Parameter regionalization for a process-oriented distributed model dedicated to flash floods

Abstract

This contribution is one of the first studies about the regionalization of parameter sets for a rainfall-runoff model process-oriented and dedicated to flash floods. MARINE model performances are tested on a large database of 117 flash floods occurred during the last two decades in the French Mediterranean region. Given the scarcity of flash flood data, the dataset used in this study represents a large sample of hydrology and landscapes from Pyrenean, Mediterranean, Cevennes-Vivarais and Provence regions. Spatial proximity and similarity approaches with several combinations of descriptors are tested. Encouraging results are obtained with two similarity approaches based on physiographic descriptors with two and three donor catchments. There is only a small decrease of performance of 10 from cal/val to regionalization for these two methods. For 13 catchments out of 16 there is at least one flood event simulated with rather good performance. This study highlights the importance of hydrological information that is available in calibration events for a gauged catchment and from donor catchment(s) for regionalization. Moreover it is found that regionalization is easier for catchments with an apparently more regular behaviour. The most sensitive parameter of MARINE model, CZ, controlling soil volume and water balance, is rather well constrained by the two similarity approaches thanks to bedrock descriptors.