Access to this full-text is provided by PLOS.
Content available from PLOS One
This content is subject to copyright.
RESEARCH ARTICLE
Remote-sensing based approach to forecast
habitat quality under climate change
scenarios
Juan M. Requena-Mullor
1
*, Enrique Lo
´pez
1,2
, Antonio J. Castro
1,3
, Domingo Alcaraz-
Segura
1,4
, Hermelindo Castro
1,5
, Andre
´s Reyes
1
, Javier Cabello
1,5
1Andalusian Center for the Assessment and Monitoring of Global Change (CAESCG), University of Almerı
´a,
Almerı
´a, Spain, 2Didactics of Experimental Sciences Area, Department of Education, University of Almerı
´a,
La Cañada de San Urbano, Almerı
´a, Spain, 3Department of Biological Sciences, Idaho State University,
Gale Life Sciences Bldg. Rm 207, 8th Avenue, Mail Stop, Pocatello, ID, United States of America,
4Department of Botany, University of Granada, Granada, Spain, 5Department of Biology and Geology,
University of Almerı
´a, La Cañada de San Urbano, Almerı
´a, Spain
*juanmir@ual.es
Abstract
As climate change is expected to have a significant impact on species distributions, there is
an urgent challenge to provide reliable information to guide conservation biodiversity poli-
cies. In addressing this challenge, we propose a remote sensing-based approach to fore-
cast the future habitat quality for European badger, a species not abundant and at risk of
local extinction in the arid environments of southeastern Spain, by incorporating environ-
mental variables related with the ecosystem functioning and correlated with climate and
land use. Using ensemble prediction methods, we designed global spatial distribution mod-
els for the distribution range of badger using presence-only data and climate variables.
Then, we constructed regional models for an arid region in the southeast Spain using EVI
(Enhanced Vegetation Index) derived variables and weighting the pseudo-absences with
the global model projections applied to this region. Finally, we forecast the badger potential
spatial distribution in the time period 2071–2099 based on IPCC scenarios incorporating
the uncertainty derived from the predicted values of EVI-derived variables. By including
remotely sensed descriptors of the temporal dynamics and spatial patterns of ecosystem
functioning into spatial distribution models, results suggest that future forecast is less favor-
able for European badgers than not including them. In addition, change in spatial pattern of
habitat suitability may become higher than when forecasts are based just on climate vari-
ables. Since the validity of future forecast only based on climate variables is currently ques-
tioned, conservation policies supported by such information could have a biased vision and
overestimate or underestimate the potential changes in species distribution derived from cli-
mate change. The incorporation of ecosystem functional attributes derived from remote
sensing in the modeling of future forecast may contribute to the improvement of the detec-
tion of ecological responses under climate change scenarios.
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 1 / 15
a1111111111
a1111111111
a1111111111
a1111111111
a1111111111
OPEN ACCESS
Citation: Requena-Mullor JM, Lo
´pez E, Castro AJ,
Alcaraz-Segura D, Castro H, Reyes A, et al. (2017)
Remote-sensing based approach to forecast
habitat quality under climate change scenarios.
PLoS ONE 12(3): e0172107. doi:10.1371/journal.
pone.0172107
Editor: Ine
´s A
´lvarez, University of Vigo, SPAIN
Received: May 6, 2016
Accepted: January 31, 2017
Published: March 3, 2017
Copyright: ©2017 Requena-Mullor et al. This is an
open access article distributed under the terms of
the Creative Commons Attribution License, which
permits unrestricted use, distribution, and
reproduction in any medium, provided the original
author and source are credited.
Data Availability Statement: Data restrictions are
derived from a legal normative imposed by the
Ministry of Agriculture, Food and Environment of
Spain that forbids distribution of information on
wildlife presence. Readers may contact Mr. Juan
Carlos Nevado, Regional Ministry for Environment,
Government of Andalucia, Spain (E-mail: delegado.
al.cmaot@juntadeandalucia.es; Phone: (+34) 950
10 16 76) to request data. Readers may also
contact Dr. Juan Miguel Requena Mullor (E-mail:
juanmir@ual.es) for data requests and queries. The
data will be available upon request to all interested
parties.
Introduction
As climate change is expected to have a significant impact on species distributions, there is an
urgent challenge to provide policy-makers with valuable information to guide conservation
biodiversity policies. To address this challenge, modeling approaches should be enhanced with
the aim of increasing the confidence of the future forecasting [1]. In particular, mammalian
species richness will be dramatically reduced throughout the Mediterranean basin; however,
the trend will not be uniform for all taxa. For instance, Mustelidae (e.g., badger, weasel) will
decrease while Canidae (e.g., wolf), Hyaenidae (e.g., hyena) and some families of Chiroptera
(bats) will increase [2]. These findings highlight the complexity of species response to climate
change and the necessity of focusing modeling efforts to enhance the reliability of the predicted
information.
Species distribution models (SDMs) are used to guide conservation programs in anticipa-
tion of future climate change effects on species [3]. SDMs capture relationships between a spe-
cies (occurrence) and its environment. However, some authors recognize that the incomplete
coverage of the environmental niche of species is one of the important sources of error when
forecasting distributions under climate change, because it does not cover completely the spatial
distribution range of species [4], and/or only consider climatic variables as predictors [5]. In
addition, forecasts of species distribution under climate change entail an inherent uncertainty
such as residual error of models, modeling algorithm and climate scenarios selected [6]. Differ-
ent solutions have been proposed for solving these issues. For example, Gallien et al. [7] pro-
vide a methodological framework where the use of predictions based on a global model to
weight pseudo-absences in a regional model significantly improved the predictive perfor-
mance of regional SDMs. Wenger et al. [6] suggested a Monte Carlo approach that accounts
for uncertainty within generalized linear regression models. Finally, the incorporation of other
environmental variables (different than climate ones), e.g., land use or ecosystem functioning
descriptors such as ecosystem production and seasonality, would improve the reliability of the
predictions [5,8]. However, we suggest that it is also necessary to incorporate the uncertainty
derived of the future values of projected environmental variables. With this aim, we have
developed a remote-sensing based approach.
