotland.British Birds, 84: 171-187.

Ferrer, M. (1993). El águila imperial.Quercus,Madrid, Spain

Fielding, A. H. & Bell, J. F. (1997). A reviewof methods for the assessment ofprediction errors in conservationpresence/absence models. EnvironmentalConservation, 24: 38-49.

Fielding, A. H. & Haworth, P. F. (1995).Testing the generality of bird-habitatmodels. Conservation Biology, 9: 1466-1481.


158

Font Tullot, I. (1983). Climatología deEspaña y Portugal.Instituto Nacional deMeteorología, Madrid

Franklin, J. (1998). Predicting the distributionof shrub species in southern Californiafrom climate and terrain-derived variables.Journal of Vegetation Science, 9: 733-748.

García, J. T., Viñuela, J. & Sunyer, C. (1998).Geographic variation of the winter diet ofthe Red Kite Milvus milvus in the IberianPeninsula. Ibis, 140: 302-309.

George, K. (1995). Überwintering vonRotmilanen (Milvusmilvus) im nordlichenHarzvorland/Sachsen-Anhalt. Vogel andUmwelt, 8: 59-64.

González, L. M., Bustamante, J. & Hiraldo, F.(1990). Factors influencing the presentdistribution of the Spanish imperial eagleAquila adalberti. Biological Conservation,51: 311-319.

Guisan, A., Theurillat, J.-P. & Kienast, F.(1998). Predicting the potentialdistribution of plant species in an alpineenvironment. Journal of VegetationScience, 9: 65-74.

Guisan, A. & Zimmermann, N. E. (2000).Predictive habitat distribution models inecology. Ecological Modelling, 135: 147-186.

Harrell, F. E. (2001). Regression modelingstrategies.Springer, New York

Hastie, T. J. & Tibshirani, R. J. (1990).Generalized Aditive Models.Chapman &Hall, London

Heredia, B., Alonso, J. C. & Hiraldo, F.(1991). Space and habitat use by red kitesMilvus milvus during winter in theGuadalquivir marshes: a comparisonbetween resident and winteringpopulations. Ibis, 133: 374-381.

Hille, S. (1995). Nahrungswahl undJagdstrategien des Rotmilans (Milvusmilvus) im BiosphärenreservatRhon/Hessen. Vogel and Umwelt, 8: 99-126.

Hiraldo, F., Blanco, J. C. & Bustamante, J.(1991). Unspecialized exploitation ofsmall carcasses by birds. Bird Study, 38:200-207.

Hiraldo, F., Heredia, B. & Alonso, J. C.(1993). Communal roosting of winteringred kites Milvus milvus (Aves,Accipitridae): Social feeding strategies forthe exploitation of food resources.Ethology, 93: 117-124.

Hiraldo, F., Veiga, J. P. & Mañez, M. (1990).Growth of nestling black kites Milvusmigrans: effects of hatching order, weatherand time of season. Journal of Zoology(London), 222: 197-214.

Lasanta, T. (1989). Evolución reciente de laagricultura de montaña: el PirineoAragonés.Geoforma Ediciones, Zaragoza,Spain

Lavers, C. P. & Haines-Young, R. H. (1996).Using models of bird abundance to predictthe impact of current land-use andconservation policies in the Flow Countryof Caithness and Sutherland, northernScotland. Biological Conservation, 75: 71-77.

Lawton, J. H. & Woodroffe, G. L. (1991).Habitat and distribution of water voles:why are there gaps in a species´ range?Journal of Animal Ecology, 60: 79-91.

Legendre, P. (1993). Spatial autocorrelation:trouble or new paradigm? Ecology, 74(6):1659-1673.

Lillesand, T. M. & Kiefer, R. W. (1994).Remote sensing and imageinterpretation.John Wiley & Sons, NewYork

Lovegrove, R. (1990). The kite´s tale. Thestory of the red kite in Wales.The RoyalSociety for the Protection of Birds,London, UK

Mammen, U. (2000). Bestandsabnahme beimRotmilan Milvus milvus von 1994 bis1997 in Deutschland. OrnithologischeMitteilungen, 52: 4-13.

Mammen, U. & Stubbe, M. (2001). Trends inthe population of Red Kite in Germany,1999-2000. Abstracts of the 4th EurasianCongress on Raptors. (eds , pp. 113.Estación Biológica de Doñana/RaptorResearch Foundation, . Sevilla, Spain .

Manel, S., Dias, J. M. & Ormerod, S. J.(1999). Comparing discriminant analysis,neural networks and logistic regression for


159

predicting species distributions: a casestudy with a Himalayan river bird.Ecological Modelling, 120: 337-347.

Manel, S., Williams, H. C. & Ormerod, S. J.(2001). Evaluating presence-absencemodels in ecology: the need to account forprevalence. Journal of Applied Ecology,38: 921-931.

Mañosa, S. (2002). The conflict betweengamebirds hunting and raptors in Europe,European Comission, REGHAB project.

Mather, P. M. (1999). Computer processingof remotely sensed images.John Wiley &Sons, New York, USA

MathSoft, I. (1999). S-Plus 2000 User´sGuide.Data Analysis Products Division,Seattle

Merrill, T., Mattson, D. J., Wright, R. G. &Quigley, H. B. (1999). Defining landscapesuitable for restoration of grizzly bearsUrsus arctos in Idaho. BiologicalConservation, 87: 231-248.

Meyburg, B.-U. (1973). Observations surl´abondance relative des rapaces(Falconiformes) dans le nord et l'ouest del'Espagne. Ardeola, 19: 129-140.

Mittlböck, M. & Schemper, M. (1996).Explained variation for logistic regression.Statistics in Medicine, 15: 1987-1997.

Monserud, R. A. & Leemans, R. (1992).Comparing global vegetation maps withthe Kappa statistic. Ecological Modelling,62: 275-293.

Newton, I., Davis, P. E. & Moss, D. (1994).Philopatry and population growth of redkites Milvus milvus in Wales. Proceedingsof the Royal Society. London. Series B.,257: 317-323.

Osborne, P. E., Alonso, J. C. & Bryant, R. G.(2001). Modelling landscape-scale habitatuse using GIS and remote sensing: a casestudy with great bustards. Journal ofApplied Ecology, 38(2): 458-471.

Pearce, J. & Ferrier, S. (2000). Evaluating thepredictive performance of habitat modelsdeveloped using logistic regression.Ecological Modelling, 133: 225-245.

Picard, R. R. & Berk, K. N. (1990). Datasplitting. The American Statistician, 44:140-147.

Preisler, H. K., Rappaport, N. G. & Wood, D.L. (1997). Regression methods forspatially correlated data: an example usingbeetle attacks in a seed orchard. ForestScience, 43: 71-77.

Purroy, F. J., Ed. (1997). Atlas de las aves deEspaña (1975-1995)Lynx editions, .Barcelona

Rico Alcázar, L., Martínez, J. A., Morán, S.,Navarro, J. R. & Rico, D. (2001).Preferencias de hábitat del Águila-azorPerdicera (Hieraaetus fasciatus) enAlicante (E de España) a dos escalasespaciales. Ardeola, 48(1): 55-62.

Sánchez-Zapata, J. A. & Calvo, J. F. (1999).Raptor distribution in relation to landscapecomposition in semi-arid Mediterraneanhabitats. Journal of Applied Ecology, 36:254-262.

Serrano, D. (1998). Interhabitat differences inthe diet of the Eagle Owl (Bubo bubo) inthe mid Ebro river valley (NE Spain):effect of european rabbit (Oryctolaguscuniculus) availability. Ardeola, 45(1): 35-46.

Serrano, D. (2000). Relationship betweenraptors and rabbits in the diet of eagle owlsin southwestern Europe: competitionremoval or food stress? Journal of RaptorResearch, 34: 305-310.

Smith, P. A. (1994). Autocorrelation inlogistic regression modelling of species`distributions. Global Ecology andBiogeography Letters, 4: 47-61.

Stubbe, M., Mammen, U. & Gedeon, K.(1995). Erfassung des Rotmilans (Milvusmilvus) im Rahmen des monitoringsgreifvogel und eulen Europas-perspektiveneines internationalen Rotmilan-monitorings. Vogel and Umwelt, 8: 165-171.

Swets, J. A. (1988). Measuring the accuracyof diagnostic systems. Science, 240: 1285-1293.

Teixeira, J., Ferrand, N. & Arntzen, J. W.(2001). Biogeography of the golden stripedsalamander Chioglossa lusitanica: a fieldsurvey and spatial modelling approach.Ecography, 24: 618-624.


160

Thiollay, J.-M. (2001). The red kite inEurope: unprecedented decline and a callfor action. Raptor News.

Titus, K., Mosher, J. A. & Williams, B. K.(1984). Chance-corrected classification foruse in discriminant analysis: ecologicalapplications. The American MidlandNaturalist, 111(1): 1-7.

Tobalske, C. & Tobalske, B. W. (1999).Using atlas data to model the distributionof woodpecker species in the Jura, France.The Condor, 101: 472-483.

Tucker, G. M. & Heath, M. F. (1994). Birdsin Europe: their conservationstatus.Birdlife International, Cambridge,UK

Van Kleef, H. & Bustamante, J. (1999). Firstrecorded polygynous mating in the red kite(Milvus milvus). The Journal of RaptorResearch, 23(3): 254-257.

Veiga, J. P. & Hiraldo, F. (1990). Food habitsand the survival and growth of nestlings intwo sympatric kites (Milvus milvus andMilvus migrans). Holarctic Ecology, 13:62-71.

Verbyla, D. L. & Litvaitis, J. A. (1989).Resampling methods for evaluatingclassification accuracy of wildlife habitatmodels. Environmental Management, 13:783-787.

Vida, S. (1993). A computer program fornon-parametric receiver operatingcharacteristic analysis. Computer Methodsand Programs in Biomedicine, 40: 95-101.

Villafuerte, R., Viñuela, J. & Blanco, J. C.(1998). Extensive predator persecutioncaused by population crash in a gamespecies: the case of red kites and rabbits inSpain. Biological Conservation, 84: 181-188.

Viñuela, J. (1996). Situación del Milano Real(Milvus milvus) en el Mediterráneo.Biology and conservation of mediterraneanraptors. (eds SEO/Birdlife), pp. 91-100.SEO/Birdlife, . Madrid, Spain .

Viñuela, J. (1997). Road transects as a large-scale census method for raptors: the caseof the Red Kite Milvus mivus in Spain.Bird Study, 44: 155-165.

Viñuela, J. & Bustamante, J. (1992). Effect ofgrowth and hatching asynchrony on thefledging age of black and red kites. TheAuk, 109: 748-757.

Viñuela, J. & Contreras, A. (2001). Status ofthe Red Kite (Milvus milvus) in Spain.Abstracts of the 4th Eurasian Congress onRaptors. (eds , pp. 194. Estación Biológicade Doñana/Raptor Research Foundation, .Sevilla, Spain .

Viñuela, J., Martí, R. & Ruiz, A., Eds. (1999).El milano real en EspañaSEO/BirdLife, .Madrid

Viñuela, J. & Sunyer, C. (1992). Nestorientation and hatching success of BlackKites in Spain. Ibis, 134: 340-345.

Viñuela, J. & Villafuerte, R. (in press).Predators and rabbits in spain: a keyconflict for conservation of Europeanraptors. Raptors in a changingenvironment, Kinkardine, Scotland, BritishOrnithologists Union.

Voisin, J.-F. (1994). Milan royal. Nouvelatlas des oiseaux necheurs in France. (edsL. Yeatman & P. Berthelot), . SocietéOrnitologique de France, . Paris, France .

Wu, H. & Huffer, F. W. (1997). Modellingthe distribution of plant species using theautologistic regression model.Environmental and Ecological Statistics, 4:49-64.

Zweig, M. H. & Campbell, G. (1993).Receiver-operating characteristic (ROC)plots: a fundamental evaluation tool inclinical medicine. Clinical Chemistry, 39:561-577.

