VT-EncNet: A Multi-Modal convolutional autoencoder with attention for vessel trajectory representation and similarity computation

Artikel i vetenskaplig tidskrift, 2026

Vessel trajectory similarity computation is of great significance in maritime data mining, traffic safety control, and intelligent shipping management. However, real-world Automatic Identification System (AIS) data are affected by noise, irregular sampling, and highly diverse navigational behaviors, which significantly degrade the performance of traditional point-matching or geometry-based similarity measures. Furthermore, existing deep learning approaches often rely on single-modal representations—either treating trajectories as spatial images or temporal sequences—failing to explicitly capture the complementary coupling between global spatial topology and local kinematic dynamics. To address these limitations, this paper proposes the Vessel Trajectory Encoding Network (VT-EncNet), a multi-modal convolutional representation learning framework for robust trajectory similarity computation. The method employs a grid-based discretization strategy to construct multi-channel inputs integrating position, speed, and course. A dual-branch architecture, utilizing 2D and 1D Convolutional Neural Networks (CNNs), independently extracts spatial structural features and dynamic motion attributes. Subsequently, a cross-attention mechanism explicitly fuses these spatial and kinematic modalities, compressing them into discriminative low-dimensional latent vectors via an autoencoder. Extensive experiments on real-world passenger vessel trajectories from the Qiongzhou Strait demonstrate the superiority of VT-EncNet. Quantitatively, it achieves a state-of-the-art Adjusted Rand Index (ARI) of 0.7955 in unsupervised clustering, successfully disentangling complex, spatially overlapping traffic flows into 12 fine-grained behavioral patterns. Furthermore, the framework exhibits exceptional scalability, delivering significant computational acceleration over traditional exhaustive distance metrics. These findings highlight the model’s robust pattern recognition capability and profound practical value for large-scale intelligent maritime traffic management.