New application of wavelets in magnetotelluric data processing: reducing impedance bias
 Hugo Larnier†^{1}Email author,
 Pascal Sailhac†^{1} and
 Aude Chambodut†^{1}
Received: 3 December 2015
Accepted: 14 April 2016
Published: 27 April 2016
Abstract
Magnetotelluric (MT) data consist of the sum of several types of natural sources including transient and quasiperiodic signals and noise sources (instrumental, anthropogenic) whose nature has to be taken into account in MT data processing. Most processing techniques are based on a Fourier transform of MT time series, and robust statistics at a fixed frequency are used to compute the MT response functions, but only a few take into account the nature of the sources. Moreover, to reduce the influence of noise in the inversion of the response functions, one often sets up another MT station called a remote station. However, even careful setup of this remote station cannot prevent its failure in some cases. Here, we propose the use of the continuous wavelet transform on magnetotelluric time series to reduce the influence of noise even for single site processing. We use two different types of wavelets, Cauchy and Morlet, according to the shape of observed geomagnetic events. We show that by using wavelet coefficients at clearly identified geomagnetic events, we are able to recover the unbiased response function obtained through robust remote processing algorithms. This makes it possible to process even single station sites and increase the confidence in data interpretation.
Keywords
Background
The second issue we would like to address comes from the complexity of MT time series that we mentioned before. The first MT processing techniques (e.g., Sims et al. 1971) were based on leastsquares analysis which fails in the presence of outliers. Outliers may originate from several issues including nonstationarity, the source effect of MT signals and brief overwhelming noise. Robust statistics were then applied to MT processing to accommodate these outliers (Chave et al. 1987; Egbert and Booker 1986). Another approach was the use of wavelet techniques to look for MT sources or remove noise in the time series. Zhang and Paulson (1997) applied the continuous wavelet transform (CWT) on highfrequency data to select sferics events. They thus increased the signaltonoise ratio in the response function determination in a straightforward way by the selection of useful signal parts. Trad and Travassos (2000) used wavelets specifically to filter MT data in the time–frequency domain before processing with robust algorithms. Escalas et al. (2013) used wavelet analysis to study the polarization properties of cultural noise sources in MT time series.
Methods
Continuous wavelet transform
Wavelet’s choice
In this work, we have chosen two different wavelets, the Cauchy and Morlet wavelets, and each is adapted to the shape of the considered geomagnetic event. Both wavelets are progressive ones with real valued Fourier transform.
Pulsations
Sferics
Sferics are transient events that occur in the frequency band range of 1 Hz to more than 10 kHz (Garcia and Jones 2002). These signals are the most energetic part of the MT signal in these frequency bands (even if other phenomena occur such as Schumann resonances between the waveguide made by the ionosphere and Earth’s surface).
These sferics are characterized by a sudden impulse in the time domain and can be divided into two main frequency bands, namely those above 1 kHz and those below it. In the lowfrequency band of extremely lowfrequency (ELF) waves, sferics consist of a highamplitude impulse (Fig. 3) called the slow tail (Mackay and FraserSmith 2010), whereas in the highfrequency band of very lowfrequency (VLF) waves, sferics are mainly shaped as oscillations. To analyze ELF waves using the CWT, we have chosen the Cauchy wavelet. Most of the previous work on MT methods with wavelets was based on the Morlet wavelet stemming from its properties of good resolution in terms of both the time and frequency (Zhang and Paulson 1997; Garcia and Jones 2008). However, the shape of the ELF wave (Fig. 4a) is very impulsive and therefore wide in frequency, so it is closer to Cauchy’s shape than to Morlet’s one. Indeed, the ELF wave is very short in time and has very few oscillations, so we need a wavelet that has similar properties. As illustrated in Fig. 4b, c, the highest coefficients obtained with the Morlet wavelet span a wide time length, i.e., longer than the actual length of the ELF wave. Because of their impulsive nature, sferics have a wide frequency content (Mackay and FraserSmith 2010). In the time–frequency domain, sferics appear as a series of highvalued coefficients along the frequency axis (very localized in time and spread out in frequency).
