Satellite imagery as a single source of predictor variables for habitat suitability modelling: how Landsat can inform the conservation of a critically endangered lemur


  • José J. Lahoz-Monfort,

    Corresponding author
    1. Department of Life Sciences, Imperial College London, Silwood Park Campus, Buckhurst Road, Ascot, Berkshire SL5 7PY, UK
    2. National Centre for Statistical Ecology, School of Mathematics, Statistics and Actuarial Science, University of Kent, Canterbury, Kent CT2 7NF, UK
      Corresponding author. E-mail:
    Search for more papers by this author
  • Gurutzeta Guillera-Arroita,

    1. Department of Life Sciences, Imperial College London, Silwood Park Campus, Buckhurst Road, Ascot, Berkshire SL5 7PY, UK
    2. National Centre for Statistical Ecology, School of Mathematics, Statistics and Actuarial Science, University of Kent, Canterbury, Kent CT2 7NF, UK
    Search for more papers by this author
  • E. J. Milner-Gulland,

    1. Department of Life Sciences, Imperial College London, Silwood Park Campus, Buckhurst Road, Ascot, Berkshire SL5 7PY, UK
    Search for more papers by this author
  • Richard P. Young,

    1. Durrell Wildlife Conservation Trust, Les Augrès Manor, Trinity, Jersey, JE3 5BP, Channel Islands
    2. Department of Biology & Biochemistry, University of Bath, Bath BA2 7AY, UK
    Search for more papers by this author
  • Emily Nicholson

    1. Department of Life Sciences, Imperial College London, Silwood Park Campus, Buckhurst Road, Ascot, Berkshire SL5 7PY, UK
    Search for more papers by this author

Corresponding author. E-mail:


1. Statistical modelling of habitat suitability is an important tool for planning conservation interventions, particularly for areas where species distribution data are expensive or hard to collect. Sometimes however the predictor variables typically used in habitat suitability modelling are themselves difficult to obtain or not meaningful at the geographical extent of the study, as is the case for the Alaotran gentle lemur Hapalemur alaotrensis, a critically endangered lemur confined to the marshes of Lake Alaotra in Madagascar.

2. We developed a habitat suitability model where all predictor variables, including vegetation indices and image texture measures at different scales (as surrogates for habitat structure), were derived from Landsat7 satellite imagery. Using relatively few presence records, the maximum entropy (Maxent) approach and AUC were used to assess the performance of candidate predictor variables, for studying the effect of scale, model selection and mapping suitable habitat.

3. This study demonstrated the utility of satellite imagery as a single source of predictor variables for a Maxent habitat suitability model at the landscape level, within a restricted geographical extent and with a fine grain, in a case where predictor variables typically used at the macro-scale level (e.g. climatic and topographic) were not applicable.

4. In the case of H. alaotrensis, the methodology generated a habitat suitability map to inform conservation management in Lake Alaotra and a replicable protocol to allow rapid updates to habitat suitability maps in the future. The exploration of candidate predictor variables allowed the identification of scales that appear ecologically relevant for the species.

5.Synthesis and applications. This study presents a cost-effective combination of maximum entropy habitat suitability modelling and satellite imagery, where all predictor variables are derived solely from Landsat7 images. With a habitat modelling method like Maxent that shows good performance with few presence samples and Landsat images now freely available, the methodology can play an important role in rapid assessments of the status of species at the landscape level in data-poor regions, when typical macro-scale environmental predictors are of little use or difficult to obtain.


Knowledge of the ecological and geographic distribution of species is critical for prioritizing and informing conservation action (Hirzel et al. 2006), for conservation planning (Wintle, Elith & Potts 2005) and for assessing threats from a range of anthropogenic factors (Akçakaya & Atwood 1997). However, detailed distribution data sufficient for designing sound conservation measures are often lacking, and field research is typically costly and labour intensive. A lack of data is even more apparent in developing countries with high biodiversity, where it may seriously hinder conservation efforts (Gaston & Rodrigues 2003). One way of overcoming this data shortfall is to build models of a species’ suitable habitat and distribution, which can then be used to plan data collection or prioritize interventions. Habitat models are usually correlative (Guisan & Zimmermann 2000) and quantify the relationship between field observations and a set of environmental variables that are expected to reflect some key aspects of the species-habitat association (Hirzel et al. 2006), which is then used to create spatial predictions of the suitability of the habitat for that species.