Comparación de la cartografía de especies

161

APÉNDICE IAPPENDIX I

LISTA DE COLABORADORES DEL PROYECTO "SITUACION DEL MILANO REAL ENESPAÑA" (1992-1994)LIST OF FIELD VOLUNTEERS FOR THE PROJECT "SITUACION DEL MILANO REAL ENESPAÑA" (1992-1994)

COORDINADORES NACIONALES: / NATIONAL COORDINATORS:

Invernada: Javier Viñuela y Alfredo OrtegaReproduccción: Javier Viñuela

ALAVA. Ramón Arambarri, Arturo F. Rodríguez (GADEN) (coords.). Yolanda Arrondo, Gorka Belamendía, AurelioCanabal, Mario Corral, Marian Fernández, José M. Fernández García, Luis Javier Fernández García, José AntonioGainzarain, José M. Iglesias, Estitxu Irazola, Luis Lobo, Carlos López Losada, J.R. López Retamero, Asier Martínez, IñakiMartínez, José Melena, José A. Nuevo Morena, Txema Pérez Ugarriza, Arturo F. Rodríguez. ASTURIAS. Luis MarioArce, Manuel Enrique Carballal, Coordinadora Ornitoloxica d'Asturies. AVILA. José Mª García Jiménez, Gabriel Sierra,Javier Viñuela. (coords.). José Miguel Abarca Antón, Antonio J. Aldea, Segundo Arcones, Carlos Bermejo Arribas, JulioCaballero, Julio Caminero Tapiador, Juan Ramón Cuervo Martín, Aurelio Delgado Hidalgo, Jesús Encinar Muñoz, MartaFernández Pérez, Pedro L. García, Ana García García, José M García Jiménez, Alicia García Muñoz, Manuel GarcíaTornero, Jesús Gil Martín, Ana Grandal Martín, Isabel Hernández Vallejo, Mariano Hernández Vallejo, Roberto Ivars,Carlos Jarque Bañuelos, Julio Jiménez Encinar, Juan Carlos Marfull Robledo, Carlos Martín, Dionisio Martín, J. MiguelMartín, Ignacio Martín García-Sancho, Luis Martín García-Sancho, Enrique Martín Serrano, Fernando Muñoz Carrera,Manuel F. Pérez Escrivá, Aurelio Ramírez Jiménez, Rosa Rodríguez Manzano, José Luis Robledo Ranea, José RodríguezMatías, César San Segundo Ontín, David Sánchez Sáez, Francisco Sánchez Sierra, Juan Manuel Sastre González, JuliánZancajo. BADAJOZ. Francisco Gragera Díaz. (coord.). Julián Andrés Ledesma, Francisco Barrena, J. Antonio Candelario,José Enrique Capilla, Juan Carlos Delgado Expósito, Carmen Galán Novella, Alfredo Eusebio Garica, María LedesmaVázquez, Cati León, José Antonio Palomo, Enrique Pérez, José Carlos Pons García, José Elías Rodríguez, Jesús RojasGonzález, Marcos Romero González, Luis Salguero Báez, Angel Sánchez García, José María Traverso Martínez.BURGOS. Roberto J. Milara, Javier Viñuela. (coords.). Enrique Alvarez, Eduardo Angulo, Rubén Arrabal, Gorka ArtiguezGallaga, Luis Miguel Arranz, Jorge Bañuelos, Florentino Barbadillo, José Ignacio Contreras, Pedro José de la Cruz, PilarDurante, Carlos García, Carlos García Güemes, Eduardo Izquierdo, Roberto López Saiz, Alberto Martínez, IsmaelMediavilla, Fernando Moreno, Miguel Angel Pinto Cebrián, Angel Quero Miguel, Rubén Río, José Román, Vicente SanzFernández de Gobero, Ignacio Sanz Moneo. CACERES. Pilar López Avila, Sebastián Hidalgo, Javier Viñuela. (coords.).Daniel Abel Schaad, Antonio Acha, Elena Angulo, Ernesto Alonso Juárez, José Antonio Bardají Zúñiga, Juana BarrigaManzano, Socorro Cancho Corrales, Mónica del Castillo Burgos, Gregorio Castillo Fernández, Antonio Civantos, MaríaJosé Domínguez, Bryan Etheridge, Richard Evans, Isidoro Fagundo Torres, Manuel Fernández, Ian Fisher, Manuel FloresCid de Rivera, Santos Fuertes, Carmen Galán Novella, Rosana García Macías, Alfredo García Sánchez, Nieves GonzálezJarri, Jaime Iglesias Duarte, Fernando Labrador Romero, Carmen Linares Tello, Pilar López, Penélope Losa Gómez, JuanJosé Luengo Rodríguez, Paloma Martín García, José Martín Pablo, Alberto Morón Pastor, Justo M. Muñoz Mohedano,Consuelo Muñoz Molina, Adolfo S. Maestre García, Patricio Mateos Quesada, Soledad Mateos Salcedo, Antonio MurielBernal, Magín Murillo Fernández, Alberto Navalón, Ana Núñez Cansado, Francisco José Ordiales, Asunción Pacheco,Serafín Polo, Alfonso Polvorinos Ovejero, Mª José Rincón Matesanz, Francisco Serrano, Mercedes Silveira Torremocha,Innes Sim, Luis Suárez Arangüena, Maite Torres Fernández, Carmen Usin, Salvador Vaquero, Francisco Serrano. CADIZ.Jose Luis Paz de la Rocha. CIUDAD REAL. Juan Pablo Castaño (coord.).José Carlos García, Tom Gullick, José Guzmán,José Luis Hernández, José Manuel Hernández, José Jiménez, Manuel López Sánchez, Juan Ramón Morcillo, Javier Muro,Manuel Muro, Rafael Palomo, Leonardo Rodríguez, Manuel Salgado. COMUNIDAD VALENCIANA. Mario JiménezRipoll. CORDOBA. Juan Antonio Antón, Federico Cabello de Alba Jurado, Antonio Gómez Miranda, Antonio LeivaBlanco, Francisco Sánchez Tortosa. CUENCA. Marcos Costa Belinchón. (coord.). Francisco Costa Belinchón, EduardoHervás Domínguez, David Manzanares Sáez, Juan Pedro Morales Gallego, Grupo Elyomis, Grupo de Estudio de las ZonasEsteparias Conquenses. GRANADA. José Mª Gil Sánchez. (coord.). Natalia Ildefonso Huertas, Rosa María IldefonsoHuertas, Francisco Manuel Molino Garrido, Gerardo Valenzuela Serrano. GUIPUZCOA. Héctor González Arcelos.HUELVA. Andrea Gardiazábal. (coord.). F. Alcántara, Rafael Carmona, José Mª Correa, Rafael Galán, Fernando Hiraldo,Manuel Máñez, Mariano Marchena, Alberto Moya, Ignacio Olano, Mercedes Sánchez, José Manuel Sayago, Luis Urbina,Javier Vilches, Escuela-taller Guadiana. HUESCA. Alberto Bueno Mir. (coord.). Carlos Acín Canfranc, Ana Alastrué, JuanCarlos Albero, J.J. Almazán, R. Aquilué, Ignacio Arizón, Antonio Arnal Martí, Guillermo Ascaso, Juan Carlos Ascaso, M.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO IX

162

A. Aspiroz, Joaquín Barrabes, Matías Belda, Francisco Bescós Claver, J. Bastarós, M. Bellosta, V. A. Boucher, José MªCanudo, Mamen Carmena, Fernando Carmena Flores, J.M. Cereza, Guillermo Costas, Oscar Díez Sánchez, CurroDomínguez, F. Domínguez, Joaquín Español, Alfredo Espinosa, Ramón Espuña, CH. Ferrer, J. Ferrer, Francisco FerrerLerín, G. Fornies, José Antonio Garcés, Juan Antonio Gil Gallús, Gerardo Goded, David Gómez, Manuel Grasa Francés,Julián Lacosta, José Luis Lacostena, D. Lanaspa, Jesús Lavedán, Isabel López, M. Losfablos, J.J. Moreno, José DamiánMoreno Rodríguez, Clara Ortega, Jesús Peña, José Andrés Pintado, Maarten Platteeuw, C. Pradel, I. Pueyo, F. Ramos,Balbino Riobó, José Enrique Ríos Cabrera, J. Sánchez, J.L. Sasot, D. Saura, Alejandro Serrano, J.R. Sesé, Carlos Tarazona,A. Torralba, J. Carlos Usieto, Fernando Vallés, Rafael Vidaller, A.G. Villacampa, Emilio Yñigo. JAEN. Joaquín MuñozCobo. Francisco Campos, José García, José García Santiago, Alfonso Godino Ruiz, Mª Carmen López, F. Ortega. LEON.Mª Teresa Andrés Ponga, Alejandro Onrubia Baticón. (coords.). Nemesio Andrés García, Concepción Andrés Ponga,Cristina Borrego, Olga Borrego Rodríguez, Javier Calzada, Mercedes Fernández Fernández, Javier García Fernández,Benito Fuertes Marcos, Benedicto González, Carlos Gutiérrez Expósito, Concepción Ponga Martínez, Eloy Revilla, JoséLuis Robles Prieto, Mª Luisa Robles, Jacinto Román, Eva Sagües, Carlos Sánchez, Jesús Villadangos Fuertes, URZ(Blanca, Noemi, Pablo y Rafael). LERIDA. Jordi Canut. (coord.). A. Baker, J. Bonfill, T. Carulla, F. Fernández, J. García,D. García-Ferre, R. Gutiérrez, A. Margalida, J. Martí, R. Martínez, A. Martínez-Vilalta, J. Medina, J.M. Parde, J. Puig,M.D. Rodríguez, R. Torres, D. Saavedra, J. Sargatal, M. Sicart, E. Streich. Patrullas del Servei d'Agents Rurals (D.A.R.P.):Pallars Jussà (J. Bolado, L. Cruz, J. Palacín); Pallars Sobira (J. Arilla, M. Arilla, J. Jové) y Urgell (L. Tomás, J. Mª Freixes,J. Reig y J. Puig). LOGROÑO. Jesús Mª García García, Carlos Gutiérrez Expósito. (coords.). Juan Carlos Luis Gil,Santiago Peña López. MADRID. Iñigo Fajardo, Alfredo Ortega, Javier Viñuela. (coords.). Ana Barón, Miguel AngelBlanco, Miguel Angel Blasco Rodríguez, Alberto Cameo, Ana Cameo, Ramona Carpintero, Pilar Cavia Cuesta, TeresaCórcoles, Manuel Cuesta Nieto, Sandra Ducazcal, Alfredo Espinosa, Iñigo Fajardo, Ignacio Fernández Aransay, OliviaGaona Palop, Francisco García González, Jesús García González, Pedro A. García Sánchez, Ricardo Gómez Calmaestra,Laura Gómez Moreno, Francisco González, Ana I. González de Castro, Mario González Pérez, Raquel Gordo Méndez, LuisHaro, R. Hernández, Miguel Angel Herrera, Mª José Linares Tello, Isabel López, Javier Marchamalo, M. Martínez Duchel,R. Millán, Julián Moral Sánchez, M. Andrés Moreno, Juan Carlos Palermo, E. Pérez Balsalobre, Yolanda Pérez Chirinos,Marina Pérez Zahonero, Luis Picazo Casariego, M.A. Ramiro, Adolfo Rodríguez Pérez, Enrique Sánchez Airas, LinoSánchez-Mármol, David Sánchez de Ron Martínez, Eduardo Soto-Largo, Carlos Sunyer, Francisco Yago. MALAGA. JoseA. Cortés. (coord.). Carla J. Atkins, Adolfo Aguilar, Miguel A. Domínguez, José L. Medina, Juan Luis Muñoz. Francisco J.Peso, Juan J. Rodríguez, Antonio Román Muñoz. MALLORCA. Carlota Viada. (coord.). Juan Salvador Aguilar, Pere L.Dietrich, Javier Gassó, Jordi Muntaner, Toni Muñoz, Matías Rebassa, Miguel A. Reus, Juan José Sánchez, Manuel Suárez,Pep Sunyer, Evelyn Tewes. MENORCA. GOB-Menorca (Josep Capó Nin, Evaristo Coll, Santi Catchot, Guillem Orfila,Félix de Pablo, Rafel Triay). NAVARRA. Juan I. Deán, Jesús Elosegui. (coord.). Mercedes Alberdi, Iosu Alfaro, AntonioAñorga, Pedro Arratibel, Carlos Astrain, Rafael Ballano, Esteban Burusko, David Campión, José Luis Carrica, TomásCerdán, Juan José Corera, Jesús Cuairán, F. Javier Deán, Edurne Elizalde, Fermín Erdozain, Alejandro Erviti, AnaEscribano, Amaia Etxebarria, José Babil Goñi, David Guzmán, Enrique Herranz, Ramón Hortelano, Juan Carlos Iriarte,Juan Jesús Iribarren, José A. Lacunza, Ana Lezaun, Joaquín Leoz, Iosu Lerga, José Luis Lizarraga, Alfonso Llamas, Luis A.López Borobia, Santiago Maiza, Mikel Mugiro, Fermín Nieto, José F. Ochoa, Javier Olave, Pedro Ollobarren, Mikel Peña,Inés Mª Pérez Abandaño, José Antonio Pérez-Nievas Martínez, Mª Dolores Pinedo, José Ignacio Riezu, Fermí Rivero,Antonio Rodríguez Arbeloa, Angel Salcedo, Eusebio Salón, Tomás Santesteban, Ernesto Sanz, Alfonso Senosiaín,Alejandro Urmeneta, Fermín Urra, Javier Urra, José Andrés Venys, Fernando Villafranca, José Venys Villar, Fco. JavierZalba, GURELUR (Koldo Aranguren Martiarena, Marcelino Barbería, Iñigo Calonge Isturiz, Miguel Ciriza Isturiz, JesúsDomínguez Iglesias, Manuel Fernández Recio, Raúl Fernández Recio, Oscar Ibáñez Escarda, Fidel Mediavilla Ayala, IñakiMorrás del Río, Antonio Munilla, Jorge Nuble Carmona, Iosu Oroz Legal, Iñigo Petri Nabarlaz,). ORENSE. AntonioVillarino Gómez. (coord.). Isabel Alvarez Balvis, Marcos Manuel Freán Hernández, Celso Nieto Pérez, José Manuel PérezPérez. PALENCIA. Enrique Gómez Crespo, Fernando Jubete, Javier Pastor. SALAMANCA. Pedro Luis Ramos Bueno.(coord.). Antonio Acha, Mª de los Angeles Coca, Raúl Bueno Hernández, Roberto Carbonell Alanís, José Miguel Colorado,Lorenzo Corrales, Myriam Cuadrado López, Pablo C. Díaz, Casimiro Domínguez Martín, Francisco Javier Galvache,Eugenio García García, Andrés C. García Montero, J. Pablo García Montero, Luis Gómez Albanán, Jesús Gómez Gómez,Javier Gómez Labrador, Oscar J. González, Angel González Losa, Mª de los Angeles González Sánchez, José AntonioHernández, Guillermo Hernández Cordero, Luis Javier Hernández Martín, Vicente López Alcázar, Tomás MerchánSánchez, Francisco Panadero Sáez, Manuel de Pedro, Eladio Sánchez, Jesús Serradilla Rodríguez, Juan Carlos VadilloSanz, José Antonio Velasco Mendo. SANTANDER. Juan José Aja. (coord.). Manuel Bahillo, Luis Felices, Ana GarcíaGarcía, Javier García-Oliva, José Manuel González González, Angel Herrero Calva, Gonzalo Palomero, Olga Pérez, CarlosSánchez, Antonio Sanz Carro. SEGOVIA. Fernando Barrio, Javier Viñuela. (coords.). Fernando de Antonio, FernandoAparicio, Mark Avery, Angel Carballés Muñoz, Richard Cowie, Juan A. de Diego Fuentenebro, Angel Gómez-Manzaneque, Mª José Jiménez Armesto, Ross Johnston, Eduardo de Juana, Lola Manteiga, Ramón Martí, Gonzalo LlorenteAlvarez, Andrés López Moreno, Javier Oria Martín, Daniel Pinela, Marino Rodríguez de Miguel, Javier Sánchez Vaquero,Jean Sears, Eduardo Sotolargo, Carlos Sunyer, Fernando Vázquez Gallardo, A.E.LANIUS (Mercedes Alonso y AndrésLópez). SEVILLA. Francisco Parreño. (coord.). Verónica Alarcón, Ricardo Arroyo, Francisco Javier Avila Domínguez,Diego Fuentes Baquero, Hipólito Benítez León, Hipólito Benítez Morgado, José Mª Bermudez, Salvador Borrego, Sergio


