Application of an Ensemble Kalman filter to a semi-distributed hydrological flood forecasting system in alpine catchments

One of the key success factor for hydrological forecasts is the establishment of initial conditions that represent well the conditions of the simulated basin at the beginning of the forecast. Real-time Data Assimilation (DA) has been shown to allow improving these initial conditions. In this article, two DA approaches are compared with the reference scenario working without DA (Control). In both approaches, discharge data at gauging station are assimilated. In the first approach, a volume-based update (VBU) compares the simulated and observed volumes over the past 24 hours to compute a correction factor used to update the soil water saturation in the upstream part of the semi-distributed hydrological model. In the second approach, an Ensemble Kalman Filter (EnKF) is implemented to account for the uncertainty in precipitation, temperature and discharge data. The comparison is carried out over 2 sub-basins of the Upper Rhône River basin upstream of Lake Geneva, where the MINERVE flood forecasting and management system is implemented. Results differ over the two studied basins. In one basin, the two data assimilation perform better than the Control simulation with the lowest error given by the VBU up to a forecast horizon of 35 hour and by the EnKF for higher forecast horizons. In the second basin, EnKF gives the lowest error over the few first hours of forecast, but then provides the weakest performance. The lowest error is given by the Control simulation, because the model already performs very well on the event without data assimilation.