Learning location and seabed type from a moving mid-frequency source.

Area of Science:

• Acoustics
• Machine Learning
• Oceanography

Background:

• Source localization and seabed classification are typically separate research areas.
• Convolutional Neural Networks (CNNs) offer potential for integrated analysis.

Purpose of the Study:

• To develop and evaluate CNNs for simultaneous acoustic source localization and seabed classification.
• To compare single-task regression CNNs with multitask CNNs for these tasks.

Main Methods:

• Application of different CNN architectures to mid-frequency tonal data from a vertical line array (VLA).
• Training CNNs on synthetic data and evaluating performance on test cases.
• Comparison of a single regression CNN with a multitask CNN (regression for source parameters, classification for seabed type).
• Assessment of environmental factors like sound speed profile and seabed variations.

Main Results:

• Multitask CNNs demonstrate superiority over single-task CNNs.
• Environmental mismatch negatively impacts source depth estimation but other parameters are estimated with bias.
• Successful application to real-world data from the Seabed Characterization Experiment 2017.

Conclusions:

• Multitask learning is advantageous for combined acoustic source localization and seabed property detection.
• CNNs show significant potential for analyzing mid-frequency sounds to infer both source and environmental characteristics.