Inversion in an uncertain ocean using Gaussian processes.

Area of Science:

• Ocean acoustics
• Signal processing
• Geophysical inversion

Background:

• Underwater acoustic data often contains noise and lacks sufficient spatial sampling for accurate source localization and environmental inversion.
• Matched Field Inversion (MFI) is a technique used for underwater source localization and environmental parameter estimation, but its performance is sensitive to data quality.

Purpose of the Study:

• To pre-process acoustic data using Gaussian processes (GPs) to improve its quality for subsequent source localization and environmental inversion.
• To investigate the effectiveness of GPs in denoising and interpolating acoustic fields in underwater waveguides.

Main Methods:

• Application of Gaussian processes (GPs) for denoising and interpolating acoustic data to create densely populated acoustic fields at virtual arrays.
• Utilizing these enhanced acoustic fields as input data for Matched Field Inversion (MFI).
• Employing Gaussian and Matérn kernels, with hyperparameters optimized via maximum likelihood estimation, and comparing performance using exhaustive search and genetic algorithms.

Main Results:

• Gaussian processes effectively denoise and interpolate acoustic data, generating high-resolution acoustic fields at virtual arrays.
• The GP-enhanced MFI approach demonstrates superior performance compared to conventional beamformer MFI on both synthetic and real-world data.
• The Matérn kernel, with its greater flexibility due to more hyperparameters, is found to be more effective than the Gaussian kernel.

Conclusions:

• Gaussian processes offer a powerful pre-processing tool for underwater acoustic data, significantly improving MFI performance.
• The use of GPs enables more accurate source localization and environmental inversion in challenging underwater acoustic environments.
• The Matérn kernel is recommended over the Gaussian kernel for its enhanced performance in this application.