Applied Vegetation Science 14 (2011) 256–267

Evaluating models to assess the distribution of Buxus

balearica in southern Spain
R. M. Navarro-Cerrillo, J. E. Hernández-Bermejo, R. Hernández-Clemente

Keywords Abstract
Environmental gradient; Habitat distribution
model; Mediterranean shrubs Question: Which is the best model to predict the habitat distribution of Buxus
balearica Lam. in southern Spain?
Abbreviations
AMTCM = mean minimum temperature of the Location: Málaga and Granada, Spain, across an area of 38 180 km2.
coldest month; AMTWM = mean maximum
Methods: Prediction models based on 17 environmental variables were tested.
temperature of the warmest month;
Six methods were compared: multivariate adaptive regression spline (MARS),
ASP = aspect; ATCM = mean temperature of
the coldest month; ATWM = mean maximum entropy approach to modelling species’ distributions (Maxent), two
temperature of the warmest month; generic algorithms based on environmental metrics dissimilarity (BIOCLIM
AUC = area under curve; COR = correlation and DOMAIN), Genetic Algorithm for Rule-set Prediction (GARP), and super-
coefficient; EDW = Euclidean distance to the vised learning methods based on generalized linear classifiers (support vector
closest drainage; GIS = Geographic machines, SVMs). To test the predictive power of the models we used the Kappa
Information System; GARP = Genetic
index.
Algorithm for Rule-set Prediction;
MARS = Multivariate Adaptative Regression Results: Maxent most accurately predicted the habitat distribution of B.
Spline; Maxent = maximum entropy approach balearica, followed by MARS models. The other models tested yielded lower
to modelling species; distributions;
accuracy values. A comparison of the predictive power of the models revealed
MPI = insolation models; SLP = slope;
that climate variables made the highest contributions among the environmen-
SVM = support vector machines; TP2 = spring
rainfall; TP3 = summer rainfall; TP4 = autumn tal variables studied. The variables that made the lowest contributions were the
rainfall insolation models. To examine the sensitivity of the models to a reduction in
the number of variables, a test showed that accuracy of over 0.90 was
Received 11 March 2010; maintained by applying just three climatic variables (spring rainfall, mean
Accepted 1 October 2010.
temperature of the warmest month, and mean temperature of the coldest
Co-ordinating Editor: Geoffrey Henebry
month). Maps derived from the algorithms of all models tested coincided well
with the known distribution of the species.
Navarro-Cerrillo, R. M. (corresponding author,
Conclusions: Model habitat prediction is a preliminary step towards high-
[email protected]) & Hernández-Clemente,
R.: Department of Forestry, School of Agriculture lighting areas of high habitat suitability of B. balearica. These data support the
and Forestry, University of Córdoba, Edf. results of previous research, which show that MaxEnt is the best technique for
Leonardo da Vinci, Campus de Rabanales s/n, modelling species distributions with small sample sizes.
Mail Box 3048, ES–14071 Córdoba, Spain
Hernández-Bermejo, J. E.: Department of
Agricultural and Forestry Science, School of
Agriculture and Forestry, University of Cordoba,
Avda 10 Edificio Mutis (C4 2a), Campus de
Rabanales, ES–14071 Cordoba, Spain)

Introduction mountains in the southeast Iberian Peninsula (Lázaro et

As a relict of the Tertiary flora, Buxus balearica is currently al. 2006). Two allopatric populations, which were initially
relegated to calcareous soils near sea level in the western considered a distinct species, are also known from Turkey.
Mediterranean. Its modern distribution includes the Ba- Its distribution in Andalusia of southern Spain consists of
learic Islands, Sardinia, northern Morocco and the coastal a few fragmented and isolated populations distributed

Applied Vegetation Science

256 Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science
R. M. Navarro-Cerrillo et al. Model distribution of Buxus balearica

