Bayesian layer-stripping inversion of seabed reflection data

Abstract

This paper develops a Bayesian inversion technique for recovering multilayer geoacoustic profiles using seabed reflection data. The measured data originate from acoustic time series windowed for a single bottom interaction, which are processed to yield spherical reflection coefficients (i.e., a response function of frequency and angle analogous to plane-wave reflection coefficients). Replica data are computed using a wave number-integration model (OASES) to calculate the full complex acoustic pressure field, which is processed to produce a similar seabed response function. The inversion results are compared to those obtained using plane-wave reflection coefficients. To address the high computational modeling costs, the Bayesian algorithm is implemented for a massively parallel computer. Further, the data are time windowed and divided into several layer packets, wherein each packet contains the seabed response to a certain depth. This layer-stripping approach uses the results of the previous layer packet as prior information for subsequent packets. The resulting posterior probability density for the final packet is considered the full solution to the inverse problem, and is interpreted in terms of optimal parameter estimates, marginal distributions, credibility intervals, and parameter correlations.