This article aims to propose a methodology for the extraction of buildings in a GeoEye image of very high spatial resolution from a hybrid multi-level approach between particle size distribution and characterization of attributes. Building extracted become an alternative for obtaining information in areas where there is no cartographic information, or where the cartographic information related to buildings is not update. Building layer of the BDtopo © from IGN France was used as a reference. The results are compared with the extraction of the same objects with a non supervised classification, through indicators of overall accuracy and kappa coefficient. The proposed method is more accurate in the extraction of buildings that the no supervised method, improving the coefficient kappa in 0.02 units and reaching an overall accuracy of 90%.