Using remote sensing to map habitat mosaics: an integrated approach applied to the classical Karst

Remote sensing is a well-established tool for habitat mapping, but its use is still challenging in heterogeneous landscape mosaics. Novel approaches to improve classification performance include multitemporal data and multiple remotely sensed variables, but have rarely included spectral heterogeneity (SH) measures. The aim of this study was to develop an integrated approach to map the natural habitats in Classical Karst (NE Italy), by quantifying the importance of SH measures and providing a robust framework to include multi-temporal remotely sensed data. First, a habitat map was produced from field surveys. Then, a collection of 12 monthly Sentinel-2 images was retrieved using the Google Earth Engine platform. Vegetation and SH indices were computed and aggregated in four temporal configurations: (1) monthly layers of vegetation and SH indices; (2) seasonal layers of vegetation and SH indices; yearly layers of multi-temporal SH indices computed (3) across the months, and (4) across the seasons. For each temporal configuration, a Random Forest classification was performed, first with the complete set of input layers and then with a subset obtained by Recursive Feature Elimination. Training and validation points were independently extracted from field data. The maximum overall accuracy (OA = 0.72) was achieved with the seasonal temporal configuration, after the number of habitat classes was reduced by aggregation from 26 to 11. SH measures allowed to improve the accuracy of the classification and the spectral β-diversity was the most important variable in most cases. Spectral α-diversity and Rao’s Q, on the other side, had a low relative importance, possibly due to the small spatial extent of the habitats. Regarding the inclusion of multi-temporal data, the aggregation of monthly data in seasonal median composites proved to be the best approach, since it allowed to reduce the number of input layers without losing accuracy. The approach developed in this study allows to improve habitat mapping in complex landscapes in a cost- and time-effective way, suitable for monitoring applications. Moreover, the results suggest that image classification frameworks could benefit from the inclusion of SH measures, that have rarely been included before.