Interpretable Vessel Behavior Inference from AIS Trajectories via CoT-Enhanced LLMs

Understanding vessel behavior is critical for maritime surveillance, traffic monitoring, and anomaly detection. Traditional methods rely heavily on statistical rules or machine learning models that often lack interpretability. In this study, we propose a novel framework that leverages large language models (LLMs) augmented with Chain-of-Thought (CoT) reasoning to infer vessel behavior in a transparent and explainable manner. Our approach consists of three main steps: trajectory feature extraction, feature fusion with natural language prompt construction, and CoT-enhanced reasoning using instructiontuned LLMs. We classify vessel behavior into four types: sailing, anchoring, waiting, and maneuvering. Experimental results on real-world AIS segments demonstrate that high-capacity models such as DeepSeek-V3 achieve superior performance in accuracy ($\mathbf{9 8. 5 \%}$), readability, and reasoning consistency, significantly outperforming smaller-scale models. Moreover, CoT-based responses not only provide accurate predictions but also generate interpretable justifications, which are essential in maritime domains where explainability is critical. This work highlights the potential of language models for structured spatial-temporal reasoning and opens new directions for transparent behavior analysis in marine intelligence applications.