Flood simulation and forecast capability have been greatly improved thanks to advances in data assimilation. Such an approach combines in-situ gauge measurements with numerical hydrodynamic models to correct the hydraulic states and reduce the uncertainties in the model parameters. However, these methods depend strongly on the availability and quality of observations, thus necessitating other data sources to improve the flood simulation and forecast performances. Using Sentinel-1 images, a flood extent mapping method was carried out by applying a Random Forest algorithm trained on past flood events using manually delineated flood maps. The study area concerns a 50-km reach of the Garonne Marmandaise catchment. Two recent flood events are simulated in analysis and forecast modes, with a +24h lead time. This study demonstrates the merits of using SAR-derived flood extent maps to validate and improve the forecast results based on hydrodynamic numerical models with Telemac2D-EnKF. Quantitative 1D and 2D metrics were computed to assess water level time-series and flood extents between the simulations and observations. It was shown that the free run experiment without DA under-estimates flooding. On the other hand, the validation of DA results with respect to independent SAR-derived flood extent allows to diagnose a model-observation bias that leads to over-flooding. Once this bias is taken into account, DA provides a sequential correction of area-based friction coefficients and inflow discharge, yielding a better flood extent representation. This study paves the way towards a reliable solution for flood forecasting over poorly gauged catchments, thanks to available remote sensing datasets.