In southern France, flash flood episodes frequently cause fatalities and severe damage. In order to inform and warn populations, the French flood forecasting service (SCHAPI, Service Central d'Hydrométéorologie et d'Appui à la Prévision des Inondations) initiated the BVNE (Bassin Versant Numérique Expérimental, or Experimental Digital Basin) project in an effort to enhance flash flood predictability. The target area for this study is the Gardon d'Anduze basin, located in the heart of the Cévennes range. In this Mediterranean mountainous setting, rainfall intensity can be very high, resulting in flash flooding. Discharge and rainfall gauges are often exposed to extreme weather conditions, which undermines measurement accuracy and continuity. Moreover, the processes governing rainfall-discharge relations are not well understood for these steeply-sloped and heterogeneous basins. In this context of inadequate information on both the forcing variables and process knowledge, neural networks are investigated due to their universal approximation and parsimony properties. We demonstrate herein that thanks to a rigorous variable and complexity selection, efficient forecasting of up to two-hour durations, without requiring rainfall forecasting as input, can be derived using the measured discharges available from a feedforward model. In the case of discharge gauge malfunction, in degraded mode, forecasting may result using a recurrent neural network model. We also observe that neural network models exhibit low sensitivity to uncertainty in rainfall measurements since producing ensemble forecasting does not significantly affect forecasting quality. In providing good results, this study suggests close consideration of our main purpose: generating forecasting on ungauged basins.

Flash floods pose significant hazards in urbanised zones and have important implications financially and for humans alike in both the present and future due to the likelihood that global climate change will exacerbate their consequences. It is thus of crucial importance to improve the models of these phenomena especially when they occur in heterogeneous and karst basins where they are difficult to describe physically. Toward this goal, this paper applies a recent methodology (Knowledge eXtraction (KnoX) methodology) dedicated to extracting knowledge from a neural network model to better determine the contributions and time responses of several well-identified geographic zones of an aquifer. To assess the interest of this methodology, a case study was conducted in southern France: the Lez hydrosystem whose river crosses the conurbation of Montpellier (400 000 inhabitants). Rainfall contributions and time transfers were estimated and analysed in four geologically delimited zones to estimate the sensitivity of flash floods to water coming from the surface or karst. The Causse de Viols-le-Fort is shown to be the main contributor to flash floods and the delay between surface and underground flooding is estimated to be 3 h. This study will thus help operational flood warning services to better characterise critical rainfall and develop measurements to design efficient flood forecasting models. This generic method can be applied to any basin with sufficient rainfall–run-off measurements.