Watershed-Based Attribute Profiles for Pixel Classification of Remote Sensing Data

Abstract

We combine two well-established mathematical morphology notions: watershed segmentation and morphological attribute profile (AP), a multilevel feature extraction method commonly applied to the analysis of remote sensing images. To convey spatial-spectral features of remote sensing images, APs were initially defined as sequences of filtering operators on the max- and min-trees computed from the original data. Since its appearance, the notion of APs has been extended to other hierarchical representations including tree-of-shapes and partition trees such as $\alpha$-tree and $\omega$-tree. In this article, we propose a novel extension of APs to hierarchical watersheds. Furthermore, we extend the proposed approach to consider prior knowledge from training samples, leading to a more meaningful hierarchy. More precisely, in the construction of hierarchical watersheds, we combine the original data with the semantic knowledge provided by labeled training pixels. We illustrate the relevance of the proposed method with an application in land cover classification using optical remote sensing images, showing that the new profiles outperform various existing features.