Evaluating the robustness of Bayesian flood mapping with Sentinel-1 data: A multi-event validation study

The impact of recent extreme flood events has once again highlighted the importance of accurate near-real-time flood information. Consequently, a number of operational services have been established that primarily use Synthetic Aperture Radar (SAR) data to map flood extent. Among them is the Global Flood Monitoring (GFM) service that is part of the Copernicus Emergency Management Service (CEMS). Using the systematic monitoring capabilities of Sentinel-1, it is the first service to deliver flood maps fully automatic on a global scale. To automatically and reliably monitor flood extent worldwide, the strengths and weaknesses of flood mapping methods need to be known under various and sometimes challenging conditions. To examine the performance of the TU Wien Bayesian flood mapping algorithm, which is one of the scientific flood algorithms used operationally in the CEMS GFM service, we designed this validation study in which we compare our results with all compatible Sentinel-1-based flood events of the CEMS on-demand mapping (ODM) service between January 2021 and January 2022. In total, the study investigates 18 events from five continents. In addition to computing common accuracy metrics, eight representative events were analysed in detail to understand the reasons for the differences found, identify potential improvements for the method, and gain generic insights for radar-based flood mapping. Most differences are caused by the use of the VH polarization in some of the ODM reference maps, while the GFM service so far relies exclusively on VV polarization due to computational costs. The impact of using two polarizations can be seen in particular over vegetation or in case of windy conditions. Furthermore, while the post-processing strategy applied in the TU Wien algorithm helps to prevent speckle impact, it also smooths out important details in small-scale flood events. Nonetheless, the automatic TU Wien algorithm achieved a Critical Success Index (CSI) of over 70% against the semi-automatic reference in 10 of 18 flood events. It exceeds this mark for all large-scale events and in cases without vegetation close to the flooded surfaces. Overall, the median User’s Accuracy (UA) is 84.0 %, the Producer’s Accuracy (PA) is 72.9% and the Overall Accuracy (OA) is 85.3%. The results demonstrate that the GFM service would benefit for using both VV and VH polarization and relaxing filters applied in the SAR processing workflow. • Insufficient validation efforts hinder progress in SAR-based flood mapping. • Performing detailed validation of a Bayesian flood mapping algorithm. • Selected unbiased 18 flood events that were quantitatively and qualitatively analysed. • Flood events are spread over five continents and pose various conditions. • Algorithm performs at a median Overall Accuracy (OA) of 85.3%.