Selection of wavelet coefficients
We manually picked geomagnetic events from our MT time series and selected wavelet coefficients from the time–frequency plane by using the following methodology.

We compute the median value \(\alpha\) of the distribution of \(\mathcal {W}_{\psi }[s](a, \tau )\) over a length \(T_a = Nt_{a}\) around \(t_0\), which is the time position of the previously picked event. N in this work is set to 30 and \(t_a\) is the time step corresponding to the analyzed scale.

From this value, we define a threshold set as \(\beta\) times the median value \(\alpha\).

All coefficients below this threshold level are discarded for the magnetotelluric response function computation.

Coefficients are kept for the final computation when the threshold is reached on (\(e_x\), \(h_x\), and \(h_y\)) or (\(e_y\), \(h_x\), and \(h_y\)).
This criterion is sufficient for geomagnetic pulsations (below the Hz). However, for ELF waves, we add one more criteria to fulfill, that is, selected coefficients must span a large frequency range (basically the whole frequency range of ELF waves without MT and audiomagnetotelluric (AMT) dead bands). In practice, we check the Fourier spectrum of the EM time series, which is a good indicator of dead band boundaries. Depending on the signaltonoise ratio or the time of the day, this limit can be highly variable.
When a remote station is available, we add another criteria to discriminate noise from geomagnetic events. Indeed, the horizontal magnetic coefficients at the remote station have to fulfill the same properties. At the end of this stage, we obtain a group of wavelet coefficients for each scale and each selected geomagnetic event.
Inversion of wavelet coefficients
To reduce the influence of noise in the electric field, we also use robust statistics instead of classical least squares. As explained in (Chave and Jones 2012, Chapter 5), robust statistics were introduced for MT because of the properties of naturalsource electromagnetic data. Among these reasons, they state “finite duration of many geomagnetic or cultural events” and “marked nonstationarity.” Moreover, the heteroscedasticity of the noise can hide natural events along the time series. The main argument for using wavelet transform is precisely to take into account these issues by getting rid of the nonuseful part of the time series (e.g., no apparent geomagnetic signal) before computing the MT response function. In practice, in many time series the nonuseful part is the largest component. Consequently, Z computation is only based on parts of the time series where there is a significant induction occurring in the subsurface. Indeed, a significant part of the records cannot be used because of the noise level of current stateoftheart magnetic sensors (Chave and Jones 2012, Chapter 9). For example, in the AMT dead band, the signal does not rise above the noise level of the induction coils during daytime (Garcia and Jones 2002). Robust statistics are still necessary, but we considerably increase the signaltonoise ratio on all selected events.
Zhang and Paulson (1997) worked in a single station configuration and used conventional leastsquares analysis to resolve Eq. (7). Instead of using all selected wavelet coefficients to recover the MT response functions, we estimate the response function on pairs of geomagnetic events. Doing so allows us to build a distribution for each component of the impedance tensor (e.g., Fig. 5).
At each scale for each pair of geomagnetic events, the response function is estimated by using the Huber Mestimator described by Chave et al. (1987) in a MT context. From the distribution of response functions at each scale estimated from all available pairs of geomagnetic events, we represent the final estimation by taking the median value (less sensitive to outliers). The confidence interval is represented by using the interquartile range (IQR) Lestimator of the distribution (Huber and Ronchetti 2009). The IQR has a breakdown point of 50 %, which makes it robust to long tails in the distribution while still being a good measure of the dispersion or skewness in the distribution.
Real data application
Datasets
We applied the previously described technique on two real datasets to illustrate the potential of using wavelet coefficients on both geomagnetic pulsations and sferics.