An increasing array of algorithms is being developed to build habitat suitability models [see Guisan & Zimmermann (2000) and Elith et al. (2006) for a comparison of methods]. Many recent models use presence records only, which are often the only reliable data available; absences can bias results if they reflect non-detection or historical reasons instead of habitat unsuitability (Hirzel et al. 2002). The environmental data used to characterize habitat requirements frequently include climate, topography, soil geology and geographical information. Although easily available at coarse resolutions (above about 1 km2), such data can be difficult and expensive to obtain at finer resolutions, particularly in data-poor regions. Remotely sensed data have also been incorporated, most commonly through derived parameters like vegetation indices (Bellis et al. 2008) or land cover maps (Ozesmi & Bauer 2002). However, ground-truthed land cover classification is a time-consuming intermediate step in which the resulting habitat classes may not coincide with habitat requirements for the focal species or may ignore relevant variability within classes (Bellis et al. 2008; St-Louis et al. 2009), potentially introducing error in habitat suitability models.

The aim of this study was to develop an efficient methodology to model habitat suitability using only remote sensing data but avoiding the intermediate step of a ground-truthed, detailed land-cover classification. The method is particularly relevant in cases where other traditionally used environmental variables are difficult to collect or not relevant for the extent of the study. In the process, candidate predictor variables based on ecological knowledge of the species of interest were explored, including an investigation of the effect of scale on the resulting model predictions. The methodology was applied to the Alaotran gentle lemur Hapalemur alaotrensis (Rumpler), a Critically Endangered endemic primate of Madagascar (IUCN 2009). The species’ precise distribution is poorly known but it is confined to a single site, the biodiverse marshes of Lake Alaotra, Madagascar’s most productive inland fishery and a key rice-growing area (Andrianandrasana et al. 2005). To balance the needs of humans and wildlife in the management of the site, it is necessary to identify which parts of the remaining marsh contain the most suitable H. alaotrensis habitat so that appropriate conservation measures can be identified. Habitat suitability models are often performed at the macro-scale (regions or continents), where environmental factors such as climate and topography are relevant. This study was carried out within a smaller geographical extent, at a landscape level and a finer scale, where biotic interactions become dominant (Soberón 2007). In our particular case, topography and climate were not informative as we were dealing with a relatively small area of marsh, with constant elevation and little meaningful climatic variation. Most of the marsh is dense and inaccessible and cannot be surveyed by boat, making remote sensing the only logistically feasible source of information. Although land-cover classifications have been performed for wetlands with satellite imagery, a process which is generally difficult for this habitat (Ozesmi & Bauer 2002), in Alaotra the ground-truthing data currently do not exist, and their collection would be a lengthy and costly process. We do not have useful ancillary GIS data (e.g. soils) to aid in habitat classification. Candidate predictors for this research were therefore extracted directly from Landsat7 satellite images, including vegetation indices and various texture measures that relate to habitat structure at different scales. We selected variables by exploration of their relative performance and used the best model to produce a habitat suitability map for H. alaotrensis, to assist in the design of conservation interventions for Lake Alaotra (Fig. 1).

Figure 1.

 Overview of the complete modelling process, from input data (Landsat7 bands and position of lemur sightings) to the habitat suitability (HS) map and a replicable protocol.

Materials and methods

Study area and species

Lake Alaotra, the largest lake in Madagascar (approximately 20 000 ha), is an important biodiversity site (Pidgeon 1996). It harbours a large marsh dominated by papyrus (Cyperus madagascariensis, Cyperus latifolius) and reeds Phragmites communis (Pidgeon 1996). The area has a dry season from May to October, and a rainy season from November to April. A five-fold increase in human population around the lake since 1960 has resulted in heavy human pressure on both the marsh and the lake (Andrianandrasana et al. 2005).

Hapalemur alaotrensis is a territorial lemur that forms small family groups with territories ranging from 0·6 to 8 ha (Mutschler & Tan 2003), with an average of 2 ha (Mutschler & Feistner 1995). It is an elusive animal with low detectability in the dense marsh (Guillera-Arroita et al. 2010), which makes monitoring extremely difficult. The species is strictly limited to marsh vegetation (Mutschler & Feistner 1995) and is exclusively folivorous, feeding mainly on three species: papyrus, reeds and the grass Echinochloa crusgalli (Mutschler, Feistner & Nievergelt 1998). Ralainasolo (2004) observed H. alaotrensis mostly in dense mature papyrus and reed vegetation with diverse undergrowth, where marsh fires had not occurred within several years; few individuals were detected in areas with younger low thin papyrus, leading to the hypothesis that H. alaotrensis requires tall strong vegetation to support locomotion and provide protection. They move along continuous vegetation by leaping and walking on bent stems and, although they can swim, they do so very infrequently and seem to avoid open water (Mutschler, Nievergelt & Feistner 1994), with man-made channels often marking territory boundaries. No systematic investigation of their habitat preferences in terms of marsh fragmentation and plant species diversity has been undertaken.