Remotely sensed indicators of ecosystem functioning are increasingly being used in animal
research. In particular, spectral Vegetation Indices (VIs) have been used to great success in
mammal ecology [9]. VIs are conceptually and empirically linked with primary production
[10], which determines the amount of green biomass available for herbivores and is referred as
the main descriptor of ecosystem functioning [11].
Functional attributes derived from VIs are usually expressed as average temporal summa-
ries, such as the annual mean (i.e., surrogate of mean annual primary production) or the sea-
sonal coefficient of variation (i.e., indicator of seasonality or temporal variation within the
year) [12]. Of particular note are spectral Vegetation Indices (VIs), such as the Normalized
Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI). Both of these
VIs are directly related with the fraction of photosynthetically active radiation (fAPAR) inter-
cepted by green vegetation [13]. This relationship allows the derivation of regional maps of pri-
mary production from radiation use efficiency values [14]. Landscape functional heterogeneity
has also been suggested as a significant driver of species [15] and ecosystem diversity [12], par-
ticularly in the Mediterranean Region. Many animal species have proved to be especially sensi-
tive to spatial heterogeneity [16]. Recent findings suggest that this sensitivity is related more to
functional heterogeneity than to structural heterogeneity [17]. For instance, Requena-Mullor
et al. [8] found modeled spatial distribution of the European badger (Meles meles L.) in SE
Spain was significantly improved when augmenting climate variables with EVI-derived
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 2 / 15
Funding: JRM received funding from the
Andalusian Center for the Assessment and
Monitoring of Global Change (CAESCG) (http://
www.caescg.org/). Funding was also received from
the Andalusian Government (http://www.
juntadeandalucia.es/medioambiente/site/portalweb/
)(Projects GLOCHARID and SEGALERT P09–RNM-
5048), the ERDF (http://ec.europa.eu/regional_
policy/es/funding/erdf/), and the Ministry of
Science and Innovation (http://www.idi.mineco.
gob.es/portal/site/MICINN/) (Project CGL2010-
22314, subprogram BOS, National Plant I +D+ I
2010). AJC was partially funded by the National
Science Foundation Idaho EPSCoR program under
award no. IIA-1301792. (http://www.isu.edu/). The
funders had no role in study design, data collection
and analysis, decision to publish, or preparation of
the manuscript.
Competing interests: The authors have declared
that no competing interests exist.
functional attributes, (esp. the spatial variability of EVI) rather than land cover and land use
variables.
The purpose of this study is to explore the benefits of including remotely sensed descriptors
of the temporal dynamics and spatial patterns of ecosystem functioning into spatial distribu-
tion models in order to enhance the reliability of future forecasts of terrestrial mammal habitat
quality. We compared the future forecasts obtained by using models with and without incor-
porating remotely sensed descriptors. We used the European badger in an arid region of
southeastern Spain as a case study. In this region the species is not abundant and at risk of
local extinction due to climate change [18]. We built generalized linear models (using climate,
and land use variables as predictors) to project the EVI-derived variables under the Inter-gov-
ernmental Panel of Climate Change (IPCC) and land cover and land use change scenarios.
Then, we incorporated the uncertainty derived from these models into future forecast by using
the distribution of residuals. Finally, we discuss how the forecasted distribution based on our
approach can guide policy-makers to review current policies and biodiversity conservation
programs.
Materials and methods
Species model
The European badger is a medium-sized carnivore widely distributed across Europe. In the
arid southeastern limits of its range (i.e., the Mediterranean drylands of Iberian Peninsula) the
European badger prefers mosaic landscapes consisting of fruit orchards and natural vegeta-
tion, which provide shelter and food resources [19]. The potential effects of climate change on
life-history traits such as population density, social organization or population growth have
been highlighted for badger [20]. Although the species is currently considered of least concern
(LC) in Spain [21], in the arid southeastern of Spain, badgers are not abundant and densities
are usually below one badger/km
2
[22]. Since the food resources exploited by the species
directly or indirectly depend on the climate and human land use [23], it would be reasonable
expect, under future change scenarios, that some territories currently occupied might be aban-
doned with the goal of finding areas with better conditions. Therefore, European badger is at
potential risk of local extinction due to climate change [18], becoming an ideal study organism
for the purpose of this study.
Modeling approach
We first designed a set of SDMs at global scale (i.e., including the distribution range of the bad-
ger) using presence-only data and climate variables. We assembled them by a committee aver-
aging method (see below a detailed description). Then, using the global model output to
weight the pseudo-absences (PAs), we built four sets of SDMs at regional scale (i.e., in an arid
region of southeastern Spain). Two of these sets included EVI-derived, climate and topo-
graphic variables (hereafter EVI-models). The other two sets included only climate and topo-
graphic variables (hereafter Climate-models). The pseudo-absences were just weighted in one
set of EVI- and Climate-models, respectively [7]. Finally, we forecasted the potential spatial
distribution of badger for the 2071–2099 period under IPCC scenarios [24] using the regional
models with the best performance. Best performance was based on the area under the receiver
operating characteristic curve (AUC). We note that to predict the future EVI-variables used in
forecasting the potential spatial distribution, we built generalized linear models (GLMs) (using
climate, and land cover and land use variables as predictors) and projected them under IPCC
and land cover and land use change scenarios. We incorporated the uncertainty derived from
GLMs into future forecast by using the distribution of residuals.
