Levee-breach inundation forecasting with deep-learning and hydrodynamic models: the 2020 Panaro river case study

Introduction In the flood-forecasting literature, deep-learning frameworks are emerging as fast surrogates for hydrodynamic models. However, their computational performance is not often benchmarked against efficient parallel codes. Methods In this work, the deep-learning FloodSformer model for inundation forecasting and the GPU-parallel 2D hydrodynamic code PARFLOOD are applied to a real case study of breach-induced flooding, in order to compare the performance of these alternative approaches for real-time forecasting. Results and key findings Results show that both strategies provide accurate enough predictions for practical purposes and that their runtimes are comparable. Even if the implementation of data-driven surrogate models is gaining momentum due to their good computational performance, this case study provides an example that 2D hydrodynamic models can also be competitive for real-time applications thanks to the runtime reduction guaranteed by GPU parallelisation.