A repeatable change detection approach to map extreme storm-related damages caused by intense surface runoff based on optical and SAR remote sensing: Evidence from three case studies in the South of France

Abstract Most flood hazards are induced either by river overflowing or intense overland flow following heavy rainfall, causing land surface damages under many forms. Until now, fine-scale detection of damages caused by intense rainwater runoff beyond the direct vicinity of major waterways has been scarcely explored using satellite remote sensing. In this work, three extreme storms in the Aude and Alpes-Maritimes departments in the South of France were investigated based on ground truths and very high resolution optical imagery (Pleiades satellite, IGN orthophotos). Plot delineation and land use information were combined to high revisit frequency and high resolution optical (Sentinel-2) and SAR (Sentinel-1) open-source data to test a simple automatic and replicable change detection method to locate damaged plots using supervised classification. Based on a unique training sample from the Aude floods of October 2018, combinations of plot-based spectral indicators allowed reaching overall detection accuracies greater than 85% on independent validation samples for all three events. A simple land use inter-class demeaning pre-processing used to account for land-specific seasonal variations improved event and site repeatability by lowering false detection rates down to a maximum of 13%. The benefits of introducing SWIR channel in addition to visible and near-infrared indices were limited to a few percentage points. SAR-derived proxies of soil moisture and roughness in weakly vegetated areas were consistent with the presence of degradations, with VV being the most sensitive polarization. However, classification accuracy was not significantly increased with Sentinel-1 data as compared to the exclusive use of Sentinel-2. Additional tests revealed that should the closest available optical images be rather distant in time because of persistent cloud cover, the method is reasonably robust as long as stable ground conditions were observed before the event. The need for images close in time was however emphasized through cross-site training. Indeed, efficient replicability from one site to another relied on using unaffected learning plots with slightly more inherent variability in time variations of spectral indices compared to the test site. Beyond the investigation of three case studies, this work demonstrates the performance and repeatability potential of a new probabilistic change detection method to expose various kinds of extreme rainfall-related disturbances, in particular those occurring far from the main hydrographic network. Should spatially accurate rainfall products be available, comprehensive mapping of intense stormwater runoff hazards using this original plot-based approach will then allow improving the understanding of overland flow generation mechanisms in hydrological models.