On 8 and 9 September 2002, an extreme flash-flood event occurred over south-eastern France. This event was simulated with the non-hydrostatic mesoscale model Meso-NH. Several experiments were performed, differing in their initial conditions. A first experiment used the 4D-Var large-scale French operational analysis, as initial state. For the second experiment, the initial state was provided by a mesoscale initialization using surface observations, radar and satellite data. Finally, in the last experiment, only the mesoscale analysis of surface observations was used as initial state. The aim of this study is to assess the validity of the high-resolution simulations for hydrological purposes. Then simulated and observed mean areal rainfall depth, over nine watersheds with surface areas ranging from 200 to 2200 km2, were compared. This comparison highlighted the contribution of both initialization procedures to a better location of the mesoscale convective system and to larger amounts of rain, during the first ten hours of simulation. Also, a hydrological model was used to compare stream flows: the simulated and observed rainfall fields were used as input data to a single-event TOPMODEL version. The comparison of the hydrological responses shows that the mesoscale initialization leads to higher peak flows and more realistic flood timing than with the simulation starting from large-scale analysis. The full initialization allowed a simulation of peak flows higher than the ones obtained with the mesoscale surface-observation initialization. Copyright © 2006 Royal Meteorological Society.