Evaluation of Relative Entropy for Distributed Passive Detection of Weak Acoustic Signals

Passive detection of weak signals in noise is formulated without assuming a parametric distribution of the received signal statistics. A binary likelihood-ratio test as to whether the probability densities of two candidate noise populations are the same, or significantly different, leads to the comparison of relative entropy with a Neyman-Pearson detection threshold. The method is demonstrated using the acoustical intensity of narrowband signals at the hydrophones of a bottom-mounted horizontal line array, which was deployed during the Transverse Acoustic Variability Experiment. Receiver operating characteristics of the relative entropy detector are compared with those for the mean signal-to-noise ratio on a set of hydrophones. It is found that at large false-alarm rates the relative entropy detector has comparatively inferior performance while remaining competitive in the region of small false-alarm probability. Relative entropy detection is insensitive to statistical dependence among the sensors, but is sensitive to data having nonstationary intensity statistics characteristic of the ocean environment in which it was tested.