Sequential Modeling of Complex Marine Navigation

This paper presents a sequential modeling framework for complex marine navigation, addressing the challenge of predicting and optimizing vessel behavior in constrained waterways with irregular traffic patterns.

Problem

Marine navigation in complex waterways — fjords, harbors, congested shipping lanes — involves sequential decision-making under uncertainty that standard trajectory models handle poorly. Vessels must respond to other traffic, weather changes, and navigational constraints simultaneously. Existing approaches typically model position prediction in isolation, disconnected from the decisions that produced the trajectory.

Approach

We model vessel navigation as a sequential decision process and apply recurrent neural architectures (LSTM + attention) to learn behavioral patterns from AIS (Automatic Identification System) tracking data. The model learns to represent not just where a vessel is, but what navigational state it's likely in — approach, transit, maneuvering, anchoring.

Key technical contributions:

• A state-space formulation that encodes vessel type, speed, heading, and contextual features (nearby vessels, channel boundaries) into a unified sequential representation
• Attention over historical trajectory to capture long-range dependencies in route behavior
• Anomaly detection framing: deviations from predicted sequential state signal unusual behavior

Results

On held-out vessel tracks from BC coastal waters, the sequential model significantly outperforms position-only baselines on 30-minute trajectory prediction. The anomaly detection application achieves useful precision/recall for flagging unusual vessel behavior, validated against known incident logs.

This work is part of a larger research program with NRC on marine vessel optimization — see also the fuel consumption prediction paper from the same collaboration.