Artificial Intelligence and GPU Card Calculations Applied to Flood Forecasting: Feedback from the Inundation Project

This paper presents the work carried out within the framework of the InteReg SUDEO “INUNDATIO” project. Our task in this project concerns the application of artificial intelligence algorithms to flood forecasting. To fill the data gap, we produced flood maps by numerical simulation by solving the shallow water equations on the French pilot sites of the project (Nive and Gave de Pau). Five machine learning algorithms were evaluated on the Nive and the Gave de Pau, in 4 scenarios. The machine learning model based on linear regression was the most effective in all scenarios. Likewise, five deep-learning algorithms were compared to the best machine learning model on the best scenario and a variant. The best configuration is obtained with the use of a multi-head neural convolution network (MH CNN), further requiring less computing resources than algorithms based on LSTM in particular. All deep learning configurations perform better than the best machine learning configuration.