In view of improving numerical flood prediction, a variational data assimilation method (4D-var) applied to a 2D shallow water model and using distributed water level obtained from one Synthetic Aperture Radar (SAR) image is presented. The RADARSAT-1 image leads to water levels with a ±40cm average vertical uncertainty of a Mosel River flood event (1997, France). Assimilated in the 2D shallow water hydraulic model, these SAR derived spatially distributed water levels prove to be capable of enhancing model calibration. Indeed, the assimilation process can identify some optimal Manning friction coefficients. Moreover, used as a guide for sensitivity analysis, remote sensing water levels allow also in identifying some areas in the floodplain and the channel where Manning friction coefficients are homogeneous. This allows basing the spatial segmentation of roughness coefficient on floodplain hydraulic functioning.