between Malaga and Almeria (Blanca 1999). These popu- matic modelling approach (Hirzel et al. 2002; Anderson
lations have decreased substantially owing to both climate et al. 2003).
change in the Mediterranean region during the Holocene MARS has recently been applied to Mediterranean
(Yll et al. 1997) as well as human activity over the past species to predict their habitat distribution (Muñoz &
few centuries (Yll et al. 1997; Lázaro et al. 2006). The Felicı́simo 2004; Navarro Cerrillo et al. 2006). Classifica-
significant reduction in population has compelled the tion algorithms commonly used are CART and TREE NET,
IUCN to catalogue B. balearica as ‘vulnerable’, and it is which often have less accurate results than MARS
considered to be at risk of extinction in Andalusia (Blanca (Muñoz & Felicı́simo 2004; Segurado & Araujo 2004).
1999). Another group of techniques (e.g. BIOCLIM, HABITAT,
Evidence shows that environmental conditions in DOMAIN, ENFA) use generic algorithms based on envir-
southern Spain are suitable for flowering, fruiting, seed onmental dissimilarity metrics (Beaumont et al. 2005).
production, seed dispersal and germination of B. balearica More recently, a machine learning approach to predictive
(Lázaro et al. 2006). Summer drought, however, the most modelling has applied new algorithms such as GARP
critical factor affecting the regeneration of a number of (Genetic Algorithm for Rule-Set Prediction) (Stockwell &
Mediterranean species, poses an extreme restriction on its Peterson 2002; Anderson et al. 2003; Pape & Gaubert
seedling establishment. Wildfires and human pressures 2007) and Maxent (Maximum Entropy) (Phillips & Dubik
contribute to population fragmentation and local extinc- 2008), which have proven particularly successful in pre-
tion in the region. These conditions represent serious dicting potential species distributions under a wide vari-
obstacles to conservation programs for the species and ety of ecological situations.
highlight the need for repopulation activities. Distribution models are thus of paramount importance
Ecological niche modelling is important for a variety of when evaluating the conservation status of B. balearica,
applications in ecology and conservation as a result of including current levels of threat and protection, plan-
global change and the corresponding need to predict ning for at-risk populations, suitable sites for reintroduc-
range shifts caused by climate change (Thomas et al. tion, key areas for germplasm collection and
2004; Matsui et al. 2009), habitat fragmentation (Thuiller improvements in the assessment of risk status. Many
et al. 2004), and conservation risks to endangered species studies emphasize the necessity of incorporating distribu-
(Feldmeyer-Christe et al. 2007). As such, ecological niche tion data into conservation planning and quantifying
modelling has been the focus of increasing attention in species’ vulnerability to human disturbance (Araújo &
recent years (Guisan & Zimmermann 2000; Thomas et al. Williams 2000; Solano & Feria 2007). However, there is a
2004; Thuiller et al. 2004; Araújo et al. 2005; Soberón & lack of model evaluation for Mediterranean species for
Peterson 2005; Elith et al. 2006; Graham & Hijmans 2006; developing conservation strategies for species inhabiting
Guisan et al. 2007; Matsui et al. 2009) placing this human-dominated landscapes. It is essential that organi-
technique among emerging new approaches relevant to zations involved in the management of B. balearica have
ecology, biogeography, and conservation biology. the tools to enable them to identify suitable sites for
The basic approach of niche models is to combine a set conservation, seed collection, and reintroduction of the
of known incidence data (presence/absence) with predic- species.
tion variables such as topography, climate, edaphic con- Our study compares six different models for assessing
ditions and biogeography. Accurate incidence data are the habitat distribution of B. balearica in Andalusia. This
rarely available, however, especially for rare species and work provides a valuable step toward using habitat
inaccessible locations. Correlative models using species distribution models in B. balearica conservation under
presence and absence records for habitat predictions have varying levels of human-induced threat and global
been referred to as discrimination techniques, while those change impacts to set management priorities for this
using only species presence records have been referred to species.
as profile techniques (Jeschke & Strayer 2008). Examples
of discrimination techniques include models based on Methods
discriminant analysis (Rogers et al. 1996), generalized
Study area
linear models (GLM) (Cumming 2000), generalized addi-
tive models (GAM) (Leathwick & Whitehead 2001), The study was conducted in the provinces of Málaga and
Multivariate Adaptive Regression Spline (MARS) (Muñoz Granada, which contain almost all populations of B.
& Felicı́simo 2004) and decision-tree based methods balearica in the Iberian Peninsula with the exception of
(Araújo & Williams 2000). In contrast, climate envelope the Rágol site located in Almeria (361 and 381 N) (Fig. 1).
techniques (e.g. ANUCLIM, BIOCLIM, DOMAIN, FEM, The study area extends along the southern slope of the
HABITAT, and Mahalanobis distance) use a classic biocli- Almijara, Cázulas and Los Guájares mountains, part of

Applied Vegetation Science

Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science 257
Model distribution of Buxus balearica R. M. Navarro-Cerrillo et al.

which comprise the 40 600 ha Natural Park of the Sierras were selected to minimize the substantial environmental
of Tejeda, Almijara and Alhama. The study area was and geographic bias in the incidence dataset and to
restricted to basic, calcareous soils, mainly marls, distrib- sample all major relevant habitat types as uniformly as
uted over 38 180 km2. possible across the study area. Annual temperature, an-
nual rainfall, and soils were selected to represent broad
climate and substrate gradients across the study region.
Data assembly
Because physiography plays an important role in the
A total of 17 environmental layers were available for distribution of B. balearica (Lazaro et al. 2006), eight
analysis from the Environmental Information Network of topographic variables related to the species’ distribution
Andalusia (Consejerıa de Medio Ambiente http://www. at macro- and micro-scales were included. The overall
juntadeandalucia.es/medioambiente/site/web/rediam/) objective was to optimize each model by retaining the
(Table 1). Datasets containing relevant environmental largest number of presence and absence records while
predictors were available in Geographic Information Sys- minimizing the number of predictor variables.
tem (GIS) grid format at a resolution of 10 m  10 m and The topographical data, including aspect (ASP), slope
comprised eight climate layers, one soil attribute and (SLP), Euclidean distance to the closest drainage (EDW)
eight topographic variables. All environmental variables and insolation models (MPI), were derived from a Digital
were resampled according to the WGS84 geographical Elevation Model (Environment Department, Seville, ES).
coordinate system at a resolution of 10 m. Predictors Elevation was not considered because the species is

Fig. 1. Distribution of Buxus balearica in the study area of southern Spain.