163

Braña Montes, Rafael Carmona Vilches, Antonio Díaz Espejo, Manuel Díaz Palma, Francisco Domínguez, Luis MiguelEscalera, Fernando Fernández Parreño, José Manuel Galera, Miguel Galuez, José Manuel García, Juan Luis García Castaño,Manuel Giraldez, Eusebio Gómez, Manuel Gracia, Inmaculada Llach, Francisco Morales, Agustín Parejo, Manuel CarlosPérez Gómez, Francisco Pérez Montiel, Luis MIguel Platero, Juan Pedro Segura, Manuel de la Riva, José Manuel RiveroHidalgo, Manuel Raimundo Rodríguez, Manuel Vega, J. Antonio del Valle, Manuel Zayas. SORIA. Félix Martínez Olivas,Fernando Chagüaceda. (coords.). Pilar Cavia Cuesta, Inmaculada Cuchillo Ibáñez, Manuel Cuesta Nieto, Ignacio FernándezAransay, Andrés García Pérez, Pedro A. García Sánchez, Ricardo Gómez Calmaestra, Juan Luis Hernández Hernández,María Herrera, José María Lenguas Gil, Félix Martínez Olivas, Manuel Meijida Fuertes, José Mercader, Javier MuñozJiménez, Manuel F. Pérez Escrivá, Federico Roviralta García, Pilar Sánchez Sánchez. TOLEDO. Marino López de Carrión,Javier Viñuela. (coords.). José Antonio Benavides, Luis Santiago Cano Alonso, Carlos Carbajal, Luis Otero Espinosa,Fernando Fernández Morcuende, Ignacio del Moral, Mariano Vidal, Félix Zaragoza, Javier Zaragoza. VALLADOLID.Alfonso Balmori, Jesús Fernández-Gutiérrez. (coords.). Harry Bayley, Juanjo Becerril Trigueros, Luis Carlos Castellanos,José Mª Collazos, Ignacio Conde, Juan Criado, Javier Espinosa, Rosa de la Fuente Torre, Dolores García González, CarlosGil Palomo, Tomás Gómez, Santiago González Martínez, Ainda Goyenechea, Vicente Guerra, Juan Carlos Guerra Velasco,Luis Hurtado Vergara, Francisco Javier Jiménez Martín, José Mª Lorenzo García, Xavier Martín, Ana Martínez, FranciscoMartínez, Antonio Martínez Ruiz, Mercedes Medina, Francisco Merage, Orlando Parrilla, Julio Pérez Gil, Rafael Ramos,Emma Rodríguez, Blanca Ruiz Otazo, Tomás Sanz Sanz, Jesús Ucero, Pedro Villalón. VIZCAYA. Gorka Artiguez Gallaga.ZAMORA. Alfredo Ortega, Javier Viñuela. (coords.). Mark Avery, Gregorio Ballestero, Len Campbell, J.F. Carreño,Carlos de Celis, Jesús Domínguez García, V. Fernández, Jesús García González, Miguel Angel García Matellanes, HipólitoHernández Martín, Manuel José Hernández Rodríguez, José Alfredo Hernández Rodríguez, Andy Hughes, AntonioMediavilla Largo, Manuel Miñambres, Pedro Moldón, Ken Norris, Debbie Pain, Jesús Palacios Alberti, Tomás RiveroRivero, Mariano Rodríguez Alonso, Luis Fernando San José, Ken Smith, José Luis Vicente, C. Zamora, OCELLUMDURII (Miguel Angel García Matellanes, José Alfredo Hernández Rodríguez, Ana Vecilla Rodrigo). ZARAGOZA. JoséManuel Sánchez Sanz. (coord.). Elisa Andrés, Gerardo Báguena, José Miguel Baselga, Manuel Benzar, Plácido AntonioBrotons, Luis M. Bueno, Adérito Calzón, Isabel Casaña, Juan Carlos Cirera, Félix Compaired, Alejandro Cortes, Juan LuisFernández, Alvaro Gajón Bazán, Julio Giménez Laaraga, Francisco Hernández Fernández, José Ignacio Iglesias, OscarIglesias, Ana Belén Lacueva, M. Nieves Laín, César Lolumo García, Javier Gomollón, Daniel Magdalena, José M. Marco,Pedro Martínez Pérez, Manuel Mercadal, Juan Moreno Twose, Luis Palacio, Enrique Pelayo Zueco, Carlos Pérez Laberda,Alfredo Pérez Pina, Patxi Picón, José Antonio Pinzolas, José Luis Rivas González, Miguel Angel Romeo Lafuente, JavierRuiz Alba, José Luis Ruiz Cerra, Francisco Javier Sampietro, Federico Sancho Puertas, José Mª Sierra Robres, José LuisTella, Luis Tirado, Fermín Torres, José de Uña, Francisco Ventura.


164

CAPÍTULO IX: Una evaluación con modelos estadísticos de la cartografía de especies

generada mediante criterio de expertos

RESUMEN

En numerosas circunstancias, aplicadas, de investigación básica, o de conservación, se requiere

tener un conocimiento detallado de la distribución de las especies. La cartografía de especies

derivada del criterio de expertos es subjetiva, difícil de evaluar y se presenta a una baja resolución.

El criterio de expertos puede usarse para refinar la información de los atlas de especies, cuyos datos

se consiguen con un trabajo de campo intensivo. En este caso la resolución espacial de los mapas

aumenta, pero aún es insuficiente para satisfacer los objetivos que la mayoría de las aplicaciones

prácticas suelen demandar (es aún mayor que 100 km2). En contraste, la cartografía de especies

derivada de modelos estadísticos es objetiva, fácil de evaluar y puede obtenerse a muy alta

resolución. En este capítulo nos planteamos la comparación de la capacidad predictiva de modelos

estadísticos y modelos generados con criterio de experto aplicados al análisis de la distribución de

aves reproductoras en Andalucía. Las preguntas que abordamos son: (1) ¿se puede generar una

cartografía equivalente a la de un atlas con menor esfuerzo de muestreo mediante modelos

estadísticos? y (2) ¿el criterio de experto permite obtener modelos de alta resolución espacial con

una capacidad predictiva equivalente a la de un modelo estadístico


165

CHAPTER IX: Using statistical models to evaluate species cartography derived from

expert opinion

ABSTRACT

In a great number of circumstances, related to applied practices, basic research or conservation, it is

needed a detailed knowledge of species distribution. The species cartography derived from expert

opinion is subjective, difficult to evaluate and it is commonly presented with a low spatial

resolution. Expert opinion may be used to refine the information of species atlases, which data is

gathered by intensive field surveys. In this case, the spatial resolution of the resulting maps

increases, but it is not enough to satisfy the objectives desired by most of applied practices (it is

>100 km2 yet). In contrast, the species cartography that is derived from statistical models is

objective, easy to evaluate and it can be obtained at a high spatial resolution. In this chapter we

compare the predictive ability of statistical models and models generated with expert opinion,

applied to the analysis of nesting birds in Andalucía. The questions we address are: (1) could a

cartography equivalent to that derived from atlases be generated with less sampling effort by

statistical techniques?, and (2) can the expert opinion obtain models of high spatial resolution with

a predictive ability equivalent to that of statistical models?


166

1. Introducción

Como se ha mencionado en capítulosanteriores, las circunstancias en que serequiere tener un conocimiento de ladistribución de las especies son numerosas yresponden a motivaciones de investigacióncientífica básica (Mourell & Ezcurra 1996;Manel, Buckton & Ormerod 2000), deaplicaciones conservacionistas (Corsi, Duprè& Boitani 1999), de planificación yevaluación de impactos (Lavers & Haines-Young 1996), o, simplemente, recreativas(Donald & Fuller 1998). Una primeraaproximación a tal conocimiento puedehacerse con los mapas publicados en guías decampo o trabajos monográficos (por ejemplo:Beaman & Madge 1998; Wilson & Ruff1999). Esta solución no suele ser satisfactoriapara la mayoría de las necesidades debido aque los mapas se presentan a muy bajaresolución, se dibujan con una cobertura de lainformación heterogénea y, por último, sucalidad es difícil de evaluar. Los atlas dedistribución de especies pueden ser másútiles, ya que se construyen con unametodología sistemática (principalmenterespecto a su ventana temporal y, en menormedida, en cuanto a los muestreos que seemplean), aunque su resolución es todavíagrosera (por encima de los 100 km2habitualmente, ver Hagemaijer & Blair 1997;Purroy 1997; Doadrio 2001). Sin embargo, enlos atlas es difícil distinguir entre cuadrículascon distinta abundancia de las especies yentre cuadrículas prospectadas con diferenteesfuerzo de muestreo (Donald & Fuller1998). Esto es así porque la presencia de unosólo o de numerosos individuos de unaespecie identifican de igual manera a unacuadrícula de 100 km2 como área dereproducción, y porque la distinta intensidadde muestreo no se suele representar en lacartografía final.

Quizá el peor efecto del poco detalle de losmapas de manchas y los atlas es que tanto las

áreas coloreadas como las cuadrículasmarcadas dan la falsa impresión de que unaespecie en cuestión se puede encontrar entoda la superficie que destacan. ¿Cómo puedemejorarse entonces la resolución espacial delos mapas y atlas para aumentar su utilidad?.Si asumimos que las especies seleccionan loshábitats que ocupan (Cody 1985; Morrison,Marcot & Mannan 1998), se puede recurrir alcriterio de expertos que determinen loshábitats adecuados para cada especie dentrode las manchas o de las cuadrículas en que seregistraron (Scott et al. 1993; Díaz, Illera &Hedo 2001), lo que se hace definiendorelaciones cualitativas entre las especies y sushábitats (conocidas como wildlife-habitatrelationships). El ejemplo más conocido deluso de modelos de experto para refinar mapasgenerales de distribución de especies son losproyectos, denominados “Gap”, que se hanestado realizando desde hace 10 añosprincipalmente en EE.UU. (Scott et al.1993). En la primera fase de estos proyectosse reúnen todos los registros disponibles decada especie (avistamientos, censos ad hoc ymuestras de colecciones) elaborándose unacartografía de su distribución. Después secartografían los distintos hábitats de la zonade estudio y se sintetiza la informaciónexistente sobre la selección de hábitat de cadaespecie. Finalmente, un comité de expertosacaba definiendo las relaciones entre lasespecies y los hábitats, que se aplican a lasáreas del mapa creado en la primera fase enlas que se supone que están las especies(Bojórquez-Tapia et al. 1995; Caicco et al.1995; Kiester et al. 1996; Powell, Barborak& Rodriguez 2000). La crítica principal queha recibido el diseño de estos proyectos esque no se comprueba que las relacionesespecies-hábitats que se definen són válidas(es decir, los modelos cualitativos que sontales relaciones no se evalúan con un conjuntoindependiente de datos, ver Short & Hestbeck1995). Además, no se conocensuficientemente los requerimientos de hábitatde la mayoría de las especies como para queel criterio de experto genere prediccionesfiables en muchas aplicaciones (p.e. en lasevaluaciones de impacto ambiental, donde


167

este problema se ha identificado como una desus deficiencias más graves, Treweek 1996).

Otra opción para la cartografía de especiesse puede hacer, como se ha mencionado encapítulos anteriores, a través de modeloscuantitativos de distribución de especies, quese han utilizado en numerosas ocasiones paradeterminar la extensión potencial de ciertosorganismos y para valorar la adecuación delhábitat (Austin et al. 1996; Guisan, Theurillat& Kienast 1998; Corsi, Duprè & Boitani1999; Osborne, Alonso & Bryant 2001). Conestos modelos las relaciones entre las especiesy el entorno físico y biótico que las rodea sedefinen a través de técnicas estadísticas(Guisan & Zimmermann 2000). Los modelosestadísticos permiten generar mapas dedistribución interpolando los resultados de unmuestreo a las áreas no prospectadas (laextrapolación, sin embargo, presenta mayoresproblemas, Ertsen et al. 1998), lo quesupone un ahorro de los recursos humanos oeconómicos que pueden destinarse a otrastareas. Además, cuentan con las ventajasrespecto los modelos elaborados mediantecriterio de experto, de que, una vezestablecido un protocolo de construcción, sonherramientas objetivas, transmisibles yfácilmente evaluables.

