ORIGINAL ARTICLE

Nest-site preferences of Eleonora’s Falcon (Falco eleonorae)on uninhabited islets of the Aegean Sea using GIS and speciesdistribution models

Christina Kassara • Anastasios Dimalexis •

Jakob Fric • Georgios Karris • Christos Barboutis •

Spyros Sfenthourakis

Received: 28 March 2011 / Revised: 13 October 2011 / Accepted: 7 November 2011 / Published online: 22 November 2011

� Dt. Ornithologen-Gesellschaft e.V. 2011

Abstract Eleonora’s Falcon breeds colonially on small

islands of the Mediterranean Sea and Macaronesia. Despite

the wealth of papers highlighting the importance of nesting

characteristics on this species’ breeding performance, few

have addressed the issue of nest-site selection explicitly. In

this paper, we develop presence–absence and presence-

pseudoabsence models to predict nest occurrence as a

function of the topography of the nesting territory. Nest

occurrence data were available for nine uninhabited islets

of the Aegean Sea, within which the majority of the global

population of Eleanora’s Falcon is encountered. Our find-

ings suggest that the presence of conspecifics together with

certain topographic features according to the surface area

of the islet being studied can be used to predict nest

occurrence on uninhabited islets of the Aegean Sea. We

conclude that predictive models characterized by flexibility

and/or the use of absence data that also consider nest

clustering can result in robust predictions of the nest

occurrence of Eleonora’s Falcons in Greek breeding colo-

nies and eventually facilitate future monitoring schemes.

Since this is the first time nest-site preferences of Eleo-

nora’s Falcon have been analyzed using species distribu-

tion models, we encourage the application of similar

methodologies to other areas within the species’ breeding

range to further validate our findings.

Keywords Geographic information system (GIS) �Nest habitat � Spatial models � Autocorrelation �Logistic regression � Maximum entropy

Zusammenfassung

Untersuchungen zu Neststandort-Praferenzen von Ele-

onorenfalken (Falco eleonorae) auf unbewohnten klei-

nen Inseln der Agais unter Zuhilfenahme von GIS und

Artverteilungsmodellen

Eleonorenfalken sind Koloniebruter kleiner Inseln des

Mittelmeers und Makaronesiens. Trotz der Vielzahl an

Veroffentlichungen, die die Bedeutung der Nist-Charakte-

ristika fur den Bruterfolg dieser Art hervorheben, haben nur

wenige versucht, explizit die Frage nach der Nistplatzwahl

zu beantworten. In der vorliegenden Arbeit haben wir

Anwesenheits-Abwesenheits- und Nur-Anwesenheits-

Modelle entwickelt um das Nestvorkommen als eine Funk-

tion der Topographie des Brutreviers vorherzusagen. Von

neun unbewohnten kleinen Inseln der Agais, dem

Hauptverbreitungsgebiet der Weltpopulation, lagen uns

Daten zu den Nestvorkommen vor. Unsere Resultate legen

Communicated by T. Gottschalk.

All experiments and observations reported in this study comply with

the current laws of Greece.

C. Kassara (&) � S. Sfenthourakis

Section of Animal Biology, Department of Biology,

University of Patras, 26500 Patras, Greece

e-mail: ckassara@upatras.gr

A. Dimalexis � J. Fric

Hellenic Ornithological Society, 24 Vas. Herakliou,

10682 Athens, Greece

G. Karris

Department of Environmental Technology and Ecology,

Technological Educational Institution (TEI) of the Ionian

Islands, 2 Kalvos Square, 29100 Zakynthos, Greece

C. Barboutis

Department of Biology and Natural History Museum of Crete,

University of Crete, P.O. Box 2208, 71409 Herakleion,

Crete, Greece

123

J Ornithol (2012) 153:663–675

DOI 10.1007/s10336-011-0784-0

nahe, dass die Anwesenheit von Artgenossen zusammen mit

bestimmten topographischen Oberflachencharakteristika

verwendet werden kann, um die Nestvorkommen auf den

untersuchten unbewohnten Inselchen der Agais vorherzu-

sagen. Daraus schlussfolgern wir, dass flexible und/oder

Abwesenheits-Daten beinhaltende Vorhersagemodelle, die

außerdem ein mogliches Clustern von Nestern berucksich-

tigen, robuste Prognosen zu Nestvorkommen des Eleono-

renfalkens in griechischen Brutkolonien zulassen. Daruber

hinaus konnten sie zukunftige Monitoringplane erleichtern.

Da hiermit zum ersten Mal die Neststandort-Praferenz von

Eleonorenfalken mit Hilfe von Artverteilungsmodellen

analysiert wurde, rufen wir dazu auf, eine ahnliche Metho-

dologie in anderen Bereichen im Brutverbreitungsgebiet der

Art anzuwenden, nicht zuletzt um unsere Resultate weiter zu

untermauern.

Introduction

A key element in the study and management of biodiversity is

the knowledge of species distributions and the environmental

conditions that shape habitat suitability. Occurrence data are

usually sporadic, with many species exhibiting a widespread,

yet habitat-specific, pattern. Modern statistical approaches,

coupled with geographic information system (GIS) tools, have

enabled scientists to develop a variety of methodologies that

have lead to the prediction of species occurrence over broad

geographic ranges (Guisan and Zimmermann 2000).

GIS data sources, generally available free or at a low

cost, minimize sampling effort of environmental data,

whereas geostatistical tools provide rapid, automated, and

sophisticated modules for data preparation, analysis, and

map creation (De Frutos et al. 2007). However, although a

good knowledge of the species biology and ecology is

essential, so that the model-building process is based on a

robust and sound conceptual frame given the objectives of

the study (Guisan and Zimmermann 2000; Brotons et al.

2004), in many cases, the grain of the analysis is restricted

to that of the available GIS data, which do not necessarily

portray the species–environment relationships in sufficient

detail (Guisan et al. 2007; Gottschalk et al. 2011). For

example, the recent advent of high-resolution land-use

maps extracted from remote-sensing images has resulted in

predictive models of higher performance for 13 bird spe-

cies (Gottschalk et al. 2011). Thus, the creation of addi-

tional thematic layers, such as Digital Elevation Models, of

similar resolution could enhance our understanding of

species–habitat interactions in the near future.

