Using empirical mode decomposition to process marine magnetotelluric data

Abstract

A major step in processing magnetotelluric (MT) data is the calculation of an impedance tensor as function of frequency from recorded time-varying electromagnetic fields. Common signal processing techniques such as Fourier transform based procedures assume that the signals are stationary over the record length, which is not necessarily the case in MT, due to the possibility of sudden spatial and temporal variations in the naturally occurring source fields. In addition, noise in the recorded electric and magnetic field data may also be non-stationary. Many modern MT processing techniques can handle such non-stationarities through strategies such as windowing of the time-series. However, it is not completely clear how extreme non-stationarity may affect the resulting impedances. As a possible alternative, we examine a heuristic method called empirical mode decomposition (EMD) that is developed to handle arbitrary non-stationary time-series. EMD is a dynamic time series analysis method, in which complicated data sets can be decomposed into a finite number of simple intrinsic mode functions. In this paper, we use the EMD method on real and synthetic MT data. To determine impedance tensor estimates we first calculate instantaneous frequencies and spectra from the intrinsic mode functions and apply the impedance formula proposed by Berdichevsky to the instantaneous spectra. We first conduct synthetic tests where we compare the results from our EMD method to analytically determined apparent resistivities and phases. Next, we compare our strategy to a simple Fourier derived impedance formula and the frequently used robust processing technique bounded-influence remote reference processing (BIRRP) for different levels of stochastic noise. All results show that apparent resistivities and phases which are calculated from EMD derived impedance tensors are generally more stable than those determined from simple Fourier analysis and only slightly worse than those from the robust processing. These results show that EMD has the potential to handle noisy data. Finally, as a test on real data, we apply our processing scheme to data measured from the Costa Rica subduction zone, and obtain similar impedance estimates as the BIRRP processing method.