Las escasas comparaciones que se hanrealizado sugieren que los modelosestadísticos generan mapas más predictivosque los generados mediante criterio deexperto (Pearce et al. 2001), al menos paralas especies que son muestreadas conmediana intensidad (los modelos estadísticosno parecen funcionar bien con especies muyraras). No obstante, la opinión de experto ylas técnicas estadísticas pueden considerarsecriterios complementarios (en vez dealternativos) en las aplicaciónes prácticas queprecisan generar cartografía de especies. Así,Díaz et al. (2001) elaboraron una metodologíade evaluación estratégica ambiental con laque valoran los efectos de planes dedesarrollo sobre la avifauna a través de lasafecciones que tales planes ejercían sobre ladistribución y abundancia de las especies. Las

relaciones entre las especies y los hábitasafectados por los planes se determinarontambién mediante criterio de expertos (Tucker& Evans 1997). Nosotros compartimos laopinión expresada por Díaz et al. (2001) deque la validación de esas relaciones especies-hábitat, mediante modelos estadísticosbasados en datos de campo, es el método másconvincente y efectivo para evaluar laspredicciones de impacto generadas pormetodologías como la suya.

Dadas estas opciones de cartografía deespecies, en este capítulo nos planteamos lacomparación de la capacidad predictiva demodelos estadísticos y modelos generadoscon criterio de experto, aplicados al análisisde la distribución de aves reproductoras enAndalucía. Los modelos estadísticos son deltipo que se ha venido desarrollando en loscapítulos previos de esta tesis doctoral. Losmodelos de experto son las mejoresaproximaciones posibles a la utilización decriterios de experto, con la informacióndisponible actualmente, en los estudios delmedio físico propios de las evaluacionesambientales o los planes de ordenaciónterritorial. Tratando de reflejar las dossituaciones más típicas en que ambos tipos demodelos podrían ser usados, se escogen dostipos de sujeto de estudio que implican dosescalas muy diferentes de análisis. Primero,se comparan modelos para paseriformes yespecies afines, cuyos mapas tienen unaextensión local (1400 km2) y una granresolución (<1ha) y, segundo, la comparaciónse realiza para cuatro especies de rapaces,cuyos mapas tienen una extensión regional (laComunidad Autónoma de Andalucía) y unabaja resolución (100 km2). Las preguntas queabordamos son: (1) ¿se puede generar unacartografía equivalente a la de un atlas conmenor esfuerzo de muestreo? y (2) ¿el criteriode experto permite obtener modelos de altaresolución espacial con una capacidadpredictiva equivalente a la de un modeloestadístico?


168

2. Áreas de estudio y métodos

2.1. Paseriformes y afines

Para el estudio de paseriformes y afines seanalizaron los resultados de 1137 estacionesde escucha de 15 minutos de duración y sinlímite de banda, que se realizaron en lasprimaveras de 1999 y 2000 en los doscuadrados de 70x70 km en Andalucíaoccidental mencionados en anteriorescapítulos (centros: 6° 21’ W 37° 39’ N, y 5°28’ W 36° 44’ N, ver figura 1 del capítuloIV). Del conjunto de especies detectadas seseleccionaron 10 representativas de distintostipos de hábitats y de distinta frecuencia enlas muestras (excluyéndose en particular lasespecies muy raras o las propias de hábitatspoco extensos). Estas fueron: Alectoris rufa(frecuencia 19%), Carduelis cannabina(28%), Certhia brachydactyla (27%),Erithacus rubecula (14%), Galerida theklae(12%), Melanocorypha calandra (7%), Paruscaeruleus (29%), Sitta europaea (15%),Sylvia melanocephala (44%) y Troglodytestroglodytes (15%).

2.2. Rapaces

En la comparación de los modelosrealizados para rapaces se dividió laComunidad Autónoma de Andalucía encuadrículas UTM 10x10 km, de las que seseleccionaron aquellas correspondientes a lasáreas de montaña (n=383, que cubren sólo37.700 km2 porque algunas son costeras oestán en el borde de los husos 29-30 o 30-31).De este conjunto se escogieron 88 pararealizar censos de rapaces por carretera, talcomo se explica en el capítulo VII (ver figura1 de ese capítulo). Las especies que seconsideran en este apartado son Buteo buteo(que aparece en el 45% de las cuadrículascensadas por carretera), Circaetus gallicus(43%), Hieraaetus pennatus (42%) y Milvusmigrans (39%). Todas ellas tienen en comúnel ser rapaces forestales de medianaabundancia para las que los censos decarretera ofrecen resultados satisfactorios.

3. Generación de los modelos

3.1. Modelos estadísticos

3.1.1. Paseriformes y afines

Las variables predictoras usadas en losmodelos estadísticos se extrajeron ymodificaron a partir del mapa digital de usosy coberturas de Andalucía (Moreira &Fernández-Palacios 1995), de un modelodigital de elevaciones y del análisis deimágenes de satélite (todas con resoluciónespacial inferior o igual a 50 metros). Elconjunto de variables predictoras potencialescomprende variables descriptoras de lavegetación (p.e., porcentaje de bosque defrondosas), del paisaje (p.e., longitud deborde entre áreas de bosques y matorral) y dela topografía (p.e., altitud). Este conjunto esmuy numeroso (n=65) e incluye variablesmuy correlacionadas entre sí, por lo que serealizó un análisis de componentesprincipales (ACP) con el que se identificaronlas 13 variables que, estandarizadas,obtuvieron mayores pesos en los primeroscomponentes del ACP (tabla 1) (algunaspruebas previas indicaron que 13 variablesera el número máximo manejable para losanálisis estadísticos que siguen). Estospredictores se promediaron en círculos de 350y 1250 metros de diámetro, centrados en lospuntos de muestreo, cuya superficie equivale,respectivamente, a las áreas de campeo de lamayoría de las aves de pequeño tamaño y a laescala que generó los modelos máspredictivos en análisis anteriores (ver capítuloV).

Se utilizaron modelos aditivosgeneralizados (GAM, Hastie & Tibshirani1990) para modelar la presencia/ausencia delas especies escogidas, empleándose erroresbinomiales y enlace logit. Los modelos sehicieron por separado para cada diámetro enel que se promediaron las variablespredictoras potenciales. Éstas se escogieronmediante un procedimiento automático por


169

Variable predictora Descripción

350

mts

1250

mts

% Forestal a Proporción de píxeles que pertenecen a cualquier categoríaforestal (incluye áreas forestales dispersas como dehesaspero no olivares)

?

% Forestalplanifolio a

Proporción de píxeles que pertenecen a cualquier categoríaforestal de árboles planifolios

?

% Matorralesriparios a

Proporción de píxeles que pertenecen a cualquier categoríariparia con vegetación subarbórea (p.e., Rubus, Phragmites)

? ?

% Agrícola a Proporción de píxeles que pertenecen a cualquier categoríaagrícola (olivares y otros cultivos tanto en secano como enregadío)

? ?

% Cultivosleñosos a

Proporción de píxeles que pertenecen a categorías de cultivoleñoso (principalmente olivares)

? ?

% Urbano a Proporción de píxeles en áreas urbanas o industriales ?Arbolado denso a Presencia/ausencia de cobertura arbolada densa (p.e.,

incluída en un área heterogénea de matorral)?

Matorral denso a Presencia/ausencia de cobertura densa de matorral (p.e.,incluída en un área dominada por masas forestales)

?

Pastizal a Presencia/ausencia de pastizal (p.e., incluída en un áreadominada por masas forestales)

?

Longitud de borde a Longitud de bordes entre la categoría forestal y el resto decategorías

?

Heterogeneidadespacial b

Dimensión fractal de un Índice Normalizado de Vegetación(NDVI) obtenido de imagen de satélite como un índice deheterogeneidad en cultivos

? ?

Distancia a cultivo a Distancia al área de cultivo más próxima mayor de 10 ha ? ?Distancia a matorral a Distancia al área de matorral más próxima mayor de 10 ha ? ?Distancia a

coníferas aDistancia al área forestal de coníferas (incluye eucaliptos)más próxima mayor de 2 ha

? ?

Distancia a cultivoleñoso a

Distancia al área de cultivo leñoso más próxima mayor de10 ha

?

Distancia avegetación riparia b, c

Distancia al área de vegetación riparia más próxima mayorde 2 ha

?

Distancia a cultivoherbáceo a

Distancia al área de cultivo herbáceo más próxima mayor de2 ha

?

Pendiente c Pendiente media (en grados) ? ?

Tabla 1. Conjuntos de variables probadas como predictores en los modelos estadísticos de distribución de aves depequeño tamaño a cada resolución espacial. Fuentes:a, mapa de usos y coberturas de Andalucía (1995) (SinambA:Consejería de Medio Ambiente, Junta de Andalucía); b, Imagen del sensor LISS III del satélite IRS (fechas: 19/07/99 y16/07/99 para ambas áreas de estudio). La dimensión fractal se calculó con IDRISI 32 sobre una imagen de NDVI; c,Modelo digital de elevaciones de Andalucía a 50 metros de resolución. Los puntos indican, para cada variable, si éstafue utilizada en los modelos con 350 y 1250 mts de resolución.


170

pasos (con la función step.gam de S-PLUS2000, MathSoft 1999) que usa unaaproximación del Criterio de Información deAkaike (Akaike´s Information Criterion,Sakamoto, Ishiguro & Kitagawa 1986) comocriterio para introducir o eliminar variables.Puesto que la forma de la relación entre lavariable respuesta y los predictores puede serno lineal, estos últimos se probaron aintroducir en los modelos como splines desuavizado (smoothing splines) con 3 gradosde libertad (Harrell 2001). Sin embargo, losmodelos resultantes no son paramétricos yresultan difíciles de implementar en un SIG(Guisan & Zimmermann 2000), por lo que seaproximaron a modelos lineares pararepresentar espacialmente sus predicciones(una técnica similar se empleó en el capítuloIII).

3.1.2. Rapaces

Las fuentes de información utilizadas paraobtener predictores ambientales fueron lasmismas que para el trabajo de paseriformes yafines: el mapa de usos y coberturas delSinambA, el modelo digital de elevaciones eimágenes de satélite (ver tabla 1 del capítuloVII). Se calculó también el índiceexperimental calibrado de vegetación global(Experimental Calibrated Global VegetationIndex) derivado del sensor AVHRR delsatélite NOAA (NOAA 1992). Para cadacuadrícula 10x10 se obtuvo un valorpromedio de los predictores.

La presencia/ausencia de cada especie enlas cuadrículas censadas por carretera seanalizó con modelos GLM y GAM por pasos,que incluían errores binomiales y enlace logit,según se detalla en el capítulo VII. En estecaso las relaciones no lineales se prefirieronincorporar con polinomios ortogonales, quetienen la ventaja de generar modelosparamétricos fáciles de pasar a un entorno deSIG. Además, se generaron predicciones parael resto de cuadrículas serranas andaluzas queno fueron censadas por carretera (n=295).

3.2. Modelos de experto


La elaboración de los modelos de opiniónde experto se obtuvo mediante lacolaboración de 3 ornitólogos profundosconocedores de la avifauna local. Se les pidióque estimaran la probabilidad de aparición decada especie en el entorno del área de estudio.Para ello se emplearon dos esquemas declasificación de hábitats: uno complejo, queincluía todas las 112 categorías del mapadigital de usos y coberturas del suelo deAndalucía (SinambA) y otro esquema, simple,que era una reclasificación de las categoríasanteriores en sólo 8 (cultivos herbáceos ypastizales, cultivos leñosos, áreas forestalesplanifolias, áreas forestales acutifolias,matorrales, vegetación riparia o lacustre yáreas urbanas e infraestructuras). Losexpertos estimaron la probabilidad depresencia en cada categoría de hábitat segúncuatro clases (0, la especie nunca estápresente en el hábitat considerado; 1,presencia ocasional; 2, presencia común; 3, laespecie siempre está en ese hábitat). Laestima de probabilidades según el esquemacomplejo es más lenta de realizar que laestima según el esquema simple y necesita unmayor conocimiento previo de la avifaunalocal. Estos dos esquemas tratan de reproducirdos posibilidades extremas que existen anteuna clasificación de la idoneidad de loshábitats para la fauna según un criterio deexperto, que es el procedimiento más comúnque se emplea en las evaluaciones de impactoambiental o en los estudios de ordenaciónterritorial. Se obtuvo además un promedio dela probabilidad predicha por los tres expertospara cada hábitat (lo que llamaremos enadelante el experto promedio), con laintención de comprobar si una forma tansencilla de sintetizar la información es útil.Finalmente, estos valores se reclasificaron aun intervalo [0,1] de probabilidad (0=0,1=0.1, 2=0.6, 3=1), y la información se pasó aun entorno SIG donde se generaron, para cadaespecie, 8 mapas de probabilidad de presenciacon una resolución de 50 mts (2


171

clasificaciones del hábitat x [3 expertos +experto promedio]). La probabilidad predichapor cada modelo de experto en cada punto demuestreo se estimó como el promedio de laprobabilidad de presencia en círculos de 350y 1250 metros de diámetro centrados en lospuntos de muestreo.

3.2.2. Rapaces

La dificultad de valorar mediante elcriterio de expertos la adecuación del hábitaten unidades de área extensas (las cuadrículasde 100 km2 que se usan en este trabajo), hizoque abandonáramos la estrategia seguida en elestudio de los paseriformes y afines. En estecaso, el criterio de experto se tomó de losdatos del nuevo atlas de aves reproductoras deEspaña (en elaboración) para Andalucía(SEO/BirdLife 2002). En éste, numerososcolaboradores (generalmente expertoslocales) han recorrido las distintas cuadrículasUTM 10x10 de España durante 1998-2001dedicando un esfuerzo heterogéneo a labúsqueda de indicios de cría de las especiesque las habitan. Las observaciones de aves secategorizan en tres clases principales deevidencia de reproducción creciente(reproducción posible, probable y segura,Purroy 1997), que fueron transformadas enprobabilidades de presencia entre 0 y 1 (tabla

2). Los atlas no se generan con modelos decriterio de experto pero su interpretacióntípica los asemeja a tales modelos de experto(la distribución de las especies se sueleconsiderar continua).

