Region-Aware 3D Tensor Decomposition Exploiting Spectral Symmetry for Hyperspectral Image Denoising

Spectral fidelity is critical for accurate hyperspectral image (HSI) processing. A key characteristic of HSI data is the strong correlation between spectral bands, which manifests as structured symmetry in spectral covariance matrices. While global low-rank tensor decompositions leverage this spectral structure, they often neglect the significant spatial heterogeneity present in real-world scenes. To address this limitation, we propose a Region-Aware 3D Tensor Decomposition (RA-3DTD) framework that balances global spectral consistency with local spatial adaptation. Our approach first performs residual energy-based region detection to identify complex regions within the hyperspectral cube, and then applies localized Higher-Order Orthogonal Iteration (HOOI) specifically to those regions requiring enhanced detail preservation. This two-phase design incorporates global low-rank constraints with local spatial processing, improving denoising accuracy. Extensive experiments on four benchmark datasets (Pavia_80, Indian Pines, Salinas, and Pavia University) demonstrate the effectiveness of our method compared to five leading model-based baselines including BM3D, LRMR, NLR, LRTD, and FastHyDe. Our approach achieves a 1.33 dB increase in PSNR over a leading model-based competitor (FastHyDe) in complex urban scenes while maintaining strong structure fidelity as measured by SSIM and SAM metrics.