Flash floods are important hazards in urbanised zone and constitute important human and financial stakes. This paper applies a novel methodology, KnoX, dedicated to extract knowledge from a neural network model. It was shown that KnoX method could help to better characterize processes of both surface and underground floods. A case study is chosen in France: the Lez karst hydrosystem whose river crosses the city of Montpellier (400 000 inhabitants). Results will help flood warning services.

During a flood event over a karst watershed, the karst may attenuate surface floods by absorbing water or contribute to the surface flood by direct contribution of karst waters in the rivers and by diffuse resurgence along the hillslopes. If it is possible to monitor each known outlet of a karst system, the diffuse contribution is yet difficult to assess. We present here a new, original method to do it, based on chemical analysis and a hydraulic study.