Analysis of Maritime Loitering Patterns With Discrete AIS Sequences Using Transformer Model Reconstruction and Classification

Tuan Manh Tao, Changha Lee, Thanh Lam Tran, Chan-Hyun Youn

Abstract: Loitering vessel behavior—characterized by prolonged, irregular, or circling maneuvers within a confined area—often indicates illicit maritime activities such as illegal fishing, covert rendezvous, or surveillance operations. Detecting such behaviors remains challenging due to the discrete, noisy, and intermittent nature of Automatic Identification System (AIS) transmissions, as well as the scarcity of labeled training data. This paper proposes a Transformer-based framework that unifies AIS trajectory reconstruction and loitering behavior detection at the segment level. The proposed approach addresses two critical limitations of prior studies: 1) the inability of lightweight filtering models to recover meaningful motion continuity under long AIS gaps, and 2) the dependence on manually engineered features (e.g., spectral entropy, heading variance) and static classifiers (e.g., Random Forest) that lack temporal context. Our proposed model, based on a cross-attention–enhanced Transformer architecture, learns vessel motion dynamics directly from raw, fragmented AIS sequences, enabling both long-gap interpolation and automatic identification of loitering segments without handcrafted heuristics. Each detected loitering segment is further classified into six interpretable categories based on learned spatiotemporal attention patterns. By leveraging self-attention and self-supervised learning, the model robustly recognizes loitering even from discrete or partially missing AIS sequences. Experiments on real-world AIS data demonstrate that the proposed framework achieves high accuracy in loitering classification and maintains stable detection performance under severe AIS sparsity.

External IDs:dblp:journals/access/TaoLTY25