AISFlow: Boundary-Informed Flow Matching for Long-Term AIS Trajectory Imputation

Automatic Identification System (AIS) trajectories are fundamental to maritime traffic monitoring, vessel behavior analysis, and safety-critical decision support. However, AIS records often contain long missing intervals due to communication failures, limited coverage, harsh maritime environments, or intentional signal disruption. Such long gaps are difficult to reconstruct because sparse boundary observations can admit multiple plausible vessel motions, limiting deterministic interpolation and conventional sequence models. To address this problem, we propose AISFlow, a boundary-informed flow matching framework for long-term vessel trajectory imputation. Unlike diffusion-based approaches that generate trajectories from unstructured Gaussian noise, AISFlow starts from a structured source distribution derived from pre-gap and post-gap vessel states and learns an ODE-based transport toward the missing gap. Boundary-derived kinematic cues, including endpoint displacement, speed/course information, gap length, and heading variation, are further used as conditional information to improve trajectory continuity. AISFlow also supports uncertainty-aware imputation through ensemble sampling and conformal calibration. Experiments on a real-world AIS dataset show that AISFlow achieves state-of-the-art imputation accuracy while reducing inference latency by up to $317\times$ under fixed-step solvers and $9.25\times$ under the adaptive Dopri5 solver, compared with diffusion-based baselines.