Composite Periodic Kernel Gaussian Process Regression for Matched-Field Processing Localization

Gaussian process regression (GPR) models the spatial correlations of the acoustic field through kernel functions, densifying and denoising the received signals for matched-field processing (MFP) localization. The commonly used Gaussian kernel in GPR accounts only for correlations between nearby points when modeling the acoustic field covariance and neglects differences in correlation structures across different receiver points, leading to insufficient stability and accuracy. In this work, a composite periodic kernel (CPK) is proposed as an improved kernel function, whose form more closely resembles the covariance structure of the acoustic field. The method focuses on reconstructing individual modal components and incorporates seabed boundary conditions to constrain the hyperparameter search space, thereby enhancing the stability and prediction accuracy of GPR. Consequently, MFP localization accuracy is further improved. The method is validated under various signal-to-noise ratios, environmental parameter mismatches, source locations, and waveguides, demonstrating superior performance compared to both the Gaussian kernel and original signals.