Hyperspectral band selection via region-wise latent feature fusion and graph filter embedded subspace clustering.

Hyperspectral band selection plays a crucial role in reducing dimensionality, extracting relevant features, and improving computational efficiency in hyperspectral data analysis. Although numerous band selection methods have emerged in recent years, there remains a significant gap in exploring spatial structures and the diversity of ground objects. In this paper, we propose a region-wise latent feature fusion and graph filter embedded subspace clustering approach to address the band selection problem. Specifically, we segment the original hyperspectral image into diverse homogeneous regions using entropy rate superpixel segmentation. Next, we fuse the features from these regions into a consensus low-dimensional latent space, effectively capturing spatial information. To explore the spectral correlation among all bands, we employ a self-representation subspace clustering model on the fused latent features. Additionally, we apply a graph filter to the raw region-wise features to reduce redundant and noisy information present in the original data. By integrating these concerns into a unified framework, we facilitate mutual reinforcement among the learning sub-tasks. We conduct experiments on four public hyperspectral datasets, demonstrating a 1.93%, 0.50%, 0.76%, and 0.94% accuracy enhancement over current state-of-the-art methods. Additionally, our proposed optimization method enables the model to converge rapidly in under 15 iterations. [Display omitted] • We propose a novel approach for hyperspectral band selection. • Features from different regions are merged into a spatially-aware space. • Improves data quality by recover a smooth representation. • Demonstrates substantial accuracy boosts and rapid convergence. [ABSTRACT FROM AUTHOR]