Magnetotelluric Data Analysis Using Advances in Signal Processing Techniques

.
Magnetotelluric data analysis using

advances in signal processing
techniques
C. MANOJ
National Geophysical Research Institute, Hyderabad 500 007,

India
THESIS SUBMITTED TO THE OSMANIA UNIVERSITY FOR THE

AWARD OF THE DEGREE OF DOCTOR OF PHILOSOPHY IN
GEOPHYSICS 2003
.
ii
Declaration
I, hereby declare that the thesis submitted for the award of the Degree of Doctor of
Philosophy in Geophysics of the Osmania University, Hyderabad, India is original in its
contents and has not been submitted before, either in parts or in full to any University
for any research degree.
(C. Manoj)
Candidate
Dr. V. P. Dimri
Dr. Nandini Nagarajan
Director
Research Supervisor
National Geophysical Research Institute
Scientist
Hyderabad
NGRI
Hyderabad
iii
Certificate
This is to certify that the thesis, entitled Magnetotelluric data analysis using advances
in signal processing techniques, which is submitted for the award of the Degree of Doctor
of Philosophy in Geophysics to the Osmania University, Hyderabad, India, is the bonafide
research work carried out by Mr. C. Manoj at National Geophysical Research Institute,
Hyderabad, India during the years 1998 to 2003 under my supervision. The work is
original and has not been submitted for any Degree at this or any other University
20-08-2003
(Nandini Nagarajan)
Research Supervisor
Scientist
NGRI
Hyderabad 500007
India
iv
Acknowledgement
I sincerely thank Dr. Nandini Nagarajan for the support and guidance she extended
to my research program as a research supervisor, over the past five years of my work at
National Geophysical Research Institute. As a mentor and group head of Magnetotellurics
Division, Dr S. V. S. Sarma gave me moral and institutional support in the early phases
of my Ph.D program. The present group head Dr. T. Harinarayana also adopted the
same approach. Dr. H.K. Gupta and Dr. V.P. Dimri, the former and present Directors
of NGRI, kindly encouraged me to pursue the research work.
Prof. (late) P.S. Moharir and Mr. G. Virupakshi are profusely thanked for the hours
of discussion they had with me. Understanding my limitations, they fixed many bugs
in my computer codes as well as knowledge. The staff of the offices of Dean, Faculty of
Science and Head, Department of Geophysics, Osmania University is thanked for their
co-operation throughout the research program. The Senior colleagues of magnetotellurics
group: Mr. D.N. Murthy, Dr. R.S. Sastry, Dr. M. Someswara Rao, Mr. M.V.C. Sarma,
Dr. Madhusudan Rao, Dr. K. Veeraswamy and Mr. S. Prabhakar E. Rao are thanked for
initiating me in magnetotellurics and their help during field campaigns. As immediate
colleagues, I acknowledge the support of Dr. B.P.K. Patro, Dr. K. Naganjaneyulu, Dr.
K. Begum, and Mr. K.K. Abdul Azeez for their cooperation and support. Financial
support from the following agencies is also acknowledged.
The magnetotelluric data for the present studies were acquired in a project funded by
Department of Science & Technology, Government of India (No. ESS/16/118/1997). I
received the Junior Research Fellowship of Council of Scientific and Industrial Research
(No. 2-31/97(i)-E.U.II) during the initial phase of the research program. Suggestions
by Dr. Martyn Unsworth, Dr. John Stodt and Dr. Xavier Garcia greatly improved
a part of the thesis. I take this opportunity to thank my friends Dr. Shyam Chand,
Dr. Jimmy Stephen, Dr. R. S. Rajesh and Mr. Tomson J.K for their comradeship that
helped me in different stages of the research program. I thank my wife Mrs Namitha Vaz,
daughter Meenakshi, father Mr. M.K.C. Nair and mother Mrs. M.N. Sarojini Amma for
the patience and support they extended to me during my research program.
Magnetotelluric data analysis using

advances in signal processing techniques
SYNOPSIS
Electrical conductivity is one of the most important physical parameters directly indicating the Earths subsurface nature. Rocks exhibit a wide range ( 106 ) of conductivity
and the electrical conductivity of rocks is sensitive to temperature, presence of fluid,
volatiles, melt as well as its bulk composition. These qualities make electrical conductivity an appropriate method to delineate Earths subsurface features with a suitable
measurement at surface. Over the past century a suite of electrical and electromagnetic
methods was established to probe electrical conductivity structure of the Earth. Among
them, the natural source electromagnetic method - Magnetotellurics has many advantages
over all the other methods: with the skin depth relation of EM waves it can virtually
probe any depth and the natural electromagnetic signals have enough power over a wide
range of frequencies to penetrate the subsurface (Cagniard [1953]). Thus magnetotelluric
(MT) method has become a promising technique to probe deep earth structure. The
magnetotelluric induction, though diffusive in nature like potential fields, is not inherently non- unique like potential fields methods. The uniqueness of this method in cases
where conductivity varies vertically (1-Dimensional) is established (Bailey [1970], Weidelt [1972]) in theory. Non-uniqueness in magnetotellurics may results from 1) errors in
measurement 2) finite frequency range of data and 3) sparse distribution of measurement
sites. MT practitioners, therefore, strive to obtain more precise data in wider bandwidth
and with more spatial density. As a result remarkable progress has been made to instrumentation (Clarke et al. [1983], Ritter et al. [1998]) and time series processing (Gamble
et al. [1979], Egbert and Booker [1986], Chave and Thomson [1989]) in the past two
decades that resulted in better estimation of full tensor (transfer function) elements and
their variance over a wide band of frequency. Together with the progress of decomposing
the impedance tensor into regional and residual (Groom and Bailey [1989], McNeice and
Jones [2001]) and the advances in inversion (Constable et al. [1987], Smith and Booker
[1991], Siripunvaraporn and Egbert [2000], Rodi and Mackie [2001]), magnetotellurics
has become a standard tool to determine Earths electrical conductivity structure.
The first step in interpretation of magnetotelluric data is to estimate the MT
impedance tensor in frequency domain from measured magnetotelluric time series. Magnetotelluric time series consist of five simultaneously measured components of earths
electromagnetic field viz. two orthogonal components of horizontal electric fields (Ex,
Ey in mV/ km) and three orthogonal components of magnetic field (Hx, Hy, Hz in nT).
Usually the measurements are done in wide bands of overlapping frequency ranges, with
different sampling intervals. Measured time series are the compounded effect of various
signal and noise processes in the frequency range of interest to MT. The objective of MT
processing is to discriminate signal and minimize the effect of noise in the estimation of
MT transfer function tensor Z . The use of traditional spectral analysis together with
least squares (LS) estimation is warranted only if the input channels (magnetic fields)
are noise free, the output channel noise has a Gaussian distribution and the MT time
series is stationary (Banks [1998]). In reality most data usually show gross departures
from the above idealistic model. The main causes are geomagnetic phenomena, thunder-
vi
storms, cultural interference and instrument problems. This results in highly oscillating
and biased estimates of MT transfer functions. Remote reference technique (Gamble
et al. [1979]) deals with the noise in magnetic fields. Reference magnetic field is recorded
at a site, which is far outside the coherency range of the noise. Using the cross spectrum
with the remote site instead of the autopower eliminates the bias in the magnetic field
power. Problems due to non-stationarity of time series are addressed by subdividing the
time series (Egbert and Booker [1986], Banks [1998]) into small segments, estimating
transfer functions for each one and averaging in a way that discriminates against noisy
data segments. Robust estimation of MT transfer functions (Egbert and Booker [1986],
Chave et al. [1987], Chave and Thomson [1989]) down weights the data sections with
such large non Gaussian noise. But this technique still gives erroneous results, if strongly
correlated noise is present during most of the recording time. To deal with such coherent
noise, robust multivariate errors-in-variables (RMEV) estimate was developed by Egbert [1997]. Correlated and uncorrelated noise are separated iteratively, using data from
multiple stations. Variants of this technique viz. Signal-Noise Separation (SNS) method
and SNS-remote-reference method are discussed by Larsen et al. [1996] and Oettinger
et al. [2001]. In a very recent work, Chave and Thomson [2003] proposed a bounded
influence function to robustly estimate magnetotelluric data contaminated with extreme
noises. Over the past decade, conventional robust methods have revolutionized the application of magnetotellurics in geophysics (Jones et al. [1989], Egbert [1997]) and are now
applied routinely and automatically producing reliable magnetotelluric responses in most
instances. The success of robust procedures may be attributed to three factors. First,
its superiority to other data processing techniques is established (Jones et al. [1989]).
Second, these procedures can be justified rigorously (Egbert and Livelybrooks [1996])
and third it can be easily implemented using iterative-weighted LS procedures and extended to remote reference processing (Chave and Thomson [1989]). However at sites
located near auroral region (Garcia et al. [1997]), major cultural noise centres, electric
railway lines etc where source field non-stationarity /noise contamination is severe, robust
methods frequently break down. The possible reasons may be attributed to:1. Failure in identifying noise source for a survey often prompts the processing of all
sites in a similar fashion. Whereas, stations that were affected with noise should
undergo special data treatment.
2. Any MT time series processing algorithm performs better, if presented with a data
set that is as clean as possible. This is often done by manual inspection of time
series. A failure/omission at this stage may contribute to break down of robust
processing.
3. Conventional robust processing uses an initial estimate of MT transfer functions
usually derived from an LS estimator. When majority of observations deviate from
the true one, the initial LS estimate becomes too far (biased) for the robust iterative
process to improve upon.
4. The cross and auto spectra between electric and magnetic field elements are usually
smoothed by a combination of band and section averaging. While effect of outliers
in section averaging is well known, the same about band averaging is overlooked.
vii
The present thesis concentrates on these problems of MT time series analysis, with
an aim to improve the estimation of transfer functions. The wide band magnetotelluric
data collected over Southern Granulite Terrain, offers scope for development of applications that may resist extreme noise contamination. There are two factors that make this
region important in this context. 1) The majority of upper crustal rocks in SGT belong
to Archaean and Proterozoic age (Naqvi and Rogers [1987]) and exhibit high electrical
resistivity as other shield regions in the world (Mareschal et al. [1994]). Highly resistive upper crust offers very little attenuation to EM signals and in principle noise can
propagate over larger distance in SGT as compared to regions where younger rocks are
exposed. 2) The high population density in the southern states of India, especially in
Tamil Nadu, where most of the MT sites are located give rise to various cultural noises.
The industrial belt along the two banks of Cauvery River is another noise source for MT.
In this context, the objectives of the present thesis are,
1. To characterize the signal/noise in magnetotelluric data collected over SGT, in
spatial, temporal and frequency domain and to locate the major sources of noise in
the data.
2. To evolve an efficient and automated method to discriminate noisy segments of time
series. The enrichment of s/n ratio in the data provided by such a process will help
the robust processing to better estimate MT transfer functions.
3. To improve the robust processing methods of MT transfer functions particularly
concentrating on the weak points (reasons stated as (iii) and (iv) earlier) in the
method, while dealing with noisy data.
4. Establish the efficacy of the processing methods proposed by application to sufficiently large amount of data collected from SGT.
A suite of advanced signal processing algorithms will be used for achieving the above
objectives. Typically MT time series is a large volume of multi-channel data (106 values) that are usually stored in compressed, binary format with a header file describing
location, sensor geometry and filter setting, and the data itself. A variety of MT equipment is currently in use, which delivers MT data in wide bands of frequencies. Standards
have been evolved in industry for the delivery and exchange of electromagnetic data, MT
in particular, as Electrical Data Interchange (EDI) format by Society of Exploration Geophysicists (E. [1988]). Though it can support time series data as well, it is not being used
so due to a various reasons. As time series is one-step closer to actual data generation
at the instrument than the MT transfer functions, the variation in MT hardware also
constraints the adoption of uni-format for MT time series. The major academic software
for MT time series processing, currently available free of charge are codes from Egbert
[1997], RRMT by Chave A.D, LiMS by Jones A.G. (available at http://mtnet.info) and
EMERALD by Ritter et al. [1998]. These codes are written specifically for certain types
of MT instruments and for use on particular operating systems. The commercial manufacturers of MT equipment also give processing software as part of the system (for e.g.
r and MAPROS
r from Metronix GmBH) which optimize user requirements,
ProcMT
viii
though not as versatile as the academic software. Moreover these software strictly cater
to the needs of one MT equipment, and mostly their source codes are not open. This
scenario puts pressure on researchers working on MT time series processing. In order to
modify processing routines, it is necessary to access the time series, calibrate the data
etc, which have been specifically evolved for particular equipment. Any new processing
codes, need to be attached to a time series reading, calibrating and storing utility, which
also has to be developed. Thus time series processing algorithm should have end to end
utilities, which involves reading compressed time series data, reading the sensor geometry
and filter settings, calibrating for system responses, the main processing and finally the
exporting the results in EDI format. In one way this facilitates easy interchange of data
and flexibility of applications, compared to a framework, wherein one relies on a set of
imported utilities to perform the peripheral tasks. For the current thesis, the processing
codes were written on end to end basis, where it performs all the peripheral processing
tasks as well.
As remote reference processing has been increasingly used by MT practitioners, in
tandem with a robust processing algorithm, a question can be asked Why should we be
concerned with single station processing at all?. There are several good reasons. First,
because of instrumental problem (as happened during data acquisition in SGT), it is
common to have single station recording in many surveys. If the reference site is very
noisy and the local site is not, single station estimation could be better than remote
reference estimates (Egbert and Livelybrooks [1996]). These factors often necessitate
single station processing of MT data. However the procedures developed in this thesis
can easily be extended to remote reference processing as well (an example in this regard is
discussed in Chapter 6). The thesis spreads in seven chapters. In the following sections,
an overview of the different chapters in the thesis is given.
Chapter I
Maxwells equations describe the properties of electromagnetic waves. The relationship between electric and magnetic filed within a conductive Earth can be expressed in
terms of wave equations by combining Maxwells equations. A conductive Earth responds
to the electromagnetic illumination at the surface by allowing the refracted components
to diffuse according to their frequency content. In consequence the magnetotelluric fields
contain information regarding conductivity distribution as a function of frequency. Instrumentation and field procedures play an important part in the quality of measured
magnetotelluric data. A discussion on field procedures, sensors and measuring device
employed to measure magnetotelluric data is given towards the end of Chapter 1.
Chapter II
The magnetotelluric transfer functions are obtained by the simultaneous measurement
of Earths time varying electrical and magnetic field. The sources of the natural electromagnetic fields in the frequency range 104 to 104 Hz have quasi-planar properties at
the surface of the Earth. Natural signal sources for MT measurements in the frequency
range 100 to 104 Hz are worldwide thunderstorm activity. At frequencies below 1 Hz the
source signals are generated due to the interaction of magnetosphere with solar wind.
Observed magnetotelluric fields at the surface of Earth contain additive noises from various manmade and natural sources, which do not behave as plane waves. A discussion on
different signal and noise sources and their effect in MT data is included in Chapter II.
ix
Chapter III
This chapter introduces the concepts of random data analysis. Magnetotelluric time
series like many other natural processes can be considered as wide band random data.
And thus the estimation of its properties becomes statistical (Bendat and Piersol [1971]).
A brief review of some of its properties viz. probability distribution functions, autocorrelation and power spectral density function are discussed. As magnetotelluric signals
constitute a multivariate system, the joint properties of individual components like cross
correlation, cross spectra and coherencies are also important. Reliable estimation of
magnetotelluric transfer function depends on the amount of noise in the time series measurements (Orange, 1989). It is often difficult to obtain noise free measurements as the
method relies on highly variable natural electromagnetic variation. Severe problems are
caused by civilization, which produce all kinds of electromagnetic noises. These noises
manifest themselves in the computed magnetotelluric transfer function as both statistical and bias shifts. Sims et al. [1971] proposed the classical least square solution for MT
transfer functions from noisy measured data. The statistical errors can be minimized by
averaging large amount of observations, if the noise distribution is Gaussian. As least
square solutions envisages a noise free input field, noise in input field can severely bias
such estimates. The bias errors can be tackled to an extent by measuring a remote
reference (Gamble et al. [1979]). But as shown by Chave et al. [1987], Chave and Thomson [1989], Egbert and Booker [1986], violation of the Gaussian assumption for errors
may lead to severe problems in least square estimation of MT transfer functions. The
properties of a least square estimator is discussed towards the end of Chapter III.
Chapter IV
The data used in the present study were measured over South India, as a part of an
integrated geological and geophysical study on the Southern Granulite Terrain by NGRI
under a DST project. The magnetotelluric measurements were carried out in two field
campaigns from 1998 to 2000. The sites were located in a 300-km long NS corridor. The
profile traverses major metamorphic and tectonic elements of the region. Harinarayana
et al. [2003] discusses the result of modeling of the data in terms of subsurface electrical
conductivity distributions and their importance in local tectonic set up. Five components
of natural EM variations were measured in the frequency range 0.0001 - 4000 s. The data
were acquired in four frequency bands. The spatial distribution and quantity of data
collected is described in detail in Chapter IV. Most of the data were collected in single
station reference, as the measurement units had time synchronization problems. This
necessitated use of single station processing techniques (Chapter III) for majority of MT
sites measured. The measurement corridor passes through many urban areas and one
industrial belt. Major electrified rail lines passes through some parts of the corridor. More
over the southern region of India is densely populated. The cultural noises thus generated
can propagate over large distances, as the upper crust is highly resistive. The indices
of geomagnetic activity during the measurement time were compared with the averaged
long period coherence of each site. The agreement of the coherence and the geomagnetic
indices indicate the validity of such an approach. However a few disagreements were
also evident. This indicates the possible noise processes contributing to the measured
data. Spatial relation of the known cultural noise centers and the quality of MT data
was examined. The strong correlation of the high noise/(signal +noise) ratio to the
major industrial belt indicate that the MT signals may get consistently degraded due
to presence of an active noise sources, irrespective of the signal activity. However the
conclusions drawn from this chapter should be treated rather cautiously. The signal
and noise are defined in this chapter as inline and outline components of a least square
solution of MT transfer functions.
Chapter V
The techniques outlined in Chapter III for estimation of magnetotelluric transfer functions from measured imprecise data requires that the time series is presented as clean
as possible. Cleaning of MT time series is presently done by manual inspection (editing). Editing of magnetotelluric time series is subjective in nature and time consuming.
Artificial neural networks (ANN) are widely used to automate processes, which requires
human intelligence. In Chapter V Artificial Neural Network is used to discriminate good
sections of data against noisy ones. Artificial neural networks (ANN) are emerging tools
that have been applied in many areas of science and engineering where pattern recognition is involved, such as speech and character recognition. The learning and adaptive
capabilities of these models make them attractive for application to some problems in
geophysics. As ANN based techniques are computationally intensive, a novel approach
was made to the problem, which involves editing of five simultaneously measured MT
time series. Neural network training was done at two levels. Signal and noise patterns
of individual channels were taught first. Training was stopped when both errors were
below acceptable level. The neural networks sensitivity to signal to noise ratio and the
relative significance of its inputs were tested to ensure the training was correct. The application of ANN based editing to magnetotelluric time series brings out some interesting
results. In a low noise environment the network editing produces results almost similar
to blind editing (using all stacks). On such a data, a simple coherency-based estimator
can do the signal discrimination to a certain level of satisfaction. However the neural
networks ability to pick out signal from moderate to high noise environment was evident
on the other data collected from SGT. In such cases it approximates human intelligence
- established from the fact that the neural network based editing gives a result similar
to manual editing. These results satisfy the second objective of the thesis, i.e., to provide a robust alternative to manual editing of magnetotelluric time series. The signal
and noise characteristics in magnetotelluric data are very different in different frequency
ranges. This is due to the difference in signal and noise source mechanism in different
part of Earths natural electromagnetic spectrum. Sensor geometry and instrumentation
also affect the pattern of signal on which the neural network is trained. This necessitates reformulation of MT signal and noise characteristics for training of ANN, wherever
necessary. However the extra computational requirement of re-training of ANN may not
pose a burden on resources, taking into consideration the ever-increasing computational
power of microprocessors.
Chapter VI
In chapter VI, two new approaches are proposed to improve the performance of robust
statistical procedures on MT time series. Non-parametric estimators such as Jackknife
(Efron [1982]) were used to robustly compute the variance of MT transfer functions
(Chave and Thomson [1989]). Its use as an effective initial guess for robust procedures is
discussed. It is shown that in majority of the cases, the use of Jackknife for initial guess
xi
resulted in better estimation of MT transfer function as compared to LS estimations. It

