Multi-Frequency and Multi-Snapshot Gaussian Processes for Matched-Field Processing Source Localization

To interpolate the sparsely sampled vertical array signal into a denser one, Gaussian process regression is employed by considering the noise-free field as a Gaussian process. This interpolated signal is then combined with matched-field processing (MFP) for accurate localization of the source. The success of Gaussian process regression depends on the selected kernel function and the hyperparameters obtained from the type-II maximum likelihood estimation. In this paper, to improve the stability of the hyperparametric search process, two methods, using multiple frequencies and multiple snapshots, are employed. These methods increase the accuracy of the Gaussian process for MFP. By densifying and denoising the received sound field, the sidelobes of the MFP ambiguity surfaces are reduced, and the probability of correct localization is increased. Compared to conventional processing methods, the improved Gaussian process-based processor performs better to some extent at a range of signal-to-noise ratios. Moreover, the probability of correct localization is further increased as the number of frequency points and snapshots increase.