Evaluation of Reduced-Rank Adaptive Matched Field Processing
for Shallow-Water Target Detection

Nigel Lee, Lisa M. Zurk, James Ward
MIT Lincoln Laboratory
Lexington, MA
Email: nigel@ll.mit.edu


Abstract This paper examines the use of rank reduction to improve adaptive matched field processing performance for passive sonar detection of quiet targets in littoral environments. Matched field processing (MFP) uses propagation physics to compute "replica" (steering) vectors, resulting in accurate source localization in range, depth, and bearing. Adaptive MFP (AMFP) algorithms that compute weight vectors based on both the replica vector and the sample covariance matrix (SCM) are designed to suppress sidelobes and reject interference, resulting in much-improved performance over non-adaptive methods. However, when the number of snapshots used to estimate the SCM is insufficient due to large arrays or to highly nonstationary environments, AMFP may perform poorly. Rank reduction addresses this by reducing the number of snapshots needed to estimate the SCM accurately.

This paper examines rank reduction of the SCM in both eigenvector and modal bases. For eigenvector decompositions, reduced-rank subspace selection criteria include eigenvalue magnitude, correlation between eigenvectors and replica vectors, and the ratio of the two. The latter two criteria produce superior results due to their use of information about the replica vector in subspace selection. For modal decompositions, subspace selection may be based on modal component strength but is more usefully based on physical properties of the modal basis vectors themselves.

In this paper, rank reduction is applied to AMFP detection of quiet submarine targets using data from the Santa Barbara Channel Experiment (SBCX). SBCX data was taken with a 150-hydrophone volumetric array in a dense shipping environ-ment. In several cases where full-rank AMFP is ineffective because of the snapshot problem, rank-reduction techniques based on both eigenvector and modal decompositions aredemonstrated to provide significant performance gains. Comparison of algorithm performance versus adaptive degrees-of-freedom and output SINR reveals that eigenvalue decompositions incorporating information about the replica vector produce the best reduced-rank AMFP detectors.

Presentation in PDF format