An interpretable adaptive mountain flood forecasting agent model framework based on multi-source terrain data

Climate change and human impact are increasing the frequency and severity of flooding, making real-time, accurate simulation of mountainous flood inundation an urgent necessity. In this study, we developed a proxy model framework that integrates multi-source terrain data to perform mountainous flood inundation simulations. The developed method was demonstrated in Baifusi Town, Hubei Province. During the data preprocessing stage, multiple terrain data sources were fused and optimised to reconstruct high-precision underwater terrain. This study employs four machine learning models, Decision Tree, Random Forest, XGBoost, and CatBoost, as custom model hyperparameters. Bayesian optimisation was employed for model hyperparameter selection to establish relationships between various geological, hydrological, and mountainous flood events. The results showed that the model could adaptively select the optimal model and hyperparameters based on the set cash problem, thereby reducing the complexity of manually selecting and adjusting algorithms. For mountainous flood events, the model's flood extent prediction and extrapolated Probability of Detection (POV) were both above 96.9%. The flood depth prediction and extrapolated RMSE and MAE were less than 0.55. Using the selected optimal model, SHAP explanation was further employed to explore the quantitative impact of various factors on mountainous floods and provide insights for practical early warning systems. The study found that elevation is the most important factor influencing the maximum flood depth. Additionally, our model demonstrates the ability to account for reservoir regulation and the impact of downstream reservoir backflow.