Enhancing the Generalization of Flood Susceptibility Models: A Leakage-Aware Ensemble Framework for Deltaic Landscapes

Shafiq Mahmud, Golam Murad , Md. Aftabur Rahman

Flood susceptibility mapping (FSM) is a cornerstone of disaster risk reduction in low-lying deltaic regions; however, conventional machine learning (ML) applications frequently suffer from spatial data leakage, resulting in inflated performance metrics and unreliable hazard predictions. To address this critical methodological shortcoming, this study develops a robust, leakage-aware ML framework for FSM in the highly vulnerable Greater Noakhali region of Bangladesh. Utilizing a flood inventory derived from Sentinel-1 SAR imagery of the catastrophic August 2024 event, we integrated twelve spatially explicit conditioning factors representing topographic, hydrological, and anthropogenic drivers. To ensure true model generalization and mitigate the effects of spatial autocorrelation, eight ensemble algorithms were evaluated using a rigorous 10-fold blocked spatial cross-validation strategy. The spatially constrained evaluation provided realistic performance estimates, in contrast to the overly optimistic ROC-AUC values typical of conventional random splits, with Random Forest demonstrating the highest generalization capability (ROC-AUC = 0.627) and XGBoost achieving the optimal balance between precision and recall (F1 = 0.451). Furthermore, SHAP (SHapley Additive exPlanations) analysis of the ensemble models identified land use, vegetation density (NDVI), and drainage density as the primary physical drivers of flood susceptibility in this coastal tract. By demonstrating that spatially robust validation is essential for credible geospatial AI, the resulting maps identify critical high-risk zones along the Meghna estuary and provide a physically plausible, operationally relevant tool for regional disaster management.