Uncertainty-aware ship trajectory prediction via Spatio-Temporal Graph Transformer

Accurate trajectory prediction is essential for enabling the autonomous navigation of unmanned ships. Recent advancements in Deep Learning (DL) based trajectory prediction using AIS data have positioned this area as a key focus in maritime transportation research. However, existing studies often fail to address trajectory uncertainty adequately. The ability to model uncertainty is crucial, as it not only quantifies the confidence in prediction results but also enhances a model’s adaptability to complex and dynamic maritime environments. Addressing this gap requires innovative approaches to trajectory prediction that effectively account for uncertainty. This paper proposes a new trajectory prediction model, the Spatio-Temporal Graph Transformer with Probability (STGTP), which seamlessly integrates spatio-temporal features with probabilistic trajectory modelling. The proposed STGTP model introduces several innovations, including a temporal attention module to capture dynamic temporal variations in ship movements and a Transformer-based Graph Convolution (TGConv) to model spatial interactions, enhancing predictive accuracy. It employs a Gaussian heatmap representation for probabilistic trajectory modelling and a Vision Transformer to extract features that quantify prediction uncertainty effectively. These components enable STGTP to provide robust and reliable prediction while explicitly modelling uncertainty, improving the safety and adaptability of autonomous navigation systems. The model’s performance was systematically evaluated across three distinct maritime regions using established metrics: Average Displacement Error (ADE), Final Displacement Error (FDE), and Fréchet Distance (FD). A comparison with ten baseline models demonstrates that the proposed STGTP model consistently outperforms all existing approaches across all evaluation metrics. These results underscore the model’s overall superiority and effectiveness in maritime transportation. By integrating probabilistic and spatio-temporal modelling, STGTP significantly enhances the accuracy of ship trajectory forecasting, marking a key advancement toward achieving robust, real-time autonomous navigation in maritime environments. • Introduced a novel ship trajectory prediction benchmark for Maritime Autonomous Surface Ships (MASS). • Developed an advanced trajectory prediction algorithm that integrates spatio-temporal and probabilistic modelling. • Pioneered an uncertainty-aware feature modelling approach to improve prediction reliability and interpretability. • Validated the model’s performance across three benchmark datasets using three key evaluation metrics.