Flood susceptibility in a coastal backwater floodplain: a spatially validated multi-event SAR framework for Alappuzha, India

Flood susceptibility mapping (FSM) is dominated by studies of upland catchments. At the same time, near-sea-level coastal and backwater floodplains, where inundation is shaped by reclamation, land use, tidal influence and constrained drainage rather than relief, remain comparatively under-explored. Existing studies also commonly rely on single-event inventories and machine-learning models validated through random data partitioning, which can overestimate predictive skill in spatial data. This study addresses these gaps for the Alappuzha district, India, which contains the below-sea-level Kuttanad wetland. A reproducible multi-event flood inventory was developed from Sentinel-1 SAR imagery for three monsoon flood years (2018, 2019 and 2021) using change detection and automatic Otsu thresholding. Flood susceptibility was modelled using twelve hydrologically relevant conditioning factors and three approaches: Frequency Ratio, Random Forest (RF) and XGBoost, with model interpretation supported by SHapley Additive exPlanations (SHAP). RF achieved the highest conventional performance (five-fold ROC–AUC = 0.978). To assess robustness, predictive skill was stress-tested using spatial block cross-validation and a unique-location validation framework that removes both spatial leakage and repeated-pixel effects, yielding ROC–AUC values of 0.87 and 0.83, respectively, indicating good discrimination under the spatial validation schemes tested. SHAP analysis revealed that absolute elevation, land use/land cover and distance from the coast were the strongest contributors to predicted susceptibility. In contrast, slope- and wetness-based indices commonly emphasised in upland studies had limited influence, with the topographic wetness index ranking lowest. The final susceptibility map identifies 28.3% of the district (399 km2), concentrated in Kuttanad, as highly to very highly susceptible. By combining a multi-event SAR inventory, spatially rigorous validation and interpretable machine learning, the study provides a transparent framework that may inform comparable coastal–backwater settings, subject to local validation, for flood susceptibility assessment in coastal and backwater floodplains.