Geospatial analysis and machine learning for flood susceptibility mapping

Graphical Abstract Sentinel-1 SAR–based flood analysis and machine learning workflow producing a flood susceptibility map for the Koshi River Basin. This study produces a multi-year flood susceptibility assessment for the Koshi River Basin, Nepal, by integrating Sentinel-1 SAR imagery with machine-learning and geographic information systems (GIS)-based approaches. Five-year (2019–2023) temporal flood analysis from Sentinel-1 GRD data was used for the preparation of flood inventory, consisting of 375 flood and 375 non-flood points by extracting multi-year composites alongside 11 environmental, topographic, climatic, and vegetation-based conditioning factors, which served as the foundation for training and validating the susceptibility models. The four machine learning (ML) classifiers: bagging, random forest (RF), support vector machine (SVM), and AdaBoost were evaluated against the weighted sum method. Bagging achieved the highest performance (accuracy 94.52%, precision 95.45%, recall 95.45%, F1-score 95.45%, and AUC 0.990) and identified 6.185% of the basin as highly susceptible to flooding. Strong agreement between bagging and RF (r = 0.76) indicated consistent model behaviour, while weaker correlations with the weighted sum method underscored limitations of rule-based approaches in capturing nonlinear flood processes. The combination of a multi-year SAR-based flood inventory and ensemble ML techniques demonstrates clear advantages in generating reliable susceptibility maps, providing actionable evidence to support land-use planning, disaster preparedness, and targeted flood mitigation.