Recent advancements in South Korea’s space technology have increased the availability of high-resolution optical satellite imagery, enabling extensive research across diverse fields such as urban planning, environmental monitoring, and disaster response. However, optical satellite images often suffer from quality degradation caused by atmospheric aerosols, cloud coverage, and variations in solar irradiance. This study aims to address these challenges by developing a robust satellite image quality assessment method using relative edge response (RER) as the primary metric. RER quantifies the optical system’s ability to reproduce edges clearly in imagery and is integral to metrics such as modulation transfer function (MTF), ground resolved distance (GRD), and national image interpretability rating scales (NIIRS). We employed transfer learning based on a pretrained ResNet-18 model to evaluate the feasibility of high-resolution satellite image quality assessment. KOMPSAT-3A imagery was used to construct the dataset, covering diverse regions such as urban areas, rivers, oceans, and forests. Data augmentation techniques using Gaussian blur were applied to overcome the limited availability of real-world satellite images within specific RER ranges, generating a synthetic dataset of 50,400 images categorized into 21 classes based on RER values ranging from 0.20 to 0.40. Results demonstrated that the model effectively captured sharpness-related features from satellite images, even in challenging environments such as oceans and rivers where edge characteristics are ambiguous. Despite relatively limited training data for non-urban areas, predictions were successful with minimal errors. This study highlights the potential applicability of ResNet-based transfer learning for assessing high-resolution satellite image quality and lays the groundwork for incorporating additional sharpness metrics such as MTF, NIIRS, and GRD into future methodologies.



Keywords

Satellite image quality assessment, Image quality assessment, Relative edge response, KOMPSAT, High-resolution satellite, ResNet-18, Transfer learning