Deep Subspace Clustering Preserving Distribution for Hyperspectral Images

Abstract

Hyperspectral images usually lie on low-dimensional nonlinear manifolds, leading to a challenging clustering task. Deep subspace clustering–based methods have been successful in this task by converting features to linear embedding using an auto-encoder with ${\ell }_2$ norm. In this setting, the embedding of the auto-encoder just learns the implicit geometric structure of original samples and loses its distribution. However, the sample distribution alignment is a generalisation of sample alignment. To remedy these issues, in this article, we propose a promising method, named DSCOT, for improving subspace clustering. Specifically, we measure the reconstruction error of the auto-encoder leveraging optimal transport distance that explicitly embeds geometric distance between samples and preserves embedding distribution in observation space. It results in more appropriate embedding for subspace clustering. Several experiments on three widely used databases show that the proposed method is superior to most state-of-the-art methods.