Applied Vegetation Science

258 Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science
R. M. Navarro-Cerrillo et al. Model distribution of Buxus balearica

Table 1. Seventeen environmental predictors used in the models. Each predictor was available as a Geographic Information System (GIS) grid layer
encompassing the study area with a pixel resolution of 10 m  10 m. DEM: digital elevation model.
Description Code Units Source

Aspect ASP Degree Generated from DEM

Slope SLO Degree Generated from DEM
Euclidean distance to watershed EDW m Generated from DEM
Model of Potential Insolation, Solar Declination 22.51 MPI22.5 Generated from DEM
Model of Potential Insolation, Solar Declination 12.51 MPI12.5 Generated from DEM
Model of Potential Insolation, Solar Declination 01 MPI0 Generated from DEM
Model of Potential Insolation, Solar Declination 12.51 MPI-12.5 Generated from DEM
Model of Potential Insolation, Solar Declination 22.51 MPI-22.5 Generated from DEM
Total Precipitation (Jan–Mar) TP1 mm National Institute of meteorology (INM) (http://
www.marm.es)
Total Precipitation (Apr–Jun) TP2 mm National Institute of meteorology (INM) (http://
www.marm.es)
Total Precipitation (Jul–Sep) TP3 mm National Institute of meteorology (INM) (http://
www.marm.es)
Total Precipitation (Oct–Dec) TP4 mm National Institute of meteorology (INM) (http://
www.marm.es)
Average temperature of the warmest month ATWM 1C National Institute of meteorology (INM) (http://
www.marm.es)
Average maximum temperature of the warmest month AMTWM 1C National Institute of meteorology (INM) (http://
www.marm.es)
Average temperature of the coldest month ATCM 1C National Institute of meteorology (INM) (http://
www.marm.es)
Average minimum temperature of the coldest month AMTCM 1C National Institute of meteorology (INM) (http://
www.marm.es)
Lithology LITO Not applicable Department of Environment of the
Government of Andalusia

Table 2. Bivariate Pearson correlation matrix of the six spatial distribution models. Significant at the 0.01 level (two-tailed). Significant at the 0.001
level (two-tailed).
Bioclim Maxent MARS GARP SVM DOMAIN

Bioclim 1.000
Maxent 0.040 1.000
MARS 0.180 0.807 1.000
GARP 0.769 0.303 0.240 1.000
SVM 0.856 0.290 0.110 0.798 1.000
DOMAIN 0.856 0.290 0.110 0.798 1.000 1.000

uniformly distributed between approximately 0 and (AMTWM), mean temperature of the coldest month
800 m, according to elevation range registers within the (ATCM) and mean minimum temperature of the coldest
study area. Insolation was calculated with SHORTWAVC aml month (AMTCM) – were obtained from a network of 28
(Arc Macro Language) application (Felicı́simo et al. 2002). weather stations (National Institute of Meteorology, Ma-
Eight insolation models (one every 45 days throughout drid, ES, www.aemet.es). The time series data were from
the year) were calculated to estimate the total radiation 1956 to 2000. Meteorological variables were calculated on
received at the Earth’s surface over a period of time (Table a 100-m grid of the study area by interpolating data
2). Taking into account the preference of B. balearica for from each station using multiple regression (http://
areas of high ambient humidity, Euclidean distance to the www.aemet.es/, http://www.juntadeandalucia.es/med-
closest drainage area was also calculated. ioambiente/). Lithology (LITO) was obtained from a
Meteorological data – winter rainfall (TP1), spring rain- digital lithological map of the study area (scale: 1:
fall (TP2), summer rainfall (TP3), autumn rainfall (TP4), 100 000).
mean temperature of the warmest month (ATWM), mean The largest number of presence locations of B. balearica
maximum temperature of the warmest month was obtained from field surveys previously undertaken in

Applied Vegetation Science

Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science 259
Model distribution of Buxus balearica R. M. Navarro-Cerrillo et al.

the species’ range (GPSmap 60CSx, Garmin, Kansas City, environmental conditions to each point where the species
US, www.garmin.com). Additional presence records were has been recorded and selects the closest distance. In
extracted from an information database of botanical terms of potential distribution, the probability of species
surveys recorded by the University of Córdoba Herbar- occurrence can be interpreted as the probability of not
ium, together with incidental sightings recorded by forest finding any biophysical limitations to the species’ exis-
surveys. Geographic coordinates of each dataset were tence. Therefore, the greater the entropy of a system, the
determined from accompanying meta-datasets, or where more likely sites without such limitations will be encoun-
absent, obtained directly from the data provider (220 tered (Phillips & Dubik 2008).
presence data). Datasets were converted to spatial files, (c) Climate envelope techniques are based on ecological
projected to a common geographic coordinate system and niche theory. Climate envelopes can be defined as the
merged. Absence records were converted to spatial layers climatic component of the fundamental ecological niche,
in GIS, integrated based on a common projection system, or the ‘climatic niche’. Some bioclimatic models are based
and merged with presence records to form a single on empirical relationships between observed species dis-
presence and absence dataset. tributions and environmental variables (Peterson 2006).
These models interpolate a species bioclimatic envelope
Statistical models by integrating the environmental variables at locations
where the species has been recorded. Envelope models
Several approaches have been used to estimate species’
are range-based and describe a species’ environmental
ecological niche (Elith et al. 2006). In this study, six
envelope as a rectilinear volume, implying that a species
methods were compared:
can tolerate locations where values of all selected para-
(a) MARS provides an alternative regression-based meters fall within the extreme values determined by the
method for fitting non-linear responses by using piece- set of known locations. They fit a minimal envelope in a
wise linear fits rather than smooth functions (Friedman multidimensional climate space and use only presence
1991). It uses a stepwise addition/deletion strategy with data instead of presence/absence data. In this study, two
linear splines to produce a model capable of predicting the climate envelope models were applied: BIOCLIM and
value of a target variable (categorical or continuous) from DOMAIN.
a set of independent variables. To do this, MARS ap- (d) GARP (Stockwell & Noble 1992) is a computing
proaches the underlying function through a set of piece- system for delineating species niches and geographical
wise functions called basic functions (BF). The outcome is distributions (Anderson et al. 2003). The models describe
a sequence of candidate models with different BF num- environmental conditions under which a species should
bers, from which MARS selects the optimal model via be able to maintain populations. GARP uses a set of point
generalized cross-validation (GCV). The software pro- localities where the species is known to occur and a set of
gramme MARS 2.0 (http://www.salford-systems.com) geographic layers representing the environmental para-
was used for modelling. It enables the estimation of the meters that might limit the species’ ability to survive. It
importance of the independent variables by systemati- employs an iterative process of rule selection, testing and
cally dropping all terms including each variable and incorporation or rejection. Points of species occurrence
calculating the reduction of the goodness of fit. Finally, are evenly divided into training and test data. Four types
MARS references these values to the most important of rules – atomic, logistic regression, bioclimatic envelope
variables as a percentage. and negated bioclimatic envelope – are applied to the
(b) The maximum entropy approach to modelling spe- training data and a rule is developed to maximize pre-
cies’ distributions (Maxent) estimates a species’ distribu- dictivity (Anderson et al. 2003). The change in predictive
tion by finding the probability distribution with accuracy of a rule from one iteration to the next is used to
maximum entropy, subject to constraints that represent evaluate whether a particular rule should be incorporated
incomplete information about the target distribution into the model, and the algorithm is run until conver-
(Phillips 2006; Phillips et al. 2006). The mathematical gence.
approach is based on the estimation of a probability (e) Support vector machines (SVMs) are linear classifiers
function that incorporates a species’ bioclimatic and (Vapnik 2000) in the data space where a maximum
environmental limits while maintaining its distribution hyperplane separating positive from negative samples is
and maximum entropy (Phillips et al. 2006). Maxent constructed. SVMs can perform binary classification (pat-
requires presence data but not absence data, as well as tern recognition) and real valued function approximation
environmental data for the whole study area, making it (regression estimation). A special property of SVMs is that
well suited to modelling species distributions. The Maxent they minimize the empirical classification error and max-
algorithm calculates the distance between the given imize the geometric margin; hence they are also known as