is usual in MT to sub divide the time series, estimate the spectral density matrices for
each segment individually and then robustly average the spectra or transfer functions
between the sub segments (section averaging). Within a segment, it is common to use
a limited number of target frequencies, and obtain smooth spectra by averaging several
adjacent Fourier harmonics (frequency band averaging). The documented researches on
robust estimation of MT spectral densities and transfer functions concentrate on section
averaging. This arises from the assumption that, within a narrow frequency band, the
distribution of Fourier coefficients are of Gaussian nature and a simple average (LS)
gives the best estimate. It is shown that this argument often fails, and the problem of
contamination is applicable to band averaging as well. Robust weighting approach is
proposed for estimation of cross and auto spectral estimation within a band, without
making specific model assumptions concerning signal or noise. Both these proposed
procedures, while applied on a large volume of MT data collected over SGT, South India,
met with moderate to good improvement of MT transfer functions.
Chapter VII
Chapter VII gives an overall summary of the results discussed in chapter IV to VI.
The main objective of the thesis, i.e. to obtain best estimates of transfer function from
measured magnetotelluric time series, may require a combination of one or more techniques introduced and demonstrated in the chapters IV to VI. Furthermore, Chapter VII
will take a closer look at the properties of robust processing and neural networks. On
comparison of individual performances of neural networks and robust processing, it was
found that their reaction to different types of noise differs in some cases. One reason is
that the outliers / or coherent noises which are obvious in time domain, failed to produce larger outliers and thus were not down weighted by robust procedures. In another
instance, data sets with noises, which the neural network allowed to pass on, were down
weighted by robust procedures. This clearly shows the need to combine the two techniques in order to discriminate / down weight a majority of noisy data. Further more,
this points to the necessity of transforming the data more than one domain to discriminate noise from signal. In this regard, possible use of wavelet transform in identifying
transient signals that are the common form of MT signals in for frequencies > 1Hz is
suggested as a future work.
Contents
1 Magnetotellurics: Basic Theory, Sensors and Field Procedures
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.1 Source fields . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.2 Maxwells equations . . . . . . . . . . . . . . . . . . . . .
1.1.3 Skin depth. . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.4 Impedance tensor: . . . . . . . . . . . . . . . . . . . . . .
1.1.5 Amplitude and phase of impedance: . . . . . . . . . . . . .
1.2 Sensors and Field Procedures . . . . . . . . . . . . . . . . . . . .
1.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2 Electric field sensors . . . . . . . . . . . . . . . . . . . . .
1.2.3 Magnetic Sensors . . . . . . . . . . . . . . . . . . . . . . .
1.2.4 Recording systems . . . . . . . . . . . . . . . . . . . . . .
1.3 Field Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1 Survey design . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.2 Site selection . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.3 Sensor deployment . . . . . . . . . . . . . . . . . . . . . .
1.3.4 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.5 Data acquisition. . . . . . . . . . . . . . . . . . . . . . . .
2 Signal and Noise Sources for magnetotellurics
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Signal sources . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Introduction . . . . . . . . . . . . . . . . . . . .
2.2.2 Geomagnetic pulsations . . . . . . . . . . . . .
2.2.3 Thunder storm activity . . . . . . . . . . . . . .
2.3 Noise Sources . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Introduction . . . . . . . . . . . . . . . . . . . .
2.3.2 Noises from power line signals . . . . . . . . . .
2.3.3 Noise from electric traction . . . . . . . . . . .
2.3.4 Noises from instrument & sensors . . . . . . . .
2.3.5 Noises from the other sources . . . . . . . . . .
2.4 Effect of active electrical noise in magnetotelluric data
xii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
2
2
3
4
5
8
8
9
9
10
12
15
15
15
15
16
17
.
.
.
.
.
.
.
.
.
.
.
.
18
19
19
19
21
23
25
25
25
26
27
28
29
xiii
3 Magnetotelluric time series analysis

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Random data analysis . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Properties of random data . . . . . . . . . . . . . . . . . . . .
3.2.1.1 Probability density function . . . . . . . . . . . . . .
3.2.1.2 Mean square and variance . . . . . . . . . . . . . . .
3.2.1.3 Median and average absolute deviation. . . . . . . .
3.2.1.4 Autocorrelation function . . . . . . . . . . . . . . . .
3.2.1.5 Power spectral density function . . . . . . . . . . . .
3.2.2 Joint signal properties . . . . . . . . . . . . . . . . . . . . . .
3.2.2.1 Cross correlation functions . . . . . . . . . . . . . . .
3.2.2.2 Cross-spectral density functions . . . . . . . . . . . .
3.2.2.3 Coherence . . . . . . . . . . . . . . . . . . . . . . . .
3.2.3 Computational aspects . . . . . . . . . . . . . . . . . . . . . .
3.2.3.1 Trend and bias removal . . . . . . . . . . . . . . . .
3.2.3.2 Power Spectral Density function . . . . . . . . . . .
3.2.3.3 Windowing . . . . . . . . . . . . . . . . . . . . . . .
3.2.3.4 Discrete Fourier Transform . . . . . . . . . . . . . .
3.2.3.5 Smoothing of spectra by band and section averaging
3.3 Magnetotelluric transfer functions . . . . . . . . . . . . . . . . . . . .
3.3.1 Least Square Solution . . . . . . . . . . . . . . . . . . . . . .
3.3.1.1 Multi input, multi output linear system . . . . . . .
3.3.1.2 Solution with noise free data . . . . . . . . . . . . .
3.3.1.3 Solution with noise in measurement . . . . . . . . . .
3.3.2 Concept of bias . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2.1 Predicted coherence . . . . . . . . . . . . . . . . . .
3.3.3 Coherent and incoherent noises . . . . . . . . . . . . . . . . .
3.3.4 Variance & Errors . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.5 Coherence and bias of transfer functions . . . . . . . . . . . .
3.3.6 Remote reference . . . . . . . . . . . . . . . . . . . . . . . . .
4 Signal and Noise Characteristics of MT data measured over
ern Granulite Terrain
4.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Geological objectives of MT investigations in the SGT . . . .
4.3 The Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Acquisition . . . . . . . . . . . . . . . . . . . . . . . .
4.3.2 Typical MT Time series . . . . . . . . . . . . . . . . .
4.3.3 Examples of MT data collected over South India . . . .
4.4 The EEJ effect on MT data . . . . . . . . . . . . . . . . . . .
4.5 Signal Activity during the Field Campaign . . . . . . . . . . .
4.6 Spatial character of coherence . . . . . . . . . . . . . . . . . .
4.7 Geographic relation of noise . . . . . . . . . . . . . . . . . . .
4.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
32
33
33
33
34
34
35
35
36
36
37
38
38
38
38
39
39
40
41
41
41
42
43
44
45
45
46
46
48
the South.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
49
50
50
52
52
53
55
59
61
62
64
64
xiv
5 The application of the artificial neural networks to magnetotelluric time

series analysis
5.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Signal and noise in the magnetotelluric time series . . . . . . . . . . . .
5.3 Visual Inspection (editing) of magnetotelluric time series data; why automation ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Magnetotelluric noise characterization . . . . . . . . . . . . . . . . . . . .
5.4.1 Patterns of signal & noise : . . . . . . . . . . . . . . . . . . . . .
5.4.2 Amplitude of signals . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.3 Correlation between simultaneously measured channels . . . . . .
5.5 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5.1 Why artificial neural network? . . . . . . . . . . . . . . . . . . . .
5.5.2 ANN theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6 Data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.1 Network engineering . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.2 Pattern Training . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.2.1 Data used: . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.2.2 Pre-processing . . . . . . . . . . . . . . . . . . . . . . .
5.6.2.3 FANN Training . . . . . . . . . . . . . . . . . . . . . . .
5.6.2.4 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . .
5.6.3 Inter channel training . . . . . . . . . . . . . . . . . . . . . . . .
5.6.3.1 The data . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.3.2 FANN training . . . . . . . . . . . . . . . . . . . . . . .
5.6.3.3 Relative significance of input . . . . . . . . . . . . . . .
5.7 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6 Estimation of Magnetotelluric Transfer Functions: Robust
Methods
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Robust estimation of MT transfer functions . . . . . . . . . .
6.2.1 Why robust methods? . . . . . . . . . . . . . . . . . .
6.2.2 Robust M estimators . . . . . . . . . . . . . . . . . . .
6.2.3 Choice of influence functions . . . . . . . . . . . . . . .
6.2.4 Implementation for MT . . . . . . . . . . . . . . . . .
6.2.4.1 Initial guess of transfer function . . . . . . . .
6.2.4.2 Jackknife estimate as initial guess . . . . . . .
6.2.4.3 Scale estimate . . . . . . . . . . . . . . . . . .
6.2.4.4 Robust transfer function estimation . . . . . .
6.2.4.5 Tukey weights . . . . . . . . . . . . . . . . . .
6.2.4.6 Computing the variance . . . . . . . . . . . .
6.2.4.7 Quantile Quantile plots . . . . . . . . . . . .
6.2.5 Application . . . . . . . . . . . . . . . . . . . . . . . .
6.2.5.1 Flowchart . . . . . . . . . . . . . . . . . . . .
6.2.5.2 Comparison of robust processing schemes . .
67
68
68
68
70
70
71
71
71
71
72
74
74
75
75
75
76
77
78
78
80
81
82
86
Statistical
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
88
89
90
90
91
93
95
96
96
100
100
101
101
101
102
102
104
xv
6.3
6.4
Robust Band averaging . . . . . . . . . . . .

6.3.1 Introduction . . . . . . . . . . . . . .
6.3.2 Effect of frequency band width . . .
6.3.3 Robust estimation of spectra . . . . .
6.3.4 Flow chart . . . . . . . . . . . . . . .
6.3.5 Validation . . . . . . . . . . . . . . .
Results and discussion . . . . . . . . . . . .
6.4.1 VP14 . . . . . . . . . . . . . . . . . .
6.4.2 TT08 . . . . . . . . . . . . . . . . . .
6.4.3 TT04 . . . . . . . . . . . . . . . . . .
6.4.4 OK18 . . . . . . . . . . . . . . . . .
6.4.5 JN12 . . . . . . . . . . . . . . . . . .
6.4.6 VP16 . . . . . . . . . . . . . . . . . .
6.4.7 OK16 . . . . . . . . . . . . . . . . .
6.4.8 VP12 . . . . . . . . . . . . . . . . . .
6.4.9 Comparison of results from the Vellar
6.4.10 Discussion . . . . . . . . . . . . . . .
7 Discussion and conclusions

7.1 Introduction . . . . . . . . . . . . . . .
7.2 Summary . . . . . . . . . . . . . . . .
7.3 Neural Network and Robust processing
7.4 Conclusions . . . . . . . . . . . . . . .
7.5 Suggestions for future work . . . . . .
Bibliography
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
Palani profile in SGT .
. . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
105
105
109
111
114
114
116
116
116
117
117
117
118
118
118
118
119
.
.
.
.
.
131
132
132
133
136
137
List of Figures
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
1.10
1.11
2.1
2.2
2.3
2.4
2.5
2.6
2.7
Quasi planar nature of electromagnteic wave front at great distance from

source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Behaviour of plane EM wave at the surface of a conducting Earth . . . .
Diagrammatic representation of 1D, 2D and 3D situations. . . . . . . .
Planar view. Off diagonal tensor elements become unequal when geology
has a preferred direction. On rotating the tensor to the strike, diagonal
elements get minimized. See text for discussion . . . . . . . . . . . . . .
Implanting of Electrical Field sensor . . . . . . . . . . . . . . . . . . . .
Installation of induction coil magnetometer . . . . . . . . . . . . . . . .
Simplified equivalent circuit for induction coil magnetometer . . . . . . .
Theoretical response function for MFS05 magnetometer (from Pulz and
Ritter [2001]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Block diagram depicting different signal processing steps in GMS05
(adapted from Metronix [1997]) . . . . . . . . . . . . . . . . . . . . . . .
GPU05 (main unit) and HDU05 (display unit) during data acquisition .
Typical lay out for MT sensors and data acquisition (after Metronix [1997])
Earths horizontal magnetic field spectrum (Modified after Macnae et al.
[1984]) The Black and gray bars indicate the frequency range of measurement bands for GMS05. (see 1.1) . . . . . . . . . . . . . . . . . . . . . .
Plot for MT time series recorded at site VP16 showing geomagnetic pulsations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Averaged magnetic amplitude spectra (Hx) at station VP16. Peaks related to geomagnetic pulsations are labeled.Thick (gray) line represents
smoothed spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Plot of MT time series showing impulse signals related to sferics recorded
at site C12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Electric Field (Ex) spectral amplitude depicting the Schumann resonance
Distribution of electromagnetic field strength of three phase power transmission line in the immediate vicinity of power lines (adopted from Szarka
[1988]). H total magnetic field, E horizontal electric field. . . . . . . . .
Amplitude spectra of electric field computed over different data length for
the site OK14. Thin black line - 1024, Thick gray line 256 and Gray dots
- 128. See text for discussion . . . . . . . . . . . . . . . . . . . . . . . .
xvi
3
3
6
7
10
11
11
12
14
14
16
20
22
22
23
24
26
27
xvii
2.8
Current circuit electric railway systems. I is the current in the overhead

power line I1 is the return current in the rails and I2 is the return current
in the Earth ( modified after Chaize and Lavergne [1970]). . . . . . . . .
2.9 Near and far field effect of a grounded dipole on MT measurements
(adapted from Zonge and Hughes [1991]). The two panels on left and
right describes the apparent resistivity and phase values for far field and
near field measurements . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.10 Effect of a near field source on MT data, as recorded at the station OK14.
See text for discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ordered Zxy plotted against number of occurrences observed at station C13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Crosscorrelogram of two magnetotelluric time series components. . . . . .
3.3 The Parzen spectral window over the target frequency plotted on background of a sample spectra . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Magnetotelluric linear system with two horizontal magnetic components
as inputs and horizontal electric components as outputs. Adapted from
Jones et al. [1989]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 MT Transfer function (Zxy ) computed using upward and downward biased
estimators compared with the coherence functions for station VP20. The
upper part of the diagram shows real and imaginary components of upward
and downward biased estimates. The variance of Zxy is shown as solid line.
The lower part shows the predicted (multiple) and ordinary coherence
functions. See the text for discussion. . . . . . . . . . . . . . . . . . . . .
28
29
30
3.1
4.1
4.2
4.3
4.4
4.5
4.6
MT stations superimposed over the geology of the measurement corridor

in South India (Geology adapted from GSI [1995]). See text for discussion
Typical MT time series measured in the SGT for two frequency ranges a)
the frequency range 256 Hz to 32kHz (band 1) and b) the period range 8
Hz to 256 Hz (band2) .See the text for discussion. . . . . . . . . . . . . .
Typical MT time series measured in the SGT for two frequency ranges c)
the frequency range 8Hz to 0.25 Hz (band 3) and d) the period range 4
sec to 128 sec.See the text for discussion. . . . . . . . . . . . . . . . . .
Plot of apparent resistivity, phase, predicted coherence and degree of freedom (DOF) vs frequency for four stations measured over South India.
Data are computed in their measured direction. See text for discussion .
Apparent resistivity vs frequency plot for 3 stations tht are near and away
from the dip equator with day and night curves superimposed (Rao et al.
[2002]). (a) OK3 - site 300 km north of dip equator, (b) IRA site 100 km
north of dip equator and (3) KAR - site near the dip equator. . . . . . .
Averaged Kp indices [Courtesy NGDC], K indices [HYB] during the MT
measurements are compared with telluric predicted coherence [4 sec 128
sec]. The mean coherence is drawn as a line. . . . . . . . . . . . . . . .
34
37
40
42
47
51
54
56
58
60
61
xviii
4.7
4.8
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
The band averaged telluric predicted coherence plotted as a contoured map