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 3 / 15
The global model. The spatial distribution range of badger was obtained from the Digital
Distribution Maps of the IUCN Red List of Threatened Species (http://www.iucnredlist.org/)
in vector format (Fig 1A). We extracted species occurrence from the Global Biodiversity Infor-
mation Facility (http://data.gbif.org) at a minimum resolution of 2.5´ x 2.5´. We used 16279
records of occurrence after post-processing the data to remove those presences with unrealistic
coordinates (e. g., sea, incomplete coordinates). From the 19 available bioclimatic variables at a
2.5´ spatial resolution of the WorldClim database (http://www.worldclim.org/), we selected
seven with a Spearman rank-correlation lower than 0.6, i.e., annual mean temperature, mean
diurnal range, temperature seasonality, temperature annual range, mean temperature of wet-
test quarter, annual precipitation and precipitation seasonality. Following to Gallien et al. [7],
we modeled species distribution using six algorithms available in the “biomod2” package ver-
sion 3.1–64 [25] in R (http://www.R-project.org/): a generalized additive model (GAM), a clas-
sification tree analysis (CTA), a multivariate adaptive regression splines (MARS), a boosted
regression trees (BRT), an artificial neural networks (ANN) and a random forest (RF). We
kept the default models options. For each algorithm, two PAs datasets (20000 each time) were
randomly selected and four-fold cross-validations performed (48 models in total) by randomly
selecting 70% of the presence locations to train the models, and the remainder 30% to evaluate
them using the AUC [26]. Then, only those models obtaining an AUC score above 0.8 were
Fig 1. Spatial distribution of European badger. (a) Spatial distribution range of the European badger obtained from the IUCN map, (b) Regional
administrative boundaries of Andalusia, (c) Case study in arid environments of southeastern Spain (7051km
2
) defined using the Martonne aridity index and
administrative boundaries of Andalusia; location of the 73 badger presence records used for the regional model. The digital elevation model showed was
downloaded from a public database available in http://www.juntadeandalucia.es/institutodeestadisticaycartografia/prodCartografia/bc/mdt.htm
doi:10.1371/journal.pone.0172107.g001
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 4 / 15
used to build the committee averaging map. The committee averaging method is an ensemble
forecasting method [27] by which the predicted probability maps of species habitat suitability
are not averaged, but instead are transformed into binary maps (using for each model the
threshold that maximizes both sensitivity and specificity). Thus, this method gives both a pre-
diction and a measure of uncertainty. When the prediction is close to 0 or 1, it means that all
models agree to predict 0 and 1 respectively (see [25] for details).
The regional model. For the regional model we selected an arid case study in the south-
eastern Iberian Peninsula (36˚58’N, 2˚29’W) (Fig 1C). This region is the most arid in all of
Europe and presents the most extreme arid conditions in the specie range. We defined this
area based on two criteria: (1) the Martonne aridity index (I
a
), only including values between
5 and 15 [28], (2) the administrative boundaries of the Andalusia (Fig 1B), because the meth-
odologies employed to obtain environmental GIS cartography (used later to estimate envi-
ronmental variables for the modeling process) are different across regional administrative
boundaries (Spanish Autonomous communities).
The presence records for the badgers were obtained from published data [8] and personal
databases from the authors (i. e., 184 occurrence registers). We reduced locally dense sampling
by thinning the records to one per 5 x 5 km grid cell to avoid spatial autocorrelation problems.
A total of 73 presence records were used in the modeling (Fig 1C). Presence samples were dis-
tributed across a wide gradient of altitude (0–1500 meters), temperature (annual minimum
mean temperatures range: -1.6–15˚C, annual maximum mean temperatures range 17–24.5˚C),
annual precipitation range (165–419 mm/year), and annual evapotranspiration range (343–
1038 mm/year). Vegetation is mainly constituted by Mediterranean shrubland (e.g., Pistacia
lentiscus,Macrochola tenacissima,Anthyllis spp.) with more xerophytic species in lower zones
near to the coast (e.g., Thymus spp., Salsola spp.). Woodland is very scarce and is present inland
and at higher altitude (e.g., Pinus halepensis). Extensive crops (e.g., fruit orchards) are associ-
ated with the main river courses where together with natural vegetation constitute the pre-
ferred habitats of the badger [8,19]. Intensive irrigated crops are dominant in areas further
north (e.g., almond and arable crops).
We selected six variables associated to ecological requirements of the European badger and
which have high predictive power in terms of habitat suitability [29,8]. Variables were related
to climate (mean annual precipitation (PREC) and mean value of monthly maximum tempera-
tures (TMEDMAX)), relief (mean slope (SLOPE)), and spatio-temporal patterns of primary
production (EVI annual mean (EVIMEAN), intra-annual coefficient of variation of EVI
(EVIC), and spatial standard deviation of EVI annual mean (EVISTD)). To avoid collinearity
between predictors, we checked that Spearman rank-correlations were less than 0.85 [30]. The
maximum Spearman correlation value obtained was -0.37, corresponding to PREC and
TMEDMAX.
PREC,TMEDMAX (for the 1971 to 2000 period) and SLOPE were derived from spatial data
layers of the Environmental Information Network of Andalusia (http://www.juntadeandalucia.
es/medioambiente/site/web/rediam). PREC and TMEDMAX had a cell size (100 x 100 m), and
SLOPE was calculated from a 20x20 m pixel digital elevation model of Andalusia. We resampled
to the 231 x 231 m grid (pixel size of the MOD13Q1 EVI products, see below) using bilinear
resampling, in Quantum GIS (QGIS; http://www.qgis.org) which is more realistic than nearest-
neighbor interpolation [26]. This pixel resolution is suitable to capture the habitat preferences
of the European badger at a local scale [8].
Our three functional descriptors of the spatiotemporal patterns of primary production were
derived from satellite images captured by the MODIS sensor onboard the NASA TERRA satel-
lite (http://www.modis.gsfc.nasa.gov/). We used the MOD13Q1 EVI product, which consists
of 16-day maximum value composite images (23 per year) of the EVI at a 231 x 231 m pixel
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 5 / 15
size. This product has atmospheric, radiometric and geometric corrections. We used EVI
instead of NDVI because it is less influenced by soil background and saturation problems at
high biomass levels [31]. We first used the Quality Assessment (QA band) information of this
product to filter out those values affected by high content of aerosols, clouds, shadows, snow
or water. Next, we calculated the mean seasonal EVI profile (average year) for the 2001–2013
period and derived the EVI annual mean (EVIMEAN) as the mean of the 23 images of the
average year and the intra-annual coefficient of variation of EVI (EVICV) as the intra-annual
standard deviation divided by EVIMEAN. The spatial standard deviation of EVIMEAN
(EVISTD) was calculated in windows of 3 x 3 km (13x13 MOD13Q1 pixels) throughout the
study area. The size of this window was determined based on the suggested 9 km
2
home range
of the European badger for low suitability habitats [19]. Horticultural greenhouses are inten-
sively used in this area [32], and because the EVI values of greenhouses cannot be interpreted
as vegetation greenness [31], we removed all grid cells containing greenhouses (5% of the
study area) to avoid their influence of species distribution modeling.
