Efficient sequential Monte Carlo estimation of range-dependent seabed properties

Abstract

Seabed geoacoustic properties play a crucial role in shallow-water sonar applications, including the detection of unexploded ordnance. Our goal is improved efficiency of Bayesian seabed parameter and uncertainty estimation for large data volumes. While Bayesian uncertainty estimation provides important information for sonar applications, the approach is computationally expensive which limits utility for large surveys, where an abundance of range-dependent data can be collected. This work considers the efficiency of a particle filter to quantify information content of multiple data sets along the survey track by considering results from previous data along the track to inform the importance sampling at the current point. Efficiency is improved by tempering the likelihood function of particle subsets and including exchange moves (parallel tempering), and by adapting the proposal distribution for the Markov-chain steps. In particular, perturbations are proposed in principal-component space, with the rotation matrix computed via eigenvector decomposition of the unit-lag parameter covariance matrix. The algorithm is applied to 350 data sets collected along a 13-km track on the Malta Plateau, Central Mediterranean Sea. Improved efficiency from parallel tempering and principal-component proposal densities are studied. [Work supported by the Strategic Environmental Research and Development Program, U.S. Department of Defence.]