Rapid Urban Flood Inundation Forecasting Using a Physics-Informed Deep Learning Approach

Physics-based models can achieve precise flood inundation forecasts, but their real-world application is limited by their high computational cost. Deep learning (DL) models, with the capability to establish mapping relationships for complex mechanistic processes and high computational efficiency, serve as promising alternatives. However, DL models require massive amounts of training data to achieve robust performance, and such data are not available in most cases. In this study, an approach that couples a hydrodynamic model and a DL model to realize rapid forecasting of urban flood inundation is proposed. Substantial data on urban flood inundation under varying rainfall events are generated based on the hydrodynamic model. Real-time water level data from hydrological gauges are employed to establish initial conditions. Based on these data, a DL model that fully considers the physical mechanisms of flood inundation and the feature attributes of inputs and outputs is developed. The results show that 1) the hydrodynamic model effectively provides training samples for the DL model, addressing the limitations of insufficient urban flood inundation data; 2) the DL model proficiently captures the occurrence of grid-based flood inundation events, demonstrating commendable effectiveness in predicting inundation depths with a high level of accuracy; and 3) the DL model forecasts flood inundation in a region of 250,000 grids over 12 time steps within 12 seconds, meeting the requirements for real-time management. Compared to traditional hydrodynamic modeling methods, the proposed approach enhances forecasting efficiency and yields high accuracy, providing an efficient and accurate method for urban flood inundation forecasting.