A SAR-verified flash flood event dataset for the haor wetlands of northeastern Bangladesh (2014–2024)

This dataset contains 77 flash flood and non-flood events (December 2014 to October 2024) for the haor wetlands of the Sunamganj/Sylhet region, northeastern Bangladesh, a data-scarce area where pre-monsoon flash floods driven by rainfall over the upstream Meghalaya foothills regularly destroy Boro rice just before harvest. Each event carries a binary flood label compiled from documentary sources (Flood Forecasting and Warning Centre reports, UN/NAWG situation reporting, and peer-reviewed literature), assigned independently of the satellite features to avoid circularity. For each event date, 14 features were extracted from real satellite imagery and reanalysis products: Sentinel-1 VV/VH backscatter over the haor area and over an upstream Barak River reference area (Silchar, India), Sentinel-2 NDWI, and ERA5-Land antecedent rainfall (24-hour, 96-hour, and 7-day accumulations), soil moisture, air temperature, temperature anomaly, and wind speed. Flood extents were verified with Otsu-thresholded Sentinel-1 change detection, cross-checked against Sentinel-2 NDWI where cloud-free optical imagery existed. The dataset contains no synthetic records. NDWI is empty for 5 of 77 events with no cloud-free Sentinel-2 scene; this missingness is documented rather than gap-filled. Two static terrain columns (slope, TWI) are constant single-AOI values. Files: haor_flood_events_v1.csv (77 rows × 18 columns), data_dictionary.md (full column documentation), README.txt. The dataset is intended as a benchmark for machine-learning flood prediction in data-scarce basins and for hydrological analysis of backwater-driven flash flooding. Given the sample size, evaluation protocols suited to small datasets (e.g., leave-one-out cross-validation) are recommended.