A Transformer-Based Data-Driven Model for Real-Time Spatio-Temporal Flood Prediction

Among the non-structural strategies for mitigating the huge economic losses and casualties caused by floods, the implementation of early-warning systems based on real-time forecasting of flood maps is one of the most effective. The high computational cost associated with two-dimensional (2D) hydrodynamic models, however, prevents their practical application in this context. To overcome this drawback, “data-driven” models are gaining considerable popularity due to their high computational efficiency for predictions. In this work, we introduce a novel surrogate model based on the Transformer architecture, named FloodSformer (FS), that efficiently predicts the temporal evolution of inundation maps, with the aim of providing real-time flood forecasts. The FS model combines an encoder-decoder (2D Convolutional Neural Network) with a Transformer block that handles temporal information. This architecture extracts the spatiotemporal information from a sequence of consecutive water depth maps and predicts the water depth map at one subsequent instant. An autoregressive procedure, based on the trained surrogate model, is employed to forecast tens of future maps.

As a case study, we investigated the hypothetical inundation due to the collapse of the flood-control dam on the Parma River (Italy). Due to the absence of real inundation maps, the training/testing dataset for the FS model was generated from numerical simulations performed through a 2D shallow‐water code (PARFLOOD). Results show that the FS model is able to recursively forecast the next 90 water depth maps (corresponding to 3 hours for this case study, in which maps are sampled at 2-minute intervals) in less than 1 minute. This is achieved while maintaining an accuracy deemed entirely acceptable for real-time applications: the average Root Mean Square Error (RMSE) is about 10 cm, and the differences between ground-truth and predicted maps are generally lower than 25 cm in the floodable area for the first 60 predicted frames. In conclusion, the short computational time and the good accuracy ensured by the autoregressive procedure make the FS model suitable for early-warning systems.