Enhancing Satellite Images of FACSAT-1 through Generative Adversarial Networks for Super-Resolution

Abstract

Abstract Satellite images have diverse applications across scientific, commercial, and other domains. As a result, institutions are increasingly deploying Earth observation satellites to cater to their specific needs. This is the case of the Colombian Aerospace Force, which launched its nanosatellite, the FACSAT-1, to contribute to developing the space sector in Colombia. However, in some cases, captured images may need more quality and resolution for their intended purposes. Numerous image processing tools have been developed to enhance and optimize satellite imagery to address this challenge. Deep learning techniques, particularly Generative Adversarial Networks, have recently shown significant advancements in image processing. This paper investigates the widespread application of Generative Networks for satellite and aerial imagery, specifically focusing on super-resolution tasks. Super-resolution involves increasing the resolution of satellite images by up to four times their original size. The study presents the implementation and training of four different generative models and evaluates their performance using qualitative and quantitative measures. Two metrics, namely the maximum signal-to-noise ratio and the structural similarity index, are employed for comparative analysis. By assessing the output of each generative model, this research aims to determine their efficacy in enhancing satellite imagery resolution.