Trans-dimensional geoacoustic inversion and uncertainty quantification for SBCEX17 data

This paper presents an efficient and general approach to Bayesian inversion and uncertainty quantification for seabed geoacoustic profile estimation. The model-selection problem of estimating an appropriate seabed parameterization is addressed with trans-dimensional (trans-D) inversion via reversible-jump Markov-chain Monte Carlo, which samples probabilistically over the number of layers. An efficient proposal density for parameter perturbations is based on using a linearized approximation to the posterior probability density, applied in principal-component (PC) space where the (rotated) parameters are uncorrelated. The acceptance rate of perturbations and birth/death steps is improved by parallel tempering, based on a series of interacting Markov chains with successively tempered (relaxed) likelihoods. The PC proposals are adapted individually to the tempering of each Markov chain. The data-error model is based on the assumption of multivariate Gaussian errors with correlations represented by an autoregressive process. The parameters of zeroth- and first-order autoregressive error processes are sampled trans-dimensionally to avoid over- or under-parameterizing the error model. The approach is illustrated for three data sets from the 2017 Seabed Characterization Experiment (SBCEX17), including broadband seabed reflection coefficients; dispersion of water-borne acoustic modes, resolved by warping analysis; and ship noise recorded at a bottom-mounted horizontal array of hydrophones. This paper presents an efficient and general approach to Bayesian inversion and uncertainty quantification for seabed geoacoustic profile estimation. The model-selection problem of estimating an appropriate seabed parameterization is addressed with trans-dimensional (trans-D) inversion via reversible-jump Markov-chain Monte Carlo, which samples probabilistically over the number of layers. An efficient proposal density for parameter perturbations is based on using a linearized approximation to the posterior probability density, applied in principal-component (PC) space where the (rotated) parameters are uncorrelated. The acceptance rate of perturbations and birth/death steps is improved by parallel tempering, based on a series of interacting Markov chains with successively tempered (relaxed) likelihoods. The PC proposals are adapted individually to the tempering of each Markov chain. The data-error model is based on the assumption of multivariate Gaussian errors with correlations represented by an autoregr...