Smooth magnetotelluric impedance estimation by optimization

In urban and industrialised areas, magnetotelluric (MT) recordings often contain a significant amount of highpower near-field noise which, constituting non-Gaussian noise for most of the recording length, may lead to the failure of the common transfer functions (TF) estimation methods. The most evident consequence of this is the appearance of sharp points in the TF curves versus frequency, deteriorating their inherent smoothness. The TF’s smoothness is in fact one of the main criteria adopted by MT practitioners to evaluate the data quality and effectiveness of the estimation procedure; when not reached by the estimation, it may be a-posteriori introduced by splines or filtering procedures, but these approaches lack physical soundness. We propose a new heuristic algorithm aimed at obtaining the maximum TF versus frequency smoothness, through an optimization scheme rejecting EM events in frequency-domain. The algorithm searches for frequency-dependent power thresholds to be applied to the events, rejecting those making the TF sharp. The TF’s smoothness is the objective function to be minimized, and the model space is constituted by the infinite set of threshold vectors for each analysed frequency. After the process, the Ordinaly Least Squares (OLS) residual distribution is much closer to a Rayleigh one, allowing for an unbiased estimation. We applied the procedure both to synthetic and real MT data. Measured data has been collected for a geothermal exploration project in the East Africa Rift, where strong coherent noise affected the recordings. The combination of the technique with the remote-reference technique effectively reduced the bias due to the artificial signal. The physical consistency of the resulting TF has been tested by dispersion relationships.