Abstract : In karst context, sustainable management of water resource needs knowledge about karst
aquifers and rivers exchanges. Regarding specifically Mediterranean karst aquifers, it appears that
they are sometimes the unique water resource of their territory and that they are often underexploited.
The site of the Cè+ze River (Rhô+ne river tributary) is subjected to important
anthropogenic impacts linked to drinking water, irrigation, and seasonal increasing population
(tourism). Interactions between Cè+ze river and karst aquifer of Mé+jannes -le-clap are
interesting because the complexity of the aquifer can be related to the existence of a possible deep
reservoir (Messinian episode). Moreover the Mediterranean climate leads to rainfalls heterogeneous in
time and space making the behaviour difficult to characterize. Due to important cited societal stakes,
the public Water Agency (Agence de l&rsquo+Eau Rhô+ne, Mé+diterrané+e,
Corse) initiated a multi-disciplinary research project in order to better apprehend water circulations.
This project investigates the karst river exchanges thanks to several approaches - thermic infrared
imagery, analysis of interstitial invertebrates, analysis of major ions, and hydrological modelling. This
presentation aims to present this last issue using neural network modelling, that are well-known for
their ability to represent non-linear and badly known functions. Thanks to a 19 years daily database,
an original architecture was proposed for rainfall-water level modelling at the station of Tharaux. A
specific architecture was designed in order to distinguish the influence of hydrologic withdrawal of
evaporation and transpiration, upstream rainfalls and lastly local rainfalls. Using 17 years training set,
and two years for validation, modelling appeared satisfying with an average Nash criterion around 0.8
and a good representation of the drought period. Illustrations will be provided on the year 1994 that
was used to compare the efficiency of various architectures of neural networks. Using scenarii of
rainfalls, the quality of the model will thus allow to design a useful tool for helping water managers to
anticipate about water conflicts and adapt to water scarcity.