for all the measured MT sites. (a) Band1 (b) Band2 (c) Band 3 and (d)
Band4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Contoured map of telluric Noise/(Signal+Noise) ratio, superimposed on
the major geographical elements of the region. Note N/(S+N) ratio on
either sides of Cauvery river near Sankari. . . . . . . . . . . . . . . . . .
Common signal and noise patterns in long period MT time series. Samples
are collected from different sites. Note the change in amplitudes. (a) Signal
patterns; Samples of Ex and Hy shows the geomagnetic pulsations. Other
channels also show signals but at longer periods. (b) Noise patterns; Ey &
Hx shows different types of spikes. A step and its decay is shown in Hy .
Sample of random noise is shown in Hz . . . . . . . . . . . . . . . . . . .
Simple three layer feed forward neural network. The data is processed at
each neuron in the layers. Each neuron performs a summing of inputs multiplied with a weight parameter and outputs the data through its sigmoid
transfer function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data flow through the feed forward artificial neural network (FANN) based
editing scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stacked amplitude and phase spectra of the training database. The spectra
of noisy data (class 0.1) clearly different from the signals (class 0.9). . .
Results from the pattern training. (a) The SSE as a function of epoch.
The error reached the minimum floor after 400 epochs. Stability of convergence is demonstrated up to 1000 epochs. (b) Deviation between manually
classified and network predicted classes for 500 test time series segments.
The scattered points shows the major deviations. The correct picking
constitute 94 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Network output versus signal content. The network was simulated by
inputs with varying signal content. A narrow region of high variance exists
when signal content is between 65 . . . . . . . . . . . . . . . . . . . . . .
Pattern, amplitude ratios and correlation coefficients of 900 stacks which
form the database for inter channel training and testing . The thick line is
the running average over 10 points. The broken bar indicates the overall
stack quality - black good (0.9) and white bad (0.1). (a) Ex (squares) and
Ey (triangles) pattern quality predicted as a function of stack number.
(b) Ex - Hy (squares) and Ey Hx (triangles) amplitude ratio. (c) The
correlation coefficients of Ex to Hy (squares) and Ey to Hx (triangles). .
Results from inter channel training. (a) The SSE as a function of training
epoch. (b) Deviation between manually classified stack quality and network predicted for 250 stacks. 235 stacks were classified similar to manual
classification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Relative significance of various inputs to the networks, viz amplitude ratios (A1 and A2), correlation coefficients (C1 and C2) and five pattern
qualities (Ex , Ey , Hx , Hy andHz ). The error deviation against each input is
a measure of its significance to the Neural Network. . . . . . . . . . . .
63
65
69
72
75
76
77
78
79
81
82
xix
5.10 Comparison of MT apparent resistivity and phase computed from different

mode of editing of data from site G12 . Filled circles represent xy and
diamonds represent yx components. (a) Using all stacks available. (b) By
neural network editing. (c) By manual editing. . . . . . . . . . . . . . .
mode of editing of data from site VP12 . Filled circles represent xy and
mode of editing of data from site JN10 . Filled circles represent xy and
mode of editing of data from site TT08 . Filled circles represent xy and
5.14 Comparison of manual and neural signal picking for site TT8. The diamonds present the neural picking and crosses, the manual. . . . . . . . .
6.1
6.2
6.3
6.4
6.5
6.6
83
84
84
85
86
Examples where robust statistical methods are desirable: (a) A one dimensional distribution with heavy tails (b) A distribution in two dimensions
fitted to straight lines. Adapted from Flannery et al. [1992] . . . . . . . . 92
Schematic diagram showing loss, influence and weight functions for Least
Square (LS or L2) , Least Absolute (L1) Huber and Tukey estimators.
Values shown in y axis are arbitrary See text for discussion. Adapted from
Zhang [1996] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Responses of different influence functions to a set of residuals from MT
data processing. Station VP13 Shows the Ex residuals for 0.1875 Hz for
all the 105 stacks. (b) Shows the response of three influence functions to
the residuals. See text for discussion. . . . . . . . . . . . . . . . . . . . . 95
Comparison of Least Square and Jackknife estimation of MT transfer functions for station VP10 (a) Jackknife difference for the first iteration. (b)
Variance as a function of iteration number. (c) and (d) comparison of
and values from LS & JK processing . . . . . . . . . . . . . . . . . . . 98
Comparison of Least Square and Jackknife estimation of MT transfer functions for station VP13 (a) Jackknife difference for the first iteration. (b)
Variance as a function of iteration number. (c) and (d) comparison of
and values from LS & JK processing . . . . . . . . . . . . . . . . . . . 99
Comparison of Least Square and Robust processing of magnetotelluric data
station TT08 for frequency 0.0791Hz . Triangles represent LS processing,
and stars represent robust (RB) processing (a) and (b) time series of Ex
residuals. C) Quantile Quantile plot of Ex residuals d) MT apparent resistivity and phase values from LS and robust processing. See text for
discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
xx
6.7
6.8
6.9
6.10
6.11
6.12
6.13
6.14
6.15
Flow chart for robust processing scheme. The gray area represents the proposed initialization of the transfer functions using Jackknife. This routine,
concentrates on the section averaging of MT data. See text for discussion.
Data flow through the flow chart representing robust processing of MT
data. The gray area represents the proposed Jackknife initialization. (a)
Process A uses Jackknife (JK) as initial guess and (b) process B uses Least
Square (LS) as initial guess. See text for discussion. . . . . . . . . . . . .
Comparison of robust processing results using Least Square (LS) and Jackknife (JK) initialization for stations JN12, VP16, OK16 and VP12. Process A refers to robust processing with JK initialization, where as Process
B refers robust processing with LS initialization See legend for symbol
identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison of robust processing results using Least Square (LS) and Jackknife (JK) initialization for stations VP14,TT08,TT04 and OK18. Process
A refers to robust processing with JK initialization, where as Process B
refers robust processing with LS initialization. See legend for symbol identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Spectrum estimation in MT using band averaging. (a) to (d) windows in
frequency with different radii. Though the window radius seems to increase
as period increases, effective bandwidth remains the same, as spectrum in
long period contains fewer Fourier harmonics. (e) Sample Ex Ex spectrum
in the band 4, with target frequencies projected as dotted lines. It is
common to have 10 12 target frequencies per band. . . . . . . . . . . . .
Apparent resistivities as a function of frequency window length. Triangles
represent xy and circles yx . Error bars represents 95% confidence interval.
See text for discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . .
Concepts of robust band averaging. (a) Magnitude of cross spectrum between Hx and Hy for station VP13, around a target frequency 6 Hz. The
spectra are multiplied by a Parzen window. LS least square, RB Robust. (b) Quantile Quantile plot for real part of the same cross spectrum.
Inverted triangles unweighted, Circles robust weighted. See text for discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison of robust and least square band averaging for different frequencies and band (Parzen) radius for station VP13. Magnitude of Hx Hy
(nT 2 /Hz) cross spectra are plotted for all the cases. X-axis show the frequency bins. Solid line represents robust average and broken line represents
Least Square average.The left column represent the the band averaging for
different target frequencies, where as the right column represents the band
averaging with different radius length for the same target frequency. See
text for discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Flow chart representing robust band averaging. Gray area represents the
proposed processing routine (a) Shows the different steps in robustly estimating the spectral matrix from raw time series Schematic diagrams (b) to
(e) shows four processing schemes defined. SS single station, RR remote
reference, RB robust band averaging, LS Least square . . . . . . . . .
103
104
106
107
110
111
112
121
122
xxi
6.16 Results of single station (SS) and remote reference (RR) processing for
station VP10. Process C (SS) and E (RR) use least squares band averaging, whereas process D (SS) and F (RR) use robust band averaging. The
MT transfer functions were derived from a robust section averaging 6.2.4
from the spectra sets produced by Process C to F. See text for discussion.
6.17 Comparison of upward and downward biased estimates of apparent resistivities for JN12 for processes C and D (a) XY component and (b) YX
component. UP - up biased (E- reference) and DN down biased (Hreferences). Symbols are same for both plots. See text for discussion.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.18 Comparison of upward and downward biased estimates of apparent resistivities for VP16 for processes C and D (a) XY component and (b) YX
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.19 Comparison of upward and downward biased estimates of apparent resistivities for KG02 for processes C and D (a) XY component and (b) YX
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.20 Comparison of upward and downward biased estimates of apparent resistivities for OK18 for processes C and D (a) XY component and (b) YX
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.21 Comparison of telluric predicted coherency, apparent resistivity and phase
values from robust processing with and without robust band averaging for
stations VP14 and TT08. The explanation for symbols are given in legend.
stations TT04 and OK18. The explanation for symbols are given in legend.
stations JN12 and VP16. The explanation for symbols are given in legend.
stations OK16 and VP12. The explanation for symbols are given in legend.
6.25 Comparison of results from two different processing scheme for MT data:
Left panel shows the results from ProcMT and the right panel shows the
results from the robust processing and ANN editing. . . . . . . . . . . .
123
123
124
124
125
126
127
128
129
130
xxii
7.1
7.2
Comparison of robust processing (RB) with and without neural network

(NN) editing for Station JN12. (a) Apparent resistivity and phase values from both processing schemes. (b) Comparison of robust and neural
network weights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Comparison of robust processing (RB) with and without neural network
(NN) editing for Station OK18. (a) Apparent resistivity and phase values from both processing schemes. (b) Comparison of robust and neural
network weights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
List of Tables
1.1
Frequency bands for measurements, GMS05 (after Metronix [1997]). . . .
14
2.1
Classification of noise sources that affect MT measurements. . . . . . . .
25
4.1
The locations and measurement details for the MT stations
. . . . . . .
53
6.1
Few commonly used influence functions. Adapted from Zhang [1996]. . .
93
xxiii
Chapter 1
Magnetotellurics: Basic Theory,
Sensors and Field Procedures
1.1
Introduction
Magnetotelluric signals originate in Earths magnetosphere and atmosphere from different

phenomena. The constant bombardment of solar wind on Earths magnetosphere leads
to its deformation and which in turn disturb the terrestrial magnetic field. This time
varying phenomena constitutes MT signals, which have periods usually above 1 second.
The worldwide thunderstorm activity generates high frequency ( > 1 Hz) electromagnetic
signals, which propagate around the globe in the Earth-ionosphere waveguide and constitute the higher frequency part of magnetotelluric signals. These electromagnetic waves
reach Earths surface as quasi-homogeneous waves and a small part of it penetrates the
conductive Earth as quasi-planar waves. The induced electromagnetic response of Earth,
both amplitude and phase, depends on the subsurface electrical conductivity structure.
The relation between electric and magnetic fields at the surface of Earth forms frequency
domain transfer functions, which can be interpreted in terms of the subsurface structure.
Magnetotellurics is a method of electromagnetic exploration that uses natural electromagnetic waves as source field (Vozoff [1972]). In the following sections the basic ideas
of magnetotellurics viz, planar assumption for source field, properties of EM field in
conductive earth, estimation of magnetotelluric transfer functions/impedances and their
behavior over different type of media are discussed. Accurate and simultaneous measurement of the time varying electromagnetic fields is the preliminary requirement to
obtain MT transfer functions. A brief discussion about sensors, recording devices and
field procedures commonly adapted for MT data acquisition is given in 1.2 and 1.3.
1.1.1
Source fields
Magnetotelluric methods evolved out of the observation of similar variation in Earth

current and magnetic fields. Telluric methods were already in use as an exploration
tool. Tikhonov [1950] and Cagniard [1953] examined the relationship between horizontal
orthogonal electromagnetic components and developed formulae to estimate impedance
of subsurface from the simultaneous measurement of these electromagnetic field components. In essence, the formulae between the two components are valid only if the fields
do not have a lateral gradient over scale lengths that varies with frequency. For example
for an electromagnetic wave, with a frequency 103 Hz and in a media of resistivity 103
ohm.m, the scale length = 350 km (Wait [1954]). The scale length and uniformity of
harmonic source field (with frequencies < 1 Hz) in the Earths ionosphere and magnetosphere has been established by Dungey [1955]. Monitoring and study of atmospheric
electricity / lightning generation, propagation of sub-ionospheric waves provide adequate
characterization of audio-frequency variations in electromagnetic waves.
At a distance from the source, the electromagnetic wave front becomes locally planar,
in the sense the oscillation of the wave is only in a plane that is perpendicular to the
propagation direction. Such waves are called plane waves or plane polarized waves (Figure
1.1). Magnetotellurics envisages its inducing electromagnetic fields as plane polarized.
When such waves are incident on Earths surface, maximum energy is reflected. In the
context of MT, it was shown by Cagniard [1953] that the refracted wave propagates down
nearly vertically, due to the large contrast in the speed of electromagnetic waves in the
Figure 1.1: Quasi planar nature of electromagnteic wave front at great distance from
source
Figure 1.2: Behaviour of plane EM wave at the surface of a conducting Earth

atmosphere and in the Earth. (or the large contrast of conductivities between the two
media). It follows that the direction of propagation of electromagnetic fields within the
Earth does not depend on the angle at which it hit the Earths surface. Figure 1.2 depicts
this situation.
1.1.2
Maxwells equations
Once refracted the electromagnetic fields propagate through the Earth. The electromagnetic fields in isotropic and homogeneous media (of constant electric conductivity
[S/m]) of uniform electric permitivity [As/Vm] and magnetic permeability [Vs/Am]
are described by the Maxwells equations. Considering the fields with harmonic temporal
4
variation (eiwt ) these equations are;
E
H
.H
.E
=
=
=
=
iH
iE + E E
0
q/
=0
(1.1)
The electric current density j [A/m2 ] is proportional to the electric field according to
Ohms law;
j = E
(1.2)
where q [As/m3 ] is the volume density of charge, H the magnetic field [Vs/m2 ],
E [V/m] the electric field and =2f, the angular frequency. Both permittivity and
permeability in the earth are assumed to have approximately constant values.
In the Cartesian system of coordinates, the first two equations of system 1.1 can be
written as

i
j
k

iH = x y
x
Ex Ey Hx
and
(1.3)

i

j
k

E
= x y x

Hx Hy Hx
It follows that
Ex
z
= iHy
Hy
z
= Ex
(1.4)
By taking partial derivative of each equations and making appropriate substitutions,
2F
+ k2F = 0
(1.5)
2
z
Where F = E or H and with the assumption that in a homogeneous Earth is a
constant. The constant k describes the complex penetration depth 1/k [m] of the EM
field.
1.1.3
Skin depth.
In terms of the diffusion factor describing the penetration in depth of the fields, the
so called skin depth (() [m]) in a homogeneous earth is defined as
s
r
2
2
() =
=
(1.6)
|k 2 |
which represents the exponential decay of the EM field amplitude with depth. At
depth () the EM-field amplitude will drop by 1/e with respect to its value at the
surface. The penetration in depth of the EM field for a stratified Earth is defined as a
response function C() = Ex /iHy . In case of a homogeneous Earth C() = 1/k.
Case 1, varies along z.
For a 1D stratified Earth of N layers, the penetration in depth of the EM fields
measured at the surface is solved iteratively, with a recursive formula described by the
EM response function Ci (). This index refers to the EM-response measured at the
top of the layer I (Weaver [1994]);
Ci () =
1 ri e2ki ti
ki [1 + ri e2ki ti ]
(1.7)
where I = N-1,N-2,. . . .1, and

ri =
1 ki Ci+1 ()
1 + ki Ci+1 ()
(1.8)
and ti is the thickness of the layer i and k i = i (i ) 1 , the diffusion factor in the
layer and i is the resistivity of the layer i (See Figure 1.3). The bottom most layer has
the response function 1/kN .
Case 2, varies along z and either x or y.
In a 2D earth with strike along the horizontal x-axis and conductivity as a function
of z and y, the Maxwells equations are de-coupled into two polarization modes. The
decoupling is valid since the EM fields are treated as plane waves. In this context, the
so called TE- polarization mode refers to the tangential electrical field and the TMpolarization mode to the tangential magnetic field; both components are tangential with
respect to the strike (x axis) of the conductivity structure.
TE polarization : (Ex Hy );
Hz
Hy
y
z
Ex
z
Ex
y
= Ex
= iHy
= iHz
(1.9)
TM polarization: (Ey Hx );
Ey
z
Hx
z
Hx
y
z
E
+
y
1.1.4
= iHx
= Ey
= Ez
(1.10)
Impedance tensor:
The electrical impedance Z [mV/T] is the ratio between the electric and magnetic
field components, which can be represented as
E=Z H
(1.11)
Figure 1.3: Diagrammatic representation of 1D, 2D and 3D situations.

In a homogeneous media the ratio of the orthogonal components is
Z = i/k
(1.12)
However, earth cannot be approximated to perfect homogeneous media and it is usual

to assume structural inhomogeneity for subsurface. Figure 1.3 depicts the 1D, 2D and
3D inhomogeneity in conductivity within earth.
In a general 3D earth, the impedance (transfer function of Earth) is expressed in
Figure 1.4: Planar view. Off diagonal tensor elements become unequal when geology
has a preferred direction. On rotating the tensor to the strike, diagonal elements get
minimized. See text for discussion
matrix form in Cartesian coordinates,

Ex
Zxx Zxy
Hx
=
Ey
Zyx Zyy
Hy
(1.13)
Thus each tensor element is Z ij = E i /H j (i,j = x,y).

For a 1-D layered Earth, besides having diagonal elements of Z 0, the off diagonal
elements are related in the form,
Zxy = Zyx orZ1D = Zxy = Zyx .
(1.14)
In a 2D earth the diagonal elements of Z vanish (in the 2D-strike coordinate system).
And the off diagonal elements become unequal.

