Abstract
Marine Protected Areas (MPA) increasingly rely on automatic identification system (AIS) data to monitor fishing activity. Enforcement remains difficult because AIS is self-reported, labels of fishing behaviour are scarce, and traffic patterns differ strongly between regions. This study develops a hierarchical machine learning framework that estimates the probability of fishing from AIS position time series using only latitude and longitude. First, a self-supervised temporal encoder is trained on millions of unlabeled AIS messages to learn motion embeddings that summarize 75-minute vessel trajectories. Supervised classifiers are then trained on these embeddings to predict fishing vessel type, active fishing behaviour, and fishing gear type. Finally, a label-shift calibration step combines these outputs with operational assumptions about local vessel composition, fishing activity levels, and gear usage. This step adapts the globally trained models to the specific conditions of the St. Anns Bank MPA to produce locally calibrated fishing probabilities. On held-out data, the vessel-type and behaviour classifiers achieve areas under the receiver operating characteristic curve of 0.91 and 0.94, and average precision scores of 0.86 and 0.88, respectively, while the gear classifier obtained moderate overall performance, performing best for common gears. When applied to two years of AIS data from St. Anns Bank, the framework successfully highlights a sparse set of trajectories, allowing for more focused inspection. These results show that self-supervised motion embeddings, combined with probabilistic fusion and operational calibration, offer a label-efficient and adaptable approach to AIS-based monitoring in marine protected areas.