An Explainable Reinforcement Learning Approach Integrated with Rule-Based Insights for Early Urban Flood Detection

Abstract

Urban pluvial flooding remains one of the most disruptive hazards in modern cities. While machine learning surrogates now allow real-time flood simulation at street level, most data-driven systems focus only on spatial prediction accuracy. They do not support operational decision-making or provide clear physical explanations as to why a specific flood alert should be issued. This leaves operators to use fixed thresholds that do not adjust to changing conditions. The specific objectives of this study are to: (i) design an integrated framework for early urban flood detection that couples a Conv2D-LSTM spatiotemporal surrogate with a model-based reinforcement learning environment and a dual-layer interpretability module; (ii) implement the reinforcement learning agent with an asymmetric reward function to prioritize safety-critical detection decisions within the designed environment; and (iii) evaluate the agent's detection reliability using precision, recall, and F1-score, and validate its ability to generate rule-based insights through XAI-driven perturbation analysis. This study consists of a three-stage development process: (i) to create the training environment, a Conv2D-LSTM surrogate was built using open-source terrain and rainfall data from USGS Earth Explorer, ESA WorldCover, OpenStreetMap, and GPM-IMERG to produce depth predictions in milliseconds. (ii) to provide the feedback necessary for reinforcement learning, the surrogate was then used to train the agent with an 8-sensor observation space over 100,000 steps. (iii) SHAP and sensitivity layers were incorporated. The model was evaluated on 382 storm sequences from Houston collected between 2017 and 2022. The results show that the surrogate model closely matched the ground truth physics-based flood maps, achieving a Nash–Sutcliffe Efficiency of 0.9966. The reinforcement learning based decision-making process reached an F1-score of 0.9911 and missed only four flood events over the entire test period. The explainable AI layer also provided rule-based audits that matched real hydrological patterns. The study concludes that fast surrogates and smart decision logic can work together to support faster and more reliable early urban flood detection. The research recommends the operational use of this framework to support emergency responders during flash floods.

DOI: https://doi.org/10.5281/zenodo.20388429