Geoacoustic inversion via trans-dimensional sampling over seabed and error models

This paper develops an efficient Bayesian sampling approach to geoacoustic scattering and reflection inversion based on trans-dimensional (trans-D) sampling over both the seabed model (number of sediment layers) and error model (autoregressive order to represent residual correlation). Sampling is carried out using a population of interacting Markov chains employing a range of sampling temperatures (parallel tempering). The approach is applied to both simulated and measured data. The advantages of trans-D autoregressive model sampling over alternative methods of error model selection is explored in terms of the reduction in posterior uncertainty of geoacoustic parameters and evaluation of residual correlation. The seabed is modeled as a stack of homogeneous fluid sediment layers overlying an elastic basement. Including elastic (shear) parameters in the basement makes this layer distinct from the overlying sediment layers and requires a novel formulation of the partition prior distribution for trans-D sampling.