Vessel detection leveraging satellite imagery and YOLO in maritime surveillance

Abstract

Vessel detection is essential for maritime surveillance, supporting the monitoring of fishing, commercial, and transportation activities, as well as detecting suspicious or illegal vessels and aiding search and rescue operations. Satellite imagery has become increasingly valuable for Earth Observation applications due to its wide range of spatio-temporal cover. The European Space Agency offers free optical imagery from the Sentinel-2 satellite, providing a cost-effective solution for large-scale maritime surveillance. Deep learning models, particularly You Only Look Once (YOLO), have demonstrated impressive object detection performance and are widely regarded as state-of-the-art for real-time applications. This article addresses an existing research gap by comparing YOLO versions 8 and 10 for vessel and wake detection using a dataset provided by CEiiA. The models were tested to determine the optimal configuration for this task and after evaluation, and a YOLOv8 configuration was selected. Following this selection, the dataset was expanded leading to further performance improvements. The final YOLOv8 model achieved an F1-score of 88.69 % (with 90.47 % precision and 86.98 % recall), an IoU of 72.91 %, a mAP50 of 87.53 % and mAP50-95 of 61.20 %. Single class training improved bounding box localization and delineation, however it slightly reduced performance, possibly due to a loss of contextual information. Additionally, testing on high-resolution images confirmed that models trained on lower-resolution data can still perform relatively well on higher-resolution images, with potential for further improvement through fine-tuning. These results support that YOLO models are highly effective for real-time vessel detection and can be reliably applied in maritime surveillance.