3.3. Evaluación de los modelos


La evaluación de los modelos estadísticosse llevó a cabo según un procedimiento deremuestreo de datos por validación cruzadaen conjuntos (k-fold crossvalidation) . Lasobservaciones se dividieron aleatoriamente en5 conjuntos de forma que con el 80% de lasobservaciones (4 de estos conjuntos) seconstruía un modelo que generabapredicciones para el restante 20% (el conjuntoque queda), donde se evaluaba su capacidadpredictiva (Verbyla & Litvaitis 1989). Ésta secalculó como el área bajo la curva (AUC deArea Under the Curve, Hanley & McNeil1982; Swets 1988) de un gráfico de operador-receptor (gráfico ROC de ReceiverCharacteristic Operating plot, Cumming2000), que es la medida de discriminaciónmás adecuada para los modelos dedistribución (ver capítulo I, Manel, Williams& Ormerod 2001). También se calculó el

Categoría DescripciónSin reproducción (0) Ausencia de la especie en época reproductoraReproducción posible

V (0.1) Especie vista en época adecuada y hábitat de cría adecuadoReproducción probable

MC (0.2) Macho con cantos territorialesT (0.3) Ave o pareja con territorio establecidoC (0.4) Cortejo, parada nupcial, disuasión ante depredadores, ...CN (0.5) Construcción de nido, entrada en agujeros, ...

Reproducción seguraCD (0.6) Distracción o fingimiento de heridas por parte de los adultosUN (0.7) Nido usado en el año o cáscaras de huevo asignables a la especieJ (0.8) Jóvenes recién salidos del nidoAC (0.9) Adultos con cebo o saco fecal en el picoN (1) Nido ocupado

Tabla 2. Categorías de reproducción que se asignan a cada especie de rapaz en las cuadrículas 10x10muestreadas en el nuevo atlas de las aves reproductoras de España. Entre paréntesis se indica laprobabilidad de aparición de la especie que se ha considerado en cada caso.


172

porcentaje de clasificación correcta (PCC),medido como la proporción de presencias yde ausencias acertadas con un punto de corteigual a 0.5 (e.d., consideramos que el modelopredecía la presencia de una especie si elvalor de probabilidad que estimaba para unpunto era mayor o igual que 0.5). El PCC estáafectado por la prevalencia y se haconsiderado una medida engañosa decapacidad predictiva (Fielding & Bell 1997, ycapítulos anteriores de esta tesis doctoral); noobstante, decidimos calcularlo por su mayorfamiliaridad y a modo comparativo con elAUC. El proceso de remuestreo se realizó 20veces para obtener 100 estimas de capacidadpredictiva (5 por cada una de las 20 veces quese dividieron los datos).

Los modelos de experto se evaluaronmediante la estimación de la capacidadpredictiva en los mismos 100 conjuntos deobservaciones que se usaron en los modelosestadísticos. Los valores de probabilidad depresencia de cada especie para cada conjuntode los puntos de muestreo (que comprende un20% de las observaciones) se compararon conla presencia/ausencia detectada en los censos,obteniéndose las estimas de AUC y PCC quedespués se promediaron.

Por último, se analizó mediante ANOVAel efecto sobre las estimas de capacidadpredictiva de los tipos de modelo según suorigen (nueve por cada especie: unoestadístico y cuatro de expertos), del esquemade clasificación adoptado (dos por especie:esquemas sencillo y complejo), de laresolución espacial (350 vs 1250 mts) y de lavariabilidad entre especies.

3.3.2. Rapaces

En este apartado comparamos primero lacapacidad predictiva del atlas, consideradocomo un modelo de experto, con la de losmodelos estadísticos. Segundo, nosplanteamos hasta qué punto las prediccionesde los modelos estadísticos coinciden con losresultados de los atlas de distribución, pues,de existir un alto grado de solapamiento, los

modelos estadísticos podrían emplearse paraextrapolar la información recogida a áreas sinprospectar en los atlas. En el primer caso lainformación del atlas se utilizó como si setratara de un modelo de experto. En elsegundo caso esa información se considerócomo la distribución real de las especies. Portanto, se hicieron dos evaluaciones diferentessegún si los datos de referencia con los quecomparar las predicciones de los modeloseran los de los censos por carretera (n=88cuadrículas) o los de los resultados del atlas(n=383 cuadrículas, el atlas se utilizó comoverdad terreno).

Los modelos estadísticos que secompararon fueron los que incorporabanvariables de vegetación, topografía ycoordenadas espaciales (modelos TUC delcapítulo VII). En el primer tipo de evaluación(n=88) se generaron predicciones jackknifepara cada cuadrícula (e.d., se construía unmodelo estadístico con 87 cuadrículas y sepredecía para la restante, ver detalles encapítulo VII), que se comparaban con losresultados de los censos por carretera(Verbyla & Litvaitis 1989). Se prefirió usarjackknife frente a otras técnicas de remuestreode datos porque esta es la más apropiadacuando se tiene un tamaño muestral reducido.Los datos del atlas correspondientes a las 88cuadrículas censadas por carretera secompararon con los resultados de los censosde dos formas: considerando los datos básicos(cualquier indicio de reproducción se tomócomo presencia de la especie), ytransformando las clases nominales enprobabilidades de presencia (tabla 2). En elsegundo tipo de evaluación (n=383) se evaluóla capacidad del modelo estadístico parapredecir la presencia/ausencia de las especiessegún el atlas. La presencia de las especies enlas cuadrículas del atlas se consideró con trescriterios: presencia en todas las cuadrículascon indicios de reproducción, en sóloaquellas con al menos indicios probables dereproducción y, por último, en sólo las que sehabía detectado una reproducción segura. Losestadísticos utilizados en las evaluacionesfueron el AUC y el PCC (con 0.5 como punto


173

de corte). Varias cuadrículas tenían menos de100 km2 de superficie al estar ocupadas enparte por el mar o bien por disponerse en elborde de un cambio de huso. Por esta razón sedieron pesos proporcionales a la superficieterrestre que ocupaba cada cuadrícula tanto alas predicciones de los modelos estadísticoscomo a las observaciones del atlas. Loserrores estándar se estimaron mediantebootstraping, recalculando los estadísticos en250 muestras de tamaño igual al conjuntooriginal (e.d., n=88 o n=383) cuyasobservaciones se escogieron al azar conrepetición.

4. Resultados

4.1. Paseriformes y afines

El desequilibrio de nuestro diseño delestudio (existe un solo modelo estadístico porcada especie frente a dos por cada experto)nos obliga a analizar el efecto del tipo demodelo de una manera poco directa. Primerodamos los resultados de los análisis en los que

nos interesa comparar los distintos modelosde criterio de experto y los estadísticos, ydespués mostramos los análisis que serefieren sólo a los modelos de experto.

En primer lugar, la capacidad predictiva delos distintos modelos de experto y de losmodelos estadísticos, medida como AUC o elporcentaje de clasificación correcta (PCC),difirió entre especies (tablas 3 y 4). Así, elANOVA de dos vías mostró una interacciónsignificativa entre el tipo de modelo y laespecie (AUC: F=18.0, P<0.0001; PCC:F=28.5, P<0.0001). Esta interacción sugierela posibilidad de que algún tipo de modelo(según si deriva de un esquema simple, unesquema complejo, o de un modeloestadístico) o alguno de los expertos sean máspredictivos que el resto para algunas especies.Esta segunda posibilidad se consideró alrepetir el ANOVA con los modelos de cadaexperto por separado, encontrándose lamisma interacción (P<0.0001 en todos loscasos). Finalmente, se controló la variabilidadintroducida por las distintas especies para

Fuente de variación gdl Suma decuadrados

Cuadradosmedios

F P

tipo de modelo(expertos y estadístico)

8 1.98 0.25 182.8 <0.0001

especie 9 2.49 0.28 204.1 <0.0001modelo x especie 72 1.76 0.02 18.0 <0.0001residuos 90 0.12 0.001

Tabla 3. ANOVA del AUC para los modelos de especies paseriformes y afines.Nótese que en este análisis se consideraron nueve modelos para cada especie:uno estadístico y dos por cada uno de los tres expertos y el promedio deexpertos (según un esquema de clasificación simple y otro complejo).


Cuadradosmedios

F P

tipo de modelo(expertos y estadístico)

8 1.01 0.13 97.0 <0.0001

Especie 9 3.73 0.41 317.5 <0.0001modelo x especie 72 2.68 0.04 28.5 <0.0001residuos 90 0.12 0.001

Tabla 4. ANOVA del porcentaje de clasificación correcta para los modelos de especiespaseriformes y afines. Nótese que en este análisis se consideraron nueve modelos para cadaespecie: uno estadístico y dos por cada uno de los tres expertos y el promedio de expertos (segúnun esquema de clasificación simple y otro complejo).


174

separar su efecto del producido por el tipo demodelo: en este caso se encontró un efectosignificativo del tipo de modelo (tablas 5 y 6).Además, la capacidad predictiva alcanzadapara cada especie no fue diferente en losmodelos realizados a distinto radio (figuras 1y 2).

Segundo, nos planteamos si la capacidadpredictiva difería entre los modelos deexpertos realizados según un esquema simpley los que siguieron un esquema complejo. Noencontramos diferencias en las estimas deAUC (tabla 7) pero sí en las de PCC (tabla 8),siendo algo mayor la capacidad predictivaque alcanzan los modelos realizados según un

esquema complejo (media=0.74, sd=0.18 vsmedia=0.69, sd=0.24). Despuéscomprobamos que los modelos derivados delcriterio de los distintos expertos no difierenen cuanto al AUC que obtienen (tabla 9,figura 1), pero sí tienden a diferir en cuanto alPCC (tabla 10, figura 2). En concordancia, losmodelos del experto promedio sólo difirierondel resto en el PCC (AUC, diferencia en lasmedias de 0.02, test de la t apareado: t=0.83,gdl=9, P=0.43; PCC, diferencia de 0.09,t=2.81, gdl=9, P=0.02). Los modelosestadísticos alcanzaron valores de AUCsignificativamente mayores que el promediode los modelos de experto (diferencia de 0.33,t=8.79, gdl=9, P<0.0001).

Variable gdl Suma decuadrados

Cuadradosmedios

F P

Error: especie Residuos 9 2.49 0.28 - -

Error: entre variables tipo de modelo

(expertos y estadístico)8 1.98 0.25 20.24 <0.0001

radio 1 0.004 0.004 0.34 0.56 modelo x radio 8 0.006 0.001 0.06 0.99Residuos 153 1.87 0.01 - -

Tabla 5. ANOVA del AUC para los modelos de paseriformes y afines en que se ha controlado la variabilidadintroducida por las distintas especies. Nótese que en este análisis se consideraron nueve modelos para cada especie: unoestadístico y dos por cada uno de los tres expertos y el promedio de expertos (según un esquema de clasificaciónsimple y otro complejo).


Cuadradosmedios

F P



(expertos y estadístico)8 1.01 0.12 7.07 <0.0001


Tabla 6. ANOVA del porcentaje de clasificación correcta para los modelos de paseriformes y afines en que se hacontrolado la variabilidad introducida por las distintas especies. Nótese que en este análisis se consideraron nuevemodelos para cada especie: uno estadístico y dos por cada uno de los tres expertos y el promedio de expertos (según unesquema de clasificación simple y otro complejo).


175

Figura 1. AUC de los modelos de experto generados mediante un esquema simple (1) y otro complejo (2, con elexperto promedio en blanco), y de los modelos estadísticos (3) a dos resoluciones espaciales. a, Alectoris rufa; b,Certhia brachydactyla; c, Galerida theklae; d, Parus caeruleus; e, Sylvia melanocephala; f, Carduelis cannabina; g,Erithacus rubecula; h, Melanocorypha calandra; i, Sitta europaea; j, Troglodytes troglodytes. Los datos son valorespromedio (n=20) de los valores medios (n=5) obtenidos en cada remuestreo. Las barras de error son 2*SE.

En contraste, los modelos estadísticostendieron a alcanzar valores de PCC mayoresque los modelos de experto, pero lasdiferencias sólo fueron significativas enalgunos casos (experto 1: diferencia de 0.13,t=2.37, gdl=9, P=0.04; experto 2: diferenciade 0.18; t=2.35, gdl=9, P=0.04; experto 3:diferencia de 0.04, t=1.25, gdl=9, P=0.24;experto promedio: diferencia de 0.02, t=0.59,gdl=9, P=0.57).

Además, y como resulta especialmentenotable en las figuras 1 y 2, la estima de lacapacidad predictiva de los modelosestadísticos tuvo una variabilidad menor quela de los modelos generados por expertos,tanto en el AUC (diferencia en el tamaño del

intervalo de confianza aproximado [2*SE] de0.051; t=4.11, gdl=9, P=0.003), como en elPCC (experto 1: diferencia de 0.014, t=14.55,gdl=9, P<0.0001; experto 2: diferencia de0.013; t=8.83, gdl=9, P<0.0001; experto 3:diferencia de 0.011, t=6.84, gdl=9, P=0.0001;experto promedio: diferencia de 0.010,t=6.43, gdl=9, P=0.0001).

