Modelling species distributions with remote sensing data: bridging disciplinary perspectives


Over the last few decades, correlative species distribution models (SDMs) have been adopted as the most widely used approach for describing and predicting spatial patterns of relationships between species occurrence and environmental conditions (Elith & Leathwick, 2009). Over this same period, the discipline of remote sensing (RS) has produced a breadth of novel geospatial datasets and analytical algorithms for mapping biogeographical heterogeneity. It is not surprising that RS data are now commonly used in SDMs: space- and airborne RS data provide a low cost means to map environmental changes across multiple spatio-temporal scales and are attractive for their ability to measure spatial factors often impossible to quantify otherwise (e.g. landscape connectivity). Looking into this trend further, our literature search – with keywords ‘remote sensing’ and ‘species distribution’ or ‘habitat suitability’ – returned 210 articles where remote sensing was integrally used in SDMs, 60% of which were published just in the past 5 years (ISI Web of Science, 2 May 2013). This development should be a good thing, right? However, several scientists have critically pointed out that the content and spatial scale of RS predictors do not often match species' life-history strategies (e.g. Bradley et al., 2012; Lechner et al., 2012). Our objective here is not to disagree with their recently proposed methodological guidelines, but rather to reflect on the various facets of using RS in SDMs in an effort to bridge disciplinary perspectives.

Despite the rapidly increasing number of interdisciplinary approaches, we believe that some remaining discrepancies between geographical and ecological perspectives still influence current SDM studies. To illustrate this diversity of disciplinary viewpoints, we build on the proposition of Carl Troll, who coined the term ‘landscape ecology’ in an attempt to integrate the ‘spatial’ approach of geographers and the ‘functional’ approach of ecologists. In SDM studies, at one extreme, RS researchers may tend to focus on mapping spatial patterns of species distributions, while ecologists may be as much concerned with the explanations of these distributions. By this generalized point of view, we do not wish to draw boxes around disciplines but believe that inherited disciplinary perspectives, differing research priorities and associated methodological advancements remain the core reasons why the integration of RS in SDMs has not yet reached its full potential. Below, we discuss how the application of RS data in SDMs can be improved by (1) a greater awareness of the differing sample size and characteristics of RS and ecological data, (2) the combination of different RS predictors in multi-scale modelling frameworks, (3) the use of RS data to infer information on species absence, and (4) a clearer definition of the modelling purpose.

First, remotely sensed and field-based ecological data typically vary in their sample sizes and functional relationships to the focal species. Ecological perspectives may sometimes lead to using relatively small samples of data compared with RS; however, those field samples are often relatively accurate and typically capture information that is ecologically relevant to the distribution of the species (direct predictors, e.g. temperature or soil type). In contrast, RS data provide a wall-to-wall census of biophysical factors with relatively low accuracies (due to errors introduced during data acquisition, processing or analysis) which may serve only as indirect surrogates for functional variables (see Elith & Leathwick, 2009). Recognizing and accepting these differences between the two data sources is crucial. In particular, relationships between species occurrence and such indirect surrogates may be non-stationary in space and time. For example, the same values of the normalized difference vegetation index (NDVI) or other indices can be observed for habitat patches with completely different plant community compositions. While the importance of this spatio-temporal non-stationarity of RS signals along elevational or climatic gradients has been recognized in vegetation mapping (see Guyon et al., 2011), we do not see careful contextualization for the use of RS data in SDMs so far.

It is therefore essential to develop approaches that will allow us to extract ecological meaning from the censuses provided by RS data and design new RS indicators that capture direct environmental drivers. While this has been accomplished for modelling animal species by approximating habitat heterogeneity from RS data (Goetz et al., 2010), there is an apparent research gap for plant species. We specifically suggest continuing to explore the potential of functionally relevant biophysical parameters, such as leaf area index (LAI) or fraction of absorbed photosynthetically active radiation (fPAR), instead of using the more common vegetation indices (e.g. NDVI or enhanced vegetation index) that consist of combinations of spectral bands and are only indirectly linked to biophysical properties of vegetation. In addition, remotely sensed land surface temperature (LST), which is one of the most important parameters for quantifying surface energy and water balances (Quattrochi & Luvall, 2004), is a valuable, yet largely untapped source of data in SDMs.

