A deep learning framework for real-time forecasting of 2-D inundation maps: the FloodSformer model

This thesis aims to contribute to the advancement of data-driven models for flood forecasting, with a specific focus on developing and implementing a deep-learning model for real-time prediction of the spatiotemporal evolution of two-dimensional (2D) inundation maps over complex large topographies. Unlike traditional flood forecasting methods, which are often computationally intensive and time-consuming, the proposed model, named FloodSformer (FS), leverages state-of-the-art deep learning techniques, integrating autoencoder and transformer architectures. This combination enables the FS model to effectively extract and process complex spatiotemporal information from a sequence of consecutive water depth maps and upstream inflow discharges, to predict inundation maps of subsequent instants through an autoregressive procedure. The FS model has been rigorously tested across a diverse range of scenarios, including both dam-break scenarios and river flood events, considering a variety of synthetic and real-world case studies. To assess the model’s applicability to practical flood forecasting, its performance has been measured in terms of both predictive accuracy and computational efficiency. In addition, extensive sensitivity analyses have been conducted to explore the influence of key hyperparameters: the types of flood events used for training and the spatial resolution of the maps. Datasets were numerically generated using a 2D hydrodynamic model. The results indicate that the FS model is capable of accurately forecasting long-duration flood events over complex and large-scale bathymetries, comprising up to millions of computational cells. Remarkably, FS achieves these results with computational times on the order of minutes, which is significantly lower than the physical time of the simulated events (with physical-to-computational time ratios up to 10,000). Furthermore, the model demonstrates predictive accuracy comparable to that of traditional physically based hydrodynamic models, achieving average root mean square errors in the range of 10-20 cm across different real-world test cases. The sensitivity analyses further indicate that larger and more diverse datasets significantly enhance the model’s accuracy and generalization capabilities. Additionally, when compared to a state-of-the-art convolutional neural network architecture, the FS model consistently outperformed in terms of predictive accuracy of river flood scenarios over complex terrain topographies. In conclusion, the FS model is a versatile and high-performing tool for flood prediction, with significant potential for applications in flood risk management and emergency response planning.