The dramatic reduction of marsh extent in Alaotra due to conversion to rice paddies, from 60 000 to 80 000 ha to below 30 000 ha in about 50 years (Mutschler & Feistner 1995), has historically been the most important factor affecting lemur numbers (Mutschler & Tan 2003). Poaching occurs, but more recently anthropogenic burning of the marsh is thought to have been a major driver of population reduction (Ralainasolo et al. 2006), from perhaps 11 000 individuals in 1994 (Mutschler & Feistner 1995) to possibly less than 2500 in 2002 (Ralainasolo et al. 2006).

Field data

Fieldwork to survey for H. alaotrensis was carried out from April to June 2008. Four villages in distant areas of the marsh (Anororo, Andilana-Atsimo, Ambodivoara and Andreba-Gare; Fig. 2) were used as bases for data collection, with transects totalling approximately 50 km radiating from these. Transects followed existing channels (typically 1–3 m wide) through the marsh, normally used by fishermen. Surveys were repeated between 3 and 12 times (6 in average), with at least half a day separation between repetitions, recording the GPS position of each direct sighting of H. alaotrensis.

Figure 2.

 Habitat suitability map for H. alaotrensis as a continuous index (logistic output), obtained from model M01 in Table 2. Empty patches correspond to clouds in the Landsat image.

Landsat7 imagery

We used a Landsat7 satellite image (path 158/row 72) taken on 22 March 2007, acquired from the U.S. Geological Survey as gap-filled Level 1G (radiometrically and geometrically corrected) georeferenced GeoTIFF files. The gaps created by the Landsat7 scan-line correction problem had been filled using images from the rainy season of the same year. The 2008 Landsat7 images of the Alaotra area during the rainy season were heavily obscured by cloud and the March 2007 images were considered the best available representation of the conditions of the marsh at the time of fieldwork. We used several spectral bands from the Enhanced Thematic Mapper Plus (ETM+), the onboard earth-observing sensor: TM1 (Blue), TM2 (Green), TM3 (Red), TM4 (near-IR), TM5 (mid-IR) and TM7 (mid-IR), with a pixel resolution of 30 m and pixel values (reflectance) between 0 and 255. The average H. alaotrensis territory (2 ha) would correspond approximately to 5 × 5 pixels and the maximum (8 ha) to around 9 × 9 pixels. To avoid small-scale spatial autocorrelation (Guisan & Zimmermann 2000) only one observation was retained for the analysis if several fell within the same or neighbouring pixels, resulting in a total of 46 H. alaotrensis usable locations of the original 88. The average nearest neighbour and inter-location distances were 341 m and 10 581 m, respectively. The raw bands were pre-processed to correct a spatial error of around 100 m, estimated manually using GPS tracks recorded along narrow channels. As H. alaotrensis lives exclusively in marsh vegetation, the marsh area was delimited visually in TM5 and extracted. We identified and masked out a few scattered clouds using an unsupervised classification.

Habitat suitability modelling and validation

The maximum entropy approach (Maxent) is used for characterizing probability distributions from incomplete information (Jaynes 1957). In habitat suitability modelling it attempts to find the distribution of maximum entropy (i.e. least constrained) that still agrees with all observed data: the value of the environmental variables at the locations where the species has been observed. The method does not require absence data, although randomly selected information from the environmental variables is used to characterize the features. It has been implemented as the free software ‘Maxent’ (Phillips, Anderson & Schapire 2006). The method is vulnerable to bias in the input data (Phillips, Dudik & Schapire 2004), however it benefits from the advantages inherent in a presence-background method, such as immunity to false absences. Maxent performs well compared to other modelling methods (Elith et al. 2006), including when few presence data are available (Hernández et al. 2006), making it especially attractive in data-poor regions.