Another important issue in species distribution models is

the presence of spatial structure in the model residual

errors, but mainly in the response variable being modelled

(i.e. species occurrence). Presence records are likely to be

positively correlated in space either due to exogenous

reasons, such as sampling bias and the spatial autocorre-

lation of environmental variables, or due to endogenous

reasons, such as aggregation patterns inherent to the spe-

cies biology (Legendre 1993; Dormann et al. 2007). In the

former case, spatial autocorrelation is usually addressed by

appropriate subsampling of the data (e.g. Graf et al. 2006;

Kaliontzopoulou et al. 2008), while in the latter case the

observed spatial pattern of species presence is incorporated

into the model with the inclusion of spatial terms (e.g. De

Frutos et al. 2007).

The aim of the current study was to reveal patterns of

nest-site selection by a colonial migratory raptor, Eleono-

ra’s Falcon (Falco eleonorae, Gene 1839), using GIS data

and commonly used predictive modelling techniques.

Eleonora’s Falcon breeds exclusively in the Mediterranean

basin, the Canary Islands and the northwestern coast of

Africa. The species delays egg-laying until late in the

summer in order to benefit from the abundant source of

food provided by migrating birds in the autumn, during the

nestling-raising period. In line with the guidelines cited in

the International Species Action Plan (Ristow 1999), the

Eleonora’s Falcon global population has been recently

reassessed and is now estimated at approximately 15,000

pairs of which Greece, and more specifically the Aegean

Sea, constitutes the centre of the species’ range, hosting

more than 80% of the total breeding pairs (12,300 pairs;

Dimalexis et al. 2008). Given the wide distribution of its

breeding colonies Eleonora’s Falcon is listed as ‘‘Least

Concern’’ in the IUCN 2010 list (Birdlife International

2010), but its highly uneven population concentration in

the Aegean Sea necessitates special attention. The typical

habitat of Eleonora’s Falcon colonies in Greece consists of

sea-cliffs and rocky uninhabited islets. The mean colony

size is 54 pairs (standard deviation [SD] 45; Dimalexis

et al. 2008), while the nests are mainly located in crevices,

under bushes and boulders, offering protection from

intense solar radiation and strong winds, factors which

have been proven to be crucial for breeding success across

its breeding colonies (Wink et al. 1982; Ristow and Wink

1985; Badami 1995; Bonnın 2004).

Despite the great number of publications focusing on the

breeding performance of Eleonora’s Falcon and the factors

that influence it, little is known about the criteria that this

species uses at the time of nest-site selection prior to the

breeding period. In particular, to the best of our knowledge

only one study has attempted to relate the nest distribution

pattern of Eleonora’s Falcon with environmental parame-

ters, although in a spatially restricted area (Urios and

Martınez-Abraın 2006). Moreover, previous studies have

demonstrated that the breeding pairs return to the same nest

in the following breeding period, a feature that seems to be

664 J Ornithol (2012) 153:663–675

123

more persistent in older pairs (Ristow et al. 1979), which is

indicative of an active nest-site selection process. Given

the species widespread distribution and the difficulties in

accessing its breeding colonies in terms of weather con-

ditions and geomorphology of the areas in question, pre-

dictive modelling can help in understanding the driving

forces of nest-site selection, thus providing a valuable tool

for future monitoring and management projects of Eleo-

nora’s Falcon breeding colonies.

To this end, we used nest occurrence records gathered

during a monitoring project implemented on uninhabited

islets in the Aegean Sea during the summers of 2004–2007

(Dimalexis et al. 2008). Our objective was to investigate

nest-site selection at the nesting territory level; conse-

quently, we focused on the topography of the nesting site

rather that on climatic or other environmental aspects that

typically vary at broader scales (Guisan and Thuiller 2005).

Until recently, species distribution models were classi-

fied into two broad classes, i.e. presence–absence and

presence-only models, depending on the availability of

absence records (for review, see Guisan and Zimmermann

2000). However, a more precise classification that takes

into account the treatment of absence data in the modelling

process recognizes four classes that differ mainly in the

estimation of potential or the realized distribution of the

species (Jimenez-Valverde et al. 2011): presence–absence,

presence–pseudoabsence, presence–background and pres-

ence-only methods. Presence-only, presence–background

and presence–pseudoabsence methods are currently

receiving special attention (Elith et al. 2006), since data on

absence are most often lacking or ambiguous in ecological

studies. If such methods can produce comparable results to

presence–absence methods, then they could constitute a

powerful tool in species conservation and management by

minimizing sampling effort in terms of time, logistics and

cost. For the ease of result interpretation, we chose to

compare the performance of two frequently cited methods,

one belonging to the presence–pseudoabsence class (com-

monly cited as a presence-only method) and one to the

presence–absence class, by modelling Eleonora’s Falcon

nest occurrence on islets for which both presence and

absence records are available. We also assessed the effect

of considering nest clustering in the model-building pro-

cess and discussed its role in the species’ breeding ecology.

Materials and methods

Nest records and topographic data

Fieldwork took place in 23 uninhabited islets of the Aegean

Sea during the summer periods between 2004 and 2007

(Dimalexis et al. 2008). In the case of 18 islets, not only

breeding parameters and characteristics of the nesting

environment were reported in the field protocols, but also

the geographic location of each nest. These geographic

coordinates formed the data pool used herein. After a

preliminary data screening that also took the availability of

topographic data and secure absence records into account,

we restricted the data pool to nine islets for the analyses

described below.

The nine islets in question lie between 35�190N and

39�120N and 23�270E and 26�480E (Fig. 1). Their substrate

is mainly formed by calcareous leptosols (Panagos and Van

Liedekerke 2004) and their elevation does not exceed

110 m a.s.l. During the breeding seasons of 2004–2007,

mean air temperature was 23.8�C and the prevailing winds

were in a northwestern direction, with a mean speed of

23.1 km/h at 2 m a.s.l., based on meteorological data

provided by the National Observatory of Athens.

We modelled nest occurrence as a function of the

topography of the nesting territory and used a Digital

Elevation Model (Hellenic Military Geographic Service,

http://www.gys.gr) to generate topographic parameters

(Table 1) with the Spatial Analyst toolbox in ArcGIS 9.2

(Esri 2006). Prior to the model-building process, we

superimposed a grid matrix of rectangular cells that mat-

ched the resolution size of the topographic parameters (i.e.

