AIS Trajectory-Based Semi-SupervisedVessel Type Recognition via Entropy-Guided Adaptive Thresholding and Contrastive Feature Alignment
Accurate vessel type recognition from Auto-matic Identification System (AIS) data is critical for mar-itime traffic management, anomaly detection, and security monitoring. While AIS trajectories encode rich kinematic information, the scarcity of annotated samples poses a major bottleneck for purely supervised approaches. This paper proposes a semi-supervised framework that jointly addresses label scarcity and class imbalance. The system extracts complementary features via a dual-branch archi-tecture: a ResNet-18 encoder for spatial trajectory patterns and a Temporal Convolutional Network for sequential motion dynamics, adaptively fused through a gated multi-modal module. For semi-supervised learning, we integrate three strategies: trajectory-aligned Mixup augmentation to enrich the training manifold, contrastive alignment with NT-Xent loss to enforce intra-class compactness, and an entropy-guided adaptive thresholding mechanism that dynamically calibrates pseudo-label confidence. Extensive experiments on a large-scale AIS dataset demonstratethat our method consistently outperforms state-of-the-art semi-supervised baselines across all annotation ratios, achieving 90.04% accuracy with 30% labels and surpassing fully supervised models at every evaluated ratio.