A reinforcement learning approach with explainable AI for spatial flood susceptibility analysis

Study region This study focuses on Chaharmahal and Bakhtiari Province, a semi-arid, mountainous area in western Iran. This region is recognized as one of the country's most hydrologically complex and climatically extreme zones, making it a compelling setting for investigating flood susceptibility and water resource challenges. Study focus This study aims to expand the application of reinforcement learning (RL) from flood management to flood susceptibility mapping by integrating RL algorithms with geographic information systems (GIS). We implemented Q-Learning (OL), Proximal Policy Optimization (PPO)-based Proximal Updating (PU), and Deep Q-Learning (DQL), and introduced RL-Stack. To ensure interpretability, we applied SHAP for variable attribution. Validated through a real-world case study, the methodology delivered accurate, actionable maps to support resilient flood risk management. New hydrological insights The PU model most effectively captured flood susceptibility, balancing sensitivity to flood-prone areas with stability across heterogeneous landscapes. DQL showed overestimation bias and unstable local predictions. RL-Stack improved the detection of highly susceptible zones, while traditional QL performed poorly in continuous, imbalanced settings. The SHAP-based analysis identified snow depth as the primary hydrological control, exhibiting a dual role: deep snowpacks can intensify floods during rapid melt or attenuate runoff when retained. Short-duration, intense rainfall strongly interacted with snow depth and flow accumulation, generating disproportionately large floods through rain-on-snow processes.