Multidynamic Graph Convolutional Networks for Vessel Trajectory Prediction

Abstract

With the increasing frequency of global maritime activities, the importance of accurate vessel trajectory prediction becomes increasingly prominent, playing a vital role in maritime safety, risk assessment, military strategic deployment, and maritime traffic management. However, current vessel trajectory prediction methods often fail to fully consider the complex interactions between vessels and the dynamic relationship with the maritime environment, limiting the accuracy and adaptability of the predictions. To address this, this study proposes a spatiotemporal convolutional trajectory prediction network framework based on multidynamic graph inference, named distance-geometric–two-stage convolutional network (named DG-TSCN). This framework utilizes a multidynamic graph inference module to simulate diverse vessel interactions, fully revealing potential social interaction links. Building on this, the spatiotemporal graph convolutional network module effectively models the temporal and spatial dependencies in vessel data. Finally, it concentrates on extracting key temporal features through the temporal feature extraction convolutional network module, significantly improving prediction accuracy. The experimental results show that DG-TSCN has high predictive accuracy and adaptability in multivessel and multistep trajectory prediction, effectively overcoming the shortcomings of traditional methods that neglect the interactions between vessels and the impact of environmental dynamics, providing strong technical support for military surveillance and civil maritime traffic management.