Pan-sharpening method based on unsupervised and divide-and-conquer strategies

Pan-sharpening is a remote sensing image super-resolution technique that aims to enhance the resolution of Low-resolution Multispectral (LRMS) images by fusing them with Panchromatic (PAN) images, resulting in High-resolution Multispectral (HRMS) images. While deep learning-based Pan-sharpening methods are currently popular, they encounter several challenges, including unstable training, lack of real reference HRMS images, convergence difficulties, and slow training speed. To overcome these issues, we propose an innovative unsupervised Pan-sharpening network that leverages attention mechanisms. Our method adopts a Generative Adversarial Network (GAN) structure, incorporating attention mechanisms within the generator. This facilitates the extraction of valuable information from both PAN and LRMS images while prioritizing important regions during the fusion process. Furthermore, we introduce a novel framework based on a divide-and-conquer strategy to expedite the Pan-sharpening process. The initial step involves segmenting the image into sub-images with varying restoration difficulties using gradient classification and a classification model. Subsequently, distinct Pan-sharpening models are applied to fuse sub-images from different categories. Our experiments demonstrate the efficacy of the proposed method in terms of both quantitative metrics and visual results.