Applied Vegetation Science

260 Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science
R. M. Navarro-Cerrillo et al. Model distribution of Buxus balearica

maximum margin classifiers. The model depends on a Results

subset of the training data because the cost function for
The AUC scores were 0.97 for Maxent and 0.95 for MARS
building the model excludes training points that lie
(Fig. 2). A clearer difference was obtained with the kappa
beyond the margin. Intuitively, this makes the classifica-
coefficient (k = 0.81 and k = 0.66, respectively) and the
tion correct for testing data that are similar but not
correlation coefficient (COR = 0.99 and COR = 0.87, re-
identical to the training data (Vapnik 2000).
spectively) (Fig. 2). The other four models tested pro-
Variable reduction was carried out by backwards selec-
duced lower accuracy values. The lowest accuracy using
tion using goodness of fit criteria measured by deviance
the kappa coefficient was ranked by DOMAIN.
reduction and based on the chi-square test of significance
The highest correlations were obtained in two separate
(P o 0.01). The final set of variables was then tested for
groups. First, BIOCLIM exhibited high correlation with
collinearity to ensure that no retained variables were
both DOMAIN and SVM (r = 0.856, Table 2). Second,
highly correlated (i.e. r2Z0.85). Models were run using
Maxent and MARS were highly correlated (r = 0.807).
the optional jackknife test. Variables contributing least to
Although Maxent and MARS follow different computa-
the model fit were removed until a maximum of eight
tional routines, both produced similar distribution mod-
variables were included or a variable’s contribution was
els. The relative contributions of the environmental
o 2%. Once the six models were evaluated, Maxent
variables were also similar.
models were developed using the same presence/absence
The highest contributions were made by the climate
dataset. A jack-knife test was performed to examine the
variables: spring rainfall (TP2), mean temperature of the
sensitivity of the models to the reduction in the number of
warmest month (ATWM), and mean temperature of the
variables and presence input data.
coldest month (ATCM) with proportional weights (Fig. 3).
The variables with lowest contributions were the insola-
Model evaluation tion models. The environmental variable with the highest
To assess the agreement between the incidence data and gain when used alone was ATCM, which therefore
environmental predictors, three statistics were used: area appeared to have the most useful information by itself.
under curve (AUC), correlation coefficient (COR) and the The environmental variable that decreases the gain the
maximum kappa index (k). The AUC ranges between 0 most when it was omitted was ATCM, which therefore
and 1. Cohen’s kappa (k) is one measurement that can be appears to have the largest amount of information that
derived from the confusion matrix. A confusion matrix was not present in the other variables (Fig. 4).
contains information about actual and predicted values Once Maxent was identified as the most accurate model
done by a classification system (Kohavi & Provost 1998). based on kappa and correlation coefficients, a test of
As a validation tool (i.e. when the ‘truth’ is known), it sensitivity to the reduction in the number of variables and
states the overall accuracy of a prediction once the number of presence input data was performed. Figure 5
element of chance has been removed (Liu et al. 2005). shows the curve described by the AUC depending on the
Otherwise, kappa can serve as a tool to assess reliability of number of variables introduced in the analysis. The num-
prediction in terms of relative agreement; a value near 0 ber of variables introduced in each model is explained in
indicates no discrimination (agreement by chance), a Table 3. Results show a decreasing tendency of the AUC
value of 1 represents perfect discrimination (agreement), with the reduction of the number of variables. However,
a value 4 0.6 is considered ‘good’ and a value 4 0.8 as
1
‘excellent’ (Graham & Hijmans 2006). Kappa is relatively
0.9
tolerant to zeros in the confusion matrix and considers
0.8
Coefficient values

both omission and commission errors in one parameter.

