Ensemble approach to deep learning seabed classification using multichannel ship noisea).

Area of Science:

• Ocean acoustics
• Machine learning
• Seabed characterization

Background:

• Hydrophone recordings of shipping noise contain information about the seabed in shallow-water environments.
• Previous methods for seabed classification are limited in complex acoustic conditions.

Purpose of the Study:

• To train neural networks for predicting seabed classes using multichannel hydrophone spectrograms of shipping noise.
• To evaluate the impact of different numbers of hydrophone channels on prediction accuracy.
• To apply ensemble modeling for improved performance and confidence assessment.

Main Methods:

• Synthetic data was used to train ResNet-18 neural networks on one, two, four, and eight hydrophone channels.
• Trained networks were applied to measured ship spectrograms from the Seabed Characterization Experiment 2017 (SBCEX 2017).
• Data preprocessing and ensemble modeling techniques were employed to enhance results.

Main Results:

• The neural networks successfully predicted seabed classes from measured ship spectrograms.
• Predictions converged towards two seabed classes with similar shallow sediment properties.
• Results were consistent with independent geoacoustic inversion findings from SBCEX 2017.
• Ensemble modeling provided a measure of confidence and precision for the predictions.

Conclusions:

• Neural networks trained on shipping noise are effective for seabed classification.
• Multi-channel hydrophone data and ensemble modeling enhance prediction accuracy and reliability.
• This approach shows promise for seabed characterization in various oceanographic conditions.