Particle smoothers in sequential geoacoustic inversion

Abstract

Sequential Bayesian methods such as particle filters have been used to track a moving source in an unknown and space/time-evolving ocean environment. These methods treat both the source and the ocean parameters as non-stationary unknown random variables and track them via the multivariate posterior probability density function. Particle filters are numerical methods that can operate on nonlinear systems with non-Gaussian probability density functions. Particle smoothers are a natural extension to these filters. A smoother is appropriate in applications where data before and after the time of interest are readily available. Both past and "future" measurements are exploited in smoothers, whereas filters just use past measurements. Geoacoustic and source tracking is performed here using two smoother algorithms, the forward-backward smoother and the two-filter smoother. Smoothing is demonstrated on experimental data from both the SWellEx-96 and SW06 experiments where the parameter uncertainty is reduced relative to just filtering alone.