A deep U-Net framework for flood hazard mapping using hydraulic simulations of the Wupper catchment, in the North Rhine-Westphalia region, Germany

Study region: The research was performed for the Wupper Catchment, in the North Rhine-Westphalia region, Germany. Study focus : The increasing frequency and severity of global flood events highlights the need for the development of rapid and reliable flood prediction tools. This process traditionally relies on hydraulic models, which are used to generate flood hazard maps. However, the process of generating hydraulic simulations is computationally expensive, and has a significant latency. Due to the latency, these models are used as offline analysis tool for flood hazard. This research proposes a framework to develop deep-learning-based surrogated models to approximate hydraulic models for the generation of flood maps with a low latency. Moreover, we detailed a data generation procedure using a patch generation and data augmentation strategies. The framework was tested using hydraulic simulations of the Wupper catchment. The trained models were compared with holdout simulations from a calibrated hydraulic model, obtaining comparable results. Moreover, due to the approach’s low latency, it can be used in (near) real-time applications. New hydrological insights for the region : The results showed a final RMSE of 0.0227 m and a significant computational time speed-up of 21.5 times compared to the available hydraulic models. Moreover, the framework can be replicated to develop surrogate models for other catchments, which is of great importance to other water associations of the region and of other countries.