Ex
0 Zxy
Hx
=
(1.15)
Ey
Hy
Zyx 0
However, if the measurement axes are not in alignment with the geological strike, the
diagonal elements may not vanish. The simplest case one can imagine is a geological
fabric in a preferred direction, or a structural orientation such as a fold or fault system
(Figure 1.4).
In such cases the tensor can be rotated to an angle with the rotation matrix R to align
it with geological strike.

cos sin
T
Zm = RZR , where R =
(1.16)
sin cos
with positive .
Once rotated optimally, the tensor may be written as,

0 ZT E
Zm =
ZT M 0
(1.17)
Where ZT E (transverse electric) is the impedance tensor element parallel to the strike
and ZT M (transverse magnetic) is impedance element perpendicular to the strike direction.
1.1.5
Amplitude and phase of impedance:
The complex impedance (transfer function) Z is usually represented by its amplitude and
phase. The electrical resistivity (inverse of ) as a function of depth can be inferred
by the EM field of the corresponding penetration depths. Resistivity obtained from the
ratio of the measured electric and magnetic fields is called apparent resistivity, which
is frequency dependent. The apparent resistivity aij [Ohm.m] (i,j = x,y) is defined in
terms of the transfer function element by the form:
aij = jZ ij j 2 /
(1.18)
In case of a homogeneous Earth, the apparent resistivity reflects the true value of
the Earths resistivity. Apparent resistivity is what we sense at the surface; the true
resistivity of objects at some depth is masked by intervening material. The phase of
the transfer function element describes the phase shift between the electric and magnetic
field components.

im(Zij )
1
(1.19)
= tan
re(Zij )
In a homogeneous Earth the impedance phase is /4 (450 ). In a 1D layered Earth,
the phase increases over 450 when the EM response penetrates into higher conductivity
media. By the same convention, phase decays below 450 for the EM response penetrating
into a less conductive media.
1.2
Sensors and Field Procedures
MT parameters such as apparent resistivity and phase values are derived from simultaneously measured time varying components of electric and magnetic fields. Magnetotelluric
time series consist of five components of the Earths electromagnetic field i.e., two orthogonal components of horizontal electric fields (Ex , Ey ) and three orthogonal components of
magnetic field (Hx , Hy , Hz ). The measurements are usually made in wide bands of overlapping frequency ranges, with different sampling intervals. In the following section, an
overview of magnetotelluric field sensors, data acquisition equipment and field procedures
given.
1.2.1
Introduction
A variety of instruments are used to measure the time varying components of Earths
natural electromagnetic fields. Primary requirement of any measuring device is that the
electric and magnetic field measurements must cover a frequency range that is appropriate
to the exploration problem and with sensitivity and accuracy, which is required for the
computation of magnetotelluric transfer functions (Kaufman and Keller [1981]). The
strength and frequency characteristics of natural electromagnetic waves are statistically
known but cannot be predicted in advance. The field equipment should have the sufficient
dynamic range and sensitivity to adequately measure the fields, even though, its strength
may vary from hour to hour at a site. For getting long period (>10 sec) information,
recordings must be carried out for time ranging from hours to days. This necessitates
the recording equipment to be stable and resistant to adverse weather conditions. The
procedure of acquiring magnetotelluric data can be divided into two parts : 1) that of
sensors, which are capable of converting the electric and magnetic field variations to
voltages and 2) recording of these voltages in a way that is suitable for recovery later on.
The modern recording devices are even capable of in site data processing and preliminary
inversion of observed data. This facilitates redefining of acquisition plans in the field itself,
which may save lot of money and time. The following sections are devoted to 1) sensors
2) recording equipment and 3) field procedures.
1.2.2
Electric field sensors
At frequencies that are measured in magnetotelluric sounding, the electric field is detected by measuring the voltage drop between pairs of electrode that are in contact with
Earth. The electric field can be thought as the limit of voltage drop when electrode
spacing tends to zero. In this way, it may be advantageous to have short electrode separation if one wants to measure electric field correctly. But in practice there are two
drawbacks for this. The current density may change at contact of rocks with different
conductivity. A large separation of electrode will average out such small gradients to
obtain a regional picture. More over, there is a chance that the voltage measured over
small electrode spacing may be below the noise level ( 1V/Hz) of the sensor. Typical
separation between electrode contact ranges from 50 to 200 m. The electrodes that are
commonly used are non-polarizing electrodes. Such electrodes consist of a metal immersed in saturated solution of one of its salts. The advantage of using non-polarizing
electrodes is that the voltage difference between a pair of electrodes is relatively stable.
If metal electrodes such as copper or steel are used, relatively large potential difference
resulting from electrochemical reactions at the metal surface can be present and these can
vary with time in an unpredictable manner. Several types of electrodes are now in use
(Petiau and Dupis [1980]) the data analysed in the thesis was collected with Cd-CdCl 4
cell. One disadvantage of this cell is the high toxicity of CdCl 4 solution. The Electric
Field Probe (EPF05 from M/s Metronix) showed good stability even during long period
measurements (1-2 days).
The electrical field sensors are planted in small pits (0.2 m) at least few hours before
they are used. The pits need to be water saturated and covered to retard evaporation
10
Figure 1.5: Implanting of Electrical Field sensor

temperature changes. In some types of soil, a few milli-Amperes of current can quickly
cause a significant polarization of the ground around the electrodes, leading to long
exponential decay of this potential. Contact resistance of the telluric electrodes should
be as low as possible; a typical value measured between two electrodes spaced at 100 m
in the highly resistive Southern Granulite Terrain ( 4.2) was between 200 - 1,000 and
10,000Ohm.m. Field photograph in Figure 1.5 shows a typical implant of electrical field
sensor.
1.2.3
Magnetic Sensors
The problem of detecting magnetic field variations with the required accuracy is much
more difficult than that of detecting electric field variations (Kaufman and Keller [1981]).
The magnetic sensors are required to detect variation in magnetic field for the frequency
range 104 Hz to 104 Hz. The dynamic range of magnetic field varies from as low as few
tenths of a nano-Tesla for micro-pulsations to as high as few nTs for diurnal variations.
Also the magnetometers should ideally have same response over such a wide range of
frequency and amplitude as well as it should be robust to the severe fluctuations in
temperature. These requirements make designing of magnetometers difficult. Some of
the commonly used magnetometers are Super Conducting Quantum Interference Devices
(SQUID see Clarke et al. [1983], Fluxgate and induction coil (Karmann [1977]). A short
description is given on the induction coil magnetometers, which were used to collect the
data analysed in this thesis.
1.2.3.1 Induction coil magnetometers
At its simplest, the induction coil is a loop of wire which produces a voltage proportional to its area multiplied by the time derivative of B across the cross sectional area
of the loop. For the low field strength and low frequencies of interest in magnetotelluric
surveys, the induction coils are fabricated with a high permeability metal core about
which the coil is wound. A very large number of turns must be used to provide a measurable voltage output from a coil. For example, an induction coil with 50,000 turns of
11
Figure 1.6: Installation of induction coil magnetometer
Figure 1.7: Simplified equivalent circuit for induction coil magnetometer

1m2 area produces 31.41 nV at 1000 sec (Kaufman & Keller, 1981). A further problem
in the design of sensitive induction coil is that if a very fine wire gauge is used, the total
resistance is very high. On the other hand a heavier gauge wire with lower resistivity will
make the induction coil heavier, limiting its mobility. One such type of magnetometer
namely MFS05 (M/s Metronix) was used to measure magnetic field in the present study
(Field photograph given in Figure 1.6). In addition to having a large resistance, an induction coil will also have an appreciable inductance and capacitance between windings. A
simplified equivalent circuit is shown in the Figure 1.7. This circuit will have a resonant
frequency above which voltage output decreases with increasing frequency.
The frequency dependence of induction coil response is ideally suited to the measurement of natural field, where the 1/f dependence of natural field strength (Figure 2.1)
12
Figure 1.8: Theoretical response function for MFS05 magnetometer (from Pulz and Ritter
[2001])
compensate for the low output from the induction coil for the longer periods (Figure 1.8,
adapted from Pulz and Ritter [2001]). In higher frequencies frequency- independent magnetometer response is obtained by feeding back into coil, a magnetic field proportional to
the current. These techniques effectively extend the dynamic range of the magnetometers.
1.2.4
Recording systems
Historically, MT studies have been concerned with the determination of the electrical
resistivity of the Earths crust or upper mantle on a regional scale. For that purpose,
signals in the period range between 10 s and 10000 s have been recorded with data
sampling rates in the order of several seconds to minutes. Accordingly, the recording
times for long period MT (LMT) data have been in the order of weeks to months at a site.
The amount of data that could be collected dependents on the progress in data storage
technology. In the 1950s and 1960s, analogue data were recorded on photographic
film or paper chart recorders and digitised manually at a later stage (For eg. See Vozoff
[1972]). From the 1970s digital data were stored directly, However, on modified analogue
audiocassette recorders (Allsopp et al. [1973]). In the 1980s, with the arrival of digital
tape and floppy disks, a new era of digital recording systems began and the most advanced
MT systems today use rugged hard drives (Ritter et al. [1998]). Robustness, compactness
and low weight are desirable features for all geophysical instruments. Because of their long
deployment times, data loggers must have good long-term stability; they must operate
very reliably and have low power consumption. The invention of the microprocessor in the
1980s opened the field to the in field data processing and modelling (Clarke et al. [1983]).
This also facilitated the measurement of high frequency magnetotelluric data (audiomagnetotelluric, AMT). AMT (Hoover et al. [1976]) data are natural electromagnetic
variation signals, typically in the frequency range 20,000 Hz - 0.01 Hz. Because a large
volume of data is produced in a relatively short time, most AMT instruments are designed
as real-time systems. Short, non-continuous segments of time series data are processed
simultaneously, enabling on-line quality control of the stacked results. The corresponding
electromagnetic fields penetrate only the first hundred meters or first few kilometres of the
Earths crust, which are economically important (Egbert and Livelybrooks [1996], Garcia
13
and Jones [2002]). The signal strength in the high frequency range varies enormously over
the spectrum shown in Figure 2.1. Artificially generated signals from the main power
supplies or electric railways exceed the natural signals by several magnitudes in narrow
frequency ranges. To eliminate these narrow band noises, MT instruments are equipped
with special hardware filters (notches). In the frequency range of the dead band around
1 Hz, the natural signal activity is at a minimum. Hence, recording in narrow frequency
bands is necessary to ensure optimum dynamic range of the signals, while LMT (>10 sec)
data are usually recorded in one broad frequency band. The remote reference technique
(Gamble et al. [1979], also see 3.3.6) that brought major improvement in the MT
impedance estimates requires synchronized recording in two MT data acquisition devices.
Clarke et al. [1983] describes a radio telemetry system for simultaneous recording of data
from two stations. Hard wiring and synchronized recording using a crystal clock are the
other options. With the advent of global positioning system (GPS), it is now possible to
have both the measuring devices synchronized to a very high precision. The two sets of
MT equipment should follow a common recording timetable, so as to facilitate remote
reference processing. Very recently the concept of computer networking is being adopted
for MT measuring devices, with each MT unit is envisaged as a node in a network,
connected by cheap coaxial network cables/telephone cables.
In the present study, the data were collected using Geophysical Measuring System
(GMS05) manufactured by M/s Metronix GmBH, Germany. On standard, the GMS 05
is equipped for the 5 channel MT and AMT method. GMS05 consists of following components.
1. Geophysical Processing Unit GPU 05
2. Sensor Connection Box SDB 05
3. Display Unit HDU06
4. 3 Magnetic Field Sensors MFS 06
5. 4 Electric field components
The sensors used by this system is already been discussed in 1.2. The first two
components are described below.
The Geophysical Processing Unit(GPU) houses the major electronic circuitry for conditioning, digitising and storing of magnetotelluric data. It can receive up to 8 channels
of data and digitises it using 16-bit A/D converter, with a dynamic range of 132 dB.
Data are stored in hard drive of 1GB capacity. The intel 486-driven mother board works
on DOS platform and performs infield data processing and storage. This also allows the
operator to inspect the time series and computed MT parameters ( 3) online. Communication is provided through serial or parallel connectors with a maximum data transfer at
115 k B/s. External power supply required by the system (12 V DC) is usually provided
with standard car batteries (100 Ah or better). For remote reference applications a highly
stable precision real time clock (PRC) synchronizes the GMS05. The clocks synchronize
automatically as soon as the signals from the GPS satellites are received. The accuracy
of the PRC is better than 1012 seconds in long term as it runs synchronously with the
14
Figure 1.9: Block diagram depicting different signal processing steps in GMS05 (adapted
from Metronix [1997])
Figure 1.10: GPU05 (main unit) and HDU05 (display unit) during data acquisition
Caesium clocks on board of the satellites. Figure 1.9 shows a schematic diagram of the
digital processing steps in GMS05. Figure 1.10 shows a field photograph of GMS05, in
operation.
The GMS05s total frequency range from (4096 sec)1 to 8192 Hz is split into a total
of 5 bands. The factor between lowest and highest frequency for one band is always 32.
In the table 1.1, an overview of the channel-sampling rates is given.
The SDB 05 is used for interconnection of different sensors with the GPU 05. The
sensor connection box has three inputs for the magnetometers and inputs to connect the
Band
1
1
3
4
5
Lower freq. Hz Upper freq. Hz Sampling rate Hz

256 Hz
8,192 Hz
32,768 Hz
8 Hz
256 Hz
1024 Hz
1/4s
8 Hz
32 Hz
1/128s
1/4s
1 Hz
1/4096s
1/128s
1/32s
Table 1.1: Frequency bands for measurements, GMS05 (after Metronix [1997]).
15
electric field probes for Ex and Ey measurement. The small waterproof box also contains
the preamplifiers for the electric fields.
1.3
1.3.1
Field Procedures
Survey design
As any modern geophysical survey, magnetotelluric surveys are also planned before fieldwork starts. Though the survey must accommodate changes due certain unforseen circumstances, the overall plan remains the same. Two major decisions to be made before
start of survey are station spacing and minimum recording time. Minimum recording
time depends on the depth of interest and the geological condition. Forward modelling
of an assumed model can provide minimum safe recording time. For deep crustal studies,
minimum recording of 2 to 3 days are required over a resistive terrain. For basement
configuration in hydrocarbon exploration, a recording times range from 21 day to 2 days.
Minimum site spacing also depends on the conductivity structure and required lateral
resolution. Spacing of 5 10 km is required for a regional survey, where as a spacing of 1
km or less is required for geothermal exploration (for eg. Harinarayana et al. [2002]). In
the current survey, the stations were spaced 8-10 km ( 4.3.1) and more densely spaced
measurements were carried out over suspected geological contacts.
1.3.2
Site selection
Site selection is one of the major factors governing the data quality. A good site can
yield quality data within a short duration of record compared to a noisy site. Cultural
noises arising from various sources such as power lines, electric railways, irrigation pumps,
vehicular traffic, radar transmitters etc may act as non planar source ( 1.1.1) field and
corrupt the computed transfer function (see 2.4 for more details). An overview of noise
sources affecting MT measurements is given in 2.3. The scenario becomes complicated
in areas like south India, where such noises travel great distances as the upper crust
is highly resistive (Harinarayana et al. [2003]). It is a practice to avoid high-tension
power lines and electric rail lines by at least 2 km and vehicular traffic by 200 m. It is
advisable to avoid any natural feature that might affect the electric field such as abrupt
topographical relief, water bodies and rock outcrops. In most surveys carried out in
India, dry and ploughed fields provide an ideal site.
1.3.3
Sensor deployment
Proper deployment of sensors is also essential for good quality MT data. Once a site is
selected, a pattern for sensor alignment is planned by surveying. Most preferable is a cross
pattern aligned with geomagnetic directions. The directions are marked with the help of
a compass and sensors are laid in alignment with the geomagnetic directions. Geography
and topography some times prevent the array from being aligned to the geomagnetic axes
or the spread for electrodes being exactly equal. These factors are taken into consideration
16
Figure 1.11: Typical lay out for MT sensors and data acquisition (after Metronix [1997])
while computing system transfer functions. A typical set-up followed for data collection
is shown in Figure 1.11.
Electrode pits are prepared few hours ahead of data acquisition to allow for stabilization. The dipole wires and connecting cables are weighted/covered with soil so as to
minimize wind caused movement. According to Clarke et al. [1983], a 1 m length of wire
vibrating at 1 Hz with amplitude of only 1 mm generates about 0.5 V, a voltage that
is comparable to the electric field signals at that frequency. Magnetometers are leveled
using a spirit level and buried in pit (30 cm depth) with a cover of plastic sheet and soil
to protect it from wind, animals and temperature transients. Induction coil Hz , requires
a vertical hole of about 1 m depth. The data acquisition crew consisting of six people,
select and prepare the site during daytime. The recording usually starts once the sensors
are stabilized and connections are properly made. More crewmembers are needed under
difficult field conditions; for example if the area inaccessible by vehicles.
1.3.4
Calibration
The natural electromagnetic fields, measured with a suite of sensors and recording
devices (system), are modified by the properties of the system itself. A proper knowledge of the measurement systems response over the entire frequency range of interest is
essential in obtaining accurate magnetotelluric transfer functions. Any modern MT data
acquisition system employs a suite of analogue filters to precondition the signal prior
to digitisation. The sensors, particularly the magnetometers, also respond differently
17
in different frequency ranges. Even the cables that carry the signals have appreciable
responses at higher frequencies. All these factors make the MT systems response a
complex non-linear function. The GMS05 system may be calibrated in two ways: 1) A
theoretical calibration table could be computed with the knowledge of different data acquisition parameters, filter setting etc and 2) By sending a Pseudo Random Binary Signal
(PRBS basically white noise) to the magnetometers and electric field preamplifiers and
by obtaining the ratio between the calibration signal and the system output in frequency
domain. In the first mode fine effects caused by tolerances in the components are not
accomodated. The accuracy that can be achieved with the model function is about 10%
or better for band 1 and 2% for band 2 to 5, whereas in the second mode accuracy
better than +/- 0.5% in magnitude and +/- 1 in phase can be achieved. The procedures
given by Ellinghaus [1997] were used to calibrate the measured MT time series, for the
present study.
1.3.5
Data acquisition.
Prior to data recording, a thorough check of sensors and instruments is a good practice.
At each site parameters such as noise levels, amplifier gain and filter performance are
checked usually using short calibration runs ( 4.3.1). During data acquisition in South
India, the electric field amplitudes were so high in the frequency range 8 Hz to 256 Hz
(band 2) for few sites that the measurement were made with lowest gain settings. As
many modern digital MT equipment, GMS05 also allows the storage of time series data
in addition to the computed MT parameters. The storage of time series was crucial in
obtaining better estimates of magnetotelluric impedance, as will be shown in the following
chapters. The quantity of data to be collected at each site depends on the requirements of
the exploration program ( 1.3.1). However, time required in collecting the data depends
on the prevailing natural signal strength as well. Once the recording begins, data quality
is monitored continuously. Monitoring consists of
1)Manual Editing: Inspection of magnetic and electric time series for obvious outliers (See 5.2 for more details regarding noise patterns of time series). A continuous
disturbance in one or more channels and dead / muted traces may be due to equipment
malfunctioning (Figure 5.1).
2)Coherency: This is a measure of the correlation between the E and H fields and
predictability of the E fields from the H fields ( 3.3.5). Coherency less than about 0.8
generally indicates problems associated with local noise sources associated with instrumentation or cultural effects (for e.g. see Figure 6.19). It can also indicate localized
natural sources such as lightning that do not conform to the MT assumption of spatially
uniform source fields.
3)Smoothness: Well-behaved MT curves will show smooth variations between estimated magnetotelluric transfer functions. Sudden offsets, rapid changes in curvature
such as slope reversals across less than 0.5 log cycle of frequency and curve slopes greater
than 45-degrees on log-frequency versus log-apparent resistivity are physically implausible and indicate a near field source ( 2.4). It is desirable to have low scatter, moderate
curvature and well-joined frequency-band curve segments.
Chapter 2
Signal and Noise Sources for
magnetotellurics
18
19
2.1
Introduction
The definition for signal and noise vary from one geophysical method to another. In
the context of magnetotellurics, Stodt (1983) defined signals as those components of
measured electric and magnetic fields, which are deterministically related through the
transfer functions of a multi-input multi-output linear system ( 3.3.1.1). Noise in MT
may be then defined as the additive components in the measured fields that are not
related in such a way. This definition carries over to the frequency domain as well as
the Fourier transform is a linear operator (Stodt [1983]). From a different point of view
Madden [1964] termed all unpredictable part of MT data as noise. One consequence of
this definition is that, deterministic noises such as 50 Hz (power line) harmonics will not
be treated as noise, even though they will not qualify by Stodt [1983] criteria for signal.
However, any unpredictable changes in amplitude and /or phase of a deterministic noise
will be treated as noise. A third way of defining signal and noise in MT would be checking
the source field morphology. However, if the signal and noise components have overlaps in
time, frequency or spatial domains, their separation is im-perfect. MT signal processing
algorithms try to differentiate the signal and noise parts by transforming the measured
data to a suitable domain and or with the help of additional independent measurements.
The derivations made in 1.1.4 for MT transfer functions assumes a planar source field.
Any process (natural or man-made) that makes the source field inhomogeneous may thus
be treated as noise source and its effect in MT records as noise. One way to better
understand the signal and noise in MT data is to study their sources. Many signal and
noise sources have its own characteristics, which may be used to separate their effects
on measured MT data. The first section of this chapter deals with the signal sources for
MT. A discussion on various noise sources is given in the second part.
2.2
2.2.1
Signal sources
Introduction
Magnetotelluric method uses naturally occurring electromagnetic field as source signal.

