Flash floods frequently hit the Mediterranean regions and cause numerous fatalities and heavy damage. Their forecast is still a challenge because of the poor knowledge of the processes involved and because of the difficulty to forecast heavy convective rainfall. In any case, early warning remains a strong need. In this study, the authors propose to build a deep artificial neural network for flash flood forecasting, allowing, by its specific architecture, to take better account of the spatial variability and the scales of the rainfall as well as the hydrological responses. The outcomes of the deep model are then compared to a classical global multilayer perceptron previously published. For this purpose, a database of 58 heavy rainfall events extracted from 16 years of hydrometeorological observations on a well-studied basin in Southern France is applied to train a deep recurrent neural network. The results are of twofold: first, the deep model improves the lead time from two hours to three hours providing then suitable forecast for an operational use. Second, the model selection process converged towards an architecture that explicitly considers spatial scales of the basin. More generally, this study shows that the implementation of a rigorous selection process mobilizing several well-known regularization methods has enabled the deep model to converge towards a parsimonious model highlighting some of the known physical processes of the basin: the roles of elevation and distance to the outlet. This work provides, thus, a very interesting piece of evidence to fuel the controversy on the interpretability of modern AI