Hypersharpening Using Intercluster Sample Equalization and Intracluster Low-Rank Constraints

Hyperspectral remote sensing plays an irreplaceable role in fine earth observation tasks, where the complete preservation of fine and rare spectral features is the core premise for subsequent quantitative applications. Hypersharpening, which enhances the spatial resolution of low-spatial-resolution hyperspectral (LH) images using the corresponding panchromatic (Pan) image over the same scene, is a critical technique to bridge the inherent spatial-spectral tradeoff of hyperspectral imaging. Dictionary learning-based hypersharpening is a powerful way to achieve this goal by optimizing an objective that minimizes the sum of representation errors for all pixels. However, it may lead to larger fusion errors for small-sized clusters, which is not conducive to subsequent applications, such as the small and weak spectral target detection or classification. Faced with this issue, a sample equalized hypersharpening method is proposed in this letter. Specifically, an intercluster sample equalization model is proposed to balance the number of spectral samples across clusters. Then, a subpixel-shift decomposition strategy is introduced to generate successive equalized subpixel-shifted sub-Pan images consistent with the LH resolution, and to transform the hypersharpening problem into a series of fusion problems between the above sub-Pan and LH images. Finally, a joint optimization process, driven by spatial continuity and spectral similarity, is applied under intracluster low-rank constraints to accurately extract the spectral information from the spectral dictionary within the LH image and the spatial information from its coefficient matrix within the LH and Pan images. Extensive experiments against related leading-edge methods indicate that our proposed method can reconstruct high-spatial-resolution hyperspectral (HH) images with superior performance.