Fishing vessel behavior pattern recognition using AIS sub-trajectory prototype learning based on Gramian Angular Field

The Automatic Identification System (AIS) collects both dynamic and static information related to vessel trajectories. The aquatic movement patterns of fishing vessels consist of multiple behavioral segments corresponding to distinct operational states. This study intends to develop a prototype learning framework inspired by computer vision, designed for the semantic interpretation of fishing vessel mobility patterns. First, we extract discriminative spatiotemporal feature sequences by conducting a systematic analysis of vessel movement characteristics. Second, we propose a prototype learning method based on deep subspace clustering; this method identifies trajectory segments that summarize the motion characteristics of each specific behavior, such as stopping, loitering, fishing, or sailing. Finally, we implement a hierarchical transformer architecture integrated with temporal attention mechanisms, which supports multi-scale trajectory prediction. This architecture facilitates position forecasting for both short-term horizons of 10 min and long-term horizons of 60 min. This technological advancement can assist maritime regulatory bodies in optimizing vessel traffic management and implementing data-driven safety protocols.