Seabed classification from merchant ship-radiated noise using a physics-based ensemble of deep learning algorithms.

Area of Science:

• Ocean acoustics
• Machine learning
• Seabed classification

Background:

• Merchant ship noise offers a potential acoustic source for passive seabed characterization.
• Deep learning (DL) models can analyze complex acoustic signals for environmental sensing.

Purpose of the Study:

• To classify seabed environments using an ensemble of deep learning algorithms trained on ship-radiated noise.
• To evaluate the robustness and generalizability of these DL models.

Main Methods:

• An ensemble of six deep learning networks (five convolutional neural networks, one residual neural network) was trained on synthetic seabed acoustic data.
• The networks were validated using fivefold cross-validation and tested against simulated environmental variations and real-world at-sea data.

Main Results:

• All trained networks achieved over 97% accuracy in seabed classification.
• Deeper DL architectures demonstrated greater robustness to sound speed and water depth mismatches.
• Testing on real-world data showed 94% agreement with previously inferred sediment types (mud over sand).

Conclusions:

• Ensemble deep learning effectively decodes sediment signatures from ship-radiated noise.
• The approach provides a unified and reliable method for seabed classification using passive acoustics.
• This technique shows promise for at-sea sediment characterization and geoacoustic inversion.