This study is dedicated to the automatic recognition and mapping of tree crops by remote sensing, using very high resolution multi-spectral satellite images (0.7 m). Our goal is to segment the images in order to perform an independent classification according to a set of pre-determined land use types: apple groves, vineyards, miscellaneous young and old groves, pastured and cropped fields, food crop, fallow lands and forests. In this article, we compare three methods of segmentation that seem to provide suitable units for the resolution of our problem: SxS, eCognition and watersheds. A set of criteria are defined to quantitatively analyze the efficiency of these segmentations. We then try to select the more relevant method in terms of subsequent classification operability.