A relative-entropy approach to distributed passive detection

There is currently much interest within the ocean acoustics community on using distributed sensor networks to monitor ocean properties. One such task is the application of distributed sensors to passive detection of weak acoustic sources. The joint likelihood detection ratio for a set of distributed sensors leads naturally to the comparison of relative entropy with a detection threshold. As a nondimensional additive measure of information, relative entropy enables data fusion among disparate kinds of sensors, e.g., acoustic and electromagnetic, and mitigates calibration issues. Furthermore, because relative entropy is an integral over probability densities of sensor outputs, it is insensitive to false alarms from transients. In this talk, the theory of relative-entropy detection will be presented, and the method illustrated using acoustic intensity data from hydrophones deployed by the Transverse Acoustic Variability Experiment (TAVEX). For the method to work, accurate estimates of the probability densities for noise and signal intensities must be obtained. For the TAVEX data, superior receiver operating characteristic curves are obtained when the noise and signal distributions are represented by log-normal distributions in comparison with gamma and nonparametric distributions. [Work supported by the Office of Naval Research.]