Super Resolution Approach Using Generative Adversarial Network Models for Improving Satellite Image Resolution

Abstract

Recently, the number of satellite imaging sensors deployed in space has experienced a considerable increase, but most of these sensors provide low spatial resolution images, and only a small proportion contribute with images at higher resolutions. This work proposes an alternative to improve the spatial resolution of Landsat-8 images to the reference of Sentinel-2 images, by applying a Super Resolution (SR) approach based on the use of Generative Adversarial Network (GAN) models for image processing, as an alternative to traditional methods to achieve higher resolution images, hence, remote sensing applications could take advantage of this new information and improve its outcomes. We used two datasets to train and validate our approach, the first composed by images from the DIV2K open access dataset and the second by images from Sentinel-2 satellite. The experimental results are based on the comparison of the similarity between the Landsat-8 images obtained by the super resolution processing by our approach (for both datasets), against its corresponding reference from Sentinel-2 satellite image, computing the Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity (SSIM) as metrics for this purpose. In addition, we present a visual report in order to compare the performance of each trained model, analysis that shows interesting improvements of the resolution of Landsat-8 satellite images.