Super‐Resolution of Planetary Images Based on Generative Adversarial Network

Abstract

Currently, satellite imagery serves as the primary means of observing terrestrial planets such as the Mars, the Moon, and Mercury. Enhancing the resolution and quality of these images can provide more detailed insights into planetary surfaces. However, improvements in image quality are often limited by the constraints of sensor technology and the high costs associated with upgrading equipment. In this study, we introduce the PLanetary Enhanced Super‐Resolution Generative Adversarial Network (PL‐ESRGAN), a super‐resolution technique designed to enhance planetary images. To mitigate common issues in planetary imagery, we developed a domain‐specific model that incorporates a tailored degradation process simulating factors such as noise, data gaps, blur, low spatial resolution, and lighting inconsistencies. Low‐resolution (LR) images were generated through this process and paired with their corresponding high‐resolution (HR) versions to train the PL‐ESRGAN. The trained model successfully achieved a four‐time increase in resolution, mitigating the degradation issues in the images. It performed consistently well across data sets from Mars, the Moon, and Mercury, surpassing traditional interpolation, CNN‐based, Transformer‐based, and hybrid‐design methods in both quantitative metrics and visual clarity. Notably, it provided clearer representations of small surface features, including impact craters, rocks, and aeolian landforms. To facilitate further research and application, the code has been open‐sourced, providing a powerful and accessible tool for enhancing the resolution and clarity of planetary images, enabling more accurate interpretation of surface features and supporting a wide range of downstream analyses.