Advancing flood risk assessment: Multitemporal SAR-based flood inventory generation using transfer learning and hybrid fuzzy-AHP-machine learning for flood susceptibility mapping in the Mahananda River Basin

Abstract

The Mahananda River basin, located in Eastern India, faces escalating flood risks due to its complex hydrology and geomorphology, threatening socioeconomic and environmental stability. This study presents a novel approach to flood susceptibility (FS) mapping and updates the region's flood inventory. Multitemporal Sentinel-1 (S1) SAR images (2020–2022) were processed using a U-Net transfer learning model to generate a water body frequency map, which was integrated with the Global Flood Dataset (2000–2018) and refined through grid-based classification to create an updated flood inventory. Eleven geospatial layers, including elevation, slope, soil moisture, precipitation, soil type, NDVI, Land Use Land Cover (LULC), geomorphology, wind speed, drainage density, and runoff, were used as flood conditioning factors (FCFs) to develop a hybrid FS mapping approach. This approach integrates the Fuzzy Analytic Hierarchy Process (FuzzyAHP) with six machine learning (ML) algorithms to create hybrid models FuzzyAHP-RF, FuzzyAHP-XGB, FuzzyAHP-GBM, FuzzyAHP-avNNet, FuzzyAHP-AdaBoost, and FuzzyAHP-PLS. Future flood trends (1990–2030) were projected using CMIP6 data under SSP2-4.5 and SSP5-8.5 scenarios with MIROC6 and EC-Earth3 ensembles. The SHAP algorithm identified LULC, NDVI, and soil type as the most influential FCFs, contributing over 60 % to flood susceptibility. Results show that 31.10 % of the basin is highly susceptible to flooding, with the western regions at greatest risk due to low elevation and high drainage density. Future projections indicate that 30.69 % of the area will remain highly vulnerable, with a slight increase under SSP5-8.5. Among the models, FuzzyAHP-XGB achieved the highest accuracy (AUC = 0.970), outperforming FuzzyAHP-GBM (AUC = 0.968) and FuzzyAHP-RF (AUC = 0.965). The experimental results showed that the proposed approach can provide a spatially well-distributed flood inventory derived from freely available remote sensing (RS) datasets and a robust framework for long-term flood risk assessment using hybrid ML techniques. These findings offer critical insights for improving flood risk management and mitigation strategies in the Mahananda River basin.