A Bayesian approach to incorporating maximum entropy-derived signal parameter statistics into the receiver operating characteristic (ROC) curves

Abstract

Our goal is to develop a probabilistic sonar performance prediction methodology that can make use of limited knowledge of random or uncertain environment, target, and sonar system parameters, but does not make unwarranted assumptions. The maximum entropy method (MEM) can be used to construct probability density functions (pdfs) for relevant environmental and source parameters, and an ocean acoustic propagation model can use those pdfs to predict the variability of received signal parameter. At this point, the MEM can be used once again to produce signal parameter pdfs. A Bayesian framework allows these pdfs to be incorporated into the signal processor to produce ROC curves in which, for example, the signal-to-noise ratio (SNR) is a random variable for which a pdf has been calculated. One output of such a processor could be a range-dependent probability of detection for fixed probability of false alarm, which would be more useful than the conventional range of the day that is still in use in some areas. [Work supported by ONR Code 321US.]