Flood modelling with a distributed event-based parsimonious rainfall-runoff model: case of the karstic Lez river catchment
Rainfall-runoff models are crucial tools for the statistical prediction of flash floods and real-time forecasting. This paper focuses on a karstic basin in the South of France and proposes a distributed parsimonious event-based rainfall-runoff model, coherent with the poor knowledge of both evaporative and underground fluxes. The model combines a SCS runoff model and a Lag and Route routing model for each cell of a regular grid mesh. The efficiency of the model is discussed not only to satisfactorily simulate floods but also to get powerful relationships between the initial condition of the model and various predictors of the initial wetness state of the basin, such as the base flow, the Hu2 index from the Meteo-France SIM model and the piezometric levels of the aquifer. The advantage of using meteorological radar rainfall in flood modelling is also assessed. Model calibration proved to be satisfactory by using an hourly time step with Nash criterion values, ranging between 0.66 and 0.94 for eighteen of the twenty-one selected events. The radar rainfall inputs significantly improved the simulations or the assessment of the initial condition of the model for 5 events at the beginning of autumn, mostly in September–October (mean improvement of Nash is 0.09; correction in the initial condition ranges from −205 to 124 mm), but were less efficient for the events at the end of autumn. In this period, the weak vertical extension of the precipitation system and the low altitude of the 0 °C isotherm could affect the efficiency of radar measurements due to the distance between the basin and the radar (~60 km). The model initial condition S is correlated with the three tested predictors ( R2 > 0.6). The interpretation of the model suggests that groundwater does not affect the first peaks of the flood, but can strongly impact subsequent peaks in the case of a multi-storm event. Because this kind of model is based on a limited amount of readily available data, it should be suitable for operational applications.