30 9 30 m) on maps illustrating the nest location on each

islet. Grid cells containing at least one nest during the

Fig. 1 The location of the islets used to model nest-site preferences

of Eleonora’s Falcon in the Aegean Sea. A North Sporades (6 islets),

B Kyklades (Cyclades, 1 islet), C Dodekanisa (Dodecanese, 1 islet),

D Crete (1 islet)

J Ornithol (2012) 153:663–675 665

123

course of the monitoring project were classified as occu-

pied, while the rest were classified as unoccupied, resulting

in 153 presence and 838 absence cells for model devel-

opment. A preliminary analysis showed that nest presence

differed among islets of substantial difference in surface

area (results not shown here). Thus, we split the data pool

into two groups (islets smaller or larger than 0.100 km2,

respectively) and proceeded with the analyses separately

for each islet group.

Species distribution models

Over the years a variety of models have emerged with the

aim of describing and predicting the current and future

distribution of the species as a function of environmental

predictors, which in turn are directly or indirectly linked to

the species presence in a given geographical region (Gui-

san and Zimmermann 2000). We chose two different

modelling approaches from the families of presence–

absence and presence–pseudoabsence methods to test their

performance in predicting Eleonora’s Falcon nest occur-

rence at the local scale.

Statistical analyses and model assessment were per-

formed in SPSS ver. 18.0 (SPSS Inc. 2009), model devel-

opment in R 2.10.1 � Development Core Team 2009) and

Maxent 3.3.1 (Phillips et al. 2004), whereas data prepara-

tion, spatial analyses and map construction were performed

in ArcGIS 9.2. Moran’s index statistics were calculated in

SAM v3.0 (Rangel et al. 2006), while the autocovariate

terms were built in R 2.10.1.

Presence–absence models

We developed presence–absence models using generalized

linear models (GLM; McCullagh and Nelder 1989) with a

binomial error distribution and a logit link function. In this

context, the probability of occurrence of an event (P) as a

function of a given set of explanatory variables is estimated

according to the formula (Eq. 1):

PðzÞ ¼ 1

1þ expðzÞ ð1Þ

where z = b0 ? b1x1 ? b2x2 ? ���bIxI, i.e. the sum of

contributions of all independent variables (xI) to the model.

By transforming the above equation into the natural log

of the odds ratio (OR) for success, the binomial probability

of the occurrence of an event can also be expressed as a

linear function of the independent variables as follows

(Eq. 2)

lnpI

1� pI

� �¼ b0 þ b1x1 þ � � � þ bixi: ð2Þ

GLMs offer greater modelling capabilities in comparison

to classic linear regression models, since they allow for the

inclusion of non-linear effects among variables and for

non-parametric distributions of the independent variables.

However, GLMs are sensitive to the presence of multi-

collinearity among the explanatory variables (Graham

2003), presence of outliers and lack of independence in the

errors, which are issues that should be addressed accord-

ingly (Fielding and Bell 1997).

Prior to GLM construction, we used Mann–Whitney

U tests to check for statistically significant differences

between occupied and unoccupied cells with regard to the

candidate topographic variables. We also investigated the

spatial pattern of Eleonora’s Falcon nest-site distribution

using Moran’s index (Moran’s I). We created correlo-

grams, in which Moran’s I values were plotted against

distances between localities (lags) in order to identify the

distance at which spatial autocorrelation was minimized

(i.e. observations were spatially independent). We tested

the statistical significance of Moran’s I values at different

lags by performing 999 Monte Carlo permutations on the

raw data set. In addition, we calculated Cook’s distance to

identify any outliers (i.e. Cook’s distance [1), in which

case they were removed from the analysis. Given the lack

of multicollinearity (i.e. variance inflation factor [VIF]

\10; Poirazidis et al. 2004), all available explanatory

variables were considered in the model-building process.

We fitted GLMs using forward stepwise logistic

regression (hereafter GLM), after a modification for vari-

able inclusion developed by Engler et al. (2004). This

technique resembles the typical forward stepwise proce-

dure, in which the model initially contains no variables and

variables are added sequentially until a final model is

obtained, but with priority given to those variables that are

most influential on the model performance based on the

reduction in residual deviance. In this way, the model-

Table 1 Description of the explanatory topographical variables used

for modeling nest-site preferences of Eleonora’s Falcon on uninhab-

ited islets of the Aegean Sea

Variable Description

Elev Ground elevation (m)

Slope Terrain slope (degrees)

Acos Cosine of terrain aspect, representing northness,a where

negative values correspond to southward-facing slopes

Asin Sine of terrain aspect, representing eastness,a where

negative values correspond to west-facing slopes

Curv Terrain curvature, where negative values represent

concave surfaces

Solar Mean incident solar radiation for August and September

2004–2007 (WH/m2)

Dist Distance to coastline (m)

a Poirazidis et al. (2004)

666 J Ornithol (2012) 153:663–675

123

building procedure is no longer sensitive to the order in

which variables enter the model; hence, under- or overfit-

ting is avoided (Pearce and Ferrier 2000).

After having verified the existence of spatial autocor-

relation, both in the response variable and in the residuals

of the nonspatial logistic regression models (GLM), we

investigated two alternative solutions commonly cited in

literature. In one case, we incorporated a third degree

polynomial of the central latitude and longitude of each

grid cell to the final logistic regression model (hereafter

GLMsp; Lichstein et al. 2002) and, in another case,

we included an autocovariate term (hereafter GLMar;

Augustın et al. 1996), estimated for each grid cell as the

average nest presence of its neighboring cells. The neigh-

bours’ effect was restricted up to that distance where spa-

tial autocorrelation of the nest occurrence was no longer

statistically significant. Therefore, for the large islet group,

the autocovariate term was calculated taking into account

49 neighboring grid cells (window size 7 9 7), while for

the small islet group, the autocovariate term was calculated

considering 25 neighboring grid cells (window size 5 9 5).

Presence-pseudoabsence model

We chose Maxent (Phillips et al. 2004), a niche-based

technique that has recently gained popularity, to predict the

habitat suitability for nesting for Eleonora’s Falcon on the

islets in question. Maxent resembles other statistical

modelling techniques, like GLMs, since it calculates the

unknown distribution of a species over a geographical

region of interest from a sample of localities of known

occurrence and spatially explicit environmental conditions.

This technique has been found to perform equally well or

even better than the presence-only and presence–absence

methods, especially when sample sizes are small (Elith

et al. 2006).

