Floods are responsible for one third of the economic losses induced by natural hazards throughout the world. To better protect the population and infrastructures, flood forecasting systems make us of weather forecasts to foresee floods several days in advance, providing more lead time for preventive measures. In the canton of Valais (Switzerland), an operational flood forecasting and management system is operational since 2013, as a result of the MINERVE project initiated in 1999. The present thesis aims at answering some of the challenges faced by this system. First, a new methodology for spatial interpolation of precipitation is implemented based on regression co-kriging using rain gauge and weather radar data. Two rain gauge networks equipped with instruments of different quality are considered. Compared to other precipitation interpolation methods, the quantitative precipitation estimates (QPE) obtained from the regression co-kriging provides the best performance over the studied area using cross-validation. The analysis highlights the need for further pre-processing of radar data, in particular to account for beam shielding by the complex topography. Integration of the above-mentioned QPE product in a snow model revealed a clear precipitation underestimation. A methodology to account for solid precipitation undercatch in QPE computation is therefore proposed. Four different QPE products are compared: the operational QPE product CombiPrecip of MeteoSwiss, the regression co-kriging QPE and two variants of it considering a correction factor for solid precipitation undercatch of 1.2 and 1.3, applied before the interpolation. The snow model is calibrated using satellite-based data from the MODIS spectroradiometer and validated using snow water equivalent measurements 11 snow monitoring sites. The best performance is obtained using the QPE product including a correction factor of 1.2. To evaluate the performance of the developed QPE products from a hydrological perspective, three sub-catchments of the MINERVE system were calibrated considering 5 different input. The GSM and SOCONT hydrological models are used to model respectively the glacial and non-glacial parts. A two-phase calibration of the model is explored, applying the MODIS-based calibration of snow-melt degree-day factors, before calibrating the other parameters using discharge data. Results suggest that the developed QPE product accounting for solid precipitation undercatch (factor 1.2) leads to the best performance over the catchment with a good radar visibility. In case of lower radar visibility, using station data provides equal or better performances. With the current implementation, the two-phase calibration did not allow to outperform the conventional calibration. Finally, an ensemble Kalman filter (EnKF) is implemented to improve initial conditions used for hydrological forecasts. Results are compared, for two high flow events, to the scenario without assimilation and to the simple assimilation scheme currently implemented in the MINERVE system, updating the soil saturation based on a discharge volume comparison over the preceding 24 hours. The Ensemble Kalman filter (EnKF) shows good performance during these events but also highlights difficulties over base flow, strengthened in presence of hydropower perturbations.