Vessel Trajectory Recognition Based on Multisource Data Multistage Fusion

Vessel trajectory recognition plays a critical role in ensuring maritime transport safety. However, most existing methods rely solely on navigation information obtained from automatic identification system (AIS) sensors, such as position, speed, and course. This reliance makes it difficult for models to accurately distinguish between different vessel behaviors when trajectories are similar or data is inaccurate. In contrast, contextual information from other sensors such as water depth can effectively enhance recognition accuracy, as it helps provide a better understanding of vessel motion characteristics, navigation strategies, and other behavioral patterns. Although contextual information from other sensors such as water depth can provide valuable complementary knowledge, fusing multisource data remains challenging due to the heterogeneous temporal-spatial resolutions, differing statistical properties, and potential redundancy or inconsistencies among sources. To address these issues, this article proposes a multisource and multistage fusion approach. Specifically, Transformer and convolutional neural network (CNN) are employed to process trajectory and contextual information separately. The learned features are then fused using an attention-based feature fusion (AFF) module to obtain preliminary soft recognition result. For vessel types with low classification accuracy, seven trajectory features (e.g., maximum speed, speed variation, and heading change) and six contextual features (e.g., max/min/average depth, nearest/farthest/average distance from shore) are extracted, based on which recognition rules are learned using a decision tree model. Finally, the recognition results obtained from the decision tree model are further fused with the preliminary soft recognition results from the previous step through the probabilistic decision fusion (PDF) module, yielding the final recognition output. Experimental comparisons in real-world scenarios demonstrate that the proposed method significantly improves the vessel trajectory recognition accuracy. Besides, several case studies further highlight its application value for maritime security through abnormal behavior detection.