Flash floods, particularly in Mediterranean regions, are a rare but destructive event, and difficult to forecast. This difficulty is due to the high rainfall intensities and total amount, the high spatial variability of rainfall, the short response time of small watershed, but also to the lack of observations during these events. Moreover, the presence of karstic terrains on the basin leads to complex and unknown interactions between the surface flood and the underground system. In this study, we calibrated radar rainfall by a known global method and a new distributed one. We used these data in a conceptual distributed model to simulate floods. The simulations results showed that the distributed method is the most accurate way to correct radar data; however the global method led to higher performances of the model. This may be due to the globalization of production parameters on the watershed, which does not represent the karst influence on surface floods. This influence seems to be significant since the piezometric level in the karst was the best indicator to initialize the model.