Learning Compressed AIS Trajectories with VQ-VAE for Activity Classification
Publication Date
1-1-2026
Document Type
Conference Proceeding
Publication Title
Proceedings IEEE International Conference on Mobile Data Management
DOI
10.1109/MDM71479.2026.00080
First Page
483
Last Page
488
Abstract
Automatic Identification System (AIS) data enables large-scale maritime monitoring but creates major challenges for storage and efficient downstream analysis because of its volume. Existing approaches typically treat trajectory compression and activity classification as separate problems, resulting in either loss of semantic information or high computational cost. To solve this issue, the proposed unified approach jointly learns compression and vessel activity classification using a Vector-Quantized Variational Autoencoder (VQ-VAE). Its model encodes each trajectory segment into a compact sequence of discrete indices, achieving up to 17.8 × compression while preserving information relevant to downstream tasks such as classification. A classifier then operates directly on the compressed representation, eliminating the need for reconstruction during inference. Experiments on AIS trajectory data from fishing vessels show that the learned discrete representation is highly informative for detecting fishing behavior. A simple logistic regression on code usage histograms generated by the VQ-VAE achieves 99.48% accuracy and 97.64% macro F1, approaching the strongest sequence-model baseline (99.61%). At the same time, the model reconstructs trajectories with a mean positional error of 7.59 km while preserving key trajectory structures and behavioral patterns and maintaining full utilization of the compression codebook (512/512 active codes). This work demonstrates that discrete latent representations provide an effective and efficient interface for jointly addressing compression and behavioral inference in maritime trajectory data.
Funding Number
101182585
Funding Sponsor
HORIZON EUROPE Framework Programme
Keywords
activity classification, discrete representation learning, trajectory compression, trajectory reconstruction, Vector-Quantized Variational Autoencoder (VQ-VAE)
Department
Computer Engineering
Recommended Citation
Christos Chronis, Spyridon Chatziargyros, Magdalini Eirinaki, Konstantinos Tserpes, and Iraklis Varlamis. "Learning Compressed AIS Trajectories with VQ-VAE for Activity Classification" Proceedings IEEE International Conference on Mobile Data Management (2026): 483-488. https://doi.org/10.1109/MDM71479.2026.00080
