Satellite-based vessel detection under turbulence and jitter: a degradation-aware comparison

Maritime domain awareness increasingly relies on satellite-based Earth observation (EO) for wide-area vessel monitoring. Low Earth orbit (LEO) platforms enable flexible acquisition, but optical imagery can be degraded by atmospheric turbulence and platform jitter. This paper presents a degradation-aware evaluation of deep learning-based maritime vessel detection that links algorithmic performance to the physical constraints of spaceborne optical imaging. Using real optical satellite imagery and physically motivated synthetic turbulence and jitter models, we compare YOLOv8, YOLOv11, RetinaNet, and Mask R-CNN under controlled degradation. Experiments assess nominal detection performance, CPU inference time, target size, and the relative out-of-distribution sensitivity of nominally trained detectors. Results show that degradation is initially dominated by recall loss: jitter causes a marked performance drop when first introduced, whereas turbulence produces a more progressive decline. YOLOv8 and YOLOv11 provide the strongest nominal performance and the best performance--latency trade-off. Under some severe conditions, Mask R-CNN retains comparatively higher recall and F1-score, although absolute performance remains insufficient for reliable standalone detection. Open-sea imagery performs substantially better than the complete dataset including coastal scenes, while restricting the analysis to vessels at least 50 pixels in size improves detection under degradation.