Generative Adversarial Synthetic Super-Resolution for Satellite Based Solar Panels Mapping

abstract

The widespread adoption of photovoltaic (PV) systems calls for scalable and cost-effective monitoring solutions. While high-resolution aerial imagery enables accurate detection, its limited coverage and high acquisition costs hinder large-scale use. In contrast, low-resolution satellite imagery like Sentinel-2 is freely available but lacks sufficient spatial detail for reliable PV segmentation. This study explores the use of generative super-resolution (SR) techniques to enhance Sentinel-2 imagery and improve downstream PV detection. Three SR models; Real-ESRGAN, Satlas, and S2DR3 were evaluated on diverse German PV sites. Enhanced outputs were assessed using perceptual quality metrics and segmentation performance with a DeepLabV3+ model. Results show that SR significantly improves segmentation accuracy, even with a lightweight model, making it practical for large-scale applications. Although high perceptual quality did not always correlate with better detection, Real-ESRGAN consistently yielded strong segmentation results scoring an IoU higher than 0.90 for certain cases. The findings emphasize the importance of task-specific evaluation in SR applications. Enhanced Sentinel-2 imagery generalizes to other remote sensing tasks that benefit from increased spatial detail, such as land use mapping and infrastructure monitoring.