BR3L: Rethinking Bilateral Reweighted Reconstruction Representation Learning for Unsupervised Hyperspectral Band Selection

Guoxia Xu, Qi Wu, Yulong Sun, Lizhen Deng, Hu Zhu

Abstract: Unsupervised band selection aims to identify the most informative spectral bands from hyperspectral images, thereby reducing spectral redundancy while preserving essential spatial–spectral structures. However, existing methods often ignore global or local structural information and noise, leading to suboptimal band subsets. In this letter, we propose a bilateral reweighted reconstruction representation model that jointly leverages spatial and spectral priors for unsupervised band selection. To simultaneously exploit both spectral correlation and spatial structure, a reconstruction residual consistency constraint is introduced into superpixels, forcing residuals of pixels inside the same superpixel to converge toward the regional mean. Additionally, spatial adaptive weight matrices $\mathbf {W}_{a}$ and spectral adaptive weight matrices $\mathbf {W}_{b}$ are designed to coordinate the weight consistency of pixels and band dimensions through bilateral weighting terms, suppress noise, and maintain edge structure. Extensive experimental results demonstrate that the proposed method outperforms several representative algorithms in terms of noise suppression and edge preservation.

External IDs:doi:10.1109/lgrs.2026.3677550