To test the utility of functional descriptors variables in forecasting future spatial distribution
of badger, we combined the six selected variables into two groups, with and without including
EVI-derived variables. Then, we built four sets of regional SMDs, two of these sets included
EVI-derived, climate and topographic variables and the other two only included climate and
topographic variables. Using the global model output, the PAs were weighted only in one set
of EVI- and Climate-models, respectively [7]. We followed the Gallien et al. approach [7]
based on six statistical algorithms, i.e. GAM, CTA, GBM, RF, ANN and GLM. A PA datasets
of 10000 points were randomly selected, followed by four cross-validation repetitions (70–30%
on the presence locations for running and testing models respectively). To deal with the poten-
tial sample bias in the presence records (i.e., some sites are more likely to be surveyed than
others), we followed the same sampling design for selecting PAs as for selecting presences
[33]. In this manner, we restricted the choice of PAs inside the buffers of 5 km (size of plots
used in the grid sampling by [8]) around any of the presence records. Ninety-six different
models were finally run, and grouped into four sets: EVI-models with and without weighted
PAs (24 models, respectively) and Climate-models with and without weighted PAs (24 models,
respectively). Currently, an incompletely coverage of the environmental niche of species is
recognized that reduces confidence of the forecasting distributions under climate change sce-
narios [4]. By focusing on a reduced area of the distribution range of badger (e.g., and arid
southeast in the Iberian Peninsula), we are ignoring that the species may resist a broader range
of environmental conditions, and thereby failing to take into account its ecological flexibility
in responding to future climate conditions. To solve this point, we employed the approach
supported by [7]. According to these authors, weighting the PAs in the regional model using
the predictions from the global model enable: decrease the influence of (regional) false ab-
sences, validate the true (regional) presences, and let the regional climate, soil and land use
refine the regional niche estimation. Thus, we used the global model projections applied to our
case study to weight each PA. Where the global model showed a high level of agreement with
the PA we attributed a high weight to the PA, and vice versa. The weights were given by the Eq
(1):
WeightðxÞ ¼ 1
1þprojGðxÞ
projGðxÞ 1
2ð1Þ
where Weight(x) is the weight attributed to the PA x, and projG is the global model prediction
at the location of x.
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 6 / 15
Regional model evaluation and future forecasting
We evaluated the performance of regional models by AUC using the remainder of the pres-
ence records (i.e., 111) after selecting the samples used to build the models (see above). With
the aim to obtain as reliable as possible future forecasting, we forecasted the badger spatial dis-
tribution for future climate scenarios in each set of models by using the best models based on
AUC developed under current environmental conditions. A one-tailed Wilcoxon signed rank
test was used to assess whether increases in overall AUC in the EVI-models relative to those in
the Climate-models were statistically significant [34].
Two scenarios proposed by the IPCC were considered: A2 and B1. The A2 scenario
assumes a continuously increasing global population, economic development that is primarily
regionally oriented and focused on economic growth and technological changes that are more
fragmented and slower than in other storylines. The B1 scenario assumes that economic struc-
tures rapidly change towards a service and information economy, and that resource-efficient
technologies are introduced. The projected PREC and TMEDMAX (100 x 100 m cell size)
variables for the 2071–2099 period were obtained from statistical downscaling methods (i.e.,
multiple linear regression) [35] of these IPCC scenarios available at the Environmental Infor-
mation Network of Andalusia (see above). The simulated daily temporal series for the pro-
jected period were summarized by the annual mean with the aim of creating a representative
year (see [36] for details). Then, we resampled to the 231 x 231 m grid using bilinear resam-
pling. Although slope has been described as a relevant factor for sett digging, in low density
areas of Mediterranean landscapes, setts were located almost everywhere [22]. Since slope
seems not to be so limiting and was not available for future scenarios due to the complex driv-
ers and predictors behind this process, we considered it static through time. The three pro-
jected functional descriptors derived from EVI (i.e., EVIMEAN,EVICV and EVISTD) were
predicted by GLMs. Although precipitation and temperature are the main two climate factors
that drive primary production of the biosphere, at the regional scale, the response of primary
production to these climate drivers can vary both spatially and temporality, modulated by dif-
ferent vegetation types [37]. Therefore, EVIMEAN and EVICV were predicted by a GLM that
used as explanatory variables the current PREC and TMEDMAX variables, land cover and land
use and slope. Then, future EVIMEAN and EVICV were projected using the fitted models with
future climate variables of each A2 and B1 scenario. In addition, we simulated three future sce-
narios of land cover-land use change: (1) irrigated crop scenario: the 5% and 10% of the natu-
ral vegetation and rainfed crop patches, respectively, converted into irrigated crop patches; (2)
crop abandonment scenario: the 7.5% and 7.5% of rainfed and irrigated crop patches, respec-
tively, converted into natural vegetation parches; (3) no land change scenario (see S1 Fig).
These scenarios were based on the future projections obtained by Piquer-Rodrı
´guez et al. [38]
for the southeastern Iberian Peninsula and were simulated using the MOLUSCE plugin in
QGIS (http://hub.qgis.org/projects/molusce). Finally, the future EVISTD variable was obtained
by using a 3 x 3 km moving window on the projected EVIMEAN variable (as explained above).
