Sparsity Regularized Deep Subspace Clustering for Multicriterion-Based Hyperspectral Band Selection

Hyperspectral images provide rich spectral information corresponding to visible and near-infrared imaging regions, facilitating accurate classification, object identification, and target detection. However, the high volume of data creates a computational challenge in processing. The band selection process identifies specific informative and discriminative spectral bands from the data to speed up the processing without impeding the performance. This article presents an application-independent band selection framework that utilizes improved sparse deep subspace clustering and introduces an efficient multicriteria-based representative band selection (BS). The proposed sparse deep subspace clustering approach efficiently identifies the underlying nonlinear subspace structure of the data and organizes the data accordingly. The work introduces a novel, robust sparsity measure to obtain a powerful self-representation and ameliorated performance compared to the prevalent subspace clustering methods. The work subsequently selects the representative bands from each cluster by combining structural information of the band images with the statistical similarity measure. We evaluate the BS performance on standard real images using information-theoretic criterion, classification, and unmixing performance. The comparative performance assessment demonstrates that our proposed method identifies the informative bands and outperforms the other approaches in terms of the subsequent tasks.