The Area Under the ROC Curve (AUC) was calculated for model validation, providing a threshold-independent evaluation statistic of model performance and avoiding the intermediate step of selecting a threshold over the continuous prediction index. AUC, the standard method to assess accuracy (see Lobo, Jiménez-Valverde & Real 2008 for criticisms) is derived from the Receiver Operating Characteristic (ROC) curves, created by plotting sensitivity (proportion of observed occurrences correctly predicted) against 1-specificity (proportion of observed absences correctly predicted) for all possible thresholds (Pearson 2008). The AUC value is the area under the resulting curve: 0·5 indicates a prediction no better than random while the closer the values are to 1 the better the model is able to predict. With presence-only methods like Maxent, specificity cannot be calculated so the standard is to use pseudo-absences (random background points) instead of absences. The classification problem now consists of distinguishing presence from random, instead of from absence. The maximum AUC value is therefore less than 1 (Phillips, Dudik & Schapire 2004).

Derived predictor variables

We extracted candidate predictor variables from Landsat images (Fig. 1). Although raw Landsat bands can convey habitat information (e.g. open water is easily differentiated from vegetation in infrared bands), derived variables can be better predictors than raw ones (Wintle, Elith & Potts 2005). H. alaotrensis uses marsh vegetation for food, shelter, and movement over water and some areas of marsh may be too fragmented or not dense enough, after fires and floodings, to sustain a population. The derived variables attempted to describe vegetation characteristics including habitat structure. Our hypothesis was that high habitat suitability for H. alaotrensis would be positively related with plant productivity but negatively with marsh fragmentation.

Vegetation indices are among the few satellite-derived variables commonly used in habitat suitability models (St-Louis et al. 2009). They are based on the fact that photosynthetically active green vegetation has a typical spectral pattern that differentiates it from other common materials on Earth. For this study we calculated the commonly used NDVI (Normalized Difference Vegetation Index (Mather 2004, p. 142) related to plant productivity) and two indices that attempt to improve NDVI’s performance in terms of resistance to atmospheric effects (ARVI: Atmospherically Resistant Vegetation Index; Kaufman & Tanré 1992) and nonlinear relationships with surface parameters (NLI: NonLinear vegetation Index; Chen 1996). Two linear transforms of raw bands, albedo (total reflectance) and tasseled cap transform KT3 that reflects wetness were computed (see Appendix S1, Supporting Information). Our Landsat7 image was collected during the rainy season, ensuring marsh vegetation was actively growing and we expected NDVI-related indices and albedo to be positively correlated with habitat suitability (low values reflecting lack of healthy vegetation) but to show a negative relationship for KT3 (high values indicative of water or flooded disturbed marsh).

Image texture measures quantify the spatial variability in pixel brightness values and have been used as surrogates of different aspects of habitat heterogeneity (St-Louis et al. 2009). We explored first-order (occurrence of pixel values) and second-order textures (relative spatial distribution of pixel values) as predictors for H. alaotrensis habitat suitability. These variables were calculated with ArcGIS9 and MATLAB based on a moving window (kernel) with sizes ranging from 3 to 11 pixels, applied on either NDVI or an unsupervised classification (Mather 2004, p. 207) of the marsh in three broad classes that correspond to open water (3); flooded fragmented marsh and fields (2); and compact marsh vegetation (1). This intermediate variable was created with ArcGIS9, using NDVI and the raw ETM+ bands. We verified visually that the classification accuracy was adequate based on our knowledge of the areas surveyed. We created the following texture measures:

  • 1 Shannon’s diversity index (Shannon 1948) was calculated for the classified image as a first-order texture that reflects the relative abundance of the three broad classes. This index has been successfully applied to habitat suitability modelling (Mestre, Ferreira & Mira 2007). We expected habitats with high values (the three classes in similar proportions within the window) to be of low suitability for H. alaotrensis.
  • 2 The percentage of pixels within the kernel belonging to each class, a measure of first-order texture. Proportions of different relevant variables within a kernel have been used in habitat studies (Wintle, Elith & Potts 2005). In our case, areas with a high percentage of marsh were expected to have higher suitability.
  • 3 Semivariogram of the NDVI raster, a second-order measure which describes the spatial correlation of the pixels that lie at a certain distance from each other, calculated as half the average difference squared between the values of all combination of pixels located at distance d. The variable was calculated for adjacent pixels (‘d = 1’, distances of 1 and 1·4 pixels) and also for ‘d = 2’ (distances of 2 and 2·2 pixels). The semivariogram has been applied to remotely sensed NDVI and habitat studies (Bayliss, Simonite & Thompson 2005). We expected high values of the semivariogram, characteristic of fragmented marsh, to appear less suitable for H. alaotrensis.
  • 4‘Power content at high frequencies’, based on the fact that an image with rapid changes in neighbouring pixels will have a larger amount of high frequency information on its 2D frequency spectrum (Mather 2004, p. 169). For each pixel on the NDVI raster, the power in the higher part of the 2D spectrum for a kernel around it was calculated using a 16-point Fast Fourier Transform. This spectral variable represents a measure akin to a spatial domain second-order texture like the semivariogram, but is calculated in the frequency domain. We therefore expected high values to be associated with low suitability, as they would represent kernels with frequent transitions between high and low NDVI pixels, that is, areas of fragmented habitat at the pixel grain. Mather (2004, p. 162) provides an accessible explanation of the application of the frequency domain operations to remote sensing, while Proakis & Manolakis (2006) offer a more theoretical account. The Fourier transform has been applied to remote sensing data for several purposes, including image enhancement, detection of periodicities and the derivation of texture measures (Mather 2004, p. 169, 239) but to our knowledge this is the first application of this predictor of habitat structure in habitat modelling.