During the model-building process, presence data were

added as samples, while absence data were added as

background points (‘‘pseudoabsence’’). We retained the

default settings for model parameterization. Maxent was

run under the ‘‘auto features’’ mode, which allows for

maximum model flexibility (Phillips and Dudık 2008), and

logistic values were chosen as output (range 0–1) to gen-

erate habitat suitability values reflecting the estimated

probability of occurrence. The contribution of each pre-

dictor to the final model was assessed by a jack-knife

analysis of the training gain, which is a measure of the

likelihood of the training samples (Phillips et al. 2006).

Model comparison and evaluation

Model performance was compared by testing the agreement

between the observed and predicted nest occurrence by

means of two different threshold-independent measures;

namely the area under the curve (AUC) score and the biserial

point correlation coefficient (COR). We did not consider

threshold-dependent metrics that are typically derived by

classification matrices, since their performance is sensitive

to model quality and species prevalence (Freeman and Mo-

isen 2008). The two chosen accuracy measures were com-

puted by partitioning the data as follows. The training set,

used to calibrate the model, consisted of 75% of the available

data, while the test set, used to evaluate the final models,

consisted of the remaining 25% (Huberty 1994 in Fielding

and Bell 1997). Presence data were disproportionally less

abundant than absence data in each training and test set

(prevalence 0.376 and 0.105 for the small and large islet

group, respectively). Data partitioning was conducted ten

times at random, ensuring that prevalence remained the same

across the resulting data sets. Model development and

evaluation were conducted with the same training and test

data sets to ensure comparable results for Maxent and the

three GLMs. Altogether, 40 models were built for six small

islets and 40 for three large islets.

As a measure of model predictive power, we used the

area under the receiver operating characteristic (ROC)

curve (AUC; Hanley and McNeil 1982). A ROC curve is

created by plotting sensitivity against the omission rate

(1-sensitivity) at all possible thresholds of presence–

absence classification, and the area beneath the curve

corresponds to AUC. AUC ranges from 0.5, for models that

predict no better than random, to 1.0, for models with

perfect predictive power. However, being a rank-based

measure, AUC does not account for the degree to which the

predicted values have been calibrated. In contrast, the COR

(Elith et al. 2006) considers whether large differences in

predicted values correspond to large differences in the

probability of occurrence (Phillips and Dudık 2008).

Results

A total of 425 nests were monitored on the nine uninhab-

ited islets during the summers of 2004–2007, for which

topographic data and true absence records were available

(205 and 220 nests for the large and small islet group,

respectively). On average, one out of four nests had been

reoccupied by breeding pairs at least once in the preceding

years during the monitoring project (33.12 and 45.93% for

the large and small islet group, respectively). On both islet

groups, nest-site distribution was non-random (Moran’s

I [ 0, P B 0.05; Fig. 2); mean distance to the nearest nest

was 31.94 (SD 52.58), while nest density ranged from 1 to

13 per grid cell (mean 2.66, SD 2.07).

The topography of the nesting territory differed

according to the size of the islet being studied (Mann–

J Ornithol (2012) 153:663–675 667

123

Whitney U tests, P B 0.05; Table 2). On small islets, nests

were mainly found in the interior of the islets and at

locations of high inclination, as compared to the available

sites. On the contrary, on large islets, nests were located

closer to the coastline, in crevices of southeasterly facing

cliffs receiving greater amounts of solar radiation as

compared to the available sites.

Among the presence–absence models, the spatial models

accounting for the presence of neighbouring nests, GLMar,

fitted the data better than the remaining models according

to the Akaike’s information criterion (Table 3). In addition,

the incorporation of spatial terms, either as an autocovar-

iate term or as the nest geographic position, relaxed the

spatial structure of errors of the nonspatial GLMs (Fig. 3).

Taking into consideration the performance of both

presence–absence and presence–pseudoabsence models,

the predictive power was higher for the spatial presence–

absence models as assessed by the two measures of model

performance, AUC and COR (Table 4). As a rule of thumb,

AUC scores[0.75 suggest good predictive ability (Phillips

and Dudık 2008). Consequently, for the small islet group,

we consider the GLMsp and GLMar models to have

average predictive power, and GLM together with Maxent

to have poor predictive power. For the large islet group, the

Fig. 2 Moran’s I values at different distance intervals for the

presence of nests on small and large islets. A statistically significant

aggregation of nests was observed up to a distance of 135 m for the

small islets and up to 180 m for the large islets (P B 0.05)

Table 2 Differences with regard to topographic features between the means of occupied and unoccupied grid cells according to Mann–Whitney

U tests

Predictors Small islets Large islets

Occupied Unoccupied U Occupied Unoccupied U

Elev 13.796 (±10.988) 8.263 (±7.975) 2,694** 15.969 (±14.337) 28.334 (±27.992) 24,541**

Slope 13.412 (±7.672) 10.665 (±6.697) 3,155* 18.308 (±8.356) 15.080 (±9.652) 23,552**

Acos 0.078 (±0.743) 0.202 (±0.682) 3,523 -0.592 (±0.543) 0.031 (±0.658) 14,730**

Asin -0.009 (±0.676) 0.045 (±0.707) 3,703 -0.218 (±0.560) 0.053 (±0.752) 26,307*

Curv 0.849 (±1.716) 0.488 (±1.739) 3,375 -0.341 (±0.941) 0.208 (±1.294) 17,704**

Solar 4,176.044 (±298.183) 4,149.162 (±232.822) 3,515 4,355.862 (±296.591) 4,041.545 (±413.846) 14,071**

Dist 27.620 (23.557) 19.121 (17.190) 3,077* 31.861 (25.462) 63.662 (53.227) 20,293**

* Statistically significant at a = 0.05, ** statistically significant at a = 0.01. Standard deviations are given in parenthesis

Table 3 Model performance of three GLM models before (GLM) and after having accounted for the effect of nest clustering by either

considering the nest geographic position (GLMsp) or the presence of neighbouring nests (GLMar)

Models Total observations (presence) Adjusted-D2 Nagelkerke’s RN2 Akaike’s information criterion

Small islets

GLM 181 (68) 0.059 0.103 229.370

GLMsp 181 (68) 0.209 0.330 211.450

GLMar 181 (68) 0.144 0.236 211.220

Large islets

GLM 810 (85) 0.264 0.336 408.750

GLMsp 810 (85) 0.343 0.427 380.160

GLMar 810 (85) 0.350 0.432 363.870

GLM, Generalized linear model, i.e., nonspatial. GLMsp, model incorporating a third degree polynomial of the central latitude and longitude of

each grid cell to the final logistic regression model, i.e. considers nest-site geographic position; GLMar, model including an autocovariate

term,estimated for each grid cell as the average nest presence of its neighboring cells, i.e. presence of neighbouring nests

668 J Ornithol (2012) 153:663–675

123

AUC scores suggest relatively good model accuracy in all

cases.

