Hyperspectral Image Super-Resolution via Sparsity Regularization-Based Spatial-Spectral Tensor Subspace Representation

Hyperspectral image super-resolution (HSI-SR), which enhances the spatial resolution of hyperspectral image (HSI) by fusing a low spatial resolution HSI (LR-HSI) with a high spatial resolution multispectral image (HR-MSI), has gained increasing attention in recent years. In this article, we propose a sparsity regularization-based spatial-spectral tensor subspace representation (SR-SSTSR) approach for HSI-SR, which accurately depicts the spectral correlations and spatial self-similarities, as well as their distinctions. To fully leverage spectral correlations and spatial self-similarities, we employ the SSTSR scheme to approximate HR-HSI as three low-dimensional tensors: One spectral subspace, one spatial subspace, and one coefficient tensor. The SSTSR allows for different subspace dimensions in these subspaces, which can accurately depict the distinct low-rank characteristics in spatial and spectral modes. Since HR-MSI and LR-HSI individually contain the majority of the spatial and spectral information, the spectral and spatial subspaces are individually estimated from LR-HSI and HR-MSI via the Qatar Riyal decomposition-based tensor singular value decomposition. After identifying the subspaces, the coefficient tensor is estimated via sparsity regularization, which promotes a sparse representation of HR-HSI in both spectral and spatial subspaces, further leveraging spectral-spatial low-rank properties while effectively mitigating noise impacts from LR-HSI and HR-MSI. The coefficient optimization is solved by the proximal alternating minimization method. Experimental results from both simulated and real datasets confirm the superiority of the proposed method.