0.7
Finally, the concordance among models was evaluated
0.6
through a correlation analysis. 0.5
0.4
Distribution maps 0.3
0.2
The models generate a cumulative probability distribution
0.1
output between 0 and 100%; we selected a decision
0
threshold to distinguish presence from absence, enabling
validation and visual interpretation of model predictions Models
(Liu et al. 2005; Allouche et al. 2006). The potential Fig. 2. Accuracy values obtained by the application of six different
distribution of B. balearica was mapped based on this spatial distribution models: BIOCLIM (BIO), GARP, DOMAIN, SVM, MARS
threshold value. and Maxent (see text).

Applied Vegetation Science

Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science 261
Model distribution of Buxus balearica R. M. Navarro-Cerrillo et al.

a 25 Maxent model. The agreement of area of spatial predic-

tions varied among models. Considering MARS and Max-
Relative contributions (%)

20 ent models, approximately 22% of the area of spatial

predictions in the Maxent model was also predicted as
15 habitat in the MARS model. Only 0.08% of spatial
predictions in the GARP model are common with the
Maxent model showing that areas of habitat predicted in
10
the GARP were overestimated.

5
Discussion

0 Potential distribution model of B. balearica

Predictive habitat models have been evaluated for B.
balearica, a Mediterranean vulnerable species that has a
Environmental variables
restricted and fragmented distribution. A reasonable de-
b 25 gree of agreement in spatial predictions of suitable habitat
was achieved among the models employed in our study.
Relative contributions (%)

20 Environmental variables based on meteorological data

made the highest contribution to the potential distribu-
15 tion defined by temperature of the warmest month
(ATWM), spring precipitation (TP2) and temperature of
the coldest month (ATCM). The single variable making
10
the greatest contribution to defining the potential distri-
bution of the species was ATWM. Buxus balearica limits
5
itself to areas with average temperatures ranging from
24.61C to 25.41C during the warmest month (Fig. 2).
0 Moreover, maximum temperature of the warmest month
(AMTWM) indicates a species tolerance between 311C
Environmental variables and 321C. As a consequence, a climate change scenario of
one degree increment could greatly diminish the survival
Fig. 3. Relative contributions of the environmental variables to the (a)
possibilities of this species. It is also clear that B. balearica
Maxent and (b) MARS model.
has a low frost tolerance, as it disappears in areas where
ATCM is below 71C and minimum temperature of the
the model obtains an accuracy of over 0.90 by applying just coldest month (AMTCM) is below 31C.
three variables (TP2, ATWM and ATCM). In contrast, the Seasonal precipitation also had a large impact on the
reduction in the number of presence input data shows a distribution model, measured by total trimester precipita-
greater sensitivity (Fig. 6). The maximum number of points tion. Spring rainfall (TP2) revealed the sensitivity of B.
taken into the analysis reaches high values in the accuracy balearica to climatic irregularity during the spring and
assessment of both Maxent and MARS. A reduction of summer (Fig. 3). Buxus balearica would be restricted to
50% in the number of presence data shows a reduction in areas with annual precipitation above 500 mm, with
kappa coefficient around 14.5% for MARS and 6.7% for spring precipitation above 110 mm and summer rainfall
Maxent. Comparing both models, MARS is more sensitive (TP3) at least 25 mm. However, this interpretation may be
to the reduction in the number of presence input data. incomplete, considering the lack of data concerning cryp-
Spatial predictions of habitat revealed a consistent to-precipitations that may compensate for low rainfall.
pattern of predicted habitat probability among models Conversely, B. balearica distribution model was less af-
across the study region, with the largest areas of prime fected by topographic variables (insolation, aspect and slope).
habitat located in the mountains range parallel to the Euclidean distance to nearest drainage, based on a watershed
coastal line (see Maxent, MARS, GARP and DOMAIN, model, was an influential factor in determining the presence
Fig. 7). A second large area of lower probability predicted of the species. Buxus balearica specimens appear to take
habitat occurs around this coastal range in the south of refuge in streams and rivers, not so much because of subsoil
the study region (Fig. 7). The total area predicted as moisture but more likely because it is in such areas that the
optimal suitable habitat (habitat probability 4 0.7) ran- species is subject to less radiation exposure, fire risk and
ged from 22 333 ha in the GARP model to 1870 ha in the seedling mortality owing to summer drought.

Applied Vegetation Science

262 Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science
R. M. Navarro-Cerrillo et al. Model distribution of Buxus balearica

1.6

1.4

Regularized training gain

1.2

0.8

0.6

0.4

0.2

0
W
EM

M 2.5

M 2.5
M 2.5

5
P

4
O

AM M
AM CM

l
ta
2.

TP

TP

TP

TP
AS

SL

TW
ED

to
1

1
-2
D

AT

T
AT
PI

PI

PI
PI
M

Environmental variables

Fig. 4. Jack-knife test of variable importance in the Maxent model (tinyed = one variable, closed = all variables, diagonally hatched = without variables).

1
Area under curve (AUC)

0.95

0.9

0.85

0.8
Mod_01 Mod_03 Mod_05 Mod_07 Mod_09 Mod_11 Mod_13 Mod_15 Mod_17 Mod_19
Maxent models

Fig. 5. Sensitivity analysis of the Maxent model based on the reduction in the number of variables (see models in Table 3).