Finalmente, las estimas de capacidadpredictiva alcanzada por los modelosestadísticos para todas las especies fueronsuperiores a las que tendría un modelo nulo(figuras 1 y 2). Sin embargo, AUC y PCCdifieren algo en cuanto al escenario que nospresentan: todos los modelos pueden

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

diámetro 350 mts diámetro 1250 mts

a f

b g

c h

d i

e j

a f

b g

c h

d i

e j

1 2 3 1 2 3 1 2 3 1 2 3


176

Figura 2. Porcentaje de clasificación correcta (PCC) de los modelos de experto y de los modelos estadísticos.Abreviaturas como en la figura 1.

considerarse buenos según el AUC (>0.8 entodas las especies), pero son más desigualessegún el PCC (Certhia brachydactyla, Paruscaeruleus y Sylvia melanocephala tienenrelativamente bajos PCC entre 0.6 y 0.7). Losmodelos de experto se comportaron tambiénde manera heterogénea (figuras 1 y 2). Losmodelos para algunas especies alcanzaronvalores altos tanto de AUC (0.6-0.7) como dePCC (>0.8) (Erithacus rubecula,Melanocorypha calandra y Sitta europaea),mientras que los de otras fueron pocopredictivos (no mejores que un modelo nulo),especialmente según el AUC (Alectoris rufa,Certhia brachydacyla y Cardueliscannabina). En general, las estimas de PCCalcanzadas por los modelos de experto fueronrelativamente mayores que sus estimas de

AUC y, salvo excepciones (Alectoris rufa,Carduelis cannabina y Galerida theklae), losmodelos de los distintos expertos obtuvieronresultados similares.

Los mapas de probabilidad predicha paracada especie por los modelos de expertopromedio y por los estadísticos pueden verseen el apéndice I.

4.2. Rapaces

En el primer tipo de evaluación, donde lareferencia son los censos por carretera, losmodelos estadísticos obtuvieron una alta ysimilar capacidad predictiva para las cuatroespecies consideradas según el AUC

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

diámetro 350 mts diámetro 1250 mts

a f

b g

c h

d i

e j

a f

b g

c h

d i

e j

1 2 3 1 2 3 1 2 3 1 2 3


177


Cuadradosmedios

F P



(simple vs complejo)1 0.000 0.000 0.01 0.92


Tabla 7. Comparación de AUC entre los modelos de experto (dos tipos de modelo por cada especie, unosigue un esquema de clasificación simple y el otro un esquema complejo) para especies paseriformes yafines.

Variable Gdl Suma decuadrados

Cuadradosmedios

F P



(simple vs complejo)1 0.11 0.11 5.30 0.023


Tabla 8. Comparación del porcentaje de clasificación correcta entre los modelos de experto (dos tipos demodelos por cada especie, uno sigue un esquema de clasificación simple y el otro un esquema complejo),para especies paseriformes y afines.

(media±SD: 0.78±0.04), aunque no tantosegún el PCC (media±SD: 0.72±0.06, figura3). La capacidad predictiva de estos modelosfue mayor que las de los derivados del atlas,tanto de los generados con los datos básicoscomo de los generados con datos en 10categorías de probabilidad. Los intervalos deconfianza aproximados de las mediastendieron a no solaparse (las diferenciasfueron, respectivamente, de 0.22 y 0.17 enAUC, y de 18 y 14% en PCC respecto losdatos básicos y en categorías). Los mapasderivados de datos en categorías fueron algomás predictivos que los derivados de los datosbásicos, pero los intervalos de confianza delas medias se solapan en gran medida por loque la diferencia debería interpretarse comono significativa (figura 3).

En el segundo tipo de evaluación, donde lareferencia son los resultados del atlas, losmodelos estadísticos obtuvieron bajas

capacidades predictivas para las cuatroespecies consideradas según el AUC(media±SD: 0.59±0.03) y el PCC (media±SD:0.51±0.08). Sólo las predicciones de losmodelos estadísticos para Milvus migransparecen acercarse moderadamente a los datosdel atlas (figura 4). El PCC sugiere que laconcordancia es mayor con el criterio másrestrictivo para asumir la presencia de unaespecie en una cuadrícula (reproducciónsegura), pero el AUC no muestra ningúnpatrón (figura 4). Una comparación similarcon los datos del atlas no se hizo para elgrupo de especies paseriformes y afinesporque existía un desacoplamiento muyimportante entre la resolución de los datos decenso (1.2 km2 para el diámetro de censomayor) y la resolución de los datos del atlas(100 km2).


178

5. Discusión

La ventaja principal de los modelosestadísticos de distribución de especies sobrelos elaborados mediante el criterio deexpertos es su mayor objetividad. Sinembargo, se puede cuestionar si el esfuerzode muestreo, sus relativamente complejosmétodos matemáticos y, en ocasiones, ladifícil interpretabilidad de los modelos estánjustificados, toda vez que el criterio de unexperto podría identificar correctamente loshábitats adecuados para una especie y, porextensión, el área en que se distribuye.Nuestros resultados resuelven esta cuestión demanera muy favorables a los modelosestadísticos. En primer lugar, éstos tuvieronuna alta capacidad predictiva en los dosgrupos de especies analizados que fue mayorque la de los modelos generados mediantecriterio de experto. Las diferencias fueron

Figura 3. Estimas de capacidad predictiva para lasrapaces analizadas de los modelos estadísticos (barrade la izquierda, gris), y de los mapas derivados de losdatos del nuevo atlas de aves reproductoras de España(barra del centro, rayada, para los datos básicos depresencia/ausencia y barra de la derecha, blanca, paralos datos en 10 categorías de probabilidad). Los datosde referencia son aquí los resultados de censos porcarretera en 1996 (n=88 cuadrículas UTM 10x10). Seindican los intervalos de confianza aproximados al95%.

Figura 4. Estimas de la capacidad predictiva de losmodelos estadísticos de las rapaces cuando los datos dereferencia son los derivados del nuevo atlas de avesreproductoras de España (barra de la izquierda, gris,para cualquier indicio de reproducción; barra delcentro, rayada, para las reproducciones al menosprobables; y barra de la derecha, blanca, para lasreproducciones seguras, n=383). Se indican losintervalos de confianza aproximados al 95%.

más notables en el grupo de especiespaseriformes y afines, lo que indica que losmodelos estadísticos son especialmenteventajosos cuando el tamaño muestral de quese dispone para generarlos es grande o,alternativamente, cuando la resoluciónespacial de los mapas es detallada. Entre losmodelos de experto, los más elaboradostendieron a ser mejores que los sencillos, porlo que se podría justificar la adopción deesquemas complejos en su construcción. Noobstante, esta tendencia no fue muy clara(sólo se detectó con el PCC) y, entre lospaseriformes y afines, parecía depender decada especie en particular.

En segundo lugar, nuestros resultadossugieren que los modelos de experto son másinconsistentes que los estadísticos, ya que susestimas de capacidad discriminativa son másvariables. Esto quiere decir que la eficacia de

0

0.2

0.4

0.6

0.8

1

Buteo buteo Circaetusgallicus

Hieraaetuspennatus

Milvus migrans

AU

C

0

0.2

0.4

0.6

0.8

1


Hieraaetuspennatus

Milvus migrans

% c

lasi

fica

ció

n c

orr

ecta

0

0.2

0.4

0.6

0.8

1


Hieraaetuspennatus

Milvus migrans

AU

C

0

0.2

0.4

0.6

0.8

1


Hieraaetuspennatus

Milvus migrans

% c

lasi

fica

ció

n c

orr

ecta


179


Cuadradosmedios

F P

tipo de modelo 1(expertos)

3 0.09 0.03 0.94 0.42

tipo de modelo 2(simple vs complejo)

1 0.00 0.00 0.00 0.96

radio 1 0.00 0.00 0.14 0.71modelo 1 x modelo 2 3 0.01 0.00 0.06 0.98modelo 1 x radio 3 0.00 0.00 0.04 0.99modelo 2 x radio 1 0.00 0.00 0.01 0.93modelo 1 x modelo 2

x radio3 0.00 0.00 0.03 0.99

residuos 144 4.33 0.03

Tabla 9. ANOVA del AUC entre los distintos modelos de expertos para especiespaseriformes y afines. Aquí se consideran los cuatro modelos de experto (tipo 1)y los dos esquemas seguidos (tipo 2).


Cuadradosmedios

F P

tipo de modelo 1(expertos)

3 0.68 0.23 5.14 0.002

tipo de modelo 2(simple vs complejo)

1 0.11 0.11 2.62 0.11

radio 1 0.05 0.05 1.06 0.30modelo 1 x modelo 2 3 0.06 0.02 0.48 0.69modelo 1 x radio 3 0.01 0.00 0.06 0.98modelo 2 x radio 1 0.02 0.02 0.05 0.83modelo 1 x modelo 2

x radio3 0.00 0.00 0.01 0.99

residuos 144 6.32 0.04

Tabla 10. ANOVA del PCC entre los distintos modelos de expertos paraespecies paseriformes y afines. Aquí se consideran los cuatro modelos deexperto (tipo 1) y los dos esquemas seguidos (tipo 2).

s modelos de experto depende del conjunto deáreas en que se evalúen, lo que sólo puedeexplicarse si estos modelos predicen bien paradeterminados hábitats y mal para otros. Segúnesta interpretación, los expertos son capacesde identificar correctamente sólo algunos delos hábitats adecuados para las especies (obien consideran apropiados algunos que no loson). Tales errores podrían ser debidos o biena lagunas en el conocimiento de los expertoso bien, más probablemente, a que lascategorías de usos y coberturas del mapaambiental utilizado como fuente depredictores se diseñaron con un propósitogeneral, por lo que pueden no reflejar los

requerimientos de las especies (al menos deforma comprensible). Se podría argumentarque los modelos de experto serían másconsistentes, y tendrían mayor capacidadpredictiva, si los predictores que usarandescribieran mejor los hábitats desde el puntode vista de su interés para la avifauna. Talargumento es intuitivo, pero olvida que no esabordable crear mapas de variablesambientales medidas teniendo en cuenta lasdistintas exigencias de las diferentes especiesde aves. Los modelos de distribución deespecies aplicados a grandes áreas debenrecurrir, inevitablemente, a mapas diseñadoscon un propósito general, que describirán con


180

precisión variable los hábitats de interés encada caso. Dada esta situación, igualmentedesfavorable para modelos estadísticos y deexperto, los primeros han demostrado ser lamejor opción en las dos escalas de análisis deeste estudio.

Existe cierto contraste, en cuanto a lospatrones anteriores se refiere, entre losresultados de las dos medidas de capacidadpredictiva que hemos usado. Así, según lamás adecuada de ellas, el AUC, los modelosde experto son poco afortunados en todas lasespecies con la posible excepción deTroglodytes troglodytes y Parus caeruleus.Por el contrario, el PCC muestra menoresdiferencias entre los modelos estadísticos ylos de experto. Esto se debe probablemente alefecto de la prevalencia en las estimas dePCC. Cuando la prevalencia es cercana al50% (como en el caso de las rapacesanalizadas aquí) no se producen grandesdiferencias entre PCC y AUC, pero cuandoquedan muy por debajo de ese 50% (como enel caso de los paseriformes), se predicenmejor las ausencias que las presencias de laespecie y las estimas de PCC pueden serengañosamente altas (Manel, Williams &Ormerod 2001). Por ejemplo, los modelos deexperto para Melanocorypha calandra yErithacus rubecula alcanzan valores altos dePCC, puesto que ambas especies tienen bajaprevalencia y sus requerimientos de hábitatpueden describirse en términos de lascategorías de usos y coberturas del suelo quese han utilizado aquí (i.e., los definidos poruna cartografía temática de propósito general,Moreira & Fernández-Palacios 1995). Enconjunto, estas discrepancias nos hacendesconfiar aún más del PCC como medida decapacidad predictiva (Pearce & Ferrier 2000;Manel, Williams & Ormerod 2001).

Dos resultados secundarios de nuestrosanálisis se refieren a las especies paseriformesy afines. El primero es la ausencia dediferencias sistemáticas en los resultadosobtenidos a las dos resoluciones (350 y 1250mts). Se ha visto anteriormente que losmodelos más predictivos incluían variables

ambientales medidas a grandes diámetrosalrededor del punto de muestreo (capítulo V),lo que se ha relacionado con la importanciaque tiene el paisaje como predictor (Saab1999; Kie et al. 2002; Wolff et al. 2002).Por tanto, una de las posibles razones por lasque los modelos de experto resultan pocosatisfactorios es que no incluyen entre susvariables ninguna relacionada con el paisaje.Si esto es así, cabría esperar que los modelosque incluyeran variables medidas en unentorno más amplio (1250 mt) fueranmejores, lo que no ocurre. Aparentemente,promediar las variables en un entorno ampliono equivale a considerar el paisaje. Elsegundo resultado para mencionar aquí es quela síntesis de la opinión de varios expertos nomejora necesariamente los modelos. En lasocasiones en que se dispone de lacolaboración de expertos con criteriosdiferentes y conocimientos similares (odifíciles de comparar), es necesario sintetizarla información que aportan (Pearce et al.2001). En nuestro trabajo esto se hizotomando el promedio de las valoraciones quelos expertos hicieron para cada hábitat,aunque es posible pensar en otras formas deintegrar la información, como buscar unconsenso (lo que obliga a que loscolaboradores interactúen), o escoger losvalores mínimos o máximos en la valoraciónfinal. Un criterio sintético podría serventajoso al corregir las deficiencias quealguno de los expertos pudiera tener enespecies o hábitats que conociera peor. Deacuerdo con esto, nuestros resultados sugierenque los modelos realizados con el criteriointegrado alcanzan valores de capacidadpredictiva elevados (aunque no siempre losmayores) en relación al resto de modelos deexperto.