The advantage of using natural field is to have unlimited power available throughout the
frequency range of interest ( 104 to 104 [Hz]). This is particularly important, in the
low frequencies (< 0.01 [Hz]), where a large power source and equipment set up would be
needed otherwise to generate signals. This ensures the probing range of MT to 100s of
kilometers. But to depend on natural source means spending sufficient time in the field
to measure enough data. This is necessary as the signal level of natural electromagnetic
waves is unpredictable and interfered with man-made and other noises ( 2.3) especially
near the frequency range 101 to 1 [Hz].
Natural electromagnetic waves come from a variety of processes and from sources
ranging from Earths core to distant galaxies (Vozoff [1991]). Within the frequency range
of interest to magnetotellurics, only two sources are important. These are the atmosphere
and magnetosphere. World wide lightning activities in the lower atmosphere are causes
for MT signals from 1 Hz to 104 Hz, whereas below 1 Hz the fields originate mainly in
the magnetosphere due to its interaction with solar wind and energy exchange between
20
Figure 2.1: Earths horizontal magnetic field spectrum (Modified after Macnae et al.
[1984]) The Black and gray bars indicate the frequency range of measurement bands for
GMS05. (see 1.1)
particles and waves. Figure 2.1illustrates the components in the natural electromagnetic
spectrum (modified after Macnae et al. [1984]) the amplitude and shape shown in the
plot may vary with location and time. The bars indicate the measurement bands used
for the study (see also table 1.1).
As can be seen from the Figure2.1, the amplitude of natural magnetic fields spectra
varies by almost 5 decades within the frequency range 102 to 105 Hz. The relatively
higher amplitude in the lower frequency ranges are due to geomagnetic pulsations. The
signals from world wide thunder storm activity (sferics) is the main source of natural
signal for frequency > 1 Hz. These two phenomena do not overlap in frequency and there
exists dead band around 1 Hz, with very low signal amplitude. We may define another
dead band around 1KHz as well. However, in magnetotelluric context, the effect of first
dead band has been recognized widely and in the present thesis, the term dead band
refers to the low energy frequency band around 1 Hz. The sharp peaks in the spectra
between 10 to 1000 Hz are due to power line interference. This clearly shows how narrow
band noise can override the natural signals. In the following sections, an overview of the
two sources for magnetotelluric signals are given.
21
2.2.2
Geomagnetic pulsations
Geomagnetic pulsations are temporal variations in the Earths magnetic field that have
a quasi-periodic structure with frequencies ranging from 103 [Hz] to 2 [Hz] (Kaufman
and Keller [1981]). The magnetosphere is the region around the Earth in which the main
magnetic field is confined by solar wind. Solar wind consists of charged particles, mostly
hydrogen and helium nuclei, ejected from sun and it take four days to reach Earth. Earths
magnetosphere contains ionised oxygen and nitrogen, which make the magnetosphere
conductive. The conductive part (100 to 250 km) of magnetosphere is called ionosphere
and the resistive lower part is called atmosphere. This very complex system of Earths
magnetosphere is constantly bombarded by solar wind. The magnetosphere responds to
magnetic pressure of solar wind plasma by generating waves. The EM hydrodynamic
waves thus set up travels towards the Earth. To reach Earths surface these waves
must cross the ionosphere and resistive atmosphere. Ionosphere, with its anisotropy does
not allow the vertical components of E and H . This not only modifies the horizontal
components of the EM field, but set up horizontal currents in the ionosphere. It is believed
that geomagnetic pulsation observed on surface of the Earth is mainly due to this current
system in ionosphere. The time behavior of magnetic pulsations is episodic but includes
features localized in frequency / space as a result of local conditions (Vozoff [1991]).
Geomagnetic pulsations are divided into two classes; continuous (Pc) and irregular (Pi).
The continuous pulsations are subdivided into six classes (2 0.001 [Hz]). They are Pc1
(2 0.2 [Hz]), Pc2 (0.2 0.1[Hz]), Pc3 (0.1 to 0.022 [Hz] ), Pc4 (0.022 0.0066 [Hz]),
Pc5 (0.0066 0.00166 [Hz]) and Pc6 with frequency less than 0.00166 [Hz]. The second
class of pulsations, Pi, which have an irregular from, is divided into three types: Pi-1 (1
0.025 [Hz]), Pi-2 (0.025 0.0066 [Hz]) and Pi-3 with frequency less than 0.0066 [Hz]
(Kaufman and Keller [1981]). As discussed earlier, due to the unpredictable occurrence
character of the geomagnetic pulsation, long time of measurement is required to obtain
a rich spectrum of the signals in full bandwidth.
To obtain signals up to 0.001 [Hz] it is usual to record for more than one day at a site.
A suite of statistical methods is then employed to obtain smooth spectra in the frequency
range of geomagnetic pulsations. Such an example from MT data collected at site VP16
is presented here. The data was collected in period range 4sec to 128 sec, with a sampling
rate 1 Hz. All the 5 MT components were measured for approximately 34 hours (25 - 27
Feb.2000). Signal activity was strong in the range of Pc3 to Pc5 during last half of the
recording session. Figure 2.2 shows a record for 4096 seconds of the measurement. The
transient sinusoids, apparent in the records are geomagnetic pulsation and can easily be
identified (see 5 ). The data was subjected to spectral analysis ( 3.2.3). The total
record was divided into 30 segments of 4096 samples each. Each of the segments were
Fourier transformed and an average spectral amplitude for Hx is given in Figure 2.3. As
a whole, the spectral power decreases monotonically with increasing frequency, much in
agreement with the natural EM spectra as given in Figure 2.1. Superimposed on the
main trends, at least three lobes associated with Pc3, Pc4 and Pc5 are prominent.
22
Figure 2.2: Plot for MT time series recorded at site VP16 showing geomagnetic pulsations
Figure 2.3: Averaged magnetic amplitude spectra (Hx) at station VP16. Peaks related
to geomagnetic pulsations are labeled.Thick (gray) line represents smoothed spectra
23
Figure 2.4: Plot of MT time series showing impulse signals related to sferics recorded at
site C12.
2.2.3
Thunder storm activity
Magnetotelluric signals with frequency above 1 Hz mainly originate from worldwide thunderstorms. About 100 to 1000 lightning storms happen at any given moment worldwide.
The fields as seen by the MT system depend on the strengths, path lengths (cloud heights)
occurring frequencies etc (Vozoff [1991]). These signals, which are called Atmospherics
or sferics, die off with distance. Most of these storms are located in the tropical region.
Each lightning produces a current flow in the atmosphere with a peak intensity of 300
[Am]. The lightning excites the electric field in the resistive atmosphere that is sandwiched between relatively conductive ionosphere and Earth. While propagating from
the lighting source, electromagnetic waves get reflected at lower and upper boundaries
due to the high resistivity contrasts. They effectively travel round the glob 7.75 times a
second (A circumference of 38,400 [km] with a speed of 297600 [km/s]). The atmosphere
acts as a resonator and filters the lightning spike to multiples of resonant frequencies.
These resonance frequencies are called Schumann resonance. In 1952 W.O. Schumann
predicted (Meloy, scientific report available at http://space.tin.it/scienza/rromero) the
existence of electromagnetic resonance in the Earth - Ionosphere cavity as,
1
fn = 7.49 [n(n + 1)] 2
f or n = 1, 2, 3...
(2.1)
The equation yields f1 = 10.6 Hz, f2 = 18.4 Hz and so on. However, the first observational evidence of resonance came after the studies of Balser and Wagner [1960], when
they measured the resonance of electromagnetic pulse generated from nuclear explosions.
They found the resonance at f1 = 7.8 Hz, f2 = 14.2 Hz etc. The difference in values
24
Figure 2.5: Electric Field (Ex) spectral amplitude depicting the Schumann resonance
between prediction and observation was attributed to the ionospheric loses. Sufficiently
far from the lighting source, the electromagnetic fields behave like a plane wave - a requirement for MT (see 1.1.1). A plot of MT time series measured at a sampling rate
1024 Hz (Band 2) is given in Figure 2.4. An impulsive source and its decay are evident between data samples 100 150. When signal activity is strong, the observed MT
data contains superposition of individual sferics originating from different thunderstorm,
throughout the world. Schumann resonance frequencies become prominent in such cases
and an example is given in Figure 2.5. The amplitude spectra plot for Ex component
of electric field clearly shows the resonance frequencies related to thunder storm activity
especially at frequencies 8 Hz and 16 Hz.
As discussed in 2.2.1, the typical natural electromagnetic signals with frequencies
around 1 kHz have very low signal strength (Figure 2.1) and are below the background
noise from the measuring instruments. After analyzing the seasonal and diurnal behavior
of a large set of high frequency MT data collected from Canada and northern Germany,
Garcia and Jones [2002] conclude that, to better estimate high frequency MT transfer
functions, the measurements should be made in the nighttime. During daytime, the
atmospheric conductivity is larger, compared to night, as direct sun light ionizes the
atmosphere. The increased conductivity results in strong attenuation of sferics. The
highest observable signal occurs when the whole propagation path from the lightning to
the site is unlit.
25
Active
Electric
power
transmission,
Electrified
Rail
lines,
Electric
switching,
Factories,Vehicle
movement.
Passive
Conductors:
power
lines,
Pipelines, Fences,
Large iron structures
Resistors:
Roads,
ditches,
culverts
Natural
Magnetic storms,
Lightning,
Microseisms, Wind,
Equatorial electrojet.
Instrument
Failure in electronics,
Sensor
malfunctioning,
Dropped bits, sensor misalignment,
muted traces.
Table 2.1: Classification of noise sources that affect MT measurements.
2.3
2.3.1
Noise Sources
Introduction
Compared to signal sources, the number of mechanisms that produce noise in the measured magnetotelluric data are numerous. This make it difficult to review the various
noise sources for MT. Further the classifications of the noise sources are different in
different research documents. An attempt is made here to classify the noise sources for
magnetotellurics, within the limit of the scope of the thesis. Thus the terms geologic and
terrain noises are not described in detail. Ward [1967] divided noise in measured electromagnetic data into instrumental, terrain & disturbance field. McCraken and Hohman
[1986] classified EM noises into geologic and electromagnetic. While reviewing man
made electromagnetic noises in geophysics, Szarka [1988] classified them into passive,
active and other effects. Hatting [1989] classified them into mechanical equipment and
electromagnetic. A compilation of the noise sources is given in the table 2.1.
The various sources listed in the table above describes the major noises that affect
electromagnetic methods and magnetotelluric in particular. The list is by no means
comprehensive. For a detailed review on noise sources see Szarka [1988], Junge [1996].
2.3.2
Noises from power line signals
Man made noise, in the EM spectrum (Figure 2.1) comes mainly from the electric power.
Figure 2.6 shows the distribution of electromagnetic field strength of three phase power
transmission line in the immediate vicinity of power lines (Szarka [1988]). The interference
from power line exponentially decreases with distance from it. Several authors studied
EM harmonics in the vicinity of 50 Hz power lines (Szarka [1988]). This interference is
inductive in nature and will not harm MT data, if the site is sufficiently far. However,
there are other types of interference that causes a current flow in the Earth. An ideal and
perfectly balanced AC power transmission system does not cause any active EM noise in
the earth. In the case of unbalanced networks (ie in practice) a current component appears
having equal amplitudes and directions in each conductor. This is the so-called zerosequence current (Szarka [1988]) which usually flows through soil. The noisy interference
from this current may travel a great distance and affect MT data, depending on the
resistivity of the soil. Motor loads operate non-synchronously and can also produce
26
Figure 2.6: Distribution of electromagnetic field strength of three phase power transmission line in the immediate vicinity of power lines (adopted from Szarka [1988]). H total
magnetic field, E horizontal electric field.
side bands and sub-harmonics of the main frequencies. Added to this are the problems
arising from finite observations. This results in spectral leakage and can corrupt the
transfer function estimate in the vicinity of power line harmonics as well. Results of
such an analysis is shown in Figure 2.7. The data were acquired at OK14, in band 2
(sampling rate 1024) and was affected by noise from nearby power lines. Power spectrum
of electric field computed from three sets of data length viz. 128, 256 and 1024 are plotted.
The distinctive peaks of the spectrum are related to main power line frequency and its
harmonics. Leakage of power from main lobe is visible for spectrum calculated from
128 and 256 data samples compared to the spectrum computed from 1024 data samples.
This shows how the narrow band noise can leak into their vicinity. The transfer function
estimations near to 50 Hz and its harmonic can get affected, if they are computed from
a low-resolution spectrum.
2.3.3
Noise from electric traction
Electrified railways are another noise source for magnetotellurics. The Indian Railways
adopted 25,000 V as standard for its electric traction in 1957. The electric traction is
powered by AC power lines, which hang above the tracks. The current loop is completed
through the electric motors which are in turn connected to the wheels and thus to the
rail tracks. Rail tracks are grounded at regular intervals for safety reasons. A part of the
current will travel through the ground to the electric substation. In this way, rail related
noises have two components. While the induction effect from the power lines can be felt
at distances less than few hundred meters, the electric impulse through ground can travel
a great distance. In the context of South India, where the upper crust is highly resistive
these noises can carry great distances from the source. Chaize and Lavergne [1970] give
27
Figure 2.7: Amplitude spectra of electric field computed over different data length for
the site OK14. Thin black line - 1024, Thick gray line 256 and Gray dots - 128. See text
for discussion
a brief description of source mechanism for noise interference. As described in the Figure
2.8 current I 2 having an impulse like character due to the power entering the Earth
where the rail track is grounded and it is this erroneous current that causes distant EM
disturbances. Power for underground railways is usually supplied at the voltage of 1000
v and its loading may result in an impulse with an amplitude 7000 A (Szarka [1988]).
2.3.4
Noises from instrument & sensors
Noises from failure in electronics in equipment, quantization errors at A/D converter

(Hatting [1989]), malfunctioning of sensors and errors in aligning the sensors are common
in magnetotelluric data. In addition to this the sensors and the circuitry has its own
background noise, which may vary with time as the electronic component develop fatigue.
One way to identify such a problem is to calibrate the system ( 1.3.5). Another way is to
set up electric and magnetic sensors in parallel (Pedersen [1988]). The sensors should be
placed sufficiently far away from each other that they do not interact but sufficiently close
that the signals are same. Then by averaging over a number of frequency components we
may define the coherence 2 between two sensor outputs. A corresponding expression for
the signal to noise ratio is obtained as S/N = 2 /(1- 2 ). A very low coherence indicates
malfunctioning of one of the sensors/circuitry. Disorientation of sensors can also induce
noise to the transfer function estimation. Pedersen [1988] showed that a 50 deviation in
magnetic sensor alignment would correspond to a skew of 0.1.
28
Figure 2.8: Current circuit electric railway systems. I is the current in the overhead
power line I1 is the return current in the rails and I2 is the return current in the Earth (
modified after Chaize and Lavergne [1970]).
2.3.5
Noises from the other sources
Switching of submersible pumps causes disturbances in measured fields and is observed in

farm areas and villages. Electric channels are more prone to this type of noise. Movements
of ferrous metals or other magnetic material in the vicinity of the magnetic field sensor can
introduce noise into the magnetic channels. Vehicular traffic generates both magnetic and
seismic noise. In most cases, the magnetic effects are negligible, when the magnetometers
are more than 200 m from the road (Clarke et al. [1983]). Seismic noise transforms into
magnetic noise through the movement of magnetometer in the Earths field. Clarke et al.
[1983] reports that for the worst case, when the sensor is aligned perpendicular to earths
main magnetic field, a rotation of 0.0020 produced by seismic vibration produces a field
change of 1 nT.
Passive noise sources of EM measurement generally mean superficial resistivity inhomogeneities of man made origin. Conductive constructions (Pipelines, metal fences
etc.) may cause a redistribution of extreme natural electromagnetic phenomena such as
magnetic storms, lightning etc. Passive distortion effects of man made construction can
surely be avoided by choosing the MT site away from them ( 1.3.2). But this may not
be possible always, especially in industrial areas (as discussed in 4). The largest and
most common source of natural noise is wind, which can either move the magnetometer
directly or induce seismic noise by blowing on trees or bushes whose roots then move the
ground. Thus ideally the magnetometers should be in a flat area and buried away from
trees. The cable connecting the sensors to the data acquisition system should be secured
to ground, to protect it from moving in wind. A 1 meter length of wire vibrating at 1 z
with an amplitude of only 1 mm generates about 0.5 V, a voltage that is comparable
to the telluric signal at that frequency.
29
Figure 2.9: Near and far field effect of a grounded dipole on MT measurements (adapted
from Zonge and Hughes [1991]). The two panels on left and right describes the apparent
resistivity and phase values for far field and near field measurements
2.4
Effect of active electrical noise in magnetotelluric