Model evaluation computational efficiency. They can be implemented in

common free software like the open modeller (http://
Results of the predictive habitat models evaluated for B. openmodeller.sourceforge.net), and they use identical in-
balearica agree with those obtained from the application of put data. The user-friendly interface and the fast computa-
the same models to other plant species distribution (Elith tion of the algorithm may prove the usefulness of these
et al. 2006; Austin 2007). The generics algorithms based models at a regional and global scale. Nevertheless, at a
on environmental dissimilarity (BIOCLIM, DOMAIN, more detail scale, the results show low accuracy values
SVC and GARP) obtain poorer results compared with and poor prediction maps. Maxent and MARS significantly
those obtained by the machine learning approach (Max- improves all the predictions. Although Maxent shares all
ent and MARS). The accuracy of Maxent was higher than the computational advantages of generic algorithms,
all the prediction models evaluated (Fig. 2). The computa- MARS presents some drawbacks that require a specific
tional efficiency of each model should be evaluated con- informatics’ format (.sav) and the need to process all the
sidering three main aspects: pre-processing required for information with a GIS.
the input data; processing efficiency of the model itself; Maxent not only exhibits a high computational effi-
and post-processing required to acquire statistical and ciency, it also produces an extensive statistical report of
cartographical prediction. Generic algorithm models the model. In order to test the sensitivity of the model,
based on BIOCLIM, DOMAIN, SVC or GARP are of a great different analyses were performed. The first one

Applied Vegetation Science

Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science 263
Model distribution of Buxus balearica R. M. Navarro-Cerrillo et al.

Table 3. Sensitivity analysis of the Maxent model. From right to left, the table shows the variables introduced following this order, the code assigned to
the model, and the relative importance in terms of the model fit of the last variable introduced. Variable codes are listed in Table 1.

Model Code Importance

ATWM Mod_01 24
ATWM, ATCM Mod_02 12.9
ATWM, ATCM, TP2 Mod_03 12.2
ATWM, ATCM, TP2, AMTCM Mod_04 7.8
ATWM, ATCM, TP2, AMTCM, MPI-22.5 Mod_05 5.8
ATWM, ATCM, TP2, AMTCM, MPI-22.5 Mod_06 5.6
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3 Mod_07 5.3
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, Mod_08 5.2
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4 Mod_09 5.1
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP Mod_10 4.7
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM Mod_11 3.8
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM, EDW Mod_12 3.5
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM, EDW, TP1 Mod_13 2.1
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM, EDW, TP1, MPI0 Mod_14 0.7
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM, EDW, TP1, MPI0, MPI-12.5 Mod_15 0.7
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM, EDW, TP1, MPI0, MPI-12.5, ASP Mod_16 0.4
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM, EDW, TP1, MPI0, MPI-12.5, ASP, MPI112.5 Mod_17 0.1
ATWM, ATCM, TP2, AMTCM, MPI-22.5, TP3, TP4, TP, AMTWM, EDW, TP1, MPI0, MPI-12.5, ASP, MPI112.5, MPI122.5 Mod_18 0.1

a 100

80

Coefficient values
60

40
correlation coefficient (COR) kappa auc
20

0
250 200 150 100 50 0
Number of presence data

100
b

80
Coefficient values

60

40

%correct kappa auc 20

0
250 200 150 100 50 0
Number of presence data

Fig. 6. Accuracy functions of the Maxent (a) and MARS (b) models based on the number of presence data.

Applied Vegetation Science

264 Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science
R. M. Navarro-Cerrillo et al. Model distribution of Buxus balearica

Fig. 7. Distribution of predicted potential Buxux balearica habitat in Maxent, MARS, GARP and DOMAIN models (grey scale = habitat probability,
points = current record of the species).

consisted of the reduction of the number of prediction identify as suitable at least some of the ‘new’ environ-
variables. The high accuracy obtained with just three mental space if the conditions are closer to the species
climate variables shows the strong dependency of the habitat as well as the conditions under which it is absent
potential habitat of the species studied on the climatol- (Hoffman et al. 2010).
ogy conditions (Fig. 6). It is particularly relevant for A recent compilation of studies on the effect of pro-
habitat assessment, management and conservation of B. jected climate change indicates that an alarming number
balearica. When reducing the number of presence data, of species may lose a large part of their range and become
compared with the MARS model, accuracy reduction ‘committed to extinction’ (Thomas et al. 2004). There are
was significantly lower with Maxent. This last result is some obvious cases of species that with climate change
also very important when presence data are relatively should lose parts of their range, such as mountain-top
limited. Maxent uses presence and absence (or random endemics, for which warming would seem highly threa-
background) data. This likely makes it able to correctly tening (Theurillat & Guisan 2001). The output of Maxent

Applied Vegetation Science

Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science 265
Model distribution of Buxus balearica R. M. Navarro-Cerrillo et al.

