Decadal Reconstruction of Monsoon Flood Inundation in the Kabul River Basin Using Sentinel-1 SAR and Machine Learning

Flood inundation mapping in mountainous basins remains challenging because of complex terrain, frequent cloud cover, and limited ground observations. This study reconstructs monsoon flood inundation in the Kabul River Basin (KRB) during 2015–2025 using Sentinel-1 Synthetic Aperture Radar (SAR) and a Random Forest (RF) machine-learning framework. A multi-feature dataset comprising pre- and post-flood VV and VH backscatter, polarization metrics, GLCM texture descriptors, and topographic variables was developed, while high-confidence pseudo-labels were used for event-specific RF classification. The results reveal substantial interannual variability in flood extent, with the largest inundation recorded in 2020 (1004 km2), followed by high flood extents in 2021 and 2022. Except for 2017, monsoon-related inundations were detected in every year, and the cumulative inundated area exceeded 4000 km2, with flooding predominantly concentrated in downstream lowland areas. The RF framework achieved overall accuracies of 0.86–0.93, Cohen’s Kappa values of 0.78–0.86, and an integrated ROC-AUC of 0.924. These findings demonstrate the utility of Sentinel-1 SAR and RF for consistent reconstruction of historical monsoon inundation in data-scarce mountainous basins and provide a temporally consistent flood inventory for hydrological analysis and subsequent flood-susceptibility assessment.