Conditional Attention Improves Ship Trajectory Prediction with Context

Predicting ship trajectories over long horizons is crucial for maritime safety and autonomous navigation, but existing Transformer-based models often treat vessel motion in isolation, relying only on historical kinematic data. This overlooks critical external factors such as weather conditions and static vessel characteristics, which profoundly influence maritime navigation. Researchers developed the Conditional Informer, an innovative encoder-decoder architecture designed to address this limitation. This model frames trajectory prediction as a conditional generation task, employing a unique Conditional Attention mechanism. This mechanism allows the vessel's state to explicitly query environmental contexts, thereby encoding the physical understanding that weather modulates, rather than is generated by, vessel dynamics. Furthermore, to combat data intermittency common in real-world scenarios, a Modality Masking training strategy was introduced. This prevents performance degradation during sensor fallbacks. Extensive experiments using AIS and ERA5 data showed that the Conditional Informer improved prediction accuracy by 15.4% over baselines when context was available. Modality Masking also drastically reduced fallback error, demonstrating its effectiveness in preventing shortcut learning.