In the linear regression, the predicted values of the response variable are estimated by the
following Eq (2):
yi¼b0þb1xi1þb2xi2þ:::: þbjxij þεið2Þ
where y
i
is the predicted value of variable response for the ith observation when the explana-
tory variable X
1
equals x
i1
, X
2
equals x
i2
, X
j
equals x
ij
;β
0
is the intercept; β
1
,β
2
and β
j
are the
partial regression coefficients for the explanatory variables 1, 2 and j;ε
i
is unexplained error
associated with the ith observation.
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 7 / 15
Errors are assumed to be normally distributed with expectation 0 and variance σ
2
. We
exploited this assumption to incorporate the uncertainty derived from the projected EVI-
derived variables into future forecasts. To do this, we generated 1000 projections of these vari-
ables by adding in each iteration the errors estimated from the GLMs. We obtained these
errors using the distribution of residuals provided by the GLMs output (see S1 Table for details
of GLMs).
Then, we assembled the future forecasts obtained using the projected climate and EVI-
derived variables by the median value. All calculations were made using R software.
Comparing current and future spatial distributions
We used a distance-based method [39] to compare the current and future spatial distributions
yielded by the different regional models and scenarios. This method combines aspects of
image analysis with distance-based statistical test to extract two types of information from an
ensemble of SDM output maps: intensity and spatial difference.
Intensity index (Eq 3) measures the overall level of habitat suitability (based on the environ-
mental variables used in the SDMs) for a species within an area extent but with the spatial
information removed.
Si¼PR
r¼1PC
c¼1Mi½r;c ð3Þ
where S
i
is the intensity; M
i
is the ith map in a set of SDM outputs; R is the number of rows
indexing northing values in the map; C is the number of columns indexing easting values in
the map.
Spatial differences between current and future distribution maps was tested by using the
Hellinger distance (Eq 4) [40]. Wilson [39] found that this distance obtained the best results
when SDMs output maps were spatially compared. Hellinger distance measures the dissimilar-
ity between two probability density functions for continuous variables. Before to be used for
comparing maps, it requires normalization by intensity, so that, the map has the properties of
a bivariate probability distribution. Both measures (i.e., Hellinger distance and the corre-
sponding difference in intensity) are largely independent and can be used in a scatterplot to
show patterns of change in the distribution of habitat suitability [39,40].
Hij ¼0:5PR
r¼1Pc
c¼1ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Mi½r;c
Si
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Mj½r;c
Sj
s
" #2
2
43
5
1=2
ð4Þ
where H
ij
is the Hellinger distance between the M
i
and M
j
maps.
To check if both measures (i.e., intensity and Hellinger distance) yielded by Climate-model
and EVI-models, respectively, were significantly different, we explored the probability distri-
butions of the normalized pairwise differences for each measure by a bootstrap method
(10.000 replicates) and then, we calculated which of these differences fell above the 95th per-
centile and below the 5th percentile.
Results and discussion
Model evaluations
All global models showed a good performance (AUC >0.8) and therefore the 48 models were
used in the ensemble modeling. However, when we selected the models at regional scale, we
used different AUC values as thresholds with the aim to maximize the performances obtained.
For the Climate-models, the AUC values were 0.55 and 0.6 for with and without weighted PAs
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 8 / 15
models, respectively; in this form, 3 and 6 models were selected to forecast the badger spatial
distribution into future scenarios. For the EVI-models, the AUC values used were 0.55 and 0.7
(with and without weighted PAs) and the number of models selected was 9 and 5, respectively.
EVI-models performed better than Climate-models (both without weighted PAs)
(W= 453.5, p<0.0001) for the current conditions, although the AUC values accomplished
were moderated (the maximum value reached was 0.74 into the EVI-models set) (Fig 2A).
When we compared between models with weighted PAs, EVI-models were also better than
Climate-models (W= 439, p<0.001), but the AUCs were still lower (Fig 2B). The comparison
of the performance between models with and without weighted PAs but using the same type of
variables, revealed that weighting the PAs did not significantly improved discrimination
between areas where badger was present and where it was not recorded (PAs) (W= 473,
p<0.0001; W= 506, p<0.00001).
Comparing forecasted future distributions
Because of the performance of models (i. e., Climate-models and EVI-models) with weighted
PAs was lower than the models without weighted PAs, we only used the former to forecast
future distribution. Badger was predicted to experience an increase in habitat suitability (posi-
tive change in intensity) and considerable spatial change when we just considered climate vari-
ables (Fig 3). A2 scenario forecasted the most favourable conditions for badger with these
variables. However, when EVI-derived variables were incorporated into the distribution mod-
els, habitat suitability was predicted to decrease (negative change in intensity), and spatial pat-
tern change was greater than without EVI-derived variables. The increase of spatial changes
Fig 2. Comparison of model performance using the area under the threshold-independent receiver operating characteristic curve (AUC). (a)
Regional EVI-model performance versus regional Climate-model, both without weighted PAs. (b) Regional EVI-model performance (AUC) versus regional
Climate-model, both with weighted PAs. Values below the diagonal line mean a better performance for EVI-model, values above mean a better performance
for Climate-model and values on the diagonal line mean an identical performance between the models compared.
doi:10.1371/journal.pone.0172107.g002
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 9 / 15
could be driven by LCLC change scenarios used to predict EVI variables. LCLU changes are
considered to be a major driver of global environmental change and biodiversity loss [41]. For
this reason, future forecasts that not include them, e.g. based only on climate variables, could
harbour an important bias, particularly, at a regional scale where the landscape features, such
as vegetation and land use, are key for the species [19,23]. Although the forecasts obtained
from the different IPCC scenarios and land cover and land uses change were similar for EVI-
models, the decrease of habitat suitability was more moderate with the increase of irrigated
Fig 3. Changes in intensity (Yaxis) and Hellinger distance (Xaxis) forecasted for the European badger in southeastern Iberian Peninsula under
IPCC scenarios. Change in intensity (overall measure of habitat suitability) between the current and IPCC A2 (circles) and B1 (squares) for 2071–2099 maps
is plotted against the corresponding Hellinger distance (representing spatial changes). Filled symbols represent EVI-models and open symbols Climate-
models, both without weighted PAs. I:irrigated crop scenario; A:crop abandonment scenario; N:no change scenario.
doi:10.1371/journal.pone.0172107.g003
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 10 / 15
crop than crop abandonment. Finally, the forecasts predicted by Climate-models and EVI-
models were significantly different at a 0.5 level of confidence (see Table A, Figs A and B in
S1 File).
