Breaking the dilemma of trajectory imbalance and sample scarcity: A Multi-Behavior Prototype Siamese Network for vessel type recognition

Vessel type recognition is at the core of intelligent maritime monitoring. Real-world AIS datasets follow a long-tailed distribution, giving rise to two coupled challenges: class imbalance and sample scarcity in minority vessel categories. Samples from majority classes dominate the optimization process, whereas minority vessel categories fail to learn stable discriminative representations due to limited trajectory observations, making conventional supervised classification methods less suitable for this scenario. To address this challenge, this paper proposes a Multi-Behavior Prototype Siamese Network (MBP-SNet) that relies solely on dynamic trajectory data. The method innovatively integrates multi-scale temporal convolution with a trajectory self-attention mechanism, capturing local navigation patterns while modeling global temporal dependencies. To alleviate class imbalance, a Multi-Behavior Prototype Representation (MBPR) mechanism is proposed, which assigns multiple behavioral prototypes to each vessel category together with a prototype coverage loss to suppress majority-class bias and prevent prototype collapse in minority classes. To address sample scarcity in minority vessel categories, a siamese metric learning paradigm based on prototype-distance matching is introduced to enhance feature discriminability under limited training samples. On a real-world AIS dataset, MBP-SNet achieves accuracies of 78.96% and 81.77% at the sample level and vessel level, respectively, significantly outperforming existing state-of-the-art methods, and demonstrates particularly strong robustness for minority vessel categories. After incorporating static features, the model’s F1-score further improves to 93.81%, highlighting its distinct advantages under long-tailed and sample-scarce conditions.