The resolution or grain of Landsat images (30 × 30 m) is finer than the minimum territory size of H. alaotrensis, allowing us to explore how landscape features affect habitat suitability at different scales. We expected relevant relationships to be related to territoriality and therefore calculated predictor variables at a range of scales from pixel size (30 m) to slightly above maximum territory size (330 m, 11 pixels). We derived NDVI first-order textures by averaging it with ArcGIS9 over circular windows of increasing size. The other image textures were also produced with kernel sizes from 3 to 11 pixels. Altogether 50 variables were selected as candidate predictors for modelling (see Appendix S1).

Exploration of variables, model selection and habitat suitability map

We built habitat suitability models with Maxent software (version 3·2·1) based on each group of related candidate variables, to compare their relative predictive performance, including the effect of scale (Fig. 1). Recommended default values for the convergence threshold (0·00001), maximum number of iterations (500), background points (10 000) and regularization multiplier (β = 1·0) were used (Phillips, Anderson & Schapire 2006). H. alaotrensis presence points were randomly divided into calibration (training) and evaluation (test) sets (25% samples for evaluation), and ROC curves and AUC figures were obtained. Jackknife tests were run during model exploration: the modelling process was repeated, first excluding each variable of the group, then using only one variable in turn. Evaluation AUC is preferred over calibration AUC for model comparison since validation with independent evaluation data (which better reflects a model’s capacity to generalize) is generally preferred over verification with calibration data (Araújo & Guisan 2006). We used bootstrapping to avoid AUC values being influenced by particular lemur sightings: each modelling exercise was run 1000 times with different random partitions of evaluation and calibration sets, and evaluation AUC figures were averaged.

The best predictor variables from each group at their optimal scale were then selected. After testing for multicollinearity by examining cross-correlation and removing highly correlated variables, the remaining ones were used in a stepwise backwards model selection process. Starting from the saturated model (with all variables), we created models with Maxent by removing one variable from the saturated set at a time. The model with highest mean evaluation AUC (average over 1000 repetitions with different partitions of evaluation and calibration sets) was selected and subsequently used as the saturated model for the next step of the selection. This process was repeated until the simplified models performed worse than the saturated model for that step.

A logistic map of habitat suitability of the marsh for H. alaotrensis was produced with Maxent based on the final model, using all available samples (Fig. 1). Maxent maps are most intuitively interpretable in a logistic format, giving a continuous index of suitability (Phillips & Dudik 2008). It is generally better practice to avoid binary prediction maps (suitable/unsuitable), unless the model is well calibrated or it is essential for the application (Wintle, Elith & Potts 2005). For demonstration purposes, we created a 3-category map using two thresholds calculated by Maxent: the maximum threshold that predicts all training samples (low-to-mid) and the value that minimizes the difference between sensitivity and specificity (mid-to-high), a threshold recommended for conservation applications (Jiménez-Valverde & Lobo 2007). The boundaries of a planned protected area in Alaotra (Andrianandrasana et al. 2005) were compared to the categorical suitability map and we calculated the amount of each suitability category in the core protection zone and in the overall protected area, as well as the percentage of each suitability category in the marsh covered by these two protection levels. Threshold-dependent binomial tests based on omission and predicted area (Phillips, Anderson & Schapire 2006) were carried out for the best model at these two thresholds with 1000 random partitions of calibration and evaluation sets (25% samples for evaluation) to test if the model predicted evaluation samples better than random (with one-tailed P-values referring to the null hypothesis that the number of correctly predicted evaluation samples can be obtained by chance according to a binomial distribution).