EVI descriptors of ecosystem functioning to forecast species
distributions
Our results suggest that future forecasts based only on climate variables should be interpreted
with caution because the temporal dynamics and spatial patterns of ecosystem functioning
may play an important role in determining the species distribution and the dynamics of its
change. At a global scale, the bioclimate envelope approach (i.e., modeling the association
between climate variables and species distribution) can provide a first approximation as to the
potentially dramatic impact of climate change [5]. However, at a regional scale, factors that
determine species distributions may vary [42] and therefore, the ecological limits predicted by
climate may be misleading and the future forecasts obtained could underestimate or overesti-
mate the potential changes in species distribution derived from climate change. In fact, we
show how Climate- models and EVI-models forecasted opposed futures for the badger under
both A2 and B1 scenarios in Mediterranean arid environments. Supporting this finding,
Requena-Mullor et al. [8] showed that EVI descriptors of ecosystem functioning are useful to
describe drivers of the spatial distribution of badger in these environments, in consequence,
the future forecast derived from EVI-models may suggest more reliability.
At this point, a key question arises: how much more reliable are these future distribution
forecasts?. Addressing this question goes beyond the goal of this study, however, our results
show that EVI-derived variables could help to decrease the underlying ecological bias of the
future forecasts based just on climate variables. In this regard, EVI-models forecasted a
decreased of the habitat suitability for badger. Macdonald et al. [20] found that badger life his-
tory parameters (such as survival, fecundity or body-weight) are correlated with annual vari-
ability of both temperature and rainfall mediated by food supply. Therefore, climate trends
might influence badgers population growth directly and through correlations with food
availability [43]. Primary production is at the base of food webs, and therefore a decrease
would translate into lower food availability for higher trophic levels. Supported by findings of
Alcaraz-Segura et al. [12], we consider that EVI-derived variables could represent the interac-
tions between climate and food availability for badger as descriptors of the temporal dynamics
and spatial patterns of primary production. On a local scale, the composition of the badger
diet is influenced primarily by human land use and management [23]. More specifically, in the
southeastern Iberian Peninsula, badger diet depends on fruit orchards and other derived food
resources (e.g., insects) [44]. EVI-derived variables have been suggested as proxies of the spa-
tial and temporal variability of food resources for badgers in these environments [8]. There-
fore, future forecasts based on both EVI-derived and climate variables would capture not only
climate change but also correlations between climate and ecosystem functioning through the
dynamic of primary production and land use change. According with this, our results showed
that the decrease of habitat suitability with an increase of irrigated crops was less than with the
abandonment crops.
The regional model performances were worse with weighted PAs. This result partially dis-
agrees with findings by [7]. As we described above, at a regional scale, other factors than cli-
mate can play an important role in determining species distribution [42]. Although climate
was not predicted as a limiting factor for the occurrence of badger in the arid southeastern of
Iberian Peninsula by the global model (based only on climate variables, see S2 File), some
authors have suggested that the species in these environments select rural landscapes
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 11 / 15
consisting in a mosaic of fruit crops and orchards mixed with patches of semi-natural vegeta-
tion as a response to food shortage [19,8], which are not considered by the prediction of the
global model. In consequence, the badger occurrence predicted by the global model in south-
eastern Iberian Peninsula was overestimated (see S2 File).
Implications for biodiversity conservation policies
Since the validity of future forecast based only on climate variables is currently questioned
[45], policies and biodiversity conservation programs supported by such information could
have a biased vision and overestimate or underestimate the potential changes in the species
distribution derived from climate change. The incorporation of ecosystem functional attri-
butes derived from remote sensing in the modeling of future forecast may contribute to the
improvement of detection of ecological responses under climate change scenarios. In addition,
our approach would allow incorporating the uncertainty derived of these attributes when they
are projected to future under both climate, and land use change scenarios. In consequence, the
future forecast obtained would be more reliable and help policy-makers to review the existing
policies and biodiversity conservation programs.
Regarding global biodiversity conservation efforts, we highlight that the use of variables
related to ecosystem functional attributes derived from remote sensing data, has been shown
to conduct cost-efficient monitoring schemes for biodiversity conservation and across a vari-
ety of ecosystems worldwide [46]. Our approach can be aligned with goals of the Group on
Earth Observations Biodiversity Observation Network (GEO BON), which is seeking consen-
sus for the key variables to improve the ever-increasing need for monitoring of biodiversity
across space and time in a rapidly changing planet.
Supporting information
S1 Fig. Land cover-land use change maps. Simulated land cover and land use change scenar-
ios.
(DOCX)
S1 Table. GLMs for EVI-derived variables. Summary of the GLMs for EVIMEAN and EVICV
variables.
(DOCX)
S1 File. Differences of intensity and Hellinger distance. Normalized differences of intensity
and Hellinger distance forecasted by Climate-models and EVI-models.
(DOCX)
S2 File. Spatial distribution predicted for the European badger by the global model. Global
model predictions and specificity of EVI-models.
(DOCX)
Acknowledgments
JRM received funding from the Andalusian Center for the Assessment and Monitoring of
Global Change (CAESCG). Funding was also received from the Andalusian Government
(Projects GLOCHARID and SEGALERT P09–RNM-5048), the ERDF, and the Ministry of Sci-
ence and Innovation (Project CGL2010-22314, subprogram BOS, National Plant I +D+ I
2010). AJC was partially funded by the National Science Foundation Idaho EPSCoR program
under award no. IIA-1301792.
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 12 / 15
Author Contributions
Conceptualization: JMR-M EL AJC.
Data curation: AR.
Formal analysis: JMR-M.
Funding acquisition: HC.
Investigation: JMR-M JC.
Methodology: JMR-M DA-S AJC.
