On marine mammal acoustic detection performance bounds

Abstract

Most current research on marine mammal acoustic detection, classification, and localization does not consider optimizing detector algorithm performance for marine mammal vocalizations that have been propagated through an uncertain multipath environment. In this study, a Bayesian likelihood ratio approach, using the original time series data, is used to benchmark optimal detection performance (ROC's) for a known marine mammal source vocalization propagated through both a known and an example uncertain ocean environment. In addition, for these same ocean environments, the detection performance (ROC's) is obtained for several algorithms based on the spectrogram of the original data. Since the spectrogram does not preserve detailed phase information contained in the original data, any algorithm based on the spectrogram is not likely to be optimum for detection. The initial results show the additional detection performance gain possible over spectrogram based algorithms. Simulations and preliminary detection performance results (ROC's) are presented using a mathematical model from the literature of the North Atlantic Right Whale (NARW), along with numerical acoustic propagation software.