allowed us to estimate the potential range of B. balearica, providing the herbarium data. We thank Antonio Pulido
as would be expected given previously reported Pastor for their facilities for data collection in the Parque
variation in such predictions (Thuiller 2004; Hoffman Natural Sierra de Tejeda, Almijara y Alhama (Málaga-
et al. 2010). This model could be used to predict the Granada, Spain) and Daniel Griffith for the English review.
effects of climate change on species distributions (Araújo We are also grateful to two anonymous referees for critical
et al. 2005). However, progress in using Maxent to predict reading and helpful suggestions on an earlier version of the
the effect of climate change on B. balearica distributions manuscript.
should be include: (1) improving data for predicting
current distributions; (2) evaluating the ability of Maxent
References
to provide accurate estimates of the effect of climate
change by including data of proposed scenarios; (3) Allouche, O., Tsoar, A. & Kadmon, R. 2006. Assessing the
increasing understanding of the drivers of species distri- accuracy of species distribution models: prevalence, kappa
butions, and the extent to which these are directly related and the true skill statistic (TSS). Journal of Applied Ecology
to individual climatic variables; and (4) integrating Max- 43: 1223–1232.
ent on conservation frameworks. Anderson, R.P., Lew, D. & Peterson, A.T. 2003. ‘Evaluating
predictive models of species’ distributions: criteria for
selecting optimal models. Ecological Modelling 162: 211–232.
Conclusion
Araújo, M.B. & Williams, P.H. 2000. Selecting areas for species
Effective conservation planning for Mediterranean land- persistence using occurrence data. Biological Conservation
scapes needs to address issues of habitat prediction 96: 331–345.
through model analysis. This study has shown that spatial Araújo, M.B., Pearson, R.G., Thuiller, W. & Erhard, M. 2005.
predictions generated from habitat models developed Validation of species–climate impact models under climate
from compiled ecological data are sufficiently accurate change. Global Change Biology 11: 1504–1513.
for use in conservation planning of a fragmented rare Austin, M. 2007. Species distribution models and ecological
species. Establishing conservation areas, restoration pro- theory: a critical assessment and some possible new
grammes, priority areas for seed collections, etc., are approaches. Ecological Modelling 1–2: 1–19.
especially significant for shrubs Mediterranean species Beaumont, L.J., Hughes, L. & Poulsen, M. 2005. Predicting
species distributions: use of climatic parameters in BIOCLIM
such as B. balearica which, owing to environmental
and its impact on predictions of species’ current and future
restrictions and human impacts, will not cover large
distributions. Ecological Modelling 2: 251–270.
areas. Buxus balearica populations have been damage by
Blanca, G. (Ed.) 1999. Libro Rojo de la flora silvestre amenazada de
forest fires, commercial and residential development and
Andalucı́a. Tomo I: especies en peligro de extinción. Consejerı́a
overgrazing, suggesting failed attempts to protect relictic
de Medio Ambiente, Junta de Andalucı́a, Sevilla, ES.
locations. A challenging issue facing organizations in- Cumming, G.S. 2000. Using between-model comparisons to
volved in the conservation of B. balearica is locating fine-tune linear models of species ranges. Journal of
suitable sites for seed collection and reintroduction of Biogeography 27: 441–455.
individuals. The model prediction is a preliminary step Elith, J., Graham, C.H., Anderson, R.P., Dudik, M., Ferrier, S.,
towards highlighting areas of high habitat suitability, Guisan, A., Hijmans, R.J., Huettmann, F., Leathwick, J.R.,
which may also be effective reintroduction zones that Lehmann, A., Li, J., Lohmann, L.G., Loiselle, B.A., Manion,
minimize anthropogenic threats for new populations. G., Moritz, C., Nakamura, M., Nakazawa, Y., Overton,
Furthermore, the variables identified as better predic- J.Mc.C., Peterson, A.T., Phillips, S.J., Richardson, K.S.,
tors of B. balearica habitat are linked with global warming, Scachetti-Pereira, R., Schapire, R.E., Soberón, J., Williams,
which will permit the development of models of global S., Wisz, M.S. & Zimmermann, N.E. 2006. Novel methods
change impacts on natural populations of this species. improve prediction of species’ distributions from occurrence
Maxent, evaluated on independent presence/absence data. Ecography 29: 129–151.
test, was shown to be the best model for modelling species Feldmeyer-Christe, E., Ecker, K., Küchler, M., Graf, U. &
distributions in this context. Waser, L. 2007. Improving predictive mapping in Swiss
mire ecosystems through re-calibration of indicator values.
Applied Vegetation Science 10: 183–192.
Acknowledgements
Felicı́simo, A.M., Francés, E., Fernández, J.M., González-Dı́ez, A.
This study was supported by the coordinated Spanish & Varas, J. 2002. Modeling the potential distribution of
MEC project INTERBOS (CGL2008-04503-CO3-02) and forests with a GIS. Photogrammetric Engineering & Remote
with financial support from the Environmental Depart- Sensing 5: 455–461.
ment of Andalucia Government. We thank the COA Friedman, J.H. 1991. Multivariate adaptive regression splines.
Herbarium (ETSIAM-University of Córdoba, Spain) for Annals of Statistics 19: 1–141.

Applied Vegetation Science

266 Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science
R. M. Navarro-Cerrillo et al. Model distribution of Buxus balearica

