"Dataset for the paper: A First Systematic Assessment of the Segment Anything Model for Zero-shot SAR-based Flood Mapping"

"Timely and reliable information on flood extent isessential for flood risk management and emergency response.However, conventional deep learning approaches typically dependon task-specific training and large volumes of labeled data,which constrains their applicability during rapidly evolving floodevents. This study evaluates a zero-shot workflow for floodmapping that combines the Segment Anything Model (SAM) withthe geospatial samgeo interface applied to Sentinel-1 SyntheticAperture Radar (SAR) imagery. SAM, originally trained onoptical RGB images, is used without retraining on SAR-derivedRGB and RGB+DEM composites for two major flood eventsin New South Wales (Australia) and southern Mozambique, toevaluate its performance for flood mapping from SAR data. Segmentationoutputs are benchmarked against Copernicus GlobalFlood Monitoring (GFM) products. The results show that SAMcan produce useful flood masks from the SAR data without anycase-specific retraining, particularly when guided by interactiveforeground and background point prompts. In the most favorablescenes, the agreement with GFM reaches IoU values of about67% and 74% and Dice coefficients above 80%, whereas denseautomatic prompting, bounding boxes, and text-based promptsproduce consistently weaker and less stable masks, with frequentover- and under-segmentation. The inclusion of DEM as a thirdinput channel does not consistently improve segmentation quality.In general, the findings indicate that SAM + samgeo constitutesa promising, training-free approach for rapid, operational SARbasedflood mapping, especially under cloud-covered conditions,while highlighting the need for SAR-specific adaptations andmore robust automatic prompting strategies to fully match theperformance of dedicated optical and SAR-based systems."