A Diffusion-based Foundation Model for Irregular Spatio-Temporal Trajectories

Bohao Qu, Zhang Wanbing, Hongwei Wang, Di Lin, Ivor Tsang, Xiuju Fu, Qing Guo

Keywords: foundation model; diffusion model;irregular spatio-temporal modeling

Abstract: Maritime mobility provides a representative yet extremely challenging setting for spatio-temporal sequence modeling. Global AIS trajectories are inherently irregular and noisy, spanning intercontinental routes as well as fine-grained near-port maneuvers, and are constrained by coastlines, shipping lanes, and regulations. These characteristics not only complicate trajectory prediction but also generalize to other irregular spatio-temporal domains such as urban mobility and autonomous navigation. Moreover, maritime applications involve diverse downstream tasks, including forecasting, imputation, and route planning, where training separate task-specific models is computationally costly and undermines scalability. This motivates the development of a foundation-level generative framework that can unify irregular sequence modeling across tasks. We propose GeoDiffusion, a diffusion-based foundation model for maritime trajectory modeling. GeoDiffusion introduces three core components: a Spatio-Temporal Offset Encoding (STOE) to robustly capture irregular sampling and missing data, a Transformer-based denoising network to learn both global and local dynamics, and a training-free conditional inference strategy that enforces geo-temporal consistency and unifies multiple tasks within a single pretrained model. Experiments on large-scale global AIS datasets demonstrate that GeoDiffusion achieves state-of-the-art performance across trajectory prediction, imputation, and planning, while generalizing robustly to unseen distributions. These results highlight GeoDiffusion as both a practical solution for maritime mobility and a blueprint for irregular spatio-temporal foundation models.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 6611