data
Most of the active noise sources discussed above manifest in MT measurements as correlated and un-correlated noise that can be attributed to the near field of a grounded
dipole (Oettinger et al. [2001]). Zonge and Hughes [1991] describe the electromagnetic
field grounded in a homogeneous half space. Figure 2.9 gives a diagrammatic sketch of
such a setup. Consider the near field and far field case. The near field case applies when
the dipole measurement site distance is far smaller than the skin depth of the EM wave
(rA << ), with = (2/), where is the frequency in radian, is the magnetic
permeability of free space and is the conductivity of the half space. Here the electric
field decays at the rate of 1/r3 and magnetic field decays as 1/r2 , where r is the separation between dipole source and measurement site. The observed apparent resistivity at
this site is then proportional to the period T. This produces an artificial 450 rise in the
apparent resistivity curve. The phase between electric and magnetic field will be zero
for all the periods. The change of the ratio E/B in any lateral distance at a frequency
varies as 1/r, due to the different dependencies of the electric and magnetic field with
respect to distance. Electric field E varies with respect to resistivity and is independent
of frequency, whereas magnetic field H varies independent of resistivity and frequency
(Zonge and Hughes [1991]). With this it becomes clear that in near field situation, the
noise from dipole source dominate the electric field compared to the magnetic field. This
corroborates with the observations that, near industrial regions the electric fields are
more corrupted with noise than magnetic fields.
30
Figure 2.10: Effect of a near field source on MT data, as recorded at the station OK14.
See text for discussion
One such example from the data collected at station OK14 is presented here. The
station is located 80 km south of Erode in the measurement corridor (Figure 4.1). The
data was collected in a visibly excellent site, but was found to be affected noise from
unknown source. Time series collected in band 2 (sample rate 1024 Hz) were sub segmented into stacks of 1024 data samples and were processed using least square technique
(Sims et al. [1971]). The magnetotelluric transfer functions, especially the xy component
showed the near source effect. The apparent resistivity and phase are plotted in Figure
2.10 follows the exact description of Zonge and Hughes [1991] of near source effect. The
error bars computed have dimension less than the symbols in the plot. In the far field
(rC >> ) the electric and magnetic field decay as 1/r3 and the apparent resistivity and
phase depends on frequency and resistivity (Figure 2.9). The ratio E/H is independent
of the dipole site distance r. The phase difference between E and H is 450 and apparent
resistivity equals the resistivity of the homogeneous half space.
The effect of near source noise in MT is one such example of how noise affects the
computed MT parameters. In the presence of noise, it becomes necessary to apply statistical signal processing tools to minimize its effect on computed MT parameters. Many
of the standard spectrum analysis techniques on random data have been applied to MT
for this purpose. In the next chapter, an overview of various signal processing methods
that are commonly employed in magnetotelluric time series analysis is given.
Chapter 3
Magnetotelluric time series analysis
31
32
3.1
Introduction
Various natural and anthropogenic processes contributing to magnetotelluric signal observed at Earths surface were described in the previous chapter. Assuming a quasiuniform source field, the linear relation between measured electric and magnetic data may
yield usable information about the distribution of electrical conductivity of earth. Unfortunately not all electromagnetic signals that are measured at Earths surface comprise
the induction process or include the information we seek. Data reduction and processing
techniques are necessary to convert the measured time series to an interpretable form.
Generally interpretation of magnetotelluric data is done in frequency domain. Therefore, the first step in interpreting magnetotelluric data involves evaluating time series
to identify and reject variations that seem to be of non-inductive sources. The next is
to estimate 101 102 frequency domain complex transfer functions Z( ) from the raw
electric and magnetic field time series E(t) and H(t) with approximately 106 real numbers
/ site (Egbert and Livelybrooks [1996]). MT data processing is straight forward when
there is no noise present in the measurements and the equations relating inducing and
induced variations through the impedance tensor ( 2) are directly applicable. When
noise is present a large number of processing methods have been proposed (Sims et al.
[1971], Goubau et al. [1978], Gamble et al. [1979], Stodt [1983]); Formerly MT time series analysis was done as an extension of classical random data analysis (Bendat and
Piersol [1971]). Much of the same data reduction and transformation techniques are still
followed. However, the last two decades have seen considerable improvement in the procedural and computational aspects of magnetotelluric method (Park and Chave [1984],
Egbert and Booker [1986], Chave and Thomson [1989], Larsen [1989], Sutarno and Vozoff
[1991], Egbert [1997], Banks [1998], Ritter et al. [1998], Oettinger et al. [2001], Smirnov
[2003], Chave and Thomson [2003], Manoj and Nagarajan [2003]).
In this chapter an introduction to various signal processing steps that are used in
MT data analysis are described. Natural signals like magnetotelluric fields are usually
treated as random process and statistical methods are applied to derive their properties.
Basic concepts of random data and its properties are discussed in 3.2.1. In the context
of magnetotellurics it is also desirable to describe certain joint properties of data from
two or more random processes, namely components of horizontal electric and magnetic
fields. The concept of joint probability distribution functions, cross correlation functions
and cross-spectral density functions are described in 3.2.2. Computational aspects of
these functions from real data are described in 3.2.3. Perhaps the dual input dual
output linear system with additive noise in all components is the best way to describe the
relation between magnetotelluric fields. The classical least square method of solving for
the magnetotelluric transfer functions, its properties and associated errors are described
towards the end of the chapter ( 3.3). Violations to the assumption of noise free input
signals may bias the LS estimation. A brief discussion on certain disadvantages of LS
estimators is also included in 3.3.
33
3.2
Random data analysis
When considered individually, both signal and noise processes in magnetotelluric field
components can be considered as independent random processes. All deterministic parts
are noise and can easily be removed. Random process may be categorized as stationary
and non stationary. Statistical properties of stationary random processes do not vary
with time, where as these properties will change with time for non-stationary processes.
3.2.1
Properties of random data
The main types of statistical functions used to represent a random process are (1) probability distribution function (2) mean square value and median (3) auto-correlation function and (4) power spectral density function.
3.2.1.1
Probability density function
Probability density function or pdf describes the probability that the data will assume
certain amplitude within some defined range at any instant of sampling. The probability
that sample r(n) assumes a value between r and r+r may be obtained by taking the
ratio N r /N where N r is the total number of occurrences that r(n) had in the range (r,
r+r) from N independent field observations. In equation form,
lim Nr
(3.1)
N N
The probability that the current sample r(n) is less than or equal to some value r is
defined by P(r ), which is equal to the integral of the probability density function from
minus infinity to r. A typical plot of probability density versus instantaneous values for
magnetotelluric transfer function is presented in Figure 3.1 to demonstrate the property.
The data recorded in band 3 (table 1.1) were sub-segmented to 174 blocks of 512 data
samples each. Transfer functions were estimated for each segments using least square
technique (equation 3.37) , with minimum 5 adjacent Fourier coefficients used for each
estimate. The derived estimates were sorted into 10 bins of ranges. The main peak (near
27.5 mV/(km.nT)) is flanked both sides by near symmetric decay.
The bell shaped probability density plots indicated in the Figure 3.1 are typical of
either narrow or wide band random processes. These probability density plots would
ideally be of the classical Gaussian form as given by the equation for variable x,
Pr ob[r < r(n) r] =
p(x) =
ex /2x2
x 2
(3.2)
Where is the standard deviation of x ( 3.2.1.2). Gaussian or normal density function

is most common in natural processes as is evident from MT measurements. Central limit
theorem postulates that when a large number of processes contribute to single random
process, its pdf will tend to be a Gaussian, irrespective of probability density functions
of individual process (Menke [1984]).
34
Figure 3.1: Ordered Zxy plotted against number of occurrences observed at station
C13
3.2.1.2
Mean square and variance
The mean square value is simply the average of squared values in the time series. Consider
the horizontal magnetic field hx (t), the mean square value 2 of the time series is given
as,
lim 1
=
T T
2
ZT
h2x (t)dt
(3.3)
Many random processes have a static time invariant component and a dynamic or
fluctuating component. The static component mean () is the simple average of all the
values and the fluctuating component variance is the mean square value about mean.
Variance is given as,
lim 1
2 =
T T
ZT
[hx (t) ]2 dt
(3.4)
The positive square root of the variance is called standard deviation.

3.2.1.3
Median and average absolute deviation.
If a set of N observations are sorted into the ascending order, h(1)h(2) h(3)
. . . .h(N), where h(j) is called j th order statistic. Median x is the middle sample
35
is T is odd. The sample median is ambiguous for N even but it is typically chosen as
(h[N/2] +h[N/2+1] )/2. Another class for the estimation of dynamic fluctuation is average
absolute deviation and is given by
ZT
lim 1
=
T T
|h(t) x|dt
(3.5)
Mean and median are various types of averages. It is less commonly realized that
such averages are the result of minimizing various norms. The mean is obtained by
minimizing the L2 (or LS) norm or least squares norm ( 3.3) of the samples. Whereas
the median is obtained by minimizing the L1 norm of samples (Chave et al. [1987]).
3.2.1.4
Autocorrelation function
The autocorrelation functions of random data describes the dependence of the values of
data at one instance on the values at another time. In equation form,
ZT
lim 1
Rx ( ) =
N N
h(n)h(n + )dn
(3.6)
n=0
where is the time lag. Autocorrelograms for a random process like magnetotelluric
will be sharply peaked at lag = 0 and rapidly diminishes to zero on either side of the
correlogram. In the limiting case of hypothetical white noise, the autocorrelogram is a
dirac delta function at zero lag (Bendat and Piersol [1971]).
3.2.1.5
Power spectral density function
Frequency analysis of random data essentially involves the implementation of the Fourier
transform. If F( ) and f(t) are frequency and time domain expression of a random data
its relation given by Fourier transform are,
1
f (t) =
2
F ()eit d
(3.7)
and
Z
F () =
f (t)eit dt.
(3.8)
The transform defined in the equations 3.7 & 3.8 are only valid for functions with
finite energy and a different treatment is necessary for data which exist for all the time so
that the total energy (f2 (t) , t = -. . . .) is not finite. In such cases one must consider
the spectral properties of the energy density rather than the spectral properties of the
amplitude fluctuations. Power spectral density function of a random data describes the
frequency composition of the data in terms of the spectral density of its mean square
36
value. Its concept is well understood by considering a narrow band analog filter. The
mean squared out put of the filter will converge on average over a long period of time.
In equation form,
lim 1
2x (f, f ) =
T T
ZT
h2 (t, f, f )dt
(3.9)
where T is the length of the record, H(t)(Bendat and Piersol [1971]). For small f a
power spectral density function S xx (f ) can be defined such that,
2x (f, f ) Sxx(f )f
(3.10)
According to so-called Wiener-Khinchin theorem (Ghil et al. [2002]) the power spectral
density is equal to the Fourier transform of the autocorrelation function. Hence power
spectrum of H(t) can be estimated by the Fourier transform Shh (f) of an estimate of
autocorrelation function R x ( ).
Z
Sx (f ) = 2
Rx ( )ej2f d
(3.11)
The random part of the error in the estimation of S xx is given by the normalized
error,
=
p
2/dof,
(3.12)
where, dof is the numbers of degree of freedom. As dof = 2 for each Fourier harmonics,
>1 implying that the standard deviation of the estimate S xx is greater than the estimate
itself. There for the need of spectral smoothing is seen. Spectral smoothing is discussed
while addressing the computational aspects of power spectral density in 3.2.3.
3.2.2
Joint signal properties
In many applications such as magnetotellurics, it is desirable to study the joint signal

properties of two or more random processes. Cross-spectral density between Earths
natural electric and magnetic field gives the relation between the fields in amplitude and
phase, which in turn is useful to understand the subsurface conductivity distribution (
1.1). The joint statistical properties like joint probability distribution, cross correlation,
cross-spectral density etc are essentially the extensions of 3.2.1.
3.2.2.1
Cross correlation functions
Cross correlation function of two sets of random data describes the general dependence
of the values of one set of data on the other (Bendat and Piersol [1971]). Consider two
37
Figure 3.2: Crosscorrelogram of two magnetotelluric time series components.

random series H(t) and E(t), their cross correlation is given as
lim 1
Rhy ( ) =
T T
ZT
h(t)y(t + )dt
(3.13)
The function Rhy ( ) is always a real valued function which is symmetrical about the
ordinate when h and y are interchanged. That is,
Rhy ( ) = Ryh ( )
(3.14)
A Typical plot of the cross correlation versus time lag plot for two random processes
(Ex and Hy) is given in Figure 3.2. Note the sharp peak at lag = 0 second and few other
less defined peaks which shows the existence of correlation between two series Ex(t) and
Hy(t) at specific displacements. The value of cross correlogram at lag=0 may be used
to measure the similarity between two processes. For example, in 5, correlation of
orthogonal electric and magnetic field components were used to as a quality parameter
of that section of time series.
3.2.2.2
Cross-spectral density functions
The cross spectral density (csd ) function of two time series can be defined as the Fourier
transform of their cross correlation function, in the same manner we defined power spectral density of a random process. As the cross correlation is an odd function, csd is a
complex valued function unlike psd.
38
3.2.2.3
Coherence
When considering physical measurement of two random processes, it is often desirable

to compute a real valued function 2 (f ), called ordinary coherence function, which is a
measure of linear relation between the processes as a function of frequency and frequency
domain equivalent of cross correlation. It may defined from the psd and csd of the
processes as,

2
Shy
(f )
2

(3.15)
hy (f ) =
Shh (f )Syy (f )
The coherence function satisfies the limit 0 <= 2 <=1, with 2 =0 indicate the processes are statistically independent and a value of 1 indicate they are fully coherent at
that frequency. Its importance in magnetotellurics is discussed in 3.3. Note the terminology used for coherence throughout this thesis. Any frequency domain measure of
correlation, which is a modulus-squared quantity, is termed coherence and its plural is
coherences. The positive square root of coherence is termed coherency and coherencies respectively.
3.2.3
Computational aspects
Before considering the processing of random data, magnetotellurics in particular, it is

worthwhile reviewing some basic steps for data preparation. In 1.2 the MT field and
data acquisition procedures were discussed. With the assumption that digitization was
performed on time series which was pre-filtered to suppress the frequencies higher than
the Nyquist frequency (1/2t [Hz], where t is the sampling interval in seconds) the
following steps for digital computer processing of random data are discussed.
3.2.3.1
Trend and bias removal
A special correction may be needed to remove trend and bias in the observed data, defined
as any frequency component whose period is larger than the ensemble size. If trends are
not removed large distortions can occur in the after processing of correlation and spectral
quantities (Bendat and Piersol [1971]). So-called average slope method is realized in the
current work to remove bias and trend from measured MT time series, in equation form,
T /2
T
T
X
X
X
t
1
h(t)
h(t)
h(t)
(3.16)
h1 (t) = h(t)
T t=1
T /2
t=1
t=1+T /2
Where the time series H(t) is measured for a period of time T.

3.2.3.2
Power Spectral Density function
For stationary or quasi-stationary (Bendat and Piersol [1971]) time series the most often
applied digital method for computing power spectral density functions via direct Fourier
Transform or Cooley-Tukey method based on finite range fast Fourier transform. It is
followed in the present study and is described here.
39
3.2.3.3
Windowing
Finite observation of an infinite process can be viewed as multiplying the infinite process
with a boxcar function. Boxcar function has a value one within the observation period and
zero outside the observation period. Then the estimated power spectral density function
Sxx (f ) is the convolution of the true power spectral density with the frequency domain
transform of the boxcar function. Thus convolution causes spectral energy to leak from
the original frequency to adjacent frequencies by adding infinite number of smaller side
lobes. Near power line frequency and its harmonics as observed in magnetotelluric time
series, the leakage may totally corrupt the adjacent frequencies as discussed in 2.3.2
(see Figure 2.7). Many alternate window functions were introduced to obtain smooth
spectra (Bendat and Piersol [1971]). A simple Hanning window can be realized as,

i
(3.17)
W (i) = 0.5 1 + cos
N
Where I = 0,1,2,3. . . ..N samples of time series.
As the window function effectively suppresses the time series data at both ends,
windowing results in loss of information. In order to retrieve the data that are lost due to
windowing, it is a common practice to overlap, adjacent windows. In the magnetotelluric
time series analysis carried out in this thesis Hanning window with 50 % overlap was
used. However, the overlapping will decrease the effective degrees of freedom (dof ) and
corrections are given as,

1 1
dof = dof
+
(3.18)
2
3.2.3.4
Discrete Fourier Transform
Discrete Fourier Transform (DFT) is obtained by frequency domain sampling of the

continuous Fourier Transform presented in equations 3.7 and 3.8. DFT realized on N
discrete samples of data H(t) is given by,
T 1
kn
1X
W (t)h(t)e2i T
FDF T (fk ) =
T t=0
(3.19)
With a spectral resolution f=f s /N. Here k = 1,2,. . . T-1 and f s, is the sampling
frequency. W(t) is the windowing function given in equation 3.17. This summarization
does not change the units the raw DFT should be divided by the sampling frequency,
F () FDF T (k )/fs
(3.20)
to obtain the same result as in equation 3.8. Thus H(t) measured in the unit V will
get the V/Hz after the transformation. To remove the effects of windowing, the DFT
has to be multiplied with inverse of the integral of the windowing function, here Hanning
windows with an integral value 2. Power spectral density S xx (f k ) of time series H(t) is
40
Figure 3.3: The Parzen spectral window over the target frequency plotted on background
of a sample spectra
thus computed by,
Sxx (fk ) =
2
|FDF T (fk )|2
T fs
(3.21)
with the units V2 /Hz. With the same reason, the amplitude spectrum get the unit
V/ Hz.
3.2.3.5
Smoothing of spectra by band and section averaging
The estimate S xx (f ) itself is a random function and from a loose definition of central limit
theorem, it can be said that the S xx (f ) will have a Gaussian distribution even though
the time series may have different distribution. It is thus usual to attribute 2 degrees
of freedom to each spectral density values (except for first (mean) and last (nyquist)
harmonics, which have only 1 dof each). From equation 3.12 it is seen that the power
spectrum estimate is inconsistent, as the estimate is less than its standard deviation. In
order to reduce bias and variance in estimated power spectrum, a number of techniques
are proposed (Jenkins and Watts [1968], Ghil et al. [2002]). A combination of frequency
band smoothing and segment averaging is generally followed in magnetotellurics to reduce
variance in psd estimations (Sims et al. [1971], Chave et al. [1987]). As the MT transfer
functions are slowly varying functions of frequency, the frequency band smoothing is
performed over a window in frequency domain, here the Parzen (Jenkins and Watts
[1968]) window,
1
if
fz f = 0
sin(u) 4
W (f ) =
(3.22)
( u ) if 0 < fz f < fr
0
if
fz f fr
where f z is the target frequency and f r is the radius of the Parzen window. A
schematic diagram is presented in Figure 3.3 showing band averaging over Parzen window.
41
Jenkins and Watts [1968] discuss the use of other types of spectral windows to reduce
the variance in spectral estimation. They have shown that given a particular radius M,
Parzen window achieves the smallest variance among other types of windows.
However, frequency smoothing reduces spectral resolution. A trade off between frequency resolution and estimated variance dictates the choice of Parzen radius. Another
problem with band averaging is the bias error. Parzen window like other spectral windows, may result in biased estimates of cross and auto spectrum and can affect the
computed MT transfer functions. As will be shown in 6.3, robust weighting of spectra
within a band can vastly improve the estimates of auto and cross spectra and thereby
yield less biased MT transfer functions. Section averaging, which is defined as averaging
of spectra computed from the sub- segments of time series, may be applied to further
smooth the spectra. Also called Bartletts smoothing procedure, this will also result in
decrease in variances associated with the estimate of spectra (Jenkins and Watts [1968]).
Cross-spectral densities are also computed in the same fashion as for psd as described
above. Frequency band averaging assumes the individual spectral values have a Gaussian
distribution ( 3.2.1.1). Whereas the problem with non-Gaussian distribution in section
averaging is well recognized, the same for band averaging has been overlooked. It will be
shown in 6.3 that a better estimate of cross and auto spectral densities can be obtained
by robust band averaging.
3.3
Magnetotelluric transfer functions
3.3.1
Least Square Solution
3.3.1.1
Multi input, multi output linear system
Magnetotelluric data can be considered as a result of a system of random process, with

the signal components deterministically related between the inputs and outputs. Under
the assumption of a plane source field, concepts appropriate for multiple-input / multiple
out put linear system can be applied (Jones et al. [1989]). The estimation of weighting
response functions, or their frequency domain equivalent transfer functions for such a system is described in this section. Conventionally, the horizontal components of magnetic
field Hx(t) and Hy(y) [nT] are treated as the inputs, and the horizontal components of
electric field Ex(t) and Ey(t) [mV/km] are treated as outputs (Figure 3.4. The inputs
and outputs are related by a convolution operation to the four time domain weighting
functions Z( ). The magnetotelluric data can be interpreted in terms of its weighting response functions Z( ) (Kunetz [1972], Yee et al. [1988]) or its frequency domain
representation, Z( ) (see Vozoff [1991]). Many authors questioned the time domain approach due to unstable statistical properties of the weighting response functions (Jenkins
and Watts [1968], Egbert [1992]). With some exceptions the transfer functions are now
computed in the frequency domain.
42
Figure 3.4: Magnetotelluric linear system with two horizontal magnetic components as
inputs and horizontal electric components as outputs. Adapted from Jones et al. [1989].
3.3.1.2
Solution with noise free data
In frequency domain the complex frequency dependent relation between the horizontal
MT fields can be written as given by Swift [1986],

