Robust multiple-station magnetotelluric data processing

Although modern magnetotelluric (MT) data are highly multivariate (multiple components, recorded at multiple stations), commonly used processing methods are based on univariate statistical procedures. Here we develop a practical robust processing scheme which is based on multivariate statistical methods. With this approach we use data from all channels to improve signal-to-noise ratios, and to diagnose possible biases due to coherent noise. To illustrate our approach we use data from two- and three-station wide-band MT arrays from an area south of San Francisco, California, where contamination of the MT signal by spatially coherent cultural electromagnetic noise is severe at some periods. To deal with such coherent noise we adopt a two-stage procedure. In the first stage we focus on reducing the effects of incoherent noise, and testing for the presence of coherent noise. To this end we have developed a robust multivariate errors-in-variables (RMEV) estimator, which estimates background noise levels, cleans up outliers in all channels, and determines the ‘coherence dimension’ of the array data. In the absence of coherent noise, the coherence dimension of the data will be two (corresponding to two polarizations of the plane-wave MT source fields). In this case the RMEV estimator provides direct estimates of MT impedances and inter-station transfer functions. We show, with synthetic and real data examples, that in some cases these estimates can be significantly better than those obtained with more standard robust remote reference estimators. When MT data is severely contaminated by coherent noise (as for our example arrays for periods of 4-50 s) the coherence dimension of the data will exceed two. The RMEV estimate thus provides a clear warning of coherent noise contamination. Although there appears to be no completely general automatic way to deal with this circumstance, useful results can be obtained from severely contaminated data in some cases. We show in particular how the RMEV estimator can be adapted to separate the MT signal from coherent noise for two special cases: when at least one site is unaffected by coherent noise, and when coherent noise sources are intermittent. We give examples of significant improvements in MT impedance estimates obtained with the RMEV estimate for each of these cases.