An explainable and transferable deep learning framework for spatiotemporal urban flood prediction by integrating...

Urban flood has become an increasingly frequent and severe hazard under intensified climate change and rapid urbanization, yet real-time prediction remains challenging. Existing data-driven models often suffer from limited physical plausibility and transferability, especially in spatially heterogeneous urban environments. Motivated by these limitations, ViTUN, a hybrid deep learning framework by combining Vision Transformer and U-Net, was proposed in this study to capture the spatiotemporal characteristics of flood propagation under varying rainfall and topographic conditions. The ViTUN model was trained and evaluated using inundation data generated from coupled hydrodynamic simulations in Yueyang, China. Compared to the U-Net, the proposed ViTUN achieved substantially higher accuracy and robustness in urban flooding water depth prediction, with improvements of about 10.2 % in Critical Success Index (CSI), 3.4 % in R² and reductions of 53.7 % in Mean Absolute Error (MAE). More importantly, ViTUN maintained transferability when directly applied to spatially untrained regions without any retraining, with an average R² of 0.940 and a MAE of 0.0022 m in predicting water depth. Further interpretability analysis using Grad-CAM provided supporting evidence that ViTUN dynamically focused on regions undergoing pronounced inundation evolution, while also highlighting low-lying areas and building edges. The proposed ViTUN model provides a fast, interpretable, and transferable framework for urban flood prediction, with strong potential to enable real-time early warning, adaptive emergency response, and flood risk management.