Uncertainty reduction in matched field inversion using Gaussian Processes

Abstract

Gaussian Processes have been shown to be an effective tool for predicting an acoustic field in the ocean at a densely populated virtual array. This property is employed in this work for pre-processing acoustic data before these are used for source localization and geoacoustic parameter estimation using matched-field inversion (MFI). The process increases the accuracy of MFI as uncertainty in the estimation process is reduced. Via the application of Gaussian Processes, the data are denoised and interpolated and the new, enhanced fields are used as input to MFI along with replicas calculated at the virtual receivers at which field predictions are made. Kernel functions, implicit in the implementation of Gaussian Processes, quantify the correlation among field values at neighboring spatial points. Employing different kernels, the approach is tested on synthetic and real data with both an exhaustive search and genetic algorithms and is found to be superior to conventional Bartlett MFI in source localization and geoacoustic inversion. [Work supported by ONR.]