Project administration: HC JC.
Software: JMR-M AR.
Supervision: HC JC.
Validation: JMR-M.
Visualization: JMR-M AJC.
Writing – original draft: JMR-M AJC.
Writing – review & editing: JMR-M EL AJC DA-S JC.
References
1. Mawdsley JR, O’Malley R, Ojima DS. A Review of Climate-Change Adaptation Strategies for Wildlife
Management and Biodiversity Conservation. Conservation Biology 2009 Oct; 23(5):1080–1089 doi: 10.
1111/j.1523-1739.2009.01264.x PMID: 19549219
2. Maiorano L, Falcucci A, Zimmermann NE, Psomas A, Pottier J, Baisero D, et al. The future of terrestrial
mammals in the Mediterranean basin under climate change. Philosophical Transactions of The Royal
Society. 2011 Aug; 366:2681–2692
3. Peterson AT, Sobero
´n J, Pearson RG, Anderson RP, Martinez-Meyer E, Nakamura M, Arau
´jo MB. Eco-
logical niches and geographic distributions. Princeton University Press Princeton, NJ; 2011
4. Wiens JA, Stralberg D, Jongsomjit D, Howell CA, Snyder MA. Niches, models, and climate change:
assessing the assumptions and uncertainties. Proceedings of the National Academy of Sciences of the
United States of America 2009 Nov; 106(2):19729–19736
5. Pearson RG, Dawson TP. Predicting the impacts of climate change on the distribution of species: are
bioclimate envelope models useful?. Global Ecology and Biogeography 2003 Sep; 12(5):361–371
6. Wenger SJ, Som NA, Dauwalter DC, Isaak DJ, Neville HM, Luce CH, et al. Probabilistic accounting of
uncertainty in forecasts of species distributions under climate change. Global Change Biology 2013
Nov; 19(11):3343–3354 doi: 10.1111/gcb.12294 PMID: 23765608
7. Gallien L, Douzet R, Pratte S, Zimmermann NE, Thuiller W. Invasive species distribution models—how
violating the equilibrium assumption can create new insights. Global Ecology and Biogeography 2012
Nov; 21(11):1126–1136
8. Requena-Mullor JM, Lo
´pez E, Castro AJ, Cabello J, Virgo
´s E, Gonza
´lez-Miras E, et al. Modeling spatial
distribution of European badger in arid landscapes: an ecosystem functioning approach. Landscape
Ecology 2014 May; 29(5):843–855
9. Cabello J, Ferna
´ndez N, Alcaraz-Segura D, Oyonarte C, Piñeiro G, Altesor A, et al. The ecosystem
functioning dimension in conservation: insights from remote sensing. Biodiversity and Conservation
2012 Dec; 21(13):3287–3305
10. Paruelo JM, Lauenroth WK, Burke IC, Sala OE. Grassland precipitation-use efficiency varies across a
resource gradient. Ecosystems 1999 Jan; 2(1):64–68
11. Alcaraz-Segura D, Paruelo JM, Cabello J. Identification of current ecosystem functional types in the Ibe-
rian Peninsula. Global Ecology and Biogeography 2006 Feb; 15:200–210
12. Alcaraz-Segura D, Paruelo JM, Epstein HE, Cabello J. Environmental and human controls of ecosys-
tem functional diversity in temperate South America. Remote Sensing 2013 Jan; 5(1):127–154
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 13 / 15
13. Ruimy A, Saugier B, Dedieu G. Methodology for the estimation of terrestrial net primary production from
remotely sensed data. Journal of Geophysical Research 1994 Mar; 99(D3):5263–5283
14. Castro AJ, Paruelo JM, Alcaraz-Segura D, Cabello J, Oyarzabal M, Lo
´pez-Carrique E. Missing gaps in
the estimation of the carbon gains service from Light Use Efficiency models. In: Alcaraz-Segura D, Di
Bella C, editors. Earth Observation of Ecosystem Services; 2013. pp. 105–124
15. Davidowitz G, Rosenzweig ML. The latitudinal gradient of species diversity among North American
grasshoppers within a single habitat: a test of the spatial heterogeneity hypothesis. Journal of Biogeog-
raphy 1998 May; 25(3):553–560
16. Fryxell JM, Wilmshurst JF, Sinclair ARE, Haydon DT, Holt RD, Abrams PA. Landscape scale, heteroge-
neity, and the viability of Serengeti grazers. Ecology Letters 2005 Feb; 8(3):328–335
17. Zaccarelli N, Li B-L, Petrosillo I, Zurlini G. Order and disorder in ecological time-series: Introducing nor-
malized spectral entropy. Ecological Indicators 2013 May; 28:22–30
18. Virgo
´s E, Revilla E, Domingo-Roura X, Mangas JG. Conservacio
´n del tejo
´n en España: sı´ntesis de
resultados y principales conclusiones. In: Virgo
´s E, Revilla E, Mangas JG, Domingo-Roura X, editors.
Ecologı
´a y conservacio
´n del tejo
´n en ecosistemas mediterra
´neos. Sociedad Española para la Conser-
vacio
´n y Estudio de los Mamı
´feros (SECEM), Ma
´laga; 2005. pp. 283–294
19. Lara-Romero C, Virgo
´s E, Escribano-A
´vila G, Mangas JG, Barja I, Pardavila X. Habitat selection by
European badgers in Mediterranean semi-arid ecosystems. Journal of Arid Environment 2012 Jan;
76:43–48
20. Macdonald DW, Newman C, Buesching CD, Nouvellet A. Are badgers ‘Under The Weather’? Direct
and indirect impacts of climate variation on European badger (Meles meles) population dynamics.
