Enhancing Spatial Resolution of Sentinel-2 Imagery through Deep Learning and Generative Adversarial Networks: GS-SRGAN

Abstract. Sentinel-2 satellites provide multi-spectral images with 13 bands at resolutions of 10, 20, and 60 m/pixel, widely used for various applications due to their cost-free access and high revisit frequency. Their open data policy has made them a key resource in remote sensing. Nonetheless, the growing need for high-resolution images has highlighted the significance of super-resolution technology (SR), which improves spatial detail through enhanced sensor precision and density. Deep learning techniques are an effective solution for enhancing Sentinel-2 images through super-resolution, improving low-resolution images by retrieving fine-grained high- frequency details. This results in high-resolution outputs from freely available data. In this research, we propose an enhancement of single-image resolution model derived from a Generative Adversarial Network, commonly abbreviated as GAN. We implemented and trained a model, named GS-SRGAN (Google Sentinel - SRGAN), built on the foundation of the Super-Resolution GAN model (SRGAN), using pairs of Google Earth and Sentinel-2 images for generating super-resolved outputs of the RGB bands from the multispectral Sentinel-2 data using a 4x scaling factor. The results from our GS-SRGAN model surpass those of current best in class models when evaluated using standard metrics such as SSIM (Structural Similarity Index) and PSNR (Peak Signal-to-Noise Ratio), enabling the super-resolved Sentinel-2 imagery for use in studies that demand very high spatial resolution.