Finalmente, las predicciones estadísticasno coincidieron con las observaciones delatlas para las rapaces, con la posibleexcepción de Milvus migrans. Ante estadisparidad, ¿qué fuente de cartografíadeberíamos suponer más fiable? Lascuadrículas del atlas fueron prospectadas demanera no sistemática y con variable esfuerzo


181

por colaboradores con experienciaheterogénea. Por el contrario, los datos conlos que se construyeron los modelos procedende un muestreo estandarizado, en el que secontroló el esfuerzo de prospección que serealizó y éste fue siempre superior a unmínimo (Bustamante, Donázar & Hiraldo1997). Además, se ha comprobado para elMilano real (Milvus milvus) que losresultados de los censos por carretera secorrelacionan con las abundancias de parejasnidificantes estimadas mediante búsquedadirecta de nidos (Viñuela 1997), lo que noshace confiar en el método cuando se aplica aotras especies. Por tanto, consideramos que lainformación más próxima a la situación realla aportan, en general, los censos por carreteray los modelos que se derivan de ellos (aunqueno hay duda de que los datos para muchasáreas del atlas son de alta calidad). Debenmencionarse, no obstante, dos circunstanciasen que la información de los atlas es másfiable que la derivada de los modelos. Se tratade las poblaciones (o parejas en el caso de lasrapaces) ligadas a ambientes atípicos, que losmodelos estadísticos no detectan pero quepueden ser muy conocidos por los ornitólogoslocales (es el caso, por ejemplo, del Milanoreal en Doñana, ver capítulo VIII). También,basta que exista una pareja aislada de unaespecie en un hábitat óptimo de pequeñaextensión para que la cuadrícula que lacontenga se identifique como con presenciade tal especie. Si la mayor parte de esacuadrícula tiene hábitats subóptimos, unmodelo estadístico tenderá a predecir una bajaprobabilidad de presencia en ella. Aúnteniendo en cuenta estas salvedades, creemosque no es prudente usar los mapas derivadosde información de atlas para análisis deextensión regional o inferior, como tememosque es común en los estudios del medio físicotípicos, p.e., de las evaluaciones de impactoambiental.

AGRADECIMIENTOS

Paco Chiclana, Javier Salcedo, JoséAntonio Lama y Jorge Garzón, del grupoornitológico SEO-Sevilla, prestaron su tiempo

para ofrecerme su criterio de expertos en laavifauna local (en esta y en otras muchasocasiones). El Ministerio de Medio Ambientecedió amablemente los datoscorrespondientes a Andalucía del nuevo atlasde aves reproductoras en España, y JuanCarlos del Moral, de SEO/BirdLife(organización encargada de la elaboración delatlas), agilizó los trámites para conseguirlos.Los muestreos del atlas y los censos derapaces por carretera fueron realizados pornumerosos colaboradores, en gran parte noprofesionales, sin cuyo trabajo no habría sidoposible ni este ni otros tantos estudios (en elapéndice I del capítulo VII se da una lista decolaboradores en los censos de rapaces porcarretera).

BIBLIOGRAFÍA

Austin, G. E., Thomas, C. J., Houston, D. C.& Thompson, D. B. A. (1996). Predictingthe spatial distribution of buzzard Buteobuteo nesting areas using a GeographicalInformation System and remote sensing.Journal of Applied Ecology, 33: 1541-1550.

Beaman, M. & Madge, S. (1998). Aves deEuropa, norte de África y próximo Oriente.Guía de identificación. Omega, Barcelona

Bojórquez-Tapia, L. A., Azuara, I., Ezcurra,E. & Flores-Villela, O. (1995). Identifyingconservation priorities in Mexico throughgeographic information systems andmodeling. Ecological Applications, 5(1):213-231.

Bustamante, J., Donázar, J. A. & Hiraldo, F.(1997). Factores que condicionan ladistribución reproductora del milano real(Milvus milvus) en Andalucía: elaboraciónde un plan de conservación. Sevilla, AMA-CSIC.

Caicco, S. L., Scott, J. M., Butterfield, B. &Csuti, B. (1995). A Gap Analysis of theManagement Status of the Vegetation ofIdaho (U.S.A.). Conservation Biology,9(3): 498-511.

Cody, M. L., Ed. (1985). Habitat selection inbirds/. Academic Press, Inc., . San Diego


182

Corsi, F., Duprè, E. & Boitani, L. (1999). ALarge-Scale Model of Wolf Distribution inItaly for Conservation Planning.Conservation Biology, 13(1): 150-159.


Díaz, M., Illera, J. C. & Hedo, D. (2001).Strategic environmental assessment ofplans and programs: a methodology forestimating effects on biodiversity.Environmental Management, 28(2): 267-279.

Doadrio, I., Ed. (2001). Atlas y libro rojo delos peces continentales de España/.Ministerio de Medio Ambiente-ConsejoSuperior de Investigaciones Científicas, .Madrid

Donald, P. F. & Fuller, R. J. (1998).Ornithological atlas data: a review of usesand limitations. Bird Study, 45: 129-145.

Ertsen, A. C. D., Bio, A. M. F., Bleuten, W.& Wassen, M. J. (1998). Comparison ofthe performance of species responsemodels in several landscape units in theprovince of Noord-Holland, TheNetherlands. Ecological Modelling,109(2): 213-223.

Fielding, A. H. & Bell, J. F. (1997). A reviewof methods for the assessment ofprediction errors in conservationpresence/absence models. EnvironmentalConservation, 24: 38-49.

Guisan, A., Theurillat, J.-P. & Kienast, F.(1998). Predicting the potentialdistribution of plant species in an alpineenvironment. Journal of VegetationScience, 9: 65-74.

Guisan, A. & Zimmermann, N. E. (2000).Predictive habitat distribution models inecology. Ecological Modelling, 135: 147-186.

Hagemaijer, E. J. M. & Blair, M. J., Eds.(1997). The EBCC Atlas of EuropeanBreeding Birds: Their distribution andAbundance/. T & A D Poyser, . London

Hanley, J. A. & McNeil, B. J. (1982). Themeaning and use of the area under a

Receiver Operating Characteristic (ROC)curve. Radiology, 143: 29-36.

Harrell, F. E. (2001). Regression modelingstrategies. Springer, New York

Hastie, T. J. & Tibshirani, R. J. (1990).Generalized Additive Models. Chapman &Hall, London

Kie, J. G., Bowyer, R. T., Nicholson, M. C.,Boroski, B. B. & Loft, E. R. (2002).Landscape heterogeneity at differingscales: effects on spatial distribution ofmule deer. Ecology, 83(2): 530-544.

Kiester, A. R., Scott, J. M., Csuti, B., Noss,R. F., Butterfield, B., Sahr, K. & White, D.(1996). Conservation Prioritization UsingGAP Data. Conservation Biology, 10(5):1332-1342.

Lavers, C. P. & Haines-Young, R. H. (1996).Using models of bird abundance to predictthe impact of current land-use andconservation policies in the Flow Countryof Caithness and Sutherland, northernScotland. Biological Conservation, 75: 71-77.

Manel, S., Buckton, S. T. & Ormerod, S. J.(2000). Testing large-scale hypothesesusing surveys: the effects of land use onthe habitats, invertebrates and birds ofHimalayan rivers. Journal of AppliedEcology, 37(5): 756-770.

Manel, S., Williams, H. C. & Ormerod, S. J.(2001). Evaluating presence-absencemodels in ecology: the need to account forprevalence. Journal of Applied Ecology,38: 921-931.

MathSoft, I. (1999). S-Plus 2000 Guide toStatistics. Data Analysis ProductsDivision, Seattle

Moreira, J. M. & Fernández-Palacios, A.(1995). Usos y coberturas vegetales delsuelo en Andalucía. Seguimiento a travésde imágenes de satélite. Junta deAndalucía. Consejería de medio ambiente,Sevilla

Morrison, M. L., Marcot, B. G. & Mannan, R.W. (1998). Wildlife-habitat relationships.Concepts and applications. The Universityof Wisconsin Press, Madison

Mourell, C. & Ezcurra, E. (1996). Speciesrichness of Argentine cacti: A test of


183

biogeographic hypotheses. Journal ofVegetation Science, 7: 667-680.

NOAA (1992). Experimental CalibratedGlobal Vegetation Index from NOAA'sAdvanced Very High ResolutionRadiometer, 1985-1991. (KGRD: 28).Global Change Data Base. Boulder,Colorado .

Osborne, P. E., Alonso, J. C. & Bryant, R. G.(2001). Modelling landscape-scale habitatuse using GIS and remote sensing: a casestudy with great bustards. Journal ofApplied Ecology, 38(2): 458-471.

Pearce, J. & Ferrier, S. (2000). Evaluating thepredictive performance of habitat modelsdeveloped using logistic regression.Ecological Modelling, 133: 225-245.

Pearce, J. L., Cherry, K., Drielsma, M.,Ferrier, S. & Whish, G. (2001).Incorporating expert opinion and fine-scalevegetation mapping into statistical modelsof faunal distribution. Journal of AppliedEcology, 38: 412-424.

Powell, G. V. N., Barborak, J. & Rodriguez,M. (2000). Assessing representativeness ofprotected natural areas in Costa Rica forconserving biodiversity: a preliminary gapanalysis. Biological Conservation, 93: 35-41.

Purroy, F. J., Ed. (1997). Atlas de las aves deEspaña (1975-1995)/. Lynx editions, .Barcelona

Saab, V. (1999). Importance of spatial scaleto habitat use by breeding birds in riparianforests: a hierarchical analysis. EcologicalApplications, 9(1): 135-151.

Sakamoto, Y., Ishiguro, M. & Kitagawa, G.(1986). Akaike Information CriterionStatistics. KTK Scientific Publishers,Tokyo

Scott, J. M., Davis, F., Csuti, B., Noss, R.,Butterfield, B., Groves, C., Anderson, H.,Caicco, S., D´erchia, F., Edwards, J., T.C.,Ulliman, J. & Wright, R.G. (1993). GAPanalysis: a geographic approach toprotection of biological diversity. WildlifeMonographs, 123: 1-41.

SEO/BirdLife (2002). El nuevo "Atlas de lasaves reproductoras de España". LaGarcilla, 112: 21.

Short, H. L. & Hestbeck, J. B. (1995).National biotic resource inventories andGAP analysis: problems of scale andunproven assumptions limit a nationalprogram. Bioscience, 45: 535-539.

Swets, J. A. (1988). Measuring the accuracyof diagnostic systems. Science, 240: 1285-1293.

Treweek, J. (1996). Ecology andenvironmental impact assessment. Journalof Applied Ecology, 33: 191-199.

Tucker, G. M. & Evans, M. I. (1997).Habitats for birds in Europe. Aconservation strategy for the widerenvironment. BirdLife International,Cambridge

Verbyla, D. L. & Litvaitis, J. A. (1989).Resampling methods for evaluatingclassification accuracy of wildlife habitatmodels. Environmental Management, 13:783-787.

Viñuela, J. (1997). Road transects as a large-scale census method for raptors: the caseof the Red Kite Milvus mivus in Spain.Bird Study, 44: 155-165.

Wilson, D. E. & Ruff, S., Eds. (1999). TheSmithsonian book of North Americanmammals/. Smithsonian Institution, .Singapore

Wolff, A., Dieuleveut, T., Martin, J.-L. &Bretagnolle, V. (2002). Landscape contextand little bustard abundance in afragmented steppe: implications forreserve management in mosaic landscapes.Biological Conservation, 107(2): 211-220.


184

20000

metros

N

Fig.1. Mapas de referenciapara las dos zonas deestudio (Aracena, arriba;Grazalema, abajo). Seseñalan, en verde, loslímites provinciales, engris, la red viaria y losnúcleos urbanos, y, ennegrita, algunos topónimos(que se han centrado en laparte superior del área alque se nombran).

SEVILLA

Aracena

Nerva

La Palmadel Condado

GuillenaAznalcóllar

Sta. Olalla del Cala

Castilblancode los arroyos

Estepona

Alcalá delos Gazules Jimena de

la Frontera

Arcos dela Frontera

Villamartín

Montellano

Olvera

RONDAGrazalema

Ubrique

Comparación de la cartografía de especies: apéndice de cartografía

185

Fondo<20%20-40-60-80-

2000

metros

N

Alectoris rufaAracena

Fig.2. Mapas de laprobabilidad de presenciade Alectoris rufa en el áreade Aracena según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CAPÍTULO IX, APÉNDICE I

186

Fondo<20%20-40-60-80-

2000

metros

N

CardueliscannabinaAracena

Fig.3. Mapas de laprobabilidad de presenciade Carduelis cannabina enel área de Aracena segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


187

Fondo<20%20-40-60-80-

2000

metros

N

CerthiabrachydactylaAracena

Fig.4. Mapas de laprobabilidad de presenciade Certhia brachydactylaen el área de Aracena segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


188

Fondo<20%20-40-60-80-

2000

metros

N

ErithacusrubeculaAracena

Fig.5. Mapas de laprobabilidad de presenciade Erithacus rubecula en elárea de Aracena según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


189

Fondo<20%20-40-60-80-

2000

metros

N

Galerida theklaeAracena

Fig.6. Mapas de laprobabilidad de presenciade Galerida theklae en elárea de Aracena según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


190

Fondo<20%20-40-60-80-

2000

metros

N

MelanocoryphacalandraAracena

Fig.7. Mapas de laprobabilidad de presenciade Melanocoryphacalandra en el área deAracena según los modelosestadísticos (arriba) y los deexperto (se representa elexperto promedio; centro:criterio complejo; abajo:criterio simple). Laresolución espacial es de 50mts.


191

Fondo<20%20-40-60-80-

2000

metros

N

Parus caeruleusAracena

Fig.8. Mapas de laprobabilidad de presenciade Parus caeruleus en elárea de Aracena según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


192

Fondo<20%20-40-60-80-

2000

metros

N

Sitta europaeaAracena

Fig.9. Mapas de laprobabilidad de presenciade Sitta europaea en el áreade Aracena según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


193

Fondo<20%20-40-60-80-

2000

metros

N

SylviamelanocephalaAracena

Fig.10. Mapas de laprobabilidad de presenciade Sylvia melanocephala enel área de Aracena segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


194

Fondo<20%20-40-60-80-

2000

metros

N

TroglodytestroglodytesAracena

Fig.11. Mapas de laprobabilidad de presenciade Troglodytes troglodytesen el área de Aracena segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


195

Fondo<20%20-40-60-80-

2000

metros

N

Alectoris rufaGrazalema

Fig.12. Mapas de laprobabilidad de presenciade Alectoris rufa en el áreade Grazalema según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


196

Fondo<20%20-40-60-80-

2000

metros

N

CardueliscannabinaGrazalema

Fig.13. Mapas de laprobabilidad de presenciade Carduelis cannabina enel área de Grazalema segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


197

Fondo<20%20-40-60-80-

2000

metros

N

CerthiabrachydactylaGrazalema

Fig.14. Mapas de laprobabilidad de presenciade Certhia brachydactylaen el área de Grazalemasegún los modelosestadísticos (arriba) y los deexperto (se representa elexperto promedio; centro:criterio complejo; abajo:criterio simple). Laresolución espacial es de 50mts.


