Enhancement of Sentinel-2A Images for Ship Detection via Real-ESRGAN Model

Abstract

Ship detection holds great value regarding port management, logistics operations, ship security, and other crucial issues concerning surveillance and safety. Recently, ship detection from optical satellite imagery has gained popularity among the research community because optical images are easily accessible with little or no cost. However, these images’ quality and quantity of feature details are bound to their spatial resolution, which often comes in medium-low spatial resolution. Accurately detecting ships requires images with richer texture and resolution. Super-resolution is used to recover features in medium-low resolution images, which can help leverage accuracy in ship detection. In this regard, this paper quantitatively and visually investigates the effectiveness of super-resolution in enabling more accurate ship detection in medium spatial resolution images by comparing Sentinel-2A images and enhanced Sentinel-2A images. A collection of Sentinel-2A images was enhanced four times with a Real-ESRGAN model that trained PlanetScope images with high spatial resolution. Separate ship detections with YOLOv10 were implemented for Sentinel-2A images and enhanced Sentinel-2A images. The visual and metric results of both detections were compared to demonstrate the contributory effect of enhancement on the ships’ detection accuracy. Ship detection on enhanced Sentinel-2A images has a mAP50 and mAP50-95 value of 87.5% and 68.5%. These results outperformed the training process on Sentinel-2A images with a mAP value increase of 2.6% for both mAP50 and mAP50-95, demonstrating the positive contribution of super-resolution.