Predictor variables selection

The raw bands with highest average evaluation AUC were TM7, TM5 and TM2 and the best vegetation indices NDVI, albedo and KT3. The best ‘averaged NDVI’ texture had a 5 pixel (150 m) diameter, similar to the average H. alaotrensis territory. The PQ1 variables (percentage of class 1: marsh) gave better predictions than PQ2 or PQ3, and the best kernel size for PQ1 was 9 × 9 pixels (270 × 270 m), approximately the maximum territory size. The Shannon diversity index performed worse than these ‘percentage of classes’ for all kernel sizes tested and was consequently not selected as predictor. The semivariogram was a better predictor at distance 1 pixel than 2 pixels, with a downwards trend in evaluation AUC for increasing kernel size. The best ‘power content at high frequencies’ was for 3 × 3 pixels. This variable had a correlation of 0·743 with the semivariogram at 1 px and 3 × 3 kernel. NDVI and TM5 were discarded for their high correlation with other variables (Pearson’s r > 0·85), leaving a final set of eight candidate predictors (Table 1).

Table 1.   Candidate predictors used in the stepwise selection, chosen by comparing with jackknife tests the variables within each of the following groups: raw ETM+ bands; all vegetation indices; averaged NDVI for different kernel sizes; ‘power at high frequencies’ for different kernel sizes; semivariogram at d = 1 and d = 2, for different kernel sizes; percentage of pixels in class x and the Shannon diversity index for different kernel sizes
Variable nameMeaningScale
TM2, TM7Raw Landsat7 ETM+ bands 2/71 pixel (30 m)
NDVI5pxNDVI averaged over circular window, radius = 75 m (Ø = 5pixels)5 pixel diameter (150 m)
ALBAlbedo (TM1 + TM2 + TM3 + TM4 + TM5 + TM7)1 pixel (30 m)
KT3Tasseled cap transformation 3: ‘Wetness’1 pixel (30 m)
PHF_3 × 3‘Power on high frequencies’ over a 3 × 3 pixel kernel on NDVI3 pixel (90 m)
PQ1_9 × 9 Percentage of class 1 (compact marsh vegetation) pixels in a 9 × 9 kernel 9 pixel (270 m)
SV1_3 × 3Semivariogram value for distance 1 pixel, on a 3 × 3 kernel on NDVI3 pixel (90 m)

Final model

The final model (Table 2) had an evaluation AUC of 0·861 and the predictors included two vegetation indices (albedo and KT3), two NDVI-derived textures as measures of habitat structure (first-order ‘averaged NDVI’ and second-order semivariogram), but did not include any of the raw ETM+ bands. Binomial tests of omission showed that the selected model predicted evaluation localities significantly better than random for both thresholds considered, minimum training presence (0·042) and equal training sensitivity and specificity (0·303). The average P-values were 0·0083 and 0·0042, respectively, with P-value significant at α = 0·05 level in at least 98% of the individual runs. The corresponding mean evaluation omission rates were 0·046 and 0·257.

Table 2.   The eight best models in the stepwise selection process, ordered by decreasing value of mean evaluation AUC (over 1000 repetitions). The variables are described in Table 1. The saturated model (with all variables) is shown for reference
ModelPredictor variablesNumber of variablesMean evaluation AUC
M01NDVI5px, ALB, KT3, SV1_3 × 340·861
M02NDVI5px, ALB, KT3, SV1_3 × 3, PHF_3 × 350·858
M03NDVI5px, ALB, KT330·856
M04NDVI5px, ALB, KT3, PQ1_9 × 9, SV1_3 × 3, PHF_3 × 360·855
M05TM2, NDVI5px, ALB, KT3, PQ1_9 × 9, SV1_3 × 3, PHF_3 × 370·854
M06NDVI5px, ALB, KT3, PQ1_9 × 9, SV1_3 × 350·854
M07TM2/7, NDVI5px, ALB, KT3, SV1_3 × 3, PHF_3 × 370·853
M08TM2/7, NDVI5px, ALB, KT3, PQ1_9 × 9, SV1_3 × 370·853
M14TM2/7, NDVI5px, ALB, KT3, PQ1_9 × 9, SV1_3 × 3, PHF_3 × 380·847

NDVI-related indices reflect vegetation productivity and as we expected the results suggested higher suitability for areas with high plant productivity, that is, healthy marsh. For the semivariogram, areas with high variations in NDVI between adjacent pixels (typical of alternating open water and marsh pixels in fragmented marsh) showed lower suitability, reasonable since a 1-pixel area would be too small for a territory. Lower wetness (KT3, which we believe associated with dense healthy marsh) contributed to suitability, and a low albedo (characteristic of open water) helped discriminate less suitable areas.

