Magnetotelluric Bayesian Inversion Under The Condition Of Observation Data With Correlation Both In Frequency And Space

In Magnetotelluric Bayesian inversion, using the diagonal covaria-nce matrix could perfectly solve inverse problems under the condition of independent observation noises. But when noises have correlation, this method will have negative effects on explanation’s accuracy because of the ignorance of the correlation between noises and overestimation of the information contained in the observational data; Meanwhile, this method has smaller inversion uncertainty, which could not truly reflect the actual situation. In current, many researches reveal that the impact of noise cor-relation on inverse problems can not be ignored.In this paper the noise correlation was added into the inversion proc-ess. At first, one-dimensional magnetotelluric models of three and five layers were built, and non-diagonal covariance matrix was established by using data residuals. The simulated annealing search helped to look for the optimal model, and every optimal model is used to make iteration of new non-diagonal matrix. In addition, the statistical analytic indicators of Bayesian inversion, such as model edge probability distribution, model covariance matrix, confidence intervals were accepted to analyze Metrop-lis sampling; moreover, the result was compared with that produced from diagonal covariance matrix. The comparison revealed that non-diagonal matrix could not only attain more accurate models but also had a more smooth probability distribution diagram and a wider confidence interval. All of these show that the non-diagonal matrix was able to eliminate the interference information out of the observational data, which led to a more correct inversion uncertainty. Then the field data COPPER1was accepted into Bayesian inversion. And this data reflected few correlations between frequencies, so there was no big differences in using non-diagon-al matrix and diagonal matrix.At last, this method above would be introduced into two-dimensional magnetotelluric space. In two-dimensional space, the finite element meth-od was used for forward modeling. Simulated annealing and Metropolis sampling were used for inversion analysis, which reflected the similar re-gulation with that in the one-demensional space.