Fishing ship trajectory prediction considering trajectory behavior characteristics

• A novel ST-Transformer model is proposed, which considers the spatio-temporal interactions and group navigation behavior characteristics of fishing ship trajectories. This model, based on the Transformer framework, incorporates the trajectory information of fishing ships in navigation states and stacks GCN layers to excavate multi-level spatial interactions with surrounding vessels. • In order to enhence the model's ability to capture the behavior characteristics of convoy navigation of fishing ships and improve the accuracy of fishing ship trajectory prediction, the concept of companion loss function is proposed. • Through the evaluation and comparative analysis of different experimental scenarios and methods, the ST-Transformer model has exhibited excellent accuracy, robustness, and generalization ability, and can achieve fishing ship trajectory prediction with relatively high precision. Understanding and accurately predicting the trajectory changes of fishing ships is of great significance for effectively preventing collisions between merchant and fishing ships and enhancing the safety supervision of fishing ships. Due to the problems that the existing methods for predicting the trajectories of fishing ships are few and have low accuracy, and the spatiotemporal interaction information and the characteristics of companion navigation of fishing ships are not fully utilized, a Space-Time-Transformer (ST-Transformer) model considering the spatiotemporal interaction and companion navigation behavior characteristics of fishing ship trajectories is proposed. This model is based on the Transformer framework and considers the trajectory information of fishing ships under navigation conditions, and through stacking Graph Convolutional Networks (GCN), it mines the multi-layer spatial interaction with surrounding ships. Meanwhile, in order to increase the model's ability to capture the behavior characteristics of convoy navigation of fishing ships and improve the accuracy of fishing ship trajectory prediction, the concept of companion loss function is proposed, and the Fréchet distance is used to measure the similarity of fishing ship trajectories, to better quantify the dynamic spatial-temporal distance between fishing ships from an overall perspective. Through the evaluation and comparative analysis of different experimental scenarios and experimental methods, the ST-Transformer model has exhibited excellent accuracy, robustness, and generalization ability, and can achieve fishing ship trajectory prediction with relatively high precision.