Super-Resolution Based Deep Learning Techniques for Panchromatic Satellite Images in Application to Pansharpening

Abstract

Pansharpening is a technique that fuses the coarser resolution of multispectral imagery (MS) with high spatial resolution panchromatic (PAN) imagery. Pansharpening is prone to spectral distortions based on the nature of the panchromatic band. If the spatial features are unclear in the panchromatic image, the pan-sharpened image will not be able to produce clear images. Super-Resolution (SR) is a technique that enhances minute details of the features in the image, thereby improving spatial information in the image. By fusing the Multispectral image with the super-resolved panchromatic image, there is a chance for producing high-quality multispectral imagery (pan-sharpened image). In this paper, ten state-of-the-art super-resolution based on deep learning techniques are tested and analyzed using ten different publicly available panchromatic datasets. On analysis, a feedback network for image super-resolution (SRFBN) technique outperforms the other algorithms in terms of sharp edges and pattern clarity, which are not visible in the input image. The proposed method is the fusion of SR applied PAN image with the MS image using a benchmarked Band Depended Spatial Detail (BDSD) pansharpening algorithm. The proposed method experiments with six datasets from different sensors. On analysis, the proposed technique outperforms the other counterpart pansharpening algorithms in terms of enhanced spatial information in addition to sharp edges and pattern clarity at reduced spectral distortion. Hence, the super-resolution based pansharpening algorithm is recommended for high spatial image applications.