FloodSformer: Transformer based surrogate model for real time forecasting of inundation maps

In this work, we present a new Transformer based data driven model, named FloodSformer , that efficiently predicts the temporal evolution of inundation maps, with the aim of providing real time flood forecasts. The water depth maps, used to train and evaluate the surrogate model, were generated using a 2D shallow water code (PARFLOOD). The case study is the hypothetical collapse of the flood control dam of the Parma River (Italy). The proposed model can forecast the water depth maps of the next 2 hours, corresponding to 60 frames, with high accuracy (mos t o f the maps have errors lower than 0.25 m), and the required computational time is less than 1 minute. The FloodSformer model can thus be useful for early warning systems.