A data variance technique for automated despiking of magnetotelluric data with a remote reference

ABSTRACT The magnetotelluric method employs co‐located surface measurements of electric and magnetic fields to infer the local electrical structure of the earth. The frequency dependent ‘apparent resistivity’ curves can be inaccurate at long periods if input data are contaminated – even when robust remote reference techniques are employed. Data despiking prior to processing can result in significantly more reliable estimates of long period apparent resistivities. This paper outlines a two‐step method of automatic identification and replacement for spike‐like contamination of magnetotelluric data; based on the simultaneity of natural electric and magnetic field variations at distant sites. This simultaneity is exploited both to identify windows in time when the array data are compromised as well as to generate synthetic data that replace observed transient noise spikes. In the first step windows in data time series that contain spikes are identified according to an intersite comparison of channel ‘activity’– such as the variance of differenced data within each window. In the second step, plausible data for replacement of flagged windows are calculated by Wiener filtering coincident data in clean channels. The Wiener filters – which express the time‐domain relationship between various array channels – are computed using an uncontaminated segment of array training data. Examples are shown where the algorithm is applied to artificially contaminated data and to real field data. In both cases all spikes are successfully identified. In the case of implanted artificial noise, the synthetic replacement time series are very similar to the original recording. In all cases, apparent resistivity and phase curves obtained by processing the despiked data are much improved over curves obtained from raw data.