Sample Latent Feature-Associated Low-Rank Subspace Clustering for Hyperspectral Band Selection

Abstract

In recent years, subspace clustering has become increasingly popular and achieved great success in band selection (BS) of hyperspectral imagery. However, current subspace clustering approaches are mostly insufficient in capturing the fine spatial structure and spectral correlation of image. Therefore, this article proposes a sample latent feature-associated low-rank subspace clustering model (SLFLRSC) tailored for BS. First, this model utilizes entropy rate segmentation to capture the rich spatial information of image. Meanwhile, Laplacian eigenmaps is employed to extract key latent features of samples in each region, enabling a joint representation of the original image that maximizes retention of key information while reducing noise and data dimensionality. Second, considering both short-range and long-range relationships of samples, a sample-spatial consistency constraint is formulated to reinforce the connections among similar samples across homogeneous and heterogeneous regions. Finally, a band-spectral local constraint is devised to rationally evaluate the global and local band adjacencies, incorporating both band similarity and spatial distance metrics. These initiatives provide a favorable condition for subspace clustering and BS tasks. The efficacy and reliability of SLFLRSC are confirmed through experiments on three datasets.