198

Fondo<20%20-40-60-80-

2000

metros

N

ErithacusrubeculaGrazalema

Fig.15. Mapas de laprobabilidad de presenciade Erithacus rubecula en elárea de Grazalema segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


199

Fondo<20%20-40-60-80-

2000

metros

N

Galerida theklaeGrazalema

Fig.16. Mapas de laprobabilidad de presenciade Galerida theklae en elárea de Grazalema segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


200

Fondo<20%20-40-60-80-

2000

metros

N

MelanocoryphacalandraGrazalema

Fig.17. Mapas de laprobabilidad de presenciade Melanocoryphacalandra en el área deGrazalema según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


201

Fondo<20%20-40-60-80-

2000

metros

N

Parus caeruleusGrazalema

Fig.18. Mapas de laprobabilidad de presenciade Parus caeruleus en elárea de Grazalema segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


202

Fondo<20%20-40-60-80-

2000

metros

N

Sitta europaeaGrazalema

Fig.19. Mapas de laprobabilidad de presenciade Sitta europaea en el áreade Grazalema según losmodelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


203

Fondo<20%20-40-60-80-

2000

metros

N

SylviamelanocephalaGrazalema

Fig.20. Mapas de laprobabilidad de presenciade Sylvia melanocephala enel área de Grazalema segúnlos modelos estadísticos(arriba) y los de experto (serepresenta el expertopromedio; centro: criteriocomplejo; abajo: criteriosimple). La resoluciónespacial es de 50 mts.


204

Fondo<20%20-40-60-80-

2000

metros

N

TroglodytestroglodytesGrazalema

Fig.21. Mapas de laprobabilidad de presenciade Troglodytes troglodytesen el área de Grazalemasegún los modelosestadísticos (arriba) y los deexperto (se representa elexperto promedio; centro:criterio complejo; abajo:criterio simple). Laresolución espacial es de 50mts.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CONCLUSIONES

SECCIÓN CUARTA

Esperanzas y desesperanzas de los modelos de distribución de especies

Facts are facts, but perception is reality:

—Conventional political wisdom

—E.J. Rykiel Jr., Relationships of scale to policy and

decision making. In: ECOLOGICAL SCALE. THEORY AND

APPLICATIONS (p.485). Columbia University Press. 1998.

¡Mírale!, y por eso le pagan...

—Alejandro (primavera de 1998), un amigo almeriense

cuyas palabras son el eco de la sociedad en la que

vivimos; un eco que yo no puedo –ni creo que deba–

quitarme de la cabeza.

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –CONCLUSIONES

206

CONCLUSIONES

CONCLUSIONES

- El modelado de la distribución de especies es una herramienta útil en numerosas aplicaciones

de la biología de la conservación (ver Introducción), aunque tanto su desarrollo como su puesta

en práctica cuentan con diversas limitaciones teóricas y prácticas (capítulo I).

- Los modelos empíricos estadísticos de la distribución de especies unidos a las herramientas de

SIG han demostrado ser una herramienta eficaz para la generación de cartografías de especies

en distintas situaciones de resolución espacial, abundancia de las especies y tipos de hábitats

que frecuentan, en los ambientes altamente heterogéneos típicos de la península Ibérica (sección

tercera).

- Los modelos estadísticos que se han desarrollado en esta tesis doctoral alcanzan una capacidad

predictiva aceptable, y sirven para generar mapas que facilitarán la comunicación entre los

investigadores y el público receptor (sección tercera).

- Parece existir un límite empírico máximo a la capacidad predictiva que alcanzan los modelos de

distribución. Tal límite podría estar determinado por procesos estocásticos de tipo de dinámica

de poblaciones, efectos históricos, influencia humana y factores ecológicos que no se han

considerado en su construcción, como los relativos a la competencia (intra e interespecífica) o

la predación (capítulo VI).

- Las especies difieren en cuanto a la capacidad predictiva que alcanzan sus modelos de

distribución, pero no hemos sido capaces de predecir en qué forma las características de cada

especie afectan a su suceptibilidad a ser modeladas (capítulos V y IX).

- Los modelos que aquí se han usado son de tipo empírico estadístico por lo que las relaciones

que muestran entre la aparición de las especies y las variables descriptoras del medio no son

causales (aunque se espera que bajo ellas subyazgan efectos causales, que se consideran muy

Conclusiones

207

difíciles de medir y cartografiar en áreas amplias) sino correlacionales. Esta es la causa más

problable de su bajo éxito al extrapolarse a otras áreas geográficas (capítulo III).

- La cartografía temática digital existente, que se ha desarrollado con otros fines distintos al

modelado de la distribución de especies, permite generar modelos de alta capacidad predictiva

aunque las variables explicativas que se derivan de ella son poco detalladas (capítulo IV). Su

resolución espacial (50x50 mts) es adecuada para los propósitos de modelado local (capítulos

IV y IX) y regional (capítulos VII y VIII).

- La cartografía temática no es satisfactoria para reflejar aspectos del territorio que se extienden

por una pequeña área, tales como los roquedos dispersos y las riberas, pero puede

complementarse en estos casos con datos de teledetección (por ejemplo, imágenes de satélite).

- Los modelos de capacidad predictiva más alta se obtienen con una combinación de la

cartografía temática con información topográfica, climática y la derivada de imágenes de

satélite (capítulo IV). Destacan por su importancia las variables que describen el paisaje en

torno a un punto de muestreo, lo que apoya la hipótesis de que las distintas especies seleccionan

sus áreas de campeo teniendo en cuenta las características de las áreas vecinas (capítulos V y

VI).

- Los modelos generados mediante un procedimiento muy automatizado (es decir, bajo una

selección de variables determinada por criterios estadísticos) alcanzan una capacidad predictiva

igual o superior a los que se generan con un protocolo supervisado, aunque existe una tendencia

a que sean menos extrapolables a otras áreas geográficas (capítulo III).

MODELOS PREDICTIVOS DE LA DISTRIBUCIÓN DE AVES TERRESTRES –AGRADECIMIENTOS

AGRADECIMIENTOS

AGRADECIMIENTOS

En primer lugar, querría agradecer a mis padres la confianza que han depositado en mí puessiempre me han apoyado aún sin comprender bien mis intereses y a qué me he dedicado estos años.A ellos les debo el haber podido dedicarme al lujo de estudiar primero e investigar después encuestiones relativas a la biología. Además, a Cristina, por todo el tiempo que no le he dado desdehace años y que siento como si yo se lo hubiera robado a nuestra juventud.

*** *** ***

Un trabajo como este, aún con sus limitaciones y errores, requiere una gran dedicación ybastante sacrificio. Sin embargo, todo el esfuerzo personal que le he dedicado no habría sidosuficiente para llevarlo a cabo de no ser por ciertas personas que quiero mencionar aquí en una listaque, me temo, está abocada a la injusticia de olvidar algún nombre.

Javier Bustamante, mi director de tesis, es el responsable de darme la oportunidad de trabajar enla Estación Biológica de Doñana (EBD) y de introducirme en la cartografía de especies. Él meproporcionó un entorno de trabajo envidiable al involucrarse profundamente en todos los aspectosde la investigación que dirigía y al estar siempre dispuesto a discutir nuevas ideas, remedar erroresy, en suma, a facilitar mi aprendizaje en materia de investigación científica. Lamento no haberlepodido corresponder con unos resultados más relevantes. Desafortunadamente, no he tratado misideas y ambiciones con mucha más gente (lo que debería ser requisito obligado para todo aquel querealizara una investigación de cualquier tipo). He podido disfrutar, sin embargo, de algunas perlasde Mario Díaz, Luisma Carrascal (que colaboraban en el proyecto del que surge esta tesis doctoral)y de Daniel López Huertas, con quienes, además, he compartido fructíferas e instructivas jornadasde campo. Ricardo Díaz-Delgado (¡aupa Atleti!) se encargó de las labores técnicas relativas al SIG,que había comenzado Manuel Ángel de la Puente, y se mostró incansablemente receptivo ante misdudas y problemas y pedagógico al resolverlos. Con Carlos Rodríguez, siempre amable, me unía ladirección, a veces el despacho y, sobre todo, un sentimiento semioculto de incertidumbre antenuestro futuro como investigadores. He aprendido mucho con todos estos compañeros de trabajo.

El Laboratorio de Teledetección de la Universidad de Valladolid y el Instituto Nacional deMeteorología aportaron amable y rápidamente datos básicos que fueron muy útiles en distintosanálisis (como se menciona en los capítulos correspondientes de esta tesis doctoral).

En un plano afectivo, importantísimo si —como yo— se desea conservar cierta cordura, deboagradecer a Julio Blas y a Sonia Cabezas que me hayan acogido estupendamente a mi llegada desdeMadrid y que con Iván Sánchez y Juan Quetglas me desvelaran las idiosincrasias de la vida en laEBD y en Sevilla. He tenido la inmensa suerte de vivir mucho tiempo con Carlos Alonso y JuanManuel Grande, sobrellevando los sabores y sinsabores de un precario investigador emigrado; dereencontrarme con Héctor Rodríguez, cuyas conversaciones siempre abrían una ventana al mundoexterior en mi cúpula de marfil; de conocer a Xim Cerdá y Jordi Figuerola, estimados compañeros aquien no sé si admiro más por su capacidad investigadora o por su inagotable manantial deinformación variada; y, en general, de disfrutar de un entorno grato de trabajo arropado, de una

Agradecimientos

209

manera indefinible, por el heterogéneo conjunto de becarios de la EBD (David Serrano, ElenaAngulo, Cristina Fuentes, José María Cani Fedriani, Marta Sánchez,...). Además, Frederique y KimJenkins suavizaron mis últimos dos meses de tesis con su paso fugaz pero amable por mi casa.Ojalá os vaya fetén a todos.

De mi cuatrienio en Sevilla guardaré también un recuerdo entrañable de Javier Balbontín y delresto de integrantes más o menos inconstantes del equipo “Calamaro” de futbito (Juan y JavierBalbontín, José caracoles Arrébola y José Sarasola, Joaquín Ayerbe, Marcelo Bertellotti, javieresde Morón y Troncoso, Antonio Manzaneda, Manolo...), con quienes atendía a la máxima de menssana in corpore magullado. Además, he aprendido y disfrutado mucho con las salidas al campo —menos frecuentes de lo que yo hubiera deseado— en las que Manolo Vázquez y Nicolás Varo (dela EBD), Jesús Fernández (de GOSUR) y, especialmente, Paco Chiclana, Javier Salcedo, JoséAntonio Lama, Jorge Garzón y Laura Plaza (de SEO-Sevilla), me hacían de cicerone por las bonitastierras andaluzas.

Quiero abrir un hueco aquí para mis antiguos compañeros de biológicas en la UnivesidadAutónoma de Madrid (UAM): Paco Martín-Azcárate, Juancho Calleja, Laura Arqueros, y losmiembros de SEO-Monticola (en especial Javier de la Puente y Ana Bermejo), con quienes crecíuna pasión por interpretar y disfrutar de la naturaleza, y para mis primeros mentores, Quico Suárezy Miguel Yanes (y también Juan Manrique) con quienes balbuceé los primeros sonidos científicos,finalmente inarticulados, en mis años de facultad.

Sospecho que el personal no investigador de un instituto de investigación como la EBD es eleterno olvidado en la sección de agradecimientos de cualquier trabajo. No quiero que sea este micaso, pues la incansable Conchita a la fotocopiadora, Mariángeles y Pedro en la biblioteca, Pepe encompras, Reyes en la secretaría y, en fin, el conjunto del personal de administración, fueronsiempre atentos ante mis peticiones de copias de artículos, material de trabajo, envío de correo ydemás entresijos oscuros pero necesarios en la actividad de un instituto como este. Además, Linallenó de humanidad los pasillos con sus saludos y ánimos para acabar el trabajo. Gracias a todos.

Por último, sería injusto no mencionar el soporte fomal de Braulio Asensio (de ANALITERS.A) y las ayudas económicas que el Ministerio de Educación y Ciencia (más tarde de Educación yCultura) y después el Consejo Superior de Investigaciones Científicas me han dado durante larealización de esta tesis. Pero vamos a dejar las cosas en su sitio: tales ayudas me dieron una ciertaestrecha independencia económica, y me permitieron realizar alguna estancia breve en el extranjeroy recibir gratuitamente los cursos de doctorado, sin embargo no contemplaban la cotización a laseguridad social (es decir, técnicamente no he estado trabajando y, por tanto, ni tuve derecho avacaciones antes ni tengo derecho a subsidio de desempleo ahora, entre otros inconvenientes) apesar de que el 1.8% de la mensualidad (entre 125 y 150.000 pts [750-900 euros] de 1998 a 2002)se pagaba como impuestos a a Hacienda (¡y, al menos en el 2002, la declaración del IRPF noresultaba negativa!). Siento, no obstante, estas becas como un privilegio no demasiado común.

This document was created with Win2PDF available at http://www.daneprairie.com.The unregistered version of Win2PDF is for evaluation or non-commercial use only.

http://www.daneprairie.com

cartografía predictiva de la distribución de aves

Documents