Watersheds are the latest segmentation tool developed in mathematical morphology. These algorithms produce a segmentation of an image into a set of basins separated by watershed pixels. The over segmentation produced by these algorithms is reduced by removing all contours with a low saliency. This contour's saliency is generally defined from the minimal height of the watershed pixels along the contour. However, such a definition does not allow to define a contour's saliency in case of thick watersheds. Moreover, the set of basins which corresponds to the intuitive notion of regions does not define an image partition. In this paper we propose a method which allows to aggregate the watershed pixels to the basins while preserving the notion of contour and the associated saliency. The model used to encode the image partition is then decimated according to the contour saliency to obtain a hierarchy of partitions.