Attention-Aware Graph Neural Network Modeling for AIS Reception Area Prediction.

Area of Science:

• Maritime technology
• Machine learning
• Signal propagation

Background:

• Accurate Automatic Identification System (AIS) reception prediction is crucial for reliable ship tracking and anomaly detection.
• Traditional propagation models struggle in dynamic environments due to complex input requirements.
• Existing methods can lead to vessel localization and behavior analysis errors.

Purpose of the Study:

• To propose a data-driven approach using graph neural networks (GNNs) for modeling AIS reception.
• To enhance the accuracy of predicting AIS reception areas by integrating environmental and geographic variables.
• To overcome limitations of traditional propagation models in dynamic maritime settings.

Main Methods:

• Utilized a GraphSAGE framework with attention mechanisms inspired by transformers.
• Integrated Bidirectional Long Short-Term Memory (BiLSTM) for attention coefficients and an attentional pooling module.
• Trained the GNN model on real-world AIS data combined with terrain and meteorological features.

Main Results:

• The GNN model effectively captured both local and long-range AIS reception patterns.
• Achieved superior performance compared to classical baselines like ITU-R P.2001 and XGBoost.
• Demonstrated significant improvements in F1-score and accuracy for AIS reception prediction.

Conclusions:

• Deep learning, specifically GNNs, offers a powerful approach for modeling sensor reception.
• The proposed method enhances the accuracy of ship tracking and anomaly detection through improved AIS prediction.
• Highlights the potential of data-driven techniques in analyzing complex sensor network data for maritime applications.