HDFormer: A transformer-based model for fishing vessel trajectory prediction via multi-source data fusion
Siyuan Lin, Yufei Jiang, Feng Hong, Lixiang Xu, Haiguang Huang, Bin Wang
Abstract
Predicting the trajectory of fishing vessels is essential for enhancing navigational safety and aiding fishery management. Compared to other vessels, trajectory prediction for fishing vessels presents a unique challenge as these vessels exhibit distinct movement patterns; they often make frequent turns while fishing and follow straight paths during steaming. This study introduces HDFormer, a Transformer-based deep learning model engineered to predict the future trajectories of fishing vessels up to one and a half hours in advance. HDFormer utilizes two innovative attention mechanisms – Trajectory Attention and Environment Attention – that integrate spatial features from historical trajectory segments, fishing effort distributions, and hydrological factor fields. These mechanisms help clarify the interactions among these elements, offering insights into potential future operational states. Tested with the VMS dataset of 418 seiners in the East China Sea and hydrological data from the Copernicus Climate Data Store, HDFormer achieves a mean absolute error of 0.773 nautical miles and a final displacement error of 1.642 nautical miles. HDFormer is readily adaptable to other oceanic regions for long-term trajectory forecasting, and its innovative Environmental Attention mechanism has broad potential applications in fishing research.