Ex
Zxx Zxy
Hx
=
(3.23)
Ey
Zyx Zyy
Hy
Where Z is the MT transfer function tensor (impedance tensor) [mV.km1 .nT1 ],
1
1
E [mV.km1 .Hz 2 ] and H [nT.Hz 2 ] are Fourier transforms ( of electric and magnetic
fields E(t) and H(t) respectively at a particular frequency in radian. The relationship
between vertical (Hz ) and horizontal magnetic field can also be expressed in the same
way as,

Hx
Hz = Tx Ty
,
(3.24)
Hy
Where T x and T y are called tipper functions (Vozoff [1991]). As the solution for
tipper functions are similar to that of impedance functions, they are not separately dealt
with in the following sections. When the measurements are noise-free, two independent
observations of field components E and H are enough to estimate the transfer functions
(Sims et al. [1971]). For example a solution for one tensor element is obtained as,

Hx1 Ex1

Hx2 Ex2

Zxy =
(3.25)

Hx1 Hy1
Hx2 Hy2
where the subscript 1 and 2 indicate data from two measurements. An additional
43
requirement is that Hx1 Hy2 6= Hx2 Hy1 , which physically mean the two field measurements
must have different source polarization (Kaufman and Keller [1981]).
3.3.1.3
Solution with noise in measurement
In the measured magnetotelluric data, we have no knowledge of the true signal components in Ex, Ey, Hx, Hy or Hz and it is desirable to make more than two independent
observations for each components. With more than two measurements equation 3.23 is
over determined and a solution for transfer functions is sought which minimizes the errors
in observations. An equivalent matrix form replaces a component of the tensor equation
3.22 as,
E = ZH + r
(3.26)
where there are N observations so that E and r are N vectors, H is an N X 2 matrix

and Z is a two vector. The last variable in equation (3.26), r is the difference between
measured and predicted output field (here, Electric) and is the residual parameter to be
minimized. Classical way of solving MT equation is by the least squares technique (Swift
[1986], Sims et al. [1971]) where in the square of the residual power in equation (3.26) is
minimized to yield a solution for Z. Writing one component of equation (3.26),
Exi = Zxx Hxi + Zxy Hyi + ri
(3.27)
Squaring this equation and summing for all the N records sets we obtain,
2
r =
N
X
(Exi Zxx Hxi Zxy Hyi ).(Exi

Zxx
Hxi
Zxy
Hyi
)
(3.28)
i=1
(Asterisk indicates complex conjugate). Setting the derivative of r2 with respect to

Zxx and Zxy in turn to zero yield two independent equations,
< Ex Hy > = Zxx < Hx Hy > + Zxy < Hy Hy >
(3.29)
< Ex Hy > = Zxx < Hx Hy > + Zxy < Hy Hy >
(3.30)
If we minimize noise in magnetic field in equation (3.27) two more equations arise.
They are,
< Ex Ex > = Zxx < Hx Ex > +Zxy < Hy Ex >
< Ex Ey > = Zxx < Hx Ey > +Zxy < Hy Ey >
(3.31)
where the quantities like H x H x and H y H y are power spectral densities (or auto
spectra) and quantities E x H x and H x H y are cross-spectral densities (or cross spectra
- 3.2.2.2) between different field components. The bars indicate segment averaging
and braces (<>) indicate an average over a small frequency band ( 3.2.3.5). From now
onwards the use of these symbols are omitted for simplicity. Any of the two equations
(3.28 through 3.29) must be solved simultaneously for Z xx and Z xy . Thus there are six
44
possible estimates for each of the tensor elements. For example, estimates of Z xy are
obtained as,
Hx Ex Ex Ey Hx Ey Ex Ex
Hx Ex Hy Ey Hx Ey Hy Ex
(3.32)
Hx Ex Ex Hx Hx Hx Ex Ex
Zxy =
Hx Ex Hy Hx Hx Hx Hy Ex
(3.33)
Zxy =
Hx Ex Ex Hy Hx Hy Ex Ex
Hx Ex Hy Hy Hx Hy Hy Ex
(3.34)
Zxy =
Hx Ey Ex Hx Hx Hx Ex Ey
Hx Ey Hy Hx Hx Hx Hy Ey
(3.35)
Zxy =
Hx Ey Ex Hy Hx Hy Ex Ey
Hx Ey Hy Hy Hx Hy Hy Ey
(3.36)
Zxy =
Hx Hx Ex Hy Hx Hy Ex Hx
Hx Hx Hy Hy Hx Hy Hy Hx
(3.37)
Zxy =
Where the hat corresponds to an estimate. Equation (3.32) is the solution for Zxy
in the equations 3.31, which assumes that all noise is in magnetic fields. Two of the
equations (3.34 and 3.35) are relatively unstable in the 1-D case where the fields are
unpolarized. Any noise in the electrical field will cause the estimate for Z xy to bias
upward. Equation 3.37 is the solution to set of equations 3.29 & 3.30, which treat the
electrical field as output, (as in Figure 3.1). Noises in magnetic field components will
down bias this estimate. However, MT transfer functions are rarely estimated using
equation (3.32), as i) The magnetic measurements are made with better accuracy than
electric field and ii) Electric fields can get polarized due to electrical resistivity anisotropy
and denominator of equation 3.32 can become very small or zero.
3.3.2
Concept of bias
A crucial factor in the Least Square method described above is the assumption of noise
free data in either the electric or magnetic data. In the equations for the estimations
Z xy (eq 3.30), this assumption is implicit in use of auto spectra. Under the assumption
that noise is uncorrelated with signal, AS B N =0, where A, B {E x , E y , H x , H y }, AS
is the signal in component A and AN is the noise component in A, one can write,
AB = (A + A )(B
+B

)=
AS B S
when A 6= B
S S
N N
A B +A B
when A = B
(3.38)
While estimates for the cross spectra are statistically distributed and yield a good
approximation to the true value, the estimates of the auto spectra are systematically too
large (M
uller [2000]). Thus noise in E x E x cause Z xy in equation (3.32) to bias upwards,
where as noise in H x H x and H y H y in the denominator of equation 3.37 biases Z xy
45
downwards. The downward and upward biased estimates give an envelope within which
true transfer functions should lie (Jones et al. [1989]). To reduce the bias effect Sims et al.
[1971] proposed to take averages over four stable estimates, the argument being that the
negative and positive bias would cancel each other. Kao and Rankin [1977] devised an
iterative approach to remove the biasing effects. Goubau et al. [1978] and Gamble et al.
[1979] used magnetic data from a remote station ( 3.3.6) in order to avoid the use of
auto powers in solutions such as equations 3.32 to 3.37.
3.3.2.1
Predicted coherence
The coherence between measured and predicted output data describes the optimality of
LS estimate such as given in equations 3.32 to 3.37. For example, squared predicted
coherence, when consider Ex as output according to equation (3.37) is given by,

p 2
|
|E
E

x
x
2
Ep Exp =
(3.39)
p p
(Ex Ex )(Ex Ex )
where,
Exp = Zxx Hx + Zx yHy
(3.40)
(modified after Swift [1986]). Predicted coherence (or multiple coherence) functions
will strongly indicate the presence or absence of linear relationship between input and
output. This should be differentiated from the ordinary coherency defined in equation
3.15 ( 3.2.2.3). A value of 1.0 indicates the derived impedances are in perfect line with
the observations of E and H. Presence of uncorrelated noise in the fields reduces the
predicted coherence from unity. The three conditions under which the coherences can
have non-unity values (Jones [1981]) are, when,:
1. noise is present in both electric and magnetic fields
2. the system relating the magnetic and electric field are non linear
3. processes other than electromagnetic induction are involved
Many approaches were made to weight MT spectra sets from different observations,
according to predicted coherence (Stodt [1983], Egbert and Livelybrooks [1996], Larsen
et al. [1996]). However, as pointed out by Dekker and Hastie [1981], the multiple coherence functions can get biased upwards when computed from fewer number of observations
and/or in the presence of correlated noise between measured fields. Correction to bias
error in coherence is discussed in 4.6.
3.3.3
Coherent and incoherent noises
Noises in MT data can be classified into coherent and in-coherent. Noises in electric or
magnetic channels (or both) that cannot be related via transfer functions (out of line)
are termed incoherent. An example is the thermal noises in sensors. Predictive coherence
46
may be used to detect them. Coherent noises are noises that affect both the input (say
magnetic) and out put (say electrical) channels in line with the linear model (impedance).
An example is a correlated spike. It may be detected by examining the time variations
of the impedance functions (6.2).
A rigid coherency gate may not be useful in dead band as the signal-to-noise ratio is
least here. Coherency Weighted Estimate (Stodt [1983]) or Coherency sorting will give
better results. However as shown by Egbert and Livelybrooks [1996], a combination of
these schemes with robust processing may significantly improves the results (6.2).
With no coherent noises, coherency gate is arguably the best choice. However, coherence cannot always be relied for the discrimination, when the measured data contains
coherent noises (Banks [1998]). Windows that are contaminated with coherent noises
can generate transfer functions significantly different from the normal ones. A realistic
approach would be a combination of the coherence and robust schemes ( Egbert and
Livelybrooks [1996]). As shown in 6, with a proper initialization, robust processing can
remove the noisy data (see 6.2.4.2).
3.3.4
Variance & Errors
Estimates of error in computed transfer function are important, as they are required at
the modeling or inversion stage to assess the adequacy of fit of a derived model to the
measured data (Eisel and Egbert [2001]) Statistical variance for each component of the
transfer function tensor (equation 3.1) may be as,
2
]Ex Ex
F (k, 2N 4, = 0.05)[1 ExExp
(Zxy ) = k
2
]Hy Hy
(2N 4)
[1 HxHy
2
(3.41)
Where = 0.05 is the significance level corresponding to a 0.95 or 95 per cent confidence limit, N is the number of Fourier coefficients used, 2N-4 is the degree of freedom,
F(m,n,) is the factor for F distribution (Bendat and Piersol [1971]) and k is 4 for confidence limit of |Zxy |2 (Schmucker [1978]). The errors for apparent resistivity and phase are
xy =
xy =
2
|Zxy |2 Zxy and
Zxy
|Zxy |
(3.42)
The estimate of variance in equation (3.41) is exact under the assumption that the
noise in output channel (r in equation 3.25) are normally distributed and statistically
independent while at the same time the inputs are error free. Chave and Thomson [1989]
describe many instances wherein these assumptions can break down resulting in biased
estimation of confidence limit.
3.3.5
Coherence and bias of transfer functions
Figure 3.5 shows MT transfer function (Z xy ) computed using equations 3.32 (up biased)
and 3.37 ( down biased) respectively, compared to their associated coherence and variance
for station VP20 (see section 4.3.3). The upward and downward biased estimates
47
Figure 3.5: MT Transfer function (Zxy ) computed using upward and downward biased
estimators compared with the coherence functions for station VP20. The upper part
of the diagram shows real and imaginary components of upward and downward biased
estimates. The variance of Zxy is shown as solid line. The lower part shows the predicted
(multiple) and ordinary coherence functions. See the text for discussion.
behave identically over most of the period range except near 8 seconds and periods >
200 sec. In these ranges, the two different estimates deviate from each other. The up
biased estimates show considerable bias from the common trend. Though the apparent
magnitude of bias seems to be less, the down biased estimates show reduced values in
these period ranges. On inspection on the coherence plots, it can be seen that these
period- ranges are associated with low predicted coherence. The ordinary coherence
function Ex Hy decreases with increase in period whereas coherence Ex Hx behaves in
exactly opposite way. The increased dependency of Ex on Hx in the longer period might
be caused by the polarization of electric fields by geological structures at depth. However,
the predicted coherence function is undisturbed by the polarization. This is due to the
fact that predicted coherence takes into consideration, the transfer function between
collinear fields as well (in this case Z xx ).
The iterative scheme proposed by Kao and Rankin [1977] averages the up and down
biased estimates to obtain bias free transfer functions. But this assumes that both electrical and magnetic channels are equal in noise content. Any deviation from this (as
practically seen, electric channels are more noisy than magnetic channels Travassos and
Beamish [1988]) will results in inferior estimate than that of conventional ones. This was
agreed by the authors while replying to a question by Hernandez and Jacobs [1979].
48
3.3.6
Remote reference
In 3.3.6, we have seen the bias error in estimated transfer functions due to auto-power
terms in equations such as 3.32 and 3.37. One effective way to reduce bias errors has been
the remote reference (RR, it is usual to denote single station processing as SS) method
(Gamble et al. [1979]), in which two independent signal channels are recorded for use as
the complex conjugate part of cross and auto powers in equation such as 3.37.
Zxy =
Hx Rx Ex Ry Hx Ry Ex Rx
Hx Rx Hy Ry Hx Ry Hy Rx
(3.43)
Equation 3.43 gives the RR estimate of Z xy . By providing a kind of synchronous

detection, the method helps compensate for noise, both internal and external to the measuring device (Vozoff [1991]). The remote channels usually designated R x and R y are
the most commonly H x and H y components measured at a remote, noise free site. In
principle the electric field also can be used as a remote reference. However, as shown by
Travassos and Beamish [1988], this may not result in better transfer function estimate, as
the electric fields are more affected (polarized) by local geology and the spatial coherence
between electric field components are often small. Combined with robust ( 6.2) estimations, this technique is widely used at present. Jones et al. [1989] proved its superiority
in a comparison of different processing techniques on MT time series. In a recent work
by Shalivahan and Bhattacharya [2002] the question of how remote can the far remote
reference site be? is addressed. They processed MT data from one permanent site with
remote fields collected at distances 80, 115 and 215 kilometers away from it. Only by
using the farthest station data, they could improve quality of MT data in all frequency
ranges. However, the data used for the thesis were collected in single station mode,
though more than one systems were used in the field. Instrument problems prevented
the synchronization data acquisition systems internal clock with that of GPS (Global
Positioning System) satellite. For this reason, the various approaches made in the thesis
are for single station processing, though it can easily be extended to dual station processing. In some cases there is scope to improve remote reference processing as shown in
the 6.3.
Chapter 4
Signal and Noise Characteristics of
MT data measured over the
Southern Granulite Terrain
49
50
4.1
Introduction.
Magnetotelluric studies were carried out as part of an integrated geophysical study of

the Southern Granulite Terrain (SGT) during 1998-2000. The region south of the 13o
N latitude marking the high-grade granulite zone was traversed by a N-S corridor of
geophysical observations (Figure 4.1). Of these, MT studies were conducted along 2 N-S
profiles extending from Kuppam, near Bangalore in the north to Palani and Kodaikanal
in the south. The Magnetotelluric measurements carried out over the Southern Granulite Terrain (SGT) in two field campaigns during September 1998 December 1998 and
January 2000 - March 2000 form the basic data used in the thesis. These MT studies
form part of a major geophysical and geological study on the SGT initiated by the Deep
Continental Studies programme of the Department of Science and Technology. There are
three factors that make this region important in terms of MT signal analysis. 1) The
majority of the upper crustal rocks in the SGT belong to the Archaean and the Proterozoic age (Naqvi and Rogers [1987]) and exhibit high electrical resistivity as other shield
regions in the world (Mareschal et al. [1994]). Highly resistive upper crust offers very little attenuation to EM signals and in principle noise can propagate larger distance on the
SGT as compared to regions where younger rocks are exposed. 2) The high population
density in the southern states of India, especially in Tamil Nadu, where most of the MT
sites are located give rise to various cultural noises ( 2.3). The industrial belt along the
two banks of Cauvery River is another noise source for MT. And 3) while studying the
effect of Equatorial Electrojet (EEJ) on magnetotelluric data, measured further south
of the main measurement corridor, Rao et al. [2002] reports perceptible decrease in daytime apparent resistivity curves compared to night-time estimates in periods excess of
100 Sec. The analysis of MT data, in context of its signal and noise characteristics in
time, frequency and space domain is discussed in this chapter. Harinarayana et al. [2003]
discusses preliminary interpretation based on 2D MT modeling along three NS profiles.
4.2
Geological objectives of MT investigations in the

