Realistic ambient noise analysis for passive surveillance algorithm design

Passive surveillance algorithms are important for continuous monitoring of the offshore assets and the coastal regions to prevent unwanted intrusions by the adversary. The strategic shift of modern navies to the littoral waters presents great challenge for the algorithm designer as the underwater channel conditions in shallow waters leads to more complicated distortion effects and ambient noise characteristics. Typically, the ambient noise is assumed to be Gaussian. However, field experiments on ambient noise recordings and their characterization have established that it is predominantly non-Gaussian due to the various noise generating sources. In this work, the performance of a passive classification algorithm has been evaluated for Gaussian as well as realistic non-Gaussian noise models of the ambient noise in an underwater channel. The input data has been processed into spectrum and cepstrum domain features and the k-nearest neighbor pattern classification algorithm has been used. The performance evaluation is done in terms of percentage correct classification for three types of propulsion in marine vessels, namely diesel, gas turbine, and steam. The data analysis has been done for synthesized received signals and real signal recordings of marine vessels in shallow water conditions. It is observed that the passive sonar classifier is quite robust to varying ambient noise statistics.