Habitat suitability map

The logistic habitat suitability map (Fig. 2) was generally consistent with the areas sampled during field surveys (e.g. degraded areas along transects), pictures taken from two vantage points around the lake (large tracts of undisturbed marsh appear as high suitability), and local expert knowledge and experience. Using the chosen thresholds to categorize suitability in three levels, 33% of the study area was classified as low suitability, 44% as mid suitability and 23% as of highest suitability (Table 3). From the overlap of the planned protected area (Fig. 3), we calculated that only about 34% of the core area is considered of high suitability for H. alaotrensis, while this percentage drops to 25% when the whole protected area is considered (Table 3).

Table 3.   Area and percentage of each of the three habitat suitability categories defined by two logistic thresholds calculated by Maxent: ‘minimum prediction that corresponds to a presence record’ (0·042) and ‘equal training sensitivity and specificity’ (0·303). We consider pixels with probability below the lowest threshold to correspond to habitat of low suitability; between the two thresholds to mid suitability; and above the highest threshold to high suitability. The last four columns show the percentage of overall or core protected area (PA) that belong to each category, and the percentage of the total suitability category area that is covered by the overall protected area and by its core protection area (see Fig. 3)
Suitability categoryTotal area (ha and %)% Overall PA area% PA core area% Total category covered by overall PA% Total category covered by core area
Low10 735(33)28·915·357·610·6
Mid14 215(44)45·950·969·226·5
High 7 403(23)25·233·873·033·8
Figure 3.

 Limits of the planned Protected Area (PA) in Alaotra displayed over a categorical map of habitat suitability based on Fig. 2, with suitability thresholds defined in Table 3. The core zone is shown separately.


This study demonstrates the feasibility of using Landsat imagery for developing habitat suitability models in data-poor situations where conventional macro-scale environmental predictors are of limited use. The model created for Alaotra had good predictive power (with random sets of evaluation presence records) and was developed with a relatively fine grain (30 m) compared to the average H. alaotrensis territory size (2 ha). The method was cost-effective; data were collected during a relatively short fieldwork season and predictors like vegetation indices and image textures were derived from Landsat images, avoiding detailed ground-truthed land cover mapping.

Model assumptions

When presence records of a species are used for modelling its niche or distribution, two factors need to be taken into account (Pearson 2008): how well the sampling represents the range of environmental conditions; and the assumption of equilibrium in the environmental conditions. Sampling bias may be an issue in this study if the values of the predictor variables along the transects were not statistically representative of the whole area considered. The selection of our transects was limited by the inaccessibility of the marsh and thus the model would benefit from a more systematic sampling along the range of predictor variable values. Secondly, equilibrium cannot be assumed in the Alaotran marsh as it is a dynamic system, both hydrologically and with parts burnt every year that over time shift into different vegetation classes, or turn into open water (Andrianandrasana 2009). The Landsat image in this study was captured 1 year before the survey, which could have an impact on the results. However, the methodology applied can be readily replicated with new data, providing updated maps of habitat suitability to reflect changes in the marsh. There is some evidence that there has been a reduction in fires over recent years (Andrianandrasana 2009); if this trend continues the medium-term validity of these maps will increase.

These maps describe currently suitable habitat rather than actual H. alaotrensis occupancy. Areas that are delineated as suitable may in fact be unoccupied due to factors like fire history, hunting pressure and human presence. The dynamics of this process, including the recolonization of recovered burnt areas by lemur groups, are beyond the scope of static habitat suitability modelling (Guisan & Zimmermann 2000). For example, during this study a group of H. alaotrensis was seen in an area of degraded habitat, presumably pushed there by recent marsh fires. In such cases, we might be training the model with sightings from a population that might not be viable, acting as a sink (Wintle, Elith & Potts 2005).

Landsat7 data are not without problems. The image gap-filling process (required since the ETM+ SLC failure in 2003) created strips in the raw bands with slightly different reflectance values. Although not readily noticeable in our habitat suitability map, this process may represent a limitation of the use of recent Landsat7 images for other studies, particularly if the study area is very large, it falls away from the satellite path or the seasonality is relevant; alternative remote sensing imagery like Landsat5 TM could be considered in such cases (Trigg, Curran & McDonald 2006).