Second, although the importance of considering scale-dependency of patterns and processes is well established in both disciplines, we see differences in implementation. Both ecologists and RS specialists widely accept that the choice of spatial scale, defined by extent (the geographical area considered) and grain (smallest measurement unit within a dataset), has to be driven by the objective and the potential application of the research. Obviously, there may be exceptions, where published studies fail to establish that their observation scale does correspond to the phenomena scale (at which the organism interacts with the environment; Lechner et al., 2012). Pearson & Dawson (2003) proposed a hierarchical framework for conceptualizing the scales at which different environmental factors affect species distributions. However, most RS-based studies that embrace multi-scale approaches compare the utility of the same index or the same biophysical parameter across various scales (e.g. from different sensors; Tarnavsky et al., 2008), rather than combining meaningful scale-dependent parameters. Future research should therefore address the use of different RS predictors in SDMs that would each match the phenomena scale or capture finer-scale factors that drive the species distribution and larger-scale controls that constrain it. For example, the LAI as an indicator of ecosystem productivity at a broad scale can be combined with remotely sensed data on vegetation structure, e.g. from LiDAR (light detection and ranging) data, at a fine scale.

Third, RS data are typically used in SDMs to inform about places in which a species is or can be present (e.g. NDVI; Pettorelli et al., 2011) but they are rarely applied to infer species absence. While absence data are often ignored in SDMs simply because they are not available for modelling, they also provide more complex information than a mere lack of suitability (Václavík & Meentemeyer, 2009; Lobo et al., 2010). For example, organisms can be absent in a location as a result of low suitability of abiotic conditions (environmental absence), but also due to factors restrictive on the pool of environmentally favourable areas such as biotic interactions, dispersal barriers or local disturbances (contingent absence). Here, we see a great opportunity for RS to help explain these two types of absences (or generate more reliable pseudo-absences) and ultimately create better SDMs. Land-cover classes derived from Landsat imagery or topographic data derived from the shuttle radar topography mission (SRTM) can easily be used to mask environment clearly unsuitable for the studied species. Continuing further in this direction, remotely sensed phenology indicators can help us incorporate contingent absences by identifying shifts in the growing season or disturbance dynamics that prevent the species from being found at otherwise environmentally suitable sites.

Fourth, the types of absence data and RS covariates also determine the capacity of SDMs to represent the potential or actual species distribution. These goals, however, are not always clearly defined in SDM studies. Those SDM studies that have a strong RS perspective typically focus on estimating actual species distributions by directly mapping dominant plant species, vegetation communities, or habitat for animal species (Kerr & Ostrovsky, 2003). In contrast, ecologists often apply SDMs to predict potential distributions, i.e. places environmentally suitable for species survival. However, both efforts require different methodological approaches and RS predictors. In the first case, SDMs should account for contingent absences and specifically use RS predictors that include the spectral signal of the target species. In the second case, they require incorporating environmental absences but should avoid capturing spectral signals of the focal species; the outcomes would otherwise underestimate species' potential ranges (Bradley et al., 2012). In addition, the overall accuracy that balances false positive and negative errors should not be the sole criterion used to include RS predictors in SDMs. While this approach may be appropriate to evaluate predictions of the actual distribution, using false positive rates makes little sense in evaluation of potential distribution because the true future distribution of a species is unknown. A rigorous approach hence requires that we clearly formulate the goals of our study, accordingly select RS predictors that contain or lack spectral signals of focal species and use adequate methods to evaluate their benefits.

We all agree that RS data are valuable complements to climatic and other environmental variables in SDMs. However, novel and creative approaches, including the combination of different RS data and techniques, are required to take full advantage of this potential. We are optimistic that increasing the awareness of the reviewed issues and bringing the disciplinary perspectives of ecology and RS together – making use of the best knowledge from both sides – will greatly advance the development and interpretation of SDMs in biogeography.