Real-time flood maps forecasting for dam-break scenarios with a transformer-based deep learning model

This paper presents a purely data-driven deep-learning approach for flood maps forecasting. For the first time in this context a Transformer-based algorithm is employed to address one of the main issues in early-warning systems for flood propagation, i.e., the long computational times required to forecast the inundation evolution in real time. The proposed model, named "FloodSformer", is trained to extract the spatiotemporal information from a short sequence of water depth maps and predict the water depth map at one subsequent instant. Then, to forecast a sequence of future maps, we employ an autoregressive procedure based on the trained surrogate model. The method was applied to both synthetic dam-break scenarios and to a real case study, specifically the ideal failure of the Parma River dam (Italy). The training and testing datasets were generated numerically from two-dimensional hydraulic simulations. In the case of the real test case, the average Root Mean Square Error was found to be equal to 10.4 cm. The short computational time (e.g., the forecast of 90 maps, representing a lead time of 3 h, takes less than 1 min) makes the FloodSformer model a suitable tool for real-time emergency applications.