Applying an improved object detection algorithm for operational oil spill detection and tracking in synthetic...

Operational oil spills (OOSs) occur mainly due to illegal discharges from ships and small-scale leakages from oil platforms. Such continuous and difficultly perceptible marine pollution is characterized by high temporal and spatial uncertainty. Most current detection methods distinguish oil pollution from seawater through binary classification. Few studies have addressed categorizing oil pollution, complicating the tracking of oil pollution sources. Therefore, this study categorized oil pollution into sources of unknown origin and OOSs. For this, an improved YOLOv8 model (YOLOv8-large selective kernel (LSK)) was proposed to facilitate the high-precision detection and classification of OOSs based on synthetic aperture radar imagery. Specific improvements include integrating the LSK attention module to enhance feature extraction, using the minimum point distance intersection over union loss function to improve the localisation accuracy, and introducing the slicing-aided hyper-inference module to process large-scale SAR images. The experimental results showed that the model's multi-class mAP50 performance reached 94.2 %, increased by 7.44 % compared to the best binary classification model that uses a binary mAP50 score. The multi-class mAP50-95 performance reached 71.6 %, increased by 3.3 % compared to the original YOLOv8 model. Furthermore, three typical cases of oil spills detected by YOLOv8-LSK are presented, and the source of the spill was successfully identified by combining the automatic identification system data. This paper provides an important reference for the intelligent detection and tracking of OOSs.