The probabilities of nest occurrence were not clearly

discriminated among occupied and unoccupied cells, as

implied by the low COR values (Table 4); therefore, AUC

scores were mainly influenced by the omission rate, i.e.

false positive fraction. A closer look at the distribution of

probability values reveals that the discrimination of true

presences was relatively poor since almost half of these

cases were assigned to probability values of \0.5 (Fig. 4).

Nonetheless, mean predicted probability values for the

occupied cells were higher than those for the unoccupied

cells in all cases (Mann–Whitney U tests, P B 0.05).

According to the topographic predictors that were

entered in the final GLMs, the probability of nest occur-

rence increases with elevation on the small islets of the

Aegean Sea (Table 5), while on the large islets, nest-site

selection depends on the curvature of the terrain, the dis-

tance to the coastline and the incident solar radiation. More

specifically, the probability of nest occurrence is higher

close to the coastline and in concave surfaces receiving

increased solar radiation (Table 5).

For the small islet group, Maxent was in agreement with

the non-spatial GLM since it identified elevation as the

most important variable, followed by slope. For the large

islets, however, Maxent highlighted, in order of impor-

tance, terrain orientation along the north–south axis, solar

radiation and distance to the coastline as the most influ-

ential factors (Table 6). A visual inspection of the response

curves of the aforementioned explanatory variables

revealed the same tendency concerning the probability of

nest occurrence as a function of these variables.

Predictive maps of nest presence for all four models

(GLM, GLMsp, GLMar, Maxent), illustrating average

predicted values, are presented for one islet per islet group

in Figs. 5, 6.

Discussion

Predictors of nest occurrence of Eleonora’s Falcon

on uninhabited islets of the Aegean Sea

Eleonora’s Falcon is a common breeder in the Aegean Sea

and mostly found on uninhabited islets, although it also

frequents larger, inhabited islands (Dimalexis et al. 2008).

Previous studies have linked human presence to reduced

breeding performance of Eleonora’s Falcon (Ristow and

Wink 1985) and to nest-site preferences in an inhabited

island of western Mediterranean sea (Urios and Martınez-

Abraın 2006). By modelling nest occurrence on uninhab-

ited islets, we were able to investigate nest occurrence

patterns of Eleonora’s Falcons in a natural setting. Based

on our results, nest-site preferences of this raptor on

uninhabited islets of the Aegean Sea are related to, among

others things, the surrounding topography, while some

variability depending on the size of the islets in question is

also exhibited. Both modelling techniques (i.e. GLM and

Fig. 3 Spatial autocorrelation of the residuals in three generalized

linear models for the small islets (a) and the large islets (b). While the

nonspatial GLM (GLM) suffered from positively autocorrelated

errors, their spatial structure was successfully relaxed with the

addition of spatial terms, i.e. the GLMar and GLMsp models (for

explanation of models, see text and footnote of Table 3)

Table 4 Average model performance results as assessed by means of

two accuracy measures, the AUC score and the biserial point corre-

lation coefficient (COR), based on 10 test data sets

Models AUC COR

Small islets

GLM 0.638* (±0.070) 0.259* (±0.099)

GLMsp 0.733 (±0.052) 0.423* (±0.091)

GLMar 0.703 (±0.084) 0.403* (±0.134)

Maxent 0.573* (±0.091) 0.113 (±0.146)

Large islets

GLM 0.836 (±0.044) 0.472 (±0.069)

GLMsp 0.897 (±0.028) 0.494 (±0.076)

GLMar 0.899 (±0.023) 0.523 (±0.083)

Maxent 0.895 (±0.039) 0.512 (±0.058)

AUC, Area under the curve

Asterisk indicates that statistically non-significant values were cal-

culated for some test data sets at a = 0.05

J Ornithol (2012) 153:663–675 669

123

Maxent) distinguished elevation as a predictor of nest

presence on the small islets. Elevation at this small scale

means some distance from the coastline, making avoidance

of wave action an important factor in nest-site selection on

islets. The small overall surface of such small islets and

their relatively smooth relief permit a larger impact of

wave action on nests than on larger islands, especially

during the summer period when strong winds are common.

In the case of large islets, different models identified

different topographic variables as being most important,

with some overlap. According to GLM, inclination of the

terrain appears to be the most influential parameter in nest-

site selection, while distance to the coastline and incident

solar radiation are also important. Taking into account the

mostly calcareous substrate of the islets in question, sites

close to the coastline are vulnerable to wave action. At the

same time, plane terrain provides easier access to terrestrial

predators, such as rats, which are more abundant on larger

Fig. 4 The cumulative frequency distribution of the probability of nest occurrence for three GLMs (GLM, GLMar, GLMsp; see text and

footnote to Table 3 for abbreviations) and a niche-based method, Maxent for the small islet group (a) and the large islet group (b)

670 J Ornithol (2012) 153:663–675

123

islands. Therefore, sites that are located at an adequate

elevation above sea level constitute suitable nesting places

since they provide better protection from both the potential

drowning of nestlings and from predators. Alternatively,

Maxent identified terrain orientation along the north–south

axis as the most significant factor, followed by solar radi-

ation and distance to the coastline. Such a discrepancy

makes sense if we consider that Maxent focuses only on the

information conveyed by positive occurrence locations

and, therefore, disregards differences among occupied and

unoccupied cells.

