DuSR-Net: a statistical- and optimization-prior-guided deep-unrolled network with cross-scale manifold alignment for hyperspectral image classification

Hyperspectral image (HSI) classification remains challenging because of its high spectral dimensionality, limited labeled samples, and spatial-spectral variability. We propose DuSR-Net, a deep-unrolled network guided by complementary statistical spectral views and an optimization-derived sparse-representation prior. The proposed method combines two statistical preprocessing views and one optimization-derived structural prior, rather than an explicit physical model of hyperspectral image formation. First, principal component analysis and variance-based band selection provide complementary low-dimensional and information-preserving statistical views of the input. Second, a Cross-Scale Manifold Alignment module uses the PCA-derived representation to guide the refinement of features extracted from the selected high-variance bands. Third, an ADMM-inspired backbone introduces an optimization-related sparse-representation prior through learnable iterative updates and soft-thresholding. These components improve the structural transparency of the network because their roles correspond to dimensionality reduction, feature alignment, and sparse regularization, respectively. We emphasize that the method does not explicitly model radiative transfer, atmospheric scattering, material reflectance, or sensor response, and therefore its interpretability is algorithmic rather than physics-based. Experiments on multiple benchmark datasets demonstrate competitive classification accuracy and spatial consistency under limited-label settings. Code: https://github.com/yongqichang202-beep/DuSR-Net