Skip to content
Open access

Low-Rank-Guided Tensor Mode Decoupling and Hierarchical Wavelet Synergy Network for Hyperspectral Unmixing

2026 · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · Vol 19, pp. 31964-31984 · 0 citations · 57 references

Abstract

Hyperspectral unmixing aims to separate endmember spectra from mixed pixels and estimate their abundance fractions. Despite remarkable advances in deep learning, accurately recovering fine-grained land-cover boundaries and textural details remains challenging. To this end, we propose a low-rank-guided tensor mode decoupling and hierarchical wavelet synergy network for hyperspectral unmixing, termed TMDWS-Net. First, an adaptive tensor-rank-guided denoising module employs a log-scaled adaptive rank selection module to suppress noise while preserving the stability and reliability of the 3-D spatial–spectral manifold structure. Second, a tensor mode-decoupled self-attention (TMDSA) module overcomes the limitations of conventional attention for tensor data by decoupling self-attention along the rows, columns, and spectral modes, efficiently capturing cross-band spectral correlations and global long-range spatial dependencies. To address the loss of local details, a wavelet-guided multiscale sparse cross-attention (WMSCA) module is constructed. By leveraging the multiresolution properties of the band sequential processing stationary wavelet transform and a low-frequency-guided sparse cross-attention mechanism, this module fully exploits multiscale frequency-domain structural information and selectively reinforces critical high-frequency responses, significantly enhancing the sensitivity to and representational capacity for fine-grained features. Finally, a global–local structure-aware fusion module adaptively fuses the global contextual features from the TMDSA module with the local frequency-domain detail features from the WMSCA module, achieving synergistic optimization and complementary gains between global structural information and local textural details. Experimental results on both synthetic and real hyperspectral datasets demonstrate that the proposed method outperforms state-of-the-art approaches in terms of unmixing accuracy and robustness.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.