Unsupervised Unmixing-based Watershed Segmentation of Hyperspectral Images

Abstract

This thesis is part of the extensive literature on the field of hyperspectral imaging (HSI). Hyperspectral images capture detailed spectral information for each spatial location of a scene, enabling the identification and analysis of materials based on their spectral signatures. Due to limited spatial resolution and the inherent heterogeneity of natural scenes, the spectrum observed at a single pixel often corresponds to a mixture of several constituent materials. Hyperspectral unmixing (HU) addresses this problem by decomposing pixel spectra into a set of pure spectral signatures, known as endmembers, and their associated fractional abundances. While HU provides physically interpretable representations of hyperspectral data, many image analysis tasks additionally require spatially coherent segmentation of the scene. Mathematical morphology provides a powerful theoretical and computational framework for image analysis, offering operators for filtering, denoising and segmentation. In particular, the watershed transform is widely used for producing spatially coherent segmentations. However, the extension of morphological processing to hyperspectral images remains challenging, as vector-valued data lack a natural total ordering and therefore do not directly admit the complete lattice structure required by classical morphological operators. This thesis presents a fully unsupervised hyperspectral image segmentation pipeline that couples hyperspectral unmixing with marker-guided watershed segmentation. Abundances are estimated using a regularized nonnegative matrix factorization HU scheme (SLRCNMF) or a multilinear rank-(L, L, 1) block-term decomposition (LL1-BTD)-based solver (GradPAPA-NN), then spatially regularized via opening/closing by reconstruction. Markers are obtained through either a tunable threshold or Otsu's automatic thresholding method applied on the smoothed abundance maps. Segmentation is then performed with two variants, namely either watershed on an abundance-aware gradient or a generalized multi-relief-based flooding mechanism. Experiments on Jasper Ridge, Salinas-A, Pavia University and Indian Pines datasets demonstrate coherent segmentations and competitive quantitative performance, including 100% overall accuracy and normalized mutual information on Salinas-A. To the best of our knowledge, this is the first fully unsupervised, optimization-based framework to bridge hyperspectral unmixing with mathematical morphology.