Graham, C.H. & Hijmans, R.J. 2006. A comparison of methods Phillips, S.J. 2006. Maxent software for species habitat
for mapping species ranges and species richness. Global modelling, ver. 2.2.0. http://www.cs.princeton.edu/
Ecology and Biogeography 15: 578–587. _schapire/maxent/ (accessed 15 June 2009)
Guisan, A. & Zimmermann, N.E. 2000. Predictive habitat Phillips, S.J. & Dubik, M. 2008. Modelling of species
distribution models in ecology. Ecological Modelling 135: distributions with Maxent: new extensions and a
147–186. comprehensive evaluation. Ecography 31: 161–175.
Guisan, A., Zimmermann, N.E, Elith, J., Graham, C.H., Phillips, Phillips, S.J., Anderson, R.P. & Schapire, R.E. 2006. Maximum
S. & Peterson, A.T. 2007. What matters for predicting entropy modelling of species geographic distributions.
spatial distributions of trees: techniques, data, or species’ Ecological Modelling 190: 231–259.
characteristics? Ecological Monographs 4: 615–630. Rogers, D.J., Hay, S.I. & Packer, M.J. 1996. Predicting the
Hirzel, A.H., Hausser, J., Chessel, D. & Perrin, N. 2002. Ecological- distribution of tsetse flies in West Africa using temporal
niche factor analysis: how to compute habitat-suitability maps Fourier processed meteorological satellite data. Annals of
without absence data? Ecology 83: 2027–2036. Tropical Medicine and Parasitology 3: 25–241.
Hoffman, J., Aguilar-Amuchastegui, N. & Tyre, A. 2010. Use of Segurado, P. & Araujo, M. 2004. An evaluation of methods for
simulated data from a process-based habitat model to modelling species distributions. Journal of Biogeography 31:
evaluate methods for predicting species occurrence. 1555–1568.
Ecography 33: 656–666. Soberón, J. & Peterson, A.T. 2005. Interpretation of models of
Jeschke, J.M. & Strayer, D.L. 2008. Usefulness of bioclimatic fundamental ecological niches and species’ distributional
models for studying climate change and invasive species. areas. Biodiversity Informatics 2: 1–10.
Annals of the New York Academy of Sciences 1134: 1–24. Solano, E. & Feria, T.P. 2007. Ecological niche modelling and
Kohavi, R. & Provost, F. 1998. Glossary of terms. Editorial for the geographic distribution of the genus Polianthes L.
special issue on application of machine learning and the (Agavaceae) in Mexico: using niche modelling to improve
knowledge of discovery process. Machine Learning 30: 271–274. assessments of risk status. Biodiversity Conservation 16:
Lazaro, A., Traveset, A. & Castillo, A. 2006. Spatial concordance 1885–1900.
at a regional scale in the regeneration process of a circum- Stockwell, D.R.B. & Noble, I.R. 1992. Induction of sets of rules
Mediterranean relict (B. balearica): connecting seed from animal distribution data: a robust and informative
dispersal to seedling establishment. Ecography 29: 683–696. method of analysis. Mathematics and Computers in Simulation
Liu, C., Berry, P.M., Dawson T, P. & Pearson, R.G. 2005. 33: 385–390.
Selecting thresholds of occurrence in the prediction of Stockwell, D.R.B. & Peterson, A.T. 2002. Effects of sample size
species. Ecography 28: 385–393. on accuracy of species distribution models. Ecological
Leathwick, J.R. & Whitehead, D. 2001. Soil and atmospheric Modelling 148: 1–13.
water deficits, and the distributions of New Zealand’s Theurillat, J.P. & Guisan, A. 2001. Potential impact of climate
indigenous tree species. Functional Ecology 15: 233–242. change on vegetation in the European Alps: a review.
Matsui, T., Takahashi, K., Tanaka, N., Hijioka, Y., Horikawa, M., Climatic Change 50: 77–109.
Yagihashi, T. & Harasawa, H. 2009. Evaluation of habitat Thomas, C.D., Williams, S.E., Cameron, A., Green, R.E.,
sustainability and vulnerability for beech (Fagus crenata) Bakkenes, M., Beaumont, L.J., Collingham, Y.C., Erasmus,
forests under 110 hypothetical climatic change scenarios in B.F.N., De Siqueira, M.F., Grainger, A., Hannah, L., Hughes,
Japan. Applied Vegetation Science 12: 328–339. L., Huntley, B., Van Jaarsveld, A.S., Midgley, G.F., Miles, L.,
Muñoz, J. & Felicı́simo, A.M. 2004. Comparison of statistical Ortega-Huerta, M.A., Peterson, A.T. & Phillips, O.L.
methods commonly used in predictive modelling. Journal 2004. Biodiversity conservation: uncertainty in predictions
of Vegetation Science 15: 285–292. of extinction risk/effects of changes in climate and land use/
Navarro Cerrillo, R.M., Lara Fernández, A., Blanco Oyonarte, P., climate change and extinction risk. Nature 430: 6995.
Calzado Martı́nez, C., López Quintanilla, J., Fernández Thuiller, W. 2004. Patterns and uncertainties of species’ range
Cancio, A., Guzmán Álvarez, J.R. & Sánchez Salguero, R. shifts under climate change. Global Change Biology 10:
2006. Aproximación a la definición del hábitat fisiográfico del 2020–2027.
Abies pinsapo Boiss. en Andalucı́a. Investigación Agraria Sistemas Thuiller, W., Araújo, M.B. & Lavorel, S. 2004. Do we need
y Recursos Forestales. Fuera de serie, 137–152. land-cover data to model species distributions in Europe?
Pape, M. & Gaubert, P. 2007. Modelling ecological niches from Journal of Biogeography 31: 353–361.
low numbers of occurrences: assessment of the Vapnik, V. 2000. The nature of statistical learning theory. 2nd ed.
conservation status of poorly known viverrids Springer-Verlag, New York.
(Mammalia, Carnivora) across two continents. Diversity Yll, E.I., Perez-Obiol, R., Pantaleon-Cano, J. & Roure, J.M.
and Distributions 13: 890–902. 1997. Palynological evidence of climatic change and
Peterson, A.T. 2006. Uses and requirements of ecological niche human activity during the Holocene on Minorca (Balearic
models and related distributional models. Biodiversity Islands). Quaternary Research 3: 339–347.
Informatics 3: 59–72.

Applied Vegetation Science

Doi: 10.1111/j.1654-109X.2010.01112.x r 2010 International Association for Vegetation Science 267