The first dataset consists of MT data that were acquired with a sampling frequency of 512 Hz in the northeastern part of France near the town of Rittershoffen as part of a geothermal monitoring experiment (Abdelfettah et al. 2014). The electric measurements were made with EPF06 electrodes and the magnetic measurements with MFS07e coils from Metronix SA. Data in this area are badly contaminated by 50/3 and 50 Hz noise and their harmonics. We filtered this noise by using a notch filter with 2048 coefficients. The remote used with these measurements is located on the site of the Welschbruch geophysical station in the Vosges mountains at about 60 km away from the measurement site.
The second dataset consists of electromagnetic horizontal measurements that were acquired on the La Fournaise volcano before, during and after the 1998 eruption (Zlotnicki et al. 2005). The data were acquired with a sampling frequency of 0.05 Hz with induction coils similar to those made by Metronix or Phoenix Geophysics (Clerc 1971), and PbPbCl\(_2\) electrodes. The remote station used here is also located on the volcano, a little less than 10 km away from the measurement station.
To assess the validity of our method, we have compared our results with the ones obtained with robust processing codes from the work of Alan D. Chave (BIRRP, Bounded Influence Remote Reference Processing, Chave et al. 1987; Chave and Thomson 2004) and Gary Egbert (EMTF, Egbert 1997). For simplicity, we consider error bars as they were calculated by the robust processing codes, even if they can be discussed in more detail (Wawrzyniak et al. 2013; Chave 2015; Wawrzyniak et al. 2015).
The threshold factor value \(\beta\) was set to 1 for the sferic application and 4 for the pulsation application, but other values can be chosen depending on the signaltonoise ratio of the processed time series.
Remote processing
We compared the MT response functions obtained by using the wavelet coefficients with those from the application of other processing codes for both datasets.
In both frequency bands (Figs. 6, 7), the response functions obtained with the wavelet coefficients were comparable to those obtained with both processing codes. For response functions in the frequency band covered by geomagnetic pulsations (lowfrequency band), wavelet inversion was also in agreement with robust processing even if the source was not as impulsive as sferics. Figure 5 also demonstrates that the use of robust statistics on the wavelet coefficients of only two events allowed for the recovery of the response functions with good accuracy.
Single site processing
For single site processing, we processed single site MT stations by removing the remote station for all processing codes. For visual comparison, only BIRRP results are shown, but conclusions remain the same for EMTF.
For highfrequency results (Fig. 8), \(\rho _{yx}\) was downwardbiased by noise over a large frequency bandwidth and \(\rho _{xy}\) was also biased around 0.08 Hz. Because of the position of the station near housing, it was difficult to assess where the noise source was (it could have been from electric fences or pipelines). In this case, the result given by robust processing in the single site configuration would lead to a wrong interpretation of the MT response. Wavelet processing allows one to drastically reduce the noise bias and recover the impedance tensor obtained by using remote processing even near the 50 Hz frequency.
For lowfrequency processing (Fig. 9), the single site \(\rho _{xy}\) was slightly downwardbiased by noise where the waveletbased result remained within the confidence bounds of the remote processing. Particularly, the \(\rho _{xy}\) component was very sensitive to the coherence parameter. Some examples of interpretable MT responses for this component are illustrated in Fig. 10. If the coherence parameter is too low (below 0.5), \(\rho _{xy}\) is distinctly biased by noise (up to a factor of three) for periods below 200 s.
Other transient applications
So far, we have illustrated only two types of transient signals, namely sferics in their lowest frequency range (e.g., slow tails) and geomagnetic pulsations. This approach can also be applied on wider MT bands than the ones presented in this paper.
For higherfrequency applications, the slow tails studied in this paper already contain frequency content above 256 Hz (up to 1 kHz). Other impulsive waves such as VLF waves have a distinctive signature in the time–frequency plane and frequency content from 1 kHz up to more than 10 kHz (Rakov and Uman 2003). In Fig. 11, we show an example of such a VLF wave and its time–frequency analysis. These waves were already used in an MT processing study by Zhang and Paulson (1997).
