Data error estimation for matched-field geoacoustic inversion

Abstract

Nonlinear Bayesian methods have been applied to geoacoustic inversion to estimate uncertainties for seabed parameters by sampling the posterior probability density. This procedure requires quantifying the errors on the acoustic data, including both measurement and theory errors, which are generally not well known. To date, point estimates for data errors have been derived using a global maximum likelihood approach. However, this is not consistent with the Bayesian formulation, and ignores the effects of uncertainty in the error estimates and interdependencies between the data errors and geoacoustic parameters. The Bayesian approach treats the data errors as random variables and includes them as additional parameters within the inversion. However, this increases significantly the number of unknowns and the computational effort. A third approach is to use a local maximum likelihood error estimate evaluated independently for each geoacoustic model considered in the sampling procedure. This has the benefit of not increasing the number of unknowns or computational effort, but includes some of the effects of the error uncertainties and interdependencies. The three approaches are compared for Bayesian matched-field geoacoustic inversion of both synthetic and experimental data.