Bathymetry data on river are crucial elements for hydro-morphology and hydroecology but difficult and costly to measure. In case of clear water river, the use of remote-sensing images might suit to produce realistic depth maps of shallow water areas [Carboneau et al., 2004]. In this study, three sites have been studied on the graved-bed river Durance (France). On these test sites, a high quantity of ground data were available. Three Very High Spatial Resolution images shoted at same time as ground data were compared: two aerial images acquired by plane (Bd-Ortho© IGN) or ULM (Avion Jaune Company) in natural colours (Red, Green, Blue) with 0.5 m resolution and one provided by the Quickbird satellite in multispectral (R, G, B, Infrared) with 2.4 m and 0.7 m resolutions. For each shot, a model of the relation between colors and depth has been fitted using linear regression. In the case of simple linear regressions, a determination coefficient of 0.4 was found in the best case and most models are very poorly relevant. On the contrary, results using a stepwise multi-linear regression according Lysenga equation are more significant with a maximum of determination coefficient up to 0.71. In most cases, only two spectral bands were identified as significant to establish a fairly good model for each shot. According to the error maps, a link is supposed on the one hand between high depths and under-assessment by the method and, on the other hand, between algae excess and over-assessment.Image pre-treatment and resolution appear to be two important parameters influencing the model fitting. Tests of model inter-crossing between each shot and images were realised in order to assess their suitability and generalization ability. These first results shows a difficult direct generalisation of regression models and highlights, for aerial images, the necessity of the images normalization using some radiometric pre-treatments. Thanks to CNES-ORFEO project (France) for providing Quick-bird images.