Underwater source ranging by Siamese network aided semi-supervised learning.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Deep learning for underwater source ranging requires large labeled datasets, which are expensive and time-consuming to acquire.
• Existing methods face challenges due to data scarcity in acoustic positioning tasks.

Purpose of the Study:

• To develop a semi-supervised learning approach to enhance underwater source ranging accuracy.
• To reduce the dependency on large amounts of labeled data for deep learning models.

Main Methods:

• Utilized a Siamese network to generate pseudo-labels for unlabeled underwater acoustic data.
• Proposed a novel confidence criterion, incorporating similarity scores and sample distribution, to assess pseudo-label reliability.
• Implemented a wrapper paradigm for semi-supervised learning to train models with expanded datasets.

Main Results:

• The proposed semi-supervised method effectively improved prediction accuracy in underwater source ranging.
• The confidence criterion demonstrated reliability in evaluating pseudo-labels, aiding model training.
• Experiments on the SwellEx-96 dataset confirmed the efficacy of the approach.

Conclusions:

• Semi-supervised learning, particularly with the proposed pseudo-labeling strategy, offers a viable solution to data scarcity in underwater source ranging.
• The method enhances the practical application of deep learning in acoustic positioning by leveraging unlabeled data effectively.