Global Change Biology 2010 Feb; 16(11):2913–2922
21. Palomo LJ, Gisbert J, Blanco JC. Atlas y Libro Rojo de los Mamı
´feros Terrestres de España. Direccio
´n
General para la Biodiversidad-SECEM-SECEMU. Madrid; 2007
22. Revilla E, Palomares F, Ferna
´ndez N. Characteristics, location and selection of diurnal resting dens by
Eurasian badger (Meles meles) in a low density area. Journal of Zoology (London) 2001 Nov; 255
(3):291–299
23. Fischer C, Ferrari N, Weber JM. Exploitation of food resources by badgers (Meles meles) in the Swiss
Jura Mountains. Journal of Zoology 2005 May; 266(02):121–131
24. IPCC (2007) Climate Change 2007: Impacts, Adaptation and Vulnerability. Parry M., Canziani O., Palu-
tikof J., van der Linden P., Hanson C. Cambridge University Press.
25. Thuiller W, Georges D, Engler R. biomod2: Ensemble platform for species distribution modeling. R
package version 3.1–64. 2014. http://CRAN.R-project.org/package=biomod2.
26. Phillips SJ, Anderson RP, Schapire RE. Maximum entropy modeling of species geographicdistribu-
tions. Ecological Modelling 2006 Jan; 190(3–4):231–259
27. Arau
´jo MB, New M. Ensemble forecasting of species distributions. Trends in Ecology and Evolution
2006 Sep; 22(1):42–47 doi: 10.1016/j.tree.2006.09.010 PMID: 17011070
28. Martonne E. Areisme et indice d’aridite
´. Geographical Review 1926; 17:397–414
29. Virgo
´s E, Casanovas JG. Environmental constraints at the edge of a species distribution, the Eurasian
badger (Meles meles L.): a biogeographic approach. Journal of Biogeography 1999 May; 26(3):559–
564
30. Booth GD, Niccolucci MJ, Schuster EG. Identifying proxy sets in multiple linear regression: an aid to bet-
ter coefficient interpretation. Intermountain Research Station. 1994. USDA Forest Service, Ogden,
Utah, USA
31. Huete AR, Didan K, Miura T, Rodriguez EP, Gao X, Ferreira LG. Overview of the radiometric and bio-
physical performance of the MODIS vegetation indices. Remote Sensing of Environment 2012 Nov; 83
(1–2):195–213
32. Quintas-Soriano C, Castro AJ, Garcı
´a-Llorente M, Cabello J, Castro H. From supply to social demand:
a landscape-scale analysis of the water regulation service. Landscape Ecology 2014 Jul; 29(6):1069–
1082
33. Phillips SJ, Dudı
´k M, Elith J, Graham CH, Lehmann A, Leathwick JR, et al. Sample selection bias and
presence-only distribution models: implications for background and pseudo-absence data. Ecological
Applications 2009 Jan; 19:181–197 PMID: 19323182
34. Buermann W, Saatchi S, Smith TB, Zutta BR, Chaves JA, Mila
´B, et al. Prediction species distributions
across the Amazonian and Andean regions using remote sensing data. Journal of Biogeography 2008
Jul; 35(7):1160–1176
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 14 / 15
35. Wilby RL, Wigley TML, Conway D, Jones PD, Hewitson BC, Main J, et al. Statistical downscaling of
general circulation model output: A comparison of methods. Water Resources Research 1998 Nov; 34
(11):2995–3008
36. Generating Climate Change scenarios in Andalusia. Foundation for Climate Research. 2006 Dec. Avail-
able on line: (http://www.juntadeandalucia.es/medioambiente/web/Bloques_Tematicos/Sostenibilidad/
Estrategia_andaluza_cambio_climatico/escenarios_de_cambio_climatico_regionalizados/informe_
escenarios_2006.pdf)
37. Cabello J, Alcaraz-Segura D, Ferrero R, Castro AJ, Liras E. The role of vegetation and lithology in the
spatial and inter-annual response of EVI to climate in drylands of Southeastern Spain. Journal of Arid
Environment 2012 Apr; 79:76–83
38. Piquer-Rodrı
´guez M, Kuemmerle T, Alcaraz-Segura D, Zurita-Milla R, Cabello J. Future land use
effects on the connectivity of protected area networks in southeastern Spain. Journal of Nature Conser-
vation 2012 Dec; 20(6):326–336
39. Wilson PD. Distance-based methods for the analysis of maps produced by species distribution models.
Methods in Ecology and Evolution 2011 Dec; 2(6):623–633
40. Legendre P, Legendre L. Numerical Ecology. 2nd ed. Amsterdam. Elsevier; 1998
41. Verburg PH, Crossman N, Ellis EC, Heinimann A, Hostert P, Mertz O, et al. Land system science and
sustainable development of the earth system: a global land project perspective 2015 Dec; 12:29–41
42. Whittaker RJ, Nogue
´s-Bravo D, Arau
´jo MB. Geographical gradients of species richness: a test of the
water–energy conjecture of Hawkins et al.(2003) using European data for five taxa. Global Ecology and
Biogeography 2007 Jan; 16(11):76–89
43. Nouvellet P, Newman C, Buesching CD, Macdonald DW. A Multi-Metric Approach to Investigate the
Effects of Weather Conditions on the Demographic of a Terrestrial Mammal, the European Badger
(Meles meles). PLoS One 2013 Jul; 8(7):e68116 doi: 10.1371/journal.pone.0068116 PMID: 23874517
44. Requena-Mullor JM, Lo
´pez E, Castro Virgo
´s E, Castro H. Landscape influence on the feeding habits of
European badger (Meles meles) in arid Spain. Mammal Research 2016 Jul; 61(3):197–207
45. Beale CM, Lennon JJ, Gimona A. Opening the climate envelope reveals no macroscale associations
with climate in European birds. Proceedings of the National Academy of Sciences of the United States
of America 2008 Sep; 105(39):14908–14912 doi: 10.1073/pnas.0803506105 PMID: 18815364
46. Pereira HM, Ferrier S, Walters M, Geller GM, Jongman HG, Scholes RJ, et al. Essential Biodiversity
Variables. Ecology 2013 Jan; 339(6117): 227–278
Modeling and forecasting habitat quality from space
PLOS ONE | DOI:10.1371/journal.pone.0172107 March 3, 2017 15 / 15
Available via license: CC BY 4.0
Content may be subject to copyright.
