SPATIO-TEMPORAL LEARNING OF VESSEL TRAJECTORIES FROM AIS DATA USING MULTI-MODAL CLUSTERING AND PHYSICS-INFORMED FORECASTING

Abstract

In recent years, vessel accidents have surged alongside increased maritime traffic, with collisions, often from human error, emerging as a primary concern. Enhancing navigators' situational awareness by predicting neighboring vessels' movements is essential to facilitate proactive collision avoidance. However, to effectively mitigate these risks, prediction models must deliver outputs that are not only accurate but also timely. This thesis addresses the shortcomings of existing approaches by formulating and investigating novel methodologies to produce accurate, computationally efficient, and physically consistent trajectory predictions across scenarios and geographic contexts.

The work begins with clustering-based strategies for future location prediction, addressing the heterogeneity and multi-modality of vessel mobility patterns. To this end, we propose a framework incorporating engineered features representing vessel behavior and environmental conditions. This improves the performance and interpretability of spatial predictions over traditional autocorrelation-based methods.

To improve fidelity to trajectories, we introduce a Physics-Informed Neural Network (PINN) solution that integrates kinematic constraints into deep learning. The model uses a discretized finite difference loss function based on Euler's first-order and Heun's second-order methods, ensuring output accuracy and physical consistency.

Further, we develop a transformer-based architecture that jointly forecasts future trajectories of both target and nearby vessels. This approach enables dynamic collision risk assessment through interaction-aware trajectory modeling, offering a comprehensive framework for maritime traffic safety, focusing on crowded waterways.

Collectively, these contributions advance the state-of-the-art in vessel trajectory prediction by integrating clustering, deep learning, and physics-informed modeling under a common purpose. Experiments and evaluation across diverse maritime contexts demonstrate significant improvement in predictive capability, robustness, and operational relevance. These findings lay a methodological foundation for next-generation decision-support systems, with potential implications for traffic management, regulatory compliance, and autonomous vessel operations.