Predicting the destination port of fishing vessels utilizing transformers

Vast databases on historical ship traffic are currently freely available in the form of AIS (Automatic Identification System) messages dating back to as early as 2002. This provides a rich source for training deep learning models for predicting various behaviors of vessels, which in this context is motivated by resource management of fisheries. In this paper, we explore the possibility for combining a transformer model’s powerful capabilities for long-term path prediction with added logic to infer probable destination harbors for fishing vessels. An additional baseline model is also developed for comparison, based on historically preferred harbors for the vessels. With AIS data from the Troms and Finnmark region of Norway, the prediction accuracy of the trained model is found to be highly dependent on the number of past tracked positions of the vessel. We foresee that a new and more precise model will need to incorporate not only dynamic AIS data, but static information about harbors and vessel types during training and inference. • Fishing vessels do not always correctly report landing harbors for catches. • Fishery management inspections rely on accurate predictions of vessels destinations. • Training a transformer model on AIS data enable inference on vessel trajectories. • Transformer path forecasting combined with logics allows target harbors prediction.