Abstract
Floods and droughts are the two main risks in France and require a special attention. In these conditions, where climate change generates increasingly frequent extreme phenomena, modeling these risks is an essential element for water resource management.Currently, discharges and water heights are mainly predicted from physical or conceptual based models. Although efficient and necessary, the calibration and implementation of these models require long and costly studies.Hydrogeological forecasting models often use data from incomplete or poorly dimensioned measurement networks. Moreover, the behavior of the study basins is in most cases difficult to understand. This difficulty is thus noted to estimate the uncertainties associated with hydrogeological modeling.In this context, this thesis, supported by IMT Mines Alès and financed by the company aQuasys and ANRT, aims at developing models based on the systemic paradigm. These models require only basic knowledge on the physical characterization of the studied basin, and can be calibrated from only input and output information (rainfall and discharge/height).The most widely used models in the environmental world are neural networks, which are used in this project. This thesis seeks to address three main goals:1. Development of a model design method adapted to different variables (surface water flows/height) and to very different types of basins: watersheds or hydrogeological basins (groundwater height)2. Evaluation of the uncertainties associated with these models in relation to the types of targeted basins3. Reducing of these uncertaintiesSeveral basins are used to address these issues: the Blavet basin in Brittany and the basin of the Southern and Central Champagne Chalk groundwater table.