Scattering statistics of nonstationary seafloors: Mixture models and choosing the number of components

Abstract

Modern high resolution synthetic aperture sonars are capable of resolving the structure of patchy seafloors, defined as seafloors having non-stationarygeoacoustic and roughness properties. Within each patch, it is often possible to model the amplitude statistics using commonly used physics based statistical models, such as the K-distribution. However, it is impossible to use on a large scale because either the parameters change from patch to patchor the imaging sonar alters the scattering statistics at different ranges. Here, we propose a method to segment different patches in a sonar image using a mixture of physicallymotivated statistical models. We detail how the number of components can be chosen using Bayesian model selection techniques, such as the deviance information criterion. The tradeoff between model accuracy and robustness is explored using a dataset provided by the Norwegian Defence Research Establishment.