The assortment of lowfrequency geomagnetic sources is large. Among transient events at periods lower than 200 s, there are Pc5 geomagnetic pulsations that have periods up to 500–600 s (Saito 1969). Geomagnetic storms are also lowfrequency sources of induction. They have already been studied with wavelet analysis (e.g., Mendes et al. 2005). We show, for example, in Fig. 12 a signal of interest at low frequency that was recorded on the Piton de La Fournaise volcano on July 16, 1997, and the signal contains significant frequency content up to more than 1000 s.
Conclusion
We have shown through these experiments that the continuous wavelet transform (CWT) is an easy and efficient way to characterize the magnetotelluric (MT) response function for transient geomagnetic events. Using this technique, we have shown that most of the information contained in the source wavelet coefficients is sufficient in datasets to enable the characterization of the MT impedance tensor. Two types of geomagnetic events were studied, geomagnetic pulsations and sferics, and thus this work increases the frequency band studied in previous publications involving CWT application during MT processing. The mother wavelet has to be adapted to each transient event to accurately recover the source information. Even irregular geomagnetic pulsations that are very localized in the time and frequency domains simultaneously can be used in MT processing. The other transient sources of induction are currently being investigated and will be the subject of future papers on wavelet analysis of magnetotelluric time series.
We have also shown that we were able to drastically reduce the noise bias of the considered datasets on the MT response function in the case of a single station configuration. This was achieved by using the high signaltonoise ratio transient events in the MT time series. By using robust statistics with only two geomagnetic events in the wavelet domain, we were able to recover accurate response functions.
Yet, we still have to develop a way to automatically detect geomagnetic events to analyze their properties (e.g., polarization) and their effect on MT impedance computation.
Notes
Declarations
Author’s contributions
The manuscript was written by HL and reviewed by all authors. All authors read and approved the final manuscript.
Acknowledgements
We would like to thank the three anonymous reviewers who greatly helped to improve this manuscript. We thank the LABEX GEauThermie of the University of Strasbourg and the field crew (Y. Abdelfettah) for the acquisition of MT measurements at Rittershoffen. We also thank J. Zlotnicki for Le Piton de la Fournaise EM time series. Finally, we also acknowledge discussions with various colleagues and students at early stages of this work.
Competing interests
The authors declare that they have no competing interests.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 Abdelfettah Y, Sailhac P, Schill E, Larnier H (2014) Preliminary magnetotelluric monitoring results at Rittershoffen. In: 3rd European geothermal workshop, Karlsruhe, 15–16 Oct 2014Google Scholar
 Bortnik J, Cutler JW, Dunson C, Bleier TE (2007) An automatic wave detection algorithm applied to Pc1 pulsations. J Geophys Res 112:A04204. doi:10.1029/2006JA011900 Google Scholar
 Chave A (2015) Comment on “Robust error on magnetotelluric impedance estimates” by P. Wawrzyniak, P. Sailhac and G. Marquis. Geophys Prospect. doi:10.1111/j.13652478.2012.01094.x Google Scholar
 Chave AD, Jones A (2012) The magnetotelluric method: theory and practice. Cambridge University Press, CambridgeView ArticleGoogle Scholar
 Chave AD, Thomson DJ (2004) Bounded influence magnetotelluric response function estimation. Geophys J Int 157(3):988–1006View ArticleGoogle Scholar