The aforementioned predictors of nest occurrence, in

combination with climatic parameters, have also been

pinpointed as important factors of nest occurrence on a

western Mediterranean island (Urios and Martınez-Abraın

2006), as well as of breeding performance in various

Mediterranean colonies (Wink et al. 1982; Ristow and

Wink 1985; Badami 1995; Bonnın 2004). In particular,

elevated sites have been related to protection from wave

action. They have also been considered to provide advan-

tageous positions for take-off and landing at the colony

site, steep slopes to lower the risk of predation and concave

sites for sufficient visual protection from predators and

conspecifics (Walter 1979; Urios and Martınez-Abraın

2006). Preference in terrain orientation is inconsistent at

other breeding areas in the Mediterranean Sea, with some

researchers linking nest-site preference to the direction of

the incoming migrant flow (Mayol 1977), which is the

main food source for Eleonora’s Falcon during the young-

raising period, while others have linked it to protection

from intense sun irradiation and increased exposure to the

prevailing winds (Urios and Martınez-Abraın 2006). In our

case, there was no apparent association to the direction of

the migrant flow, but instead, taking into consideration the

mean wind direction during the summer period in the area,

nests located in southeasterly facing terrain were more

protected against wind. The observed preference for hotter

surfaces was highlighted in an earlier study on another

Greek Eleonora’s Falcon breeding colony (Walter 1979).

Due to the strong northern winds blowing in the study areas

during the summer period, hotter surfaces could counter-

balance the chilling effect of such wind conditions during

egg incubation. The availability of climatic data at such a

fine scale could shed more light into the nesting

requirements of the species (Urios and Martınez-Abraın

2006).

Influence of nest aggregation in model performance

Nest distribution on the nine islets we studied exhibited a

clustered pattern. Although during the breeding period

Eleonora’s Falcon feeds on migrant passerines (Walter

1979), in essence it is an insectivorous species. Its socia-

bility emerged due to the need to hunt insects in the win-

tering grounds, while colonial breeding appeared later in

evolutionary times (Ristow 2004). The presence of high

nest densities may also be indicative of a high local

availability of high-quality nest-sites or low high-quality

nest-site availability regionally, as has been reported for

other breeding colonies in Morocco (Walter 1979). In such

cases, tolerance to neighbours can be compensated by the

advantages offered by communal defence of the colony

from aerial trespassers, the efficiency of bird hunting in

groups during the nestling-raising period and/or increased

food availability (Walter 1979; Ristow et al. 1982; Rosen

et al. 1999).

Regardless of the underlying reasons and processes, nest

aggregation in raptors is an endogenous characteristic (De

Frutos et al. 2007) and, as such, it should be incorporated

explicitly in species distribution models (Augustin et al.

1996; Lichstein et al. 2002). Spatial autocorrelation was

accounted for only in the case of GLMs, with the inclusion

of either the geographic coordinates of occurrence locali-

ties or an autocovariate term to the final models. If the

autocovariate term exerts a higher impact on the species

distribution, then its inclusion will overshadow the effect of

the rest of the predictors (Dormann et al. 2007), as noted in

our study (results not shown). Another source of the spatial

Table 5 Model parameters of the final nonspatial GLM developed to

model Eleonora’s Falcon nest occurrence in the Aegean Sea

Predictors b SE P

Small islets

Intercept -1.178 0.248 0.000

Elev 0.062 0.017 0.000

Large islets

Intercept 35.074 9.844 0.000

Curv -0.293 0.110 0.008

Dist -0.023 0.005 0.000

Solar -0.022 0.004 0.000

Solar2 3.360 9 10-6 6.438 9 10-7 0.000

Table 6 Average percentage contribution of the topographic vari-

ables included in the final Maxent models predicting nest occurrence

of Eleonora’s Falcon in the Aegean Sea

Predictors Small islets Large islets

Acos 13.431 31.271

Asin 14.101 4.699

Curv 10.340 5.067

Dist 6.950 19.069

Elev 38.768 10.871

Slope 15.950 3.684

Solar 0.459 25.339

J Ornithol (2012) 153:663–675 671

123

structuring of species occurrence is spatial autocorrelation

in the environmental predictors, which is likely to cause

spatially correlated residuals (Dormann et al. 2007).

However, the addition of spatial terms removed the spatial

structure of model errors and improved the predictive

power of GLMs substantially, as has been also found in

similar studies (Araujo and Williams 2000; De Frutos et al.

2007). Despite the ongoing debate on its role in species

distribution modelling, up to now, few studies have con-

sidered spatial autocorrelation in maximum entropy models

(De Marco et al. 2008; Kaliontzopoulou et al. 2008). In

view of our findings, we believe that future studies

implementing spatially explicit maximum entropy models

could provide extremely useful information, especially

given the lack of true absence records.

Model comparison: presence–absence versus

presence-pseudoabsence models

The choice between presence–absence models and models

that do not take into account recorded absences has been the

subject of many studies, and different conclusions can be

drawn according to the study species. Overall, in our study,

GLMs performed slightly better than Maxent, as suggested

by the two metrics of accuracy used, namely, the AUC score

and the biserial point correlation coefficient (COR). This

result could be attributed to the fact that although Maxent

relies only on presence data, it can reconstruct complex

interactions between the response variable and the environ-

mental predictors that reflect species–environment rela-

tionships in a more realistic way. In other words, function

Fig. 5 Representative maps of

average predicted values of nest

occurrence for the small islet

group based on four models;

logistic regression (GLM),

logistic regression with

geographic coordinates as

spatial terms (GLMsp), logistic

regression with an autocovariate

term (GLMar) and Maxent

672 J Ornithol (2012) 153:663–675

123

complexity in Maxent is balanced against a more spherical

and accurate knowledge of the species’ actual distribution as

considered in GLM. In addition, presence–absence models

are more likely to describe the realized niche of the species in

question (Jimenez-Valverde et al. 2011), i.e. that part of the

suitable environmental space (fundamental niche) which is

available to the species (potential niche) given the ‘‘envi-

ronmental conditions, biotic interactions and dispersal

limitations’’ at that particular time. On the other hand,

presence–pseudoabsence models illustrate rather the poten-

tial niche of the species (Jimenez-Valverde et al. 2011).

Thus, in our case, Maxent predicted habitat suitability as it

would have been evaluated by those breeding pairs, being the

first to establish their nesting territory in a given colony,

while GLMs, and especially the spatial GLMs, GLMsp and

GLMar, portrayed the final nest distribution, having

accounted for the additional effect of conspecifics on the

nest-site selection of newcomers.

