A Learned Reduced-Rank Sharpening Method for Multiresolution Satellite Imagery

This paper implements an unsupervised single-image sharpening method for multispectral images, focusing on Sentinel-2 and Landsat 8 imagery. Our method combines traditional model-based methods with neural network optimization techniques. Our method solves the same optimization problem as traditional model-based methods while leveraging neural network optimization techniques through a customized U-Net architecture and specialized loss function. The key innovation lies in simultaneously optimizing a low-rank approximation of the target image and a linear transformation from the subspace to the sharpened image within an unsupervised training framework. Our method offers several distinct advantages: it requires no external training data beyond the image being processed, it provides fast training speeds through a compact, interpretable network model, and most importantly, it adapts to different input images without requiring extensive parameter tuning—a common limitation of traditional methods. The method was developed with a focus on sharpening Sentinel-2 imagery. The Copernicus Sentinel-2 satellite constellation captures images at three different spatial resolutions, 10, 20, and 60 m, and many applications benefit from a unified 10 m resolution. Still, the method’s effectiveness extends to other remote sensing tasks, achieving competitive results in both sharpening and multisensor fusion scenarios. It is evaluated using both real and simulated data, and its versatility is shown through successful applications to Sentinel-2 sharpening and Sentinel-2/Landsat 8 fusion. In comparison with leading methods, it is shown to give excellent results.