 Chave AD, Thomson DJ, Ander ME (1987) On the robust estimation of power spectra, coherences, and transfer functions. J Geophys Res 92(B1):633–648View ArticleGoogle Scholar
 Clerc G (1971) Contribution à l’optimisation des capteurs à induction destinés à la mesure des variation du champ magnétique terrestre (10−3 à 104 Hz). PhD thesis, Paris, FranceGoogle Scholar
 Egbert GD (1997) Robust multiplestation magnetotelluric data processing. Geophys J Int 130(2):475–496View ArticleGoogle Scholar
 Egbert GD, Booker JR (1986) Robust estimation of geomagnetic transfer functions. Geophys J R Astron Soc 87(1):173–174View ArticleGoogle Scholar
 Escalas M, Queralt P, Ledo J, Marcuello A (2013) Polarisation analysis of magnetotelluric time series using a waveletbased scheme: a method for detection and characterisation of cultural noise sources. Phys Earth Planet Interiors 218:31–50View ArticleGoogle Scholar
 Gamble TD, Goubau WM, Clarke J (1979) Magnetotellurics with a remote magnetic reference. Geophysics 44(1):53–68View ArticleGoogle Scholar
 Garcia X, Jones AG (2002) Atmospheric sources for audiomagnetotelluric (AMT) sounding. Geophysics 67(2):448–458View ArticleGoogle Scholar
 Garcia X, Jones AG (2008) Robust processing of magnetotelluric data in the AMT dead band using the continuous wavelet transform. Geophysics 73(6):223–234View ArticleGoogle Scholar
 Holschneider M (1995) Wavelets: an analysis tool. Oxford Science Publications, OxfordGoogle Scholar
 Huber PJ, Ronchetti EM (2009) Robust statistics, 2nd edn. Wiley, New YorkView ArticleGoogle Scholar
 Jun C, Shiokawa K, Connors M, Schofield I, Poddelsky I, Shevtsov B (2014) Study of Pc1 pearl structures observed at multipoint ground stations in Russia, Japan, and Canada. Earth Planets Space 66(140):1–14Google Scholar
 Mackay C, FraserSmith AC (2010) Lightning location using the slow tails of sferics. Radio Sci 45(5). doi:10.1029/2010RS004405
 McPherson RL (2005) Magnetic pulsations: their sources and relation to solar wind and geomagnetic activity. Surv Geophys 26(5):545–592View ArticleGoogle Scholar
 Mendes O Jr, Domingues MO, da Costa M, Clúa de Gonzalez A (2005) Wavelet analysis applied to magnetograms: singularity detections related to geomagnetic storms. J Atmos Solar Terr Phys 67:1827–1836View ArticleGoogle Scholar
 Rakov VA, Uman MA (2003) Lightning: physics and effects. Cambridge University Press, CambridgeView ArticleGoogle Scholar
 Saito T (1969) Geomagnetic pulsations. Space Sci Rev 10:319–412View ArticleGoogle Scholar
 Sims WE, Bostick FXJ, Smith HW (1971) The estimation of magnetotelluric impedance tensor elements from measured data. Geophysics 36(5):938–942View ArticleGoogle Scholar
 Trad DO, Travassos JM (2000) Wavelet filtering of magnetotelluric data. Geophysics 65(2):482–491View ArticleGoogle Scholar
 Wawrzyniak P, Sailhac P, Marquis G (2013) Robust error on magnetotelluric impedance estimates. Geophys Prospect 61:533–546View ArticleGoogle Scholar
 Wawrzyniak P, Sailhac P, Marquis G (2015) Reply of the authors to AD Chave’s comment on Wawrzyniak P., Sailhac P. and Marquis G. 2013. Robust error on magnetotelluric impedance estimates. Geophys Prospect 64(1):250–251View ArticleGoogle Scholar
 Zhang Y, Paulson KV (1997) Enhancement of signaltonoise ratio in naturalsource transient magnetotelluric data with wavelet transform. Pure Appl Geophys 149:405–419View ArticleGoogle Scholar
 Zlotnicki J, Le Mouël JL, Gvishiani A, Agayan S, Mikhailov V, Bogoutdinov S, Kanwar R, Yvetot P (2005) Automatic fuzzylogic recognition of anomalous activity on long geophysical records: application to electric signals associated with the volcanic activity of La Fournaise volcano (Reunion Island). Earth Planet Sci Lett 234:261–278View ArticleGoogle Scholar