Effect of data error correlations on trans-dimensional MT Bayesian inversions

Real magnetotelluric (MT) data errors are commonly correlated, but MT inversions routinely neglect such correlations without an investigation on the impact of this simplification. This paper applies a hierarchical trans-dimensional (trans-D) Bayesian inversion to examine the effect of correlated MT data errors on the inversion for subsurface geoelectrical structures, and the model parameterization (the number of conductivity interfaces) is treated as an unknown. In the inversion considering error correlations, the data errors are parameterized by the first-order autoregressive (AR(1)) process, which is included as an unknown in the inversion. The data information itself determines the AR(1) parameter. The trans-D inversion applies the reversible-jump Markov chain Monte Carlo algorithm to sample the trans-D posterior probability density (PPD) for the model parameters, model parameterization and AR(1) parameters, accounting for the uncertainties of the model dimension and data error correlation in the uncertainty estimates of the conductivity profile. In the inversion ignoring the correlation, we neglect the correlation effect by turning off the AR(1) parameter. Then the correlation effect on the MT inversion can be examined upon comparing the posterior marginal conductivity profiles from the two inversions. Further investigation is then carried out for a synthetic case and a real MT data example. The results indicate that for strong correlation cases, neglecting error correlations can significantly affect the inversion results.[graphic not available: see fulltext]