A novel method for pattern recognition of fishing vessel behavior using VMS data

• PL-XGBoost recognizes unlabeled trajectory behaviors with limited labeled data. • VMS trajectory processing improves data quality by splitting voyages and filtering noise. • Multidimensional spatio-temporal attention LSTM based on differentiated sequence encoding features enhances behavior recognition. • Self-training pseudo-labels with staged Bayesian optimization tackle sparse labels and class imbalance in trajectory embedding. • The presented approach achieves superior metrics versus baselines under limited labeled data. Accurate behavioral pattern recognition is essential for marine traffic regulation and monitoring illegal fishing, yet challenged by offshore vessel tracking and unlabeled trajectory data. This study proposes a semi-supervised Vessel Monitoring System (VMS)-based framework to identify behavioral patterns using limited labeled samples. Firstly, the VMS trajectory processing flow is established, and the voyage splitting algorithm that integrates berthing point detection and motion features effectively solves the voyage blurring problem caused by frequent berthing. Secondly, a spatio-temporal multidimensional attention Long Short Term Memory (LSTM) model is designed for the differences in trajectory patterns, and the spatio-temporal attention mechanism is used to adaptively capture the differences in latent features of trajectories with different behavioral patterns; Finally, a machine learning-based confidence weight pseudo-labeling optimization strategy is developed to achieve accurate identification and efficient classification of fishing vessel behavioral patterns under limited labeled samples. Experimental results demonstrate superior performance in fishing/sailing behavior recognition task, and its evaluation metrics such as Accuracy (93 %), F1-Score (Class0 91.16 %, Class1 94.27 %) are significantly better than the traditional supervised learning methods. This work provides robust technical support for intelligent monitoring of far-shore fishing vessel behaviors.