A transferability study of the kernel-based reclassification algorithm for habitat delineation

Wetland mapping using Earth observation (EO) data has proved to be a challenging task for practitioners due to the complexity in the spatial structure and composition, the wide within-class spectral variability and the absence of easily distinguishable boundaries between habitat types. Furthermore, the inherent temporal water instability of these landscapes poses an obstacle to the integration of field data with remote sensing data, which also are not acquired simultaneously at all times. e with these limitations we tested the applicability of the Kernel-based reclassification (KRC) algorithm on very high spatial resolution (VHR) satellite imagery over a wetland. A composite multi-temporal (i.e. dual-date) VHR WorldView-2 image consisting of spectral bands and indices derived from two images acquired during flooded and dry water conditions were employed. This dataset stresses the seasonal variations of the habitat in response to environmental changes (i.e. flooding) occurring between the two acquisition dates. Multi-temporal imagery is an important information source for fine mapping of wetlands such are river deltas. A multi-temporal approach could reveal even more specific information during the phenology of these habitats. thodology was applied firstly to Axios and then to Aliakmonas river deltas in Northern Greece. The results revealed an overall accuracy of 53% in the first and more complex site, and 86% in the second site.