Optimizing Flood Susceptibility Detection Using Ensemble Learning Methods

Abstract

Floods are among the most devastating natural disasters, causing widespread loss of life, infrastructure damage, and long-term socio-economic disruption. Accurate flood susceptibility assessment is therefore vital for effective disaster risk reduction and environmental management. The Nekaroud watershed in Iran is particularly flood-prone due to its complex terrain and dense hydrological networks. This study employs Radial Basis Function Neural Network (RBFN)-based ensemble models to map and predict flood susceptibility in this challenging environment. A dataset comprising 133 recorded flood events was compiled, with 70% used for training and 30% for validation. Fourteen critical flood-conditioning parameters were selected and evaluated for multicollinearity to ensure model robustness. The ensemble models developed include RBFN-Attribute Selected Classifier (ASC), RBFN-Decorate (De), RBFN-Dagging (Da), RBFN-Cascade Generalization (CG), RBFN-Random Subspace (RSS), and RBFN-Function Tree (FT). Model performance was assessed using accuracy, Kappa coefficient, and Area under the ROC Curve (AUC). The models classified between 6 and 13% of the area as high and 8–11% as very high susceptibility zones. Among these, the RBFN-RSS model demonstrated the highest predictive performance, with an AUC of 0.963 and identifying approximately 10.76% of the area as highly susceptible to flooding. The most influential factors contributing to flood susceptibility were elevation (0.393), proximity to streams (0.341), and drainage density (0.319). Overall, the study demonstrates the effectiveness of RBFN-based ensemble models in improving flood prediction accuracy. It also recommends incorporating additional environmental variables to further refine susceptibility mapping under evolving hydrological conditions.