A compressed-sensing-based approach for remote sensing image fusion

Abstract

Remote sensing image pan-sharpening is an image fusion process which fuses a low-resolution multi-spectral (LRMS) image with its corresponding high-resolution panchromatic (HRP) image to create a high-resolution multi-spectral (HRMS) image. In this paper, pan-sharpening methods based on compressed sensing (CS) theory are proposed. In the proposed methods, the HRP and LRMS dictionaries are learned from the input images (HRP, LRMS). Moreover, this paper proposes a new algorithm to reconstruct the unknown HRMS image by considering remote sensing physics. The proposed algorithm extracts non-overlapping patches from input images and provides an initial estimation of HRMS dictionary. Then, the initial HRMS dictionary and LRMS image are used to reconstruct unknown HRMS image. The algorithm neither needs to extract overlapping patches, nor training dataset. So, it makes the proposed methods fast and practical. Furthermore a high-pass filter is used to preserve more details in the fusion process. The proposed methods are tested on WorldView-2 and QuickBird satellite images and these results are compared with several popular and state-of-the-art methods quantitatively and visually.