Learning Non-Local Spatial-Spectral Correlation for Hyperspectral Image Super-Resolution

Non-local spatial-spectral correlation is critical for hyperspectral image super-resolution but remains underexplored, and existing methods often neglect this correlation, resulting in suboptimal reconstruction. In this work, we propose SSCNet, a novel framework that achieves high-fidelity hyperspectral image super-resolution by effectively modeling both local and global relationships. SSCNet leverages a parameter-sharing strategy to learn spatial-spectral consistency biases, employs a progressive non-local receptive field to capture long-range dependencies across spatial and spectral dimensions, and introduces a spatial-spectral affinity loss to explicitly enforce coherence between spatial and spectral information, thereby enhancing reconstruction fidelity. Extensive experiments on multiple public datasets demonstrate that SSCNet significantly outperforms state-of-the-art methods.