Comparison of model selection techniques for seafloor scattering statistics

In quantitative analysis of seafloor imagery, it is common to model the collection of individual pixel intensities scattered by the seafloor as a random variable with a given statistical distribution. There is a considerable literature on statistical models for seafloor scattering, mostly focused on areas with statistically homogeneous properties (i.e. exhibiting spatial stationarity). For more complex seafloors, the pixel intensity distribution is more appropriately modeled using a mixture of simple distributions. For very complex seafloors, fitting 3 or more mixture components makes physical sense, but the statistical model becomes much more complex in these cases. Therefore, picking the number of components of the mixture model is a decision that must be made, using a priori information, or using a data driven approach. However, this information is time consuming to collect, and depends on the skill and experience of the human. Therefore, a data-driven approach is advantageous to use, and is explored in this work. Criteria for choosing a model always need to balance the trade-off for the best fit for the data on the one hand and the model complexity on the other hand. In this work, we compare several statistical model selection criteria, e.g., the Bayesian information criterion. Examples are given for SAS data collected by an autonomous underwater vehicle in a rocky environment off the coast of Bergen, Norway using data from the HISAS-1032 synthetic aperture sonar system.