A NEAREST-NEIGHBOR ALGORITHM FOR FAST MATCHED-FIELD PROCESSING WITH AVERTICAL LINE ARRAY

Matched-field processing (MFP) matches measured acoustic fields of und erwater acoustic sources against replica fields for a grid of source p ositions covering a search region. The search region consists of a vol ume of ocean in which the detection and localization of underwater aco ustic sources is desired. Despite the increased speed of computers, th e large number of replicas limits real time systems to small search re gions at a few frequencies. Selection of replica fields that are most like the data, i.e., the nearest neighbors (NN's) to the data, offers a way of reducing the computational search space, thereby making large r physical search spaces or a larger number of frequencies practical. To enable selection of NN's a vector basis for the search space is req uired. We use the large eigenvectors of the covariance matrix for unco rrelated sources spread over the search region. This is not only a sui table vector basis of the search space, but also results in a dimensio nal reduction from the full set of eigenvectors, with a further comput ational saving. The replica vectors for the search region are partitio ned by finding their projection on this vector basis. One can then sel ect for matching only those replicas which have similar squared projec tions on the vector basis. This selection process carries a modest cos t in computing overhead, provided that the code, the partitioning, and the replica selection parameters are optimized. The detection perform ance and false alarm probability for the Bartlett beamformer, with and without selection of the replicas, were estimated from simulations of noisy data received on a vertical line array at practical time-bandwi dth products. An order of magnitude speedup was obtained at a cost of reductions in detection and localization probabilities, but with the a dded benefit of a reduced false alarm probability. The reduction in de tection performance was equivalent to that which would have been obtai ned at a 1 dB lower signal-to-noise ratio with the Bartlett beamformer without selection. A greater speedup is to be expected for search reg ions larger than the small one of this study.