A Spectral-Preserving Zero-Shot Technique for Hyperspectral Pansharpening

Hyperspectral pansharpening is a key fusion process aimed to enhance the spatial resolution of a hyperspectral image benefiting of a simultaneously acquired higher resolution panchromatic band. Promising approaches, which leverage on deep learning models, are optimized in unsupervised manners using full-resolution real datasets with no reference. The lack of ground-truths imposes the use of a two-fold composite loss to ensure both spectral and spatial quality according to some trade-off criterion. The common solution, however, is agnostic with respect to the spectral-spatial balance, giving the same relevance to both properties. Indeed, the spectral information is the most distinctive feature to preserve in any application based on hyperspectral data. Motivated by this last observation, in this work we propose a spectral-preserving fusion approach that puts the spectral quality ahead thanks to a suitably defined unsupervised loss. Our experiments show very promising results with really low spectral distortions, on average and uniformly along the spectral axis, ensuring good sharpening levels too.