Two major assumptions in species distribution mod-

elling are related to the fact that the species being

modeled is considered to be in ‘‘equilibrium’’ with its

environment, i.e. the species is not invading new regions

nor does it occupy suboptimal habitats due to sudden

changes in habitat quality induced, for example, by cli-

mate change (Brotons et al. 2004; Elith and Leathwick

2009) and that species records have been sampled over

all possible environmental conditions within the species

range (Elith and Leathwick 2009). In such cases, absence

corresponds to low habitat suitability and, therefore,

improves model accuracy in presence–absence studies

(Hirzel et al. 2001). Given the high rate of philopatry at

these long-known colonies (Ristow et al. 1979), we are

quite confident that Eleonora’s Falcon is indeed in

equilibrium on the islets of interest. As for the second

assumption, we attempted to model nest-site preferences

based on the available data from a large geographical

region, the Aegean Sea, within which different geomor-

phological formations are observed. Yet, more studies

from the rest of the breeding regions are needed to

further corroborate our findings.

Fig. 6 Representative maps of

average predicted values of nest

occurrence for the large islet

group based on four models;

logistic regression (GLM),

logistic regression with

geographic coordinates as

spatial terms (GLMsp), logistic

regression with an autocovariate

term (GLMar) and Maxent

J Ornithol (2012) 153:663–675 673

123

Regardless of the modelling approach, previous studies

have also emphasized that the success in predicting species

distributions depends on the distribution pattern itself,

where more generalist species are less accurately modelled

than narrow-distributed species (Brotons et al. 2004).

Widespread species could either present regional variations

in habitat use (Segurado and Araujo 2004) or be influenced

by environmental factors operating at different scales

(Brotons et al. 2004). Again, in such cases presence–

absence models are expected to provide more accurate

results (Brotons et al. 2004).

To sum up, Eleonora’s Falcons tend to occupy the same

nest in consecutive years, suggesting an active selection

process. The presence of conspecifics and the topography of

the nesting territory proved to be important criteria at the

time of nest-site selection on uninhabited islets of the Aegean

Sea. Our results are in agreement with those reported in

previous studies, even in cases where human activities were

present (Urios and Martınez-Abraın 2006), suggesting that

humans and falcons can coexist harmonically as long as an

effective protection regime is established (Martınez-Abraın

et al. 2002). The comparison between GLMs and Maxent

showed that spatial models based on presence and absence

records and those allowing flexible relationships between the

response variable (i.e. nest occurrence) and the explanatory

predictors could facilitate monitoring projects. We believe

that future studies considering a wider geographical extent

and/or islands of different geomorphology within Eleonora’s

Falcon breeding distribution, as well as topographic data of

finer resolution, could further validate our findings. In this

case, more insight may be achieved if the effect of envi-

ronmental predictors on the nest-site preferences of this

raptor were to be considered non-stable across the region of

interest (Segurado and Araujo 2004; Dormann et al. 2007), as

well as in a multi-scale framework (Brotons et al. 2004).

Acknowledgments Data were collected in the framework of and

funded by the LIFE-Nature Project ‘‘Conservation measures for Falcoeleonorae in Greece’’ (LIFE 03NAT/GR/000091) coordinated by the

Hellenic Ornithological Society (HOS-Birdlife-Greece) and the

Leventis Foundation. We would like to thank Portolou Danae, Lat-

soudis Panagiotis, Bourdakis Stratis, Xirouchakis Stavros, Georgiaka-

kis Panagiotis and all field ornithologists, volunteers and boat captains

for their assistance in fieldwork. We also thank Thomas Gottschalk and

two anonymous referees for their fruitful comments on a previous

version of this manuscript. Special thanks are extended to Sinos Giokas

(University of Patras) for his advice on several statistical issues, Olga

Tzortzakaki for contributing the German translation of the abstract and

to Costas Lagouvardos (National Observatory of Athens) for the pro-

vision of meteorological data for the study areas.

References

Araujo MB, Williams PH (2000) Selecting areas for species

persistence using occurrence data. Biol Conserv 96:331–345

Augustın NH, Mugglestone MA, Buckland S (1996) An autologistic

model for the spatial distribution of wildlife. J Appl Ecol

33:339–347

Badami A (1995) Relazioni tra fattori ecologici e riproduzione nel

Falco della Regina (Falco eleonorae). Suppl Ric Biol Selvaggina

22:115–119

BirdLife International (2010) Species factsheet: Falco eleonorae.

Available at: http://www.birdlife.org. Accessed: 15 June 2010

Bonnın J (2004) Recompte I parametres reproductors de la poblacio

balear del Falco marı Falco eleonorae a l’any 2004. An Ornitol

Balears 19:1–9

Brotons L, Thuiller W, Araujo MB, Hirzel AH (2004) Presence-

absence versus presence-only modelling methods for predicting

bird habitat suitability. Ecography 27:437–448

De Frutos A, Olea PP, Vera R (2007) Analyzing and modelling spatial

distribution of summering lesser kestrel: the role of spatial

autocorrelation. Ecol Model 200:33–44

De Marco P, Diniz-Filho JAF Jr, Bini LM (2008) Spatial analysis

improves species distribution modelling during range expansion.

Biol Lett 4:577–580

Dimalexis A, Xirouchakis S, Portolou D, Latsoudis P, Karris G, Fric

J, Georgiakakis P, Barboutis C, Bourdakis S, Ivovic M et al

(2008) The status of Eleonora’s Falcon (Falco eleonorae) in

Greece. J Ornithol 149:23–30

Dormann CF, McPherson JM, Araujo MB, Bivand R, Bolliger J, Carl

G, Davies RG, Hirzel A, Jetz W, Kissling WD et al (2007)

Methods to account for spatial autocorrelation in the analysis of

species distributional data: a review. Ecography 30:609–628

Elith J, Leathwick JR (2009) Species distribution models: ecological

explanation and prediction across space and time. Annu Rev

Ecol Evol Syst 40:677–697

Elith J, Graham CH, Anderson RP, Dudık M, Ferrier S, Guisan A,

Hijmans RJ, Huettmann F, Leathwick JR, Lehmann A et al

(2006) Novel methods improve prediction of species’ distribu-

tions from occurrence data. Ecography 29:129–151

Engler R, Guisan A, Rechsteiner L (2004) An improved approach for

predicting the distribution of rare and endangered species from

occurrence and pseudo-absence data. J Appl Ecol 41:263–274

Esri (2006) ArcGIS desktop for Windows. Version 9.2. Esri, Redlands

Fielding AH, Bell JF (1997) A review of methods for the assessment

of prediction errors in conservation presence/absence models.