Habitat suitability

The contribution to habitat suitability of the predictors in the final model reinforces our understanding of the ecology of our focal species. As expected, healthy marsh vegetation was an important determinant of suitability for H. alaotrensis. Marsh fragmentation on the other hand reduced suitability, as indicated by image textures like the amount of healthy marsh (NDVI5px, a first-order measure) and its relative spatial distribution (semivariogram, a second-order measure). As seen in previous studies (Gottschalk, Huettmann & Ehlers 2005) the spatial scale of the heterogeneity variables was important in determining their predictive power, and we obtained predictors of suitability that range from pixel scale up to the average territory size. This suggests that fragmentation affects the species at different scales, from impeding the movement of the lemurs along discontinuous vegetation to reducing the amount of healthy marsh within a potential territory.

We did not investigate the impact on suitability of some aspects of vegetation, such as species composition or vertical structure, as we could not relate them to the derived predictors. This point highlights a limitation for ecological inference when using only satellite data, which produces variables that cannot necessarily be related to physical features in the absence of field measurements.

Management implications

When properly obtained and interpreted, maps of suitable habitat are invaluable tools for conservation, allowing for spatial prioritization of interventions. Our comparison of a planned protected area in Alaotra with the habitat suitability map highlighted that only 34% of the high suitability category was covered by the core protected area, and that 15% of the core protected area consisted of low suitability marsh. Although such maps can be useful, it is important to remember that conservation planning is a multi-faceted activity and other issues must also be considered. Firstly, the habitat suitability map does not reflect actual occupancy, and secondly, the boundaries of the Alaotra Protected Area are a trade-off between biodiversity and local people’s needs.

Broader applicability

Satellite images have been used in the past for modelling species distributions, often through land cover classification, at broader geographical extents and with samples located in distant sites in a macro-scale context, where they are typically accompanied by environmental layers such as climatic and topographic variables [see Gottschalk, Huettmann & Ehlers (2005) and Leyequien et al. (2007) for extensive reviews]. In our case, because H. alaotrensis has a geographically limited range in a flat and featureless marsh and we were interested in habitat occupancy at the landscape level, climatic and topographic variables were not informative. We computed vegetation indices and image textures from raw Landsat images, avoiding a ground-truthed detailed habitat classification. Textures derived from remote sensing data have been recently applied to the prediction of species diversity (St-Louis et al. 2009) and occurrence (Bellis et al. 2008) in different habitats. We show their usefulness in marsh ecosystems, adding to the growing body of evidence supporting their potential for habitat suitability modelling. We are not aware of other studies applying them in a Maxent model, or to the study of an endangered species. By describing the physical characteristics of the habitat that are captured by the predictors in our study, we provide guidance to their applicability for other habitats and species, emphasizing the generalizability of the method beyond our particular case study. Our methodology is particularly applicable to other species in similar ecological circumstances, but is also relevant when there are major logistical or budgetary constraints on field surveys and ground-truthed detailed land cover mapping is considered impractical or not relevant. It could therefore play an important role in rapid assessments for conservation planning, particularly in data-poor regions as well as when other environmental predictors are difficult to obtain or the readily available macro-scale level ones are of little use. In an encouraging move, in 2009 the U.S. Geological Survey made all their new and archive Landsat images freely downloadable through their website, offering a wealth of potentially valuable information. Free high quality satellite imagery could be an excellent foundation for cost-effective conservation prioritization and planning exercises. More testing of the strengths and limitations of these data for applied ecology in different ecosystems, as well as innovative methods for coupling data with habitat suitability models, are urgently needed.


We thank the Ministry of Environment, Water and Forests of Madagascar for granting the permit to conduct this study; the Leverhulme Trust, a Royal Society Wolfson Research Merit award to EJMG, and a Marie Curie Fellowship (‘EcoEcoMonitoring’) to EN for financial support; John Fa, Richard Lewis, Jonah Ratsimbazafy, Herizo Andrianandrasana and other Durrell Wildlife Conservation Trust personnel in Madagascar, especially Bary Jean Rasolonjatovo in Ambatondrazaka; our local guides Richard Rasolonjatovo and André Rakotonierana; and Joaquín Hortal and two anonymous reviewers for comments that greatly improved the manuscript.