Semi-supervised graph learning for underwater source localization using ship-of-opportunity spectrograms.

Area of Science:

• Acoustic signal processing
• Underwater acoustics
• Machine learning for localization

Background:

• Conventional underwater source localization methods struggle to fully exploit crucial data correlations.
• Graph-based approaches offer potential for capturing spatial relationships in acoustic data.
• Limited labeled data is a significant challenge in supervised learning for acoustic localization.

Purpose of the Study:

• To develop a novel graph learning module for accurate underwater source localization.
• To leverage spectrograms from ships-of-opportunity for acoustic signal analysis.
• To address the challenge of limited labeled data in acoustic localization tasks.

Main Methods:

• A two-step approach combining a pre-trained convolutional neural network (CNN) for feature extraction and a graph neural network (GNN) for localization.
• Self-supervised learning for CNN feature extraction and semi-supervised learning for GNN training.
• Graph construction using k-nearest neighbors on CNN-extracted features from ship noise spectrograms.

Main Results:

• The proposed graph learning framework achieves performance comparable to conventional supervised learning models.
• The approach effectively utilizes data correlations through graph representation.
• Demonstrated generalization capabilities on both synthetic and measured data.

Conclusions:

• The novel graph learning module offers an effective solution for underwater source localization, particularly with limited labeled data.
• The integration of CNNs and GNNs provides a robust framework for acoustic signal analysis.
• The method shows promise for real-world applications in underwater acoustic monitoring and navigation.