Multi-Source Remote Sensing Data Fusion and Ensemble Machine Learning for Flood Inundation Mapping

The Brahmaputra River basin in Assam experiences catastrophic annual flooding affecting 4-8 lakh hectares and displacing 3-6 million people each flood season, with the 2023 flood season resulting in 174 deaths, agricultural losses of ₹4,200 crore, and infrastructure damage to 2,847 km of roads. The combination of the Brahmaputra's high sediment load, braided channel morphology, and the monsoon's concentrated rainfall in the Eastern Himalayan catchment creates flood events of unpredictable spatial extent that challenge both the accuracy of hydrological models and the timeliness of emergency response. Satellite remote sensing — and specifically Synthetic Aperture Radar (SAR) imagery from ESA's Sentinel-1 constellation, which penetrates cloud cover and provides 6-day revisit regardless of weather — offers the only reliable source of near-real-time flood extent mapping during active monsoon flood events when optical imagery is obscured. This paper presents an ensemble machine learning framework for flood inundation mapping from multi-source remote sensing data fusion, combining Sentinel-1 SAR (VV/VH polarisation), Sentinel-2 optical (NDWI, NDVI), SRTM DEM-derived topographic features, SMAP soil moisture, and Landsat-8 historical water occurrence. A six-class land cover classification (open water, flooded vegetation, dry land, urban, agriculture, forest) is implemented using an ensemble of Random Forest, XGBoost, and a U-Net convolutional neural network, achieving overall accuracy of 97.6% and Cohen's kappa of 0.962 on the 2023 Assam flood validation dataset. An early warning component evaluates Probability of Detection (POD) and False Alarm Rate (FAR) at lead times of 1-8 days using ECMWF extended-range ensemble forecast precipitation as the upstream trigger.