Environ Conserv 24:38–49

Freeman EA, Moisen GG (2008) A comparison of the performance of

threshold criteria for binary classification in terms of predicted

prevalence and kappa. Ecol Model 217:48–58

Gottschalk TK, Aue B, Hotes S, Ekschmitt K (2011) Influence of

grain size on species-habitat models. Ecol Model 222:3403–

3412

Graf RH, Bollmann K, Sachot S, Suter W, Bugmann H (2006) On the

generality of habitat distribution models: a case study of

capercaillie in three Swiss regions. Ecography 29:319–328

Graham MH (2003) Confronting multicollinearity in ecological

multiple regression. Ecology 84:2809–2815

Guisan A, Thuiller W (2005) Predicting species distribution: offering

more than simple habitat models. Ecol Lett 8:993–1009

Guisan A, Zimmermann NE (2000) Predictive habitat distribution

models in ecology. Ecol Model 135:147–186

Guisan A, Zimmermann NE, Elith J, Graham CH, Phillips S, Peterson

AT (2007) What matters for predicting the occurrences of trees:

techniques, data, or species’ characteristics? Ecol Monogr

77:615–630

Hanley JA, McNeil BJ (1982) The meaning and use of the area under

a Receiver Operating Characteristic (ROC) curve. Radiology

143:29–36

Hirzel AH, Helfer V, Metral F (2001) Assessing habitat-suitability

models with a virtual species. Ecol Model 145:111–121

674 J Ornithol (2012) 153:663–675

123

Jimenez-Valverde A, Peterson AT, Soberon J, Overton JM, Aragon P,

Lobo JM (2011) Use of niche models in invasive species risk

assessments. Biol Invasions. doi:10.1007/s10530-011-9963-4

Kaliontzopoulou A, Brito JC, Carretero MA, Larbes S, Harris DJ

(2008) Modelling the partially unknown distribution of wall

lizards (Podarcis) in North Africa: ecological affinities, potential

areas of occurrence, and methodological constraints. Can J Zool

86:992–1001

Legendre P (1993) Spatial autocorrelation: trouble or new paradigm?

Ecology 74:1659–1673

Lichstein JW, Simons TR, Shriner SA, Franzreb KE (2002) Spatial

autocorrelation and autoregressive models in ecology. Ecol

Monogr 72:445–463

Martınez-Abraın A, Oro D, Ferris V, Belenguer R (2002) Is growing

tourist activity affecting the distribution or number of pairs in a

small colony of the Eleonora’s Falcon? Anim Biodivers Conserv

25(2):47–51

Mayol J (1977) Estudios sobre el halcon de Eleonor, Falco eleonorae,

en las islas Baleares. Ardeola 23:103–136

McCullagh P, Nelder JA (1989) Generalized linear models. Chapman

and Hall, London

Panagos P, Van Liedekerke M (2004) The European soil database

(distribution version 2). Available at: http://eusoils.jrc.ec.

europa.eu/. Accessed 2 June 2009

Pearce J, Ferrier S (2000) An evaluation of alternative algorithms for

fitting species distribution models using logistic regression. Ecol

Model 128:127–147

Phillips SJ, Dudık M (2008) Modeling of species distributions with

Maxent: new extensions and a comprehensive evaluation.

Ecography 31:161–175

Phillips SJ, Dudık M, Schapire RE (2004) A maximum entropy

approach to species distribution modeling. In: Proc 21st Int Conf

Machine Learning, ACM Press, New York, pp 655–662

Phillips SJ, Anderson RP, Schapire RE (2006) Maximum entropy

modeling of species geographic distributions. Ecol Model

190:231–259

Poirazidis K, Goutner V, Skartsi T, Stamou G (2004) Modelling

nesting habitat as a conservation tool for the Eurasian Black

Vulture (Aegypius monachus) in Dadia National Reserve,

northeastern Greece. Biol Conserv 118:148–235

R Development Core Team (2009) R: a language and environment for

statistical computing. R Foundation for Statistical Computing,

Vienna, Austria. Available http://www.R-project.org

Rangel TFLVB, Diniz-Filho JAF, Bini LM (2006) Towards an

integrated computational tool for spatial analysis in macroecol-

ogy and biogeography. Glob Ecol Biogeogr 15:321–327

Ristow D (ed) (1999) International species action plan Eleonora’s

Falcon (Falco eleonorae). Birdlife International, Council of

Europe, Cambridge

Ristow D (2004) On the insect diet of Eleonora’s Falcon Falcoeleonorae and its importance for coloniality. In: Chancellor RD,

Meyburg B-U (eds) Raptors worldwide. World Working Group

on Birds of Prey/MME-Birdlife, London, pp 705–712

Ristow D, Wink M (1985) Breeding success and conservation

management of Eleonora’s Falcon. In: Newton I, Chancellor RD

(eds) Conservation studies on raptors. ICBP Technical publica-

tion no. 5. International Centre for Birds of Prey (ICBP),

Cambridge, pp 147–152

Ristow D, Wink C, Wink M (1979) Site tenacity and pair bond of the

Eleonora’s Falcon. Il-Merill 20:16–18

Ristow D, Wink C, Wink M (1982) Biology of Eleonora’s Falcon

(Falco eleonorae): 1. Individual and social defense behavior.

Raptor Res 16:65–70

Rosen M, Hedenstrom A, Badami A, Spina F, Akesson S (1999)

Hunting flight behaviour of the Eleonora’s Falcon Falcoeleonorae. J Avian Biol 30:342–350

Segurado P, Araujo MB (2004) An evaluation of methods for

modelling species distributions. J Biogeogr 31:1555–1568

SPSS Inc. (2009) SPSS Base 18.0 for Windows user’s guide. SPSS

Inc., Chicago

Urios G, Martınez-Abraın A (2006) The study of nest-site preferences

in Eleonora’s Falcon through digital terrain models on a western

Mediterranean island. J Ornithol 147:13–23

Walter H (1979) Eleonora’s Falcon: adaptations to prey and habitat in

a social raptor. University of Chicago Press, Chicago

Wink M, Wink C, Ristow D (1982) Biologie des Eleonorenflaken

(Falco eleonorae): 10. Der Einfluß der Horstlage auf den

Bruterfolg. J Ornithol 123:401–408

J Ornithol (2012) 153:663–675 675

123