Underwater acoustic source detection using a median based estimator

Abstract

Recently a class of cross spectral density matrix (CSDM) estimators have been introduced based on the use of matrix medians to average signal dyads rather than the more conventional sample mean. Due to the medians robustness against outliers, these new estimators are intended to for use at low SNR or in highly cluttered areas. The median in this case is given a geometric interpretation, as the point which minimizes the sum of the distance to all of the samples. This geometric interpretation has been shown to be quite useful for matched field processing, and in this work it is applied to detection of unknown signals at low SNR. Several estimators are developed corresponding to different choices of non-Euclidean distance between CSDMs. These new estimators are then applied to a relative entropy based detection scheme which detects unknown signals by comparing them to insitu estimated noise and finding individual frequencies which are highly dissimilar from this noise estimate. Using data from the DSNCON-19 experiment the performance of these novel estimators is characterized by their Receiver Operating Characteristic (ROC) curve and compared to the maximum likelihood estimator (MLE) performance. [Work supported by the Office of Naval Research.]