STGVAD: Spatio-Temporal Graph-Based Vessel Behavior Anomaly Detection

Predicting vessel trajectories and detecting anomalous behaviors are fundamental yet challenging tasks in the maritime domain, primarily due to the dynamic and unstructured nature of the marine environment. While Spatio-Temporal Graph Neural Networks (ST-GNNs) have shown great promise in modeling relational and temporal dependencies, their reliance on fixed spatial anchors limits their applicability to maritime scenarios, where such anchors are inherently absent. To address this limitation, we propose a novel GNN-based framework, denoted as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Spatio-Temporal Graph-based Vessel behavior Anomaly Detection</i> (STGVAD), that redefines graph construction for vessel behavior modeling. Instead of relying on static spatial points, we represent each timestamped vessel state as a node and construct a unified multi-ship trajectory graph by linking temporally adjacent nodes and incorporating spatial proximity using the OPTICS clustering algorithm. This enables the joint modeling of temporal dynamics and inter-vessel interactions within a single graph structure. Our framework employs a GNN encoder to capture spatio-temporal patterns, followed by a time-series module. To support objective evaluation, we manually injected anomalies into real AIS dataset. Experimental results demonstrate that STGVAD significantly outperforms conventional time-series baselines, particularly under conditions with scarce anomalies. These findings highlight the importance of joint spatio-temporal modeling and validate the robustness and generalizability of our approach in complex maritime environments.