A self-learning approach for pan-sharpening of multispectral images

Due to the importance of high-resolution multi-spectral (HRM) images in many remote sensing applications, pan-sharpening techniques have been proposed to increase the spatial resolution of a low-resolution multi-spectral (LRM) image using a high-resolution panchromatic (HRP) image. In this paper, we propose a self-learning approach to pan-sharpen the LRM images. Many structures in a natural image redundantly tend to repeat in the same scale as well as different scales. These similar structures in different levels can be used to reconstruct the HRM bands with more details; in this perspective, we can construct the HRM data from the available HRP and LRM data by using self-similarity in a multi-scale procedure. The proposed method has been applied on GeoEye-1 data and DEIMOS-2 data, and then fused images compared with some popular and state-of-the-art methods in terms of several assessment indexes. The experimental results demonstrate that the proposed method can retain spectral and spatial information of the source images efficiently.