A residual convolutional neural network for pan-shaprening

Abstract

Pan-sharpening has become an important tool in remote sensing, which normally aims at fusing a multi-spectral image with high spectral resolution and a panchromatic image with high spatial resolution. However, some problems, such as spectral distortion, are facing pan-sharpening methods. Inspired by the applications of convolutional neural network (CNN) in many areas, we adopt an effective CNN model to fulfill pan-sharpening. In our method, only the sparse residuals between the interpolated MS and the pan-sharpened image are learned, which achieves fast convergence and high pan-sharpening quality. The experimental results on real-world data validate the effectiveness of the method.