Source depth estimation with feature matching using convolutional neural networks in shallow water.

Area of Science:

• Underwater acoustics
• Signal processing
• Machine learning

Background:

• Accurate underwater source localization is crucial for naval applications.
• Traditional matched-field processing (MFP) methods struggle with environmental variations.

Purpose of the Study:

• To propose a novel feature matching method based on convolutional neural networks (FM-CNN) for underwater source depth estimation.
• To evaluate the performance and robustness of FM-CNN compared to conventional MFP.

Main Methods:

• Developed a feature matching method based on convolutional neural networks (FM-CNN).
• Trained FM-CNN using acoustic field replicas from a propagation model.
• Compared FM-CNN with conventional MFP in range-independent and mildly range-dependent environments.
• Conducted sensitivity analysis on environmental mismatches (bottom parameters, sound speed profile, topography).

Main Results:

• FM-CNN demonstrated higher robustness to environmental mismatches than conventional MFP for both single and multiple source depth estimation.
• Validated FM-CNN using real-world data from the East China Sea experiment.
• FM-CNN reliably estimated source depths in complex environments where MFP showed significant failure rates.

Conclusions:

• The proposed FM-CNN is a robust and reliable method for underwater source depth estimation in challenging acoustic environments.
• FM-CNN offers significant advantages over conventional MFP, particularly in the presence of environmental variability and complex conditions.