Spatio-Temporal Transformer Network for Ship Trajectory Prediction with Adaptive Channel Frequency Enhancement

Ship trajectory prediction's precision is key to early warning of ensuring safe navigation and accidents. However, traditional prediction methods exhibit deficiencies in precision and stability when capturing the strong spatio-temporal dependence of ship trajectories and dealing with the inflexion of ships in traffic separation, becoming a technical bottleneck needing to be overcome urgently. Addressing the issues of weak proximity relationship perception and local perception fields in traditional time series prediction methods, this paper innovatively proposes a ship trajectory prediction network called Spatio-temporal Transformer Network with Adaptive Channel Frequency Enhancement(ST-Transformer with FE). By adopting the Transformer as the underlying architecture, combined with convolutional neural's strong local spatial perception capabilities networks (CNN) and the advantages of global context dependency modeling of extended long short-term memory networks(xLSTM), a spatio-temporal Transformer network is designed to improve the ability to map complex spatio-temporal relationships. Further, a frequency-enhanced channel attention mechanism (FECAM) is introduced to adaptively analyze the rich frequency information in ship AIS information, using Discrete Cosine Transform to eliminate Gibbs effects and highlight impact of strongly correlated frequency features. Ablation experiments and extensive performance comparisons are conducted, and the results show that the ST-Transformer with FE network outperforms other classical prediction network in predictive performance.