Improving Statistical Robustness of Matched-Field Source Localization via General-Rank Covariance Matrix Matching

Performance of matched-field source localization is highly dependent on the precision of the model to actual physical processes. Model mismatch, including sound propagation environmental mismatch, statistical mismatch, and system mismatch, causes severe performance degradation. Statistical mismatch occurs when an insufficient snapshot set is used to estimate the data covariance matrix. Diagonal loading can improve the widely used minimum variance distortionless response (MVDR) and minimum power distortionless response (MPDR) processors against statistical mismatch, however, the result relies on selection of the loading level. Previously, a new processor named the matched-covariance estimator (MCE) was shown to have statistically robust estimation capability under white noise conditions by matching the general-rank data covariance matrix with a covariance matrix of the modeled replicas. In this paper, the method is further developed in more realistic application scenarios with discrete point interferences and/or surface-generated noises. Simulations and analyses show that MCE can outperform classical MVDR/MPDR matched-field processors by: exploiting both signal and noise data structure at the same time; being less prone to mistaken nulling of the signal; requiring no rank-1 signal space restriction; and not needing signal-free and long duration training data.