Model-based convolutional neural network approach to underwater source-range estimation.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Conventional matched-field processing (MFP) is widely used for underwater source-range estimation.
• MFP can be sensitive to environmental variations, potentially leading to inaccurate predictions.
• Exploring alternative methods for robust underwater acoustic source localization is crucial.

Purpose of the Study:

• To evaluate a model-based convolutional neural network (CNN) as an alternative to MFP for underwater source-range estimation.
• To assess the CNN's performance under varying environmental conditions, including slight deviations from the training model.
• To validate the CNN approach using both simulated and real-world acoustic data.

Main Methods:

• Developed a model-based CNN trained on simulated underwater acoustic data.
• Compared the CNN's performance against traditional MFP using simulated data with environmental mismatches.
• Tested the trained CNN on real-world acoustic field data from the Beaufort Sea and Southern California.

Main Results:

• The CNN approach exhibited improved prediction accuracy and lower mean-absolute-error (MAE) compared to MFP when tested with data from slightly deviated environments.
• CNN performance transferred effectively to real-world data, showing consistency with expected source ranges and outperforming MFP in MAE.
• The CNN demonstrated greater constraint against highly inaccurate predictions under environmental mismatch, unlike MFP.

Conclusions:

• Model-based CNNs show significant potential as a robust alternative to MFP for underwater source-range estimation.
• The CNN approach offers improved accuracy and error reduction, particularly in scenarios with environmental variability.
• While CNNs provide better robustness to environmental mismatch, they may offer less certainty in prediction when the environment is perfectly modeled.