Evaluating the effect of satellite image resolution for predictive habitat mapping: a sensitivity analysis in the Cantabrian Mountains (NW Spain)

A current biodiversity conservation challenge is to estimate the spatial extent of habitat types as defined by the guidelines of environmental policies such as the Habitats Directive (92/43/EEC) and moving beyond landscape patterns of broader vegetation types. In the absence of field-based fine-resolution maps (Evans 2006; Vanden Borre et al. 2011), predictive models based on environmental variables can represent a valuable tool, helping the assessment of the local area of occurrence (AOO) of vegetation communities (Figure 1) (Ferrier & Guisan 2006; Elith, Kearney & Phillips 2010; Rodríguez et al. 2015). However, the implementation of these tools is still uncommon in regional conservation planning (Wiens et al, 2009). Among other factors, vegetation mapping is hampered by the complexity of natural systems due to environmental or successional gradients, plant traits and human disturbances acting altogether at different spatial and temporal scales (Álvarez-Martínez et al. 2014). Recent developments in data availability and processing methods have linked this monitoring task to remote sensing (RS) (Nagendra et al. 2013; Corbane et al. 2015). Satellite imagery temporal, spectral and spatial resolutions allow the exploration of different vegetation status and phenology aspects, which may lead to a proper habitat type discrimination (Lillesand et al., 2008). In this study, we combined data obtained from different sources: i) an intensive vegetation survey across a Special Area of Conservation of the Cantabria region (Spain), dominated by peatlands, bogs, grasslands and heathlands, ii) abiotic limiting factors (i.e. topography, climate and soil properties), iii) LiDAR data informing about vegetation height and structure, and iv) three different satellite images (Figure 2). In order to study the role played by satellite data resolutions, we first chose two Landsat8 OLI images (30-m pixel size) collected during the spring and summer of 2016, to evaluate the predictive skills of temporal resolution. Secondly, we used a Sentinel-2 MSI image from summer 2016, in order to assess the combined effect of a 10-m pixel size and enhanced near infrared spectral capabilities. Lastly, we analyzed a Deimos-2 image, with a pixel size of 3.3 meters. A Landsat summer image was used to generate baseline models, to be compared to the former three approaches. All maps were validated at different levels. We calculated ommision-commission errors and overall accuracy scores by using independent field checked points for the habitat types encountered across the study area. Subsequently, we determined how each modelling aproach (i.e. using temporally, spectrally and spatially enhanced satellite imagery) differed from the baseline approach of a single Landsat image. The results of this study represent and important step towards improving, in complex landscapes, the identification of habitat type occupancy areas under current Global Change effects.