Through-Ice Acoustic Source Tracking Using Vision Transformers with Ordinal Classification.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Underwater acoustic localization is challenging in icy environments due to multipath and non-linear signal propagation.
• Conventional methods struggle with the complexities introduced by ice cover.

Purpose of the Study:

• To evaluate and compare the performance of various deep learning networks for passive localization and tracking of on-ice acoustic sources.
• To investigate the effectiveness of ordinal classification as a localization strategy.

Main Methods:

• Comparison of deep learning models including Transformers, Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Vision Transformers (ViTs).
• Utilizing two underwater acoustic vector sensors for passive acoustic data acquisition.
• Employing ordinal classification for source localization and comparing it with standard regression methods.

Main Results:

• Vision Transformers (ViTs) demonstrated strong performance in tracking moving acoustic sources on ice.
• Ordinal classification outperformed regression for CNNs and ViTs in this localization task.
• The study successfully analyzed passive acoustic data from an anthropogenic source on ice.

Conclusions:

• Deep learning, particularly Vision Transformers, offers a viable solution for acoustic source localization in challenging ice environments.
• Classification-based localization approaches can be more effective than regression for certain deep learning architectures in this context.