For a better consideration of multi-scale behaviors of complex hydrosystems by neural network models : application to flash floods ; Vers une meilleure prise en compte des comportements multi-échelles des hydrosystèmes complexes par les modèles à réseaux de neurones : application aux crues éclair

Flash floods are among the most devastating natural hazards in the world. According to Jamali et al. (2020), they are responsible for nearly 84% of deaths from natural disasters. While flooding from a conventional flood can be anticipated with some lead time, a flash flood is much more rapid and localized, and therefore much more difficult to predict. This is the case in the Mediterranean regions of France. Faced with this problem, the institutions in charge of flood forecasting need quality information and efficient models in order to optimize their responses. Because the rainfall that generates these flash floods is very heterogeneous both in time and space, in addition to the fundamentally non-linear nature of the rainfall-flow relationship, forecasting them remains a very serious challenge. For three decades now, neural networks have proven their efficiency in solving complex and non-linear problems, in particular rainfall-flow relationships in various hydrological situations. Within these types of models, Deep Learning as a learning method that is mainly applied to deep neural networks has proven to be particularly successful in many disciplines. However, because of their black box character, which seems to us rather an advantage considering the lack of knowledge on some hydrological processes, the interest of their application is sometimes questioned.For this reason, this work has applied deep neural networks to flash flood forecasting with three main objectives: the first objective aims at interpreting the parameters of the deep layers of the three types of perceptrons generally used in hydrology: static, directed, recurrent. To do so, this work followed two steps: (i) extracting the parameter values of the optimized models using the "Knowledge eXtraction (KnoX)" method proposed by Kong A Siou et al. (2013); (ii) interpreting these parameters through a comparative analysis of this information with data characterizing some hydrological processes of the watershed; this part has been published in the ...