Active sonar clutter classification using higher order moments

Abstract

The statistics of normalized matched-filter echoes from an active sonar system operating in a myriad of oceanic environments have been studied extensively for three broad clutter classes including the use of cumulants to classify sub-regions of the data (Gelb et al. 2010, Journal of Oceanic Engineering 35(1), and references therein). That work compared empirical distributions to parametric models such as the K distribution and the generalized Pareto distribution. We report on extensions of that work including studies of the use of high order moments in the classification process. For each class, with increasingly heavy non-Rayleigh distributed tails, we present comparisons of moments computed directly from the data with analytically derived values based on parametric fits to the data. Using a feature-based classifier we compare the gains of using increasingly higher order moments with the diminishing returns of utilizing additional features (i.e., problems with training the classifier using limited data). Additionally, we developed a spatial feature based on a minimum spanning tree, and we show the efficacy of using this feature for improved clutter classification. [Sponsored by ONR.]