SGT
Deep structure of South Indian Shield Region (SISR) (Figure 4.1) has warranted the
attention of earth scientists owing to its association with various tectonic features. It is
characterized by expanses of high-grade crystalline provinces south of 13o N. Regional geology has been well mapped and studied, resulting in the identification of gradual increase
in high-grade metamorphic assemblages. It comprises of Archaean and Proterozoic terrain and exposes major crustal scale thrust faults and tectonic lineaments. The geology of
the South Indian Peninsula could be envisaged as a northward plunging structure that exposes the Archaean craton and adjacent shallow green stone belts to the north and deeper
high grade granulites to the South in an oblique section (Naqvi and Rogers [1987]. These
features seem to have played a major role in the evolution of the continental lithosphere
of the region. Several models of tectonic evolution of the region, based on metamorphic,
structural and geophysical data have been proposed (Drury et al. [1984], Radhakrishna
[1989]). Earlier geophysical studies include regional gravity (Mishra [1988]), MAGSAT
51
Figure 4.1: MT stations superimposed over the geology of the measurement corridor in
South India (Geology adapted from GSI [1995]). See text for discussion
52
(Mishra and Venkatarayudu [1985]), aeromagnetic studies (Reddi et al. [1988]), seismic
tomographic studies (Rai et al. [1993]) and electromagnetic induction studies (Nityananda
et al. [1977], Nityananda and Jayakumar [1981]). During a magnetometer array study in
South India, no large scale electrical conductivity anomaly was observed in Kuppam
Salem area, in contrast to crustal electrical anomalies observed along coastal margins in
the Southern granulite block ( Thakur et al. [1986]). A detailed discussion of the more
recent insights into crustal evolution of South Indian shield has been compiled by Mahadevan [1994]. In order to obtain more information about the deep crust of south India
and its evolution, integrated geophysical studies in the Southern Granulite Terrain (SGT)
were initiated under Deep Continental Studies program of Department of Science and
Technology, Government of India. Studies involving coincident seismic refraction and
reflection profiling, magnetotellurics, gravity and deep resistivity soundings have been
completed (For a detailed report, see Memoir of Geol. Soc. Ind. No 50, 2003).
4.3
4.3.1
The Data
Acquisition
Figure 4.1 shows the locations of the selected MT stations superimposed on the geology
of the measurement corridor. The corridor, which roughly trends NS, lies in the southern
peninsula of India, with an approximate dimension of 300 X 100 km. The MT data were
acquired mainly along three profiles, viz. Kuppam-Bommidi (KB), Omalur- Kodaikanal
(OK) and Kolattur-Palani (KP), all being roughly oriented NS, with a station spacing of
around 10 km. The corridor cuts across major geologic and tectonic elements of the region
highly resistive high grade granulite gniess, ultra-basics, granite gniesses and low-grade
amphibolite granulite facies in the south (Naqvi and Rogers [1987]). For the MT survey,
a wide-band digital MT system (GMS05 from Metronix GmbH) was used to measure
the five electromagnetic field components in single station mode. The electric fields were
measured using porous pots with Cd-CdCl2 electrodes. Four electrodes were laid in a
cross arrangement with 90m distance between the pairs. Induction coil magnetometers
with 30,000 turns were used to measure the natural magnetic fields ( 1.2.3). MT time
series were measured in four overlapping frequency bands ( 1.2.4) for a period of 24-48
hours per site, with a maximum possible frequency range, (4096 sec)1 to 8192 [Hz].
At most of the sites the data were recorded in three sessions. Out of three, one of the
session would carry out relatively longer period of measurements than other sessions.
It was typical to have around 100 stacks (1 stack = 1024 data points) for band 1 to 3
and 80 to 100 stacks for band 4. This effectively means a total recording time of more
than 24 hours. In the other sessions, the quantity of data collected would be less. As the
signal and noise in MT are time variant processes, the chief session need not always result
in best estimate of MT transfer functions. However, the present thesis concentrate on
processing the data from the chief session of each site, as this contains the largest volume
of continuous data. A selected subset of about 24 MT stations is used to demonstrate
the methodologies developed for MT time series processing. MT data from sites with
differing signal/noise distribution and geology were included in this set of 24 stations.
Table 4.1 shows location, Village names and stations codes, recording date etc.
53
4.3.2
Typical MT Time series
The observed MT time series exhibit varying amplitude, pattern and frequency content
based on the signal as well as noise sources, which in turn are time variant processes.
It is important to inspect the time series to identify obvious outliers in time series and
remove those segments from further processing (Manoj and Nagarajan [2003]). Time
series samples measured in the four frequency bands in South India are displayed, as
representative examples in Figures 4.2 and 4.3 .
Figure 4.2(a) shows a snapshot of MT time series (1024 points) measured in the band
1 (sampling rate was 32 kHz). Here the natural signal is superimposed on power line
frequency and its harmonics. The pattern of the signal in this frequency range is pseudosinusoids. Superimposed on such a background, a burst of signal activity is observed
at 0.015 sec. This activity is correlated well over four of the five measured channels.
The obvious source for this transient signal is lightning ( 2.2.2). Since the transient
envelope is rich in high frequency content, the lightning might have occurred near to the
site. But the term near has to treated cautiously. If the lighting were occurred so near
MT Station
JN02
JN07
JN10
JN12
EW01
VP01
VP03
TT06
KG01
KG02
TT08
KG03
TT09
TT10
VP11
VP10
VP13
VP12
VP16
VP15
VP17
OK15
VP19
OK18
Village
Venkatapalli
Kochchakallur
Puliampatti
Regadahalli
Vellicahndi
Vellar
Ellakutaur
Ramudaiyanur
Tottikinaru
Pakapudur
Valasiramani
Kottur
Venkatachalapuram
Viralipatti
Pudupalayam
Siliamaptti
Muttukkalivalasu
Rasipalayam
Talakkarai
Gollipatti
Kuppanavalasu
Kuthilabbai
Karadikuttam
Viralipatty
Latitude
12.900917
12.402861
12.152708
12.031389
12.335361
11.892528
11.779278
11.156917
11.795417
11.718194
11.123028
11.636722
11.244278
11.12175
11.16875
11.300222
10.968222
11.072722
0.796778
10.703222
10.635556
10.601583
10.452778
10.140778
Longitude
78.3894167
78.3060556
78.3267083
78.2559722
78.0815833
77.9628056
77.9203889
78.1988889
77.7229722
77.6813056
78.3431944
77.6697222
78.5219722
78.7151389
77.6191111
77.5708611
77.5275556
77.5706389
77.5331111
77.5446944
77.52925
77.7785556
77.4468889
77.7341389
Day Start
12-Jan-00
22-Jan-00
28-Jan-00
29-Jan-00
03-Feb-00
07-Feb-00
11-Feb-00
12-Feb-00
14-Feb-00
16-Feb-00
16-Feb-00
17-Feb-00
17-Feb-00
19-Feb-00
22-Feb-00
23-Feb-00
23-Feb-00
24-Feb-00
25-Feb-00
26-Feb-00
27-Feb-00
28-Feb-00
28-Feb-00
01-Mar-00
Day End
14-Jan-00
24-Jan-00
29-Jan-00
01-Feb-00
05-Feb-00
09-Feb-00
12-Feb-00
14-Feb-00
16-Feb-00
17-Feb-00
18-Feb-00
18-Feb-00
19-Feb-00
20-Feb-00
23-Feb-00
23-Feb-00
25-Feb-00
26-Feb-00
27-Feb-00
27-Feb-00
28-Feb-00
29-Feb-00
29-Feb-00
03-Mar-00
Table 4.1: The locations and measurement details for the MT stations
54
Figure 4.2: Typical MT time series measured in the SGT for two frequency ranges a)
the frequency range 256 Hz to 32kHz (band 1) and b) the period range 8 Hz to 256 Hz
(band2) .See the text for discussion.
55
that the em fields due to lighting behaved non planar at the site ( 1.1.1), this burst
would be treated as noise. By constructing MT transfer function between the channels
and verifying its consistency between segments, one can easily check this. The telluric
signals fluctuates within 100mV/km and magnetic fields within 0.2 nT. But in certain
instances the telluric fields can have even higher amplitudes. MT time series measured at
site TT08 in band 2 (sampling rate 1024 Hz) is presented in Figure 4.2(b). The narrow
band noise present in all the channels is noticeable. They are power line signals (50 Hz),
perfectly manifested in MT time series as sinusoids. The time series data were filtered
with analogue notch filters (50 & 150 Hz) before being digitized. The high power of the
50 Hz signals even after notch filtering shows the level of 50 Hz signals prevalent in the
area. Consider the higher amplitude for telluric and magnetic fields compared to band 1.
The visual inspection of this time series will not yield any first hand information about
the signal content, its distribution over frequency etc. It is common practice to look for
large spikes or discontinuity in the time series, to exclude such segments from further
processing. The large spike seen at 0.3 sec and is reflected in all the channels is one of
such examples. The magnetic fields show a sharp decreasing trend at the beginning of
each channel. It repeats in all the segments and could most probably be an artefact from
measurement system.
MT time series measured in Band3 (Figure 4.3(a) includes the well-known MT dead
band around 1 sec ( 2.2.1). The natural electromagnetic signal energy (Figure 2.1) is low
in this band and the time series usually look like a random process. The data displayed
in Figure 4.3(a) were measured at VP12, which is relatively a noise free site (see 4.3.3 in
this chapter). The data were measured with a sampling rate 32 Hz and the duration of
the displayed time series is 128 sec( 4096 samples). The time series fluctuates randomly
from a common mean, with some large spike activity in between. Spike activity is more
in Ex, Hy and Hz channels and is less in Ey and Hx channels. But the major spikes, for
example the one occurs at 95 sec is well correlated between the channels. This indicates
an inductive source for the noise and was probably due to switching of large power loads.
The amplitude of telluric signals are in the order of 4 mV/Km and for magnetic fields
it is 0.01 nT/s.
Visual inspection of time series ( 5) is more appropriate in the measurement band 4
(sampling rate 1 Hz). The time series measured at JN10 are presented in Figure 4.3(b),
contains 4096 samples. The main signal source for long period MT measurements are
geomagnetic pulsations. Envelopes of pulsation activity (Pc4 and Pc5 - see 2.2.2) are
observed near 1500, 1900, 2500 and 3500 seconds and is best observed in Hx. The spike
activity with much larger magnitude made the pulsation activity less obvious in other
channels. In addition to the short period pulsations, fluctuations with longer period
were also observed (and correlated) in all the channels. The magnitudes of the signals
displayed were largely controlled by the spike activity.
4.3.3
Examples of MT data collected over South India
Station VP12 (western part) and TT08 (eastern part) lie in the centre of the measurement corridor (Figure 4.1) and are near to a major EW trending shear zone in the
56
Figure 4.3: Typical MT time series measured in the SGT for two frequency ranges c) the
frequency range 8Hz to 0.25 Hz (band 3) and d) the period range 4 sec to 128 sec.See
the text for discussion.
57
region. JN10 and VP20 are from northern and southern part of the corridor, respectively.
All of the sites are on exposed crystalline gneisses of Archaean / Proterozoic age. The
time series in each station were processed in the following manner:
1. For each measurement band, the total available time series were sliced into sub
segments with the constraints of lowest frequency of interest and required degree
of freedom for each estimate.
2. The bias and trend of the time series were eliminated by procedures outlined in
3.2.3.1.
3. To reduce the bias of the spectra due to finite measurements, a hanning window
was applied to each time segments, followed by FFT and calibration.
4. Smooth cross and auto spectra of MT field components were estimated for each
time series segment by frequency band averaging ( 3.2.3.5).
5. Robust processing ( 6.2) was applied to obtain global spectral matrices from all
the segments.
Apparent resistivity, phase, coherence between predicted observed electrical field (
3.3.5) and degree of freedom for each estimates are presented for sites VP12 in the Figure
4.4(a). Data are presented in their measured co-ordinates. The xy and yx components
of apparent resistivity show split that becomes appreciable for frequencies below 0.01Hz.
As the phase components also behave differently from each other, it could be the result of
lateral resistivity contrast. Predicted coherence are above 0.8 for most of the frequency
range, but with a low in the dead band (5 Hz to 10 Sec). The degrees of freedom
(dof ) exponentially decrease with frequency in each band. Fourier transform results in
equally spaced harmonics in linear scale, where as the target frequency to compute MT
parameters are equally spaced in logarithmic scale. Due to this, the number of spectral
lines available for target frequencies exponentially reduces with decrease in frequency.
In addition to this the selective stacking processes modify the dof. But the effect of
degree of freedom seems to be minimal on the predicted coherence. In fact, near 0.1 Hz
the even with an effective dof of 2000, both the coherences are low. The steep rise in
apparent resistivity and low values for phase observed between 10,000 Hz to1,000 Hz is
clearly the effect of a near field signal. Computed error bars (xy) shows large values in
the dead band and frequency less than 0.01 Hz. Whereas the large errors for dead band
result from the poor signal strength, for lower frequency range, where the coherences
are relatively high, low dof are the reason for large error bars. Station JN10 lies in the
northern part of the measurement corridor and is relatively noisy compared to VP12.
The apparent resistivity values (Figure 4.4(b)) monotonously decrease with increase in
frequency. Phase smoothly varies except in the dead band. The coherences assume low
values between 10 Hz to 0.01 Hz. Quantity of observed data is less compared to VP12, as
seen on dof plot. TT08 lies in the eastern part of the measurement corridor and is heavily
affected with noise especially in band 3 & band 4 (frequencies below 8 Hz). Apparent
resistivity and phase values of xy (Figure 4.4(c)) component are scattered and have larger
associated errors. Though the quantity of measured data is relatively high (see the dof
58
Figure 4.4: Plot of apparent resistivity, phase, predicted coherence and degree of freedom
(DOF) vs frequency for four stations measured over South India. Data are computed in
their measured direction. See text for discussion
59
plot), the predicted coherences (E xp E x ) are less than 0.6 for majority of the frequency
range. Apparent resistivity observed in frequencies less than 10 Hz is much lower than
one would expect on a crystalline terrain. The MT transfer functions are evidently biased
in the frequency range 10 Hz to 0.01 Hz by noise. VP20 is located in the southern part
and Figure 4.4(d) presents the MT data from this station. The site was relatively noise
free and the apparent resistivity and phase values vary smoothly and are self-consistent.
Still there is a clear evidence for bias error in the xy values for frequency less than 0.01
Hz. The coherence plot also shows low values (E xp E x ) for corresponding frequencies. To
conclude, bias and random errors due to noise are evident in many of the sites and it is
not always reflected in the associated coherences and errors.
4.4
The EEJ effect on MT data
Equatorial Electrojet (EEJ) is a non-uniform east flowing current in the ionosphere,

within 50 each side of the magnetic equator. The direct incidence of suns radiation at
equatorial regions, ionizes the upper atmosphere, where geomagnetic field is essentially
horizontal. This current amplifies the northward component of slower magnetic variations. The EEJ will also affect the geomagnetic pulsation ( 2.2.2) amplitude (Sarma
et al. [1982], Sastry et al. [1983]) of MT signal below 3 Hz. The two reported studies on
the effect of equatorial electrojet on measured MT transfer functions at sites close to dip
equators are by Padilha et al. [1997] and by Rao et al. [2002]. MT data were collected
along 1000-km profile in Brazil, with the dip equator passing through the center of the
profile. The time series were measured in day and night were processed separately using
conventional as well as robust processing, to obtain apparent resistivity and phase for
the un-rotated diurnal and nocturnal tensor elements (Padilha et al. [1997]). However,
the comparison between daytime and nighttime results did not show any significant difference in the entire period band. The MT curves were nearly identical, within very low
error bars. This was true for all the measured sites along the profile, indicating that EEJ
currents did not affect the MT data. The theoretical modeling of EEJ, approximated
as a conductive layer at 110 km, the authors found EEJ may affect MT responses at
periods greater than 1000s. But they did not have enough observed data at comparable period range. They concluded that the theoretically anticipated EEJ distortions are
probably overestimated and the plane wave assumption of MT signals at equatorial regions is valid at least in the frequency range 1000 to 0.0005 [Hz]. But recent studies by
Rao [2000] and Rao et al. [2002], on analyzing MT data collected near and away from
Indian magnetic dip equator showed the effect of EEJ is appreciable above 100 sec. The
data were collected during the same field campaign in South India, the stations being
south of the main measurement corridor. The data collected at three sites viz OK3, IRA
and KAR, which lie in a NS profile, with KAR being southern most and at the center
of dip equator. MT time series measured at each site were separated into daytime and
nighttime segments. After omitting the sub-segments with obvious outliers, independent
daytime and nighttime estimates of MT impedance were made. It was observed (Figure
4.5(a) to (c) that the yx component is very similar between day and night estimates at
OK3 as well as IRA. But at station KAR, which is at the center of EEJ, the daytime
60
Figure 4.5: Apparent resistivity vs frequency plot for 3 stations tht are near and away
from the dip equator with day and night curves superimposed (Rao et al. [2002]). (a)
OK3 - site 300 km north of dip equator, (b) IRA site 100 km north of dip equator and
(3) KAR - site near the dip equator.
estimates were biased down at frequencies less than 0.01 [Hz]. This corroborate the idea
that, increased signal amplitude for Hx as a result of the eastward current flow associated
with EEJ might bias the MT transfer functions.
The effect of EEJ on MT transfer functions seem to be decreasing fast as one goes
away from the dip equator, as evidenced by the similarity in daytime and nighttime curves
of IRA and OK3, which lies to the north of dip equator and KAR. As the measurement
61
Figure 4.6: Averaged Kp indices [Courtesy NGDC], K indices [HYB] during the MT
measurements are compared with telluric predicted coherence [4 sec 128 sec]. The mean
coherence is drawn as a line.
corridor for the present study is located 300 km north of the dip equator, the effect
of EEJ on MT transfer functions may be treated as negligible, at least in the frequency
range of measurement.
4.5
Signal Activity during the Field Campaign
MT data were collected in two field campaigns in September - December 1998 and January
March 2000. In order to study the ability of different processing methods to estimate
MT transfer functions (> 1 sec) in the presence of both low and high noise levels, compared
to signal levels, the geomagnetic activity during the measurement were examined. Shown
in Figure 4.6 as histograms are the averaged daily global Kp and Hyderabad K indices,
with long period average coherence ( 4.5). K indices are a local quasi-logarithmic
measure of geomagnetic activity (Jones et al. [1989]) and have 10 classes between K =0
and K =9 (magnetic storm). K =9 correspond to a range of 300 nT for Hyderabad
observatory. Where as Kp indices are global indicators of geomagnetic activity and
are weighted average of a selection of local K indices and with weights reflecting the
geomagnetic latitude and longitude. Note that some stations are omitted. The stations
are plotted in sequential order from Jan 12 to Mar 3, 2000. The K and Kp indices
shows almost same magnitude and variation through out the measurements time, with
a maximum at VP3 on 12th February 2000. The low geomagnetic activity (indices less
than 2) is observed between station KG01 and VP11 (16-23rd February) which in turn is
flanked by two well defined highs (indices > 3). Predicted coherence functions are good
measure of the quality of MT data ( 3.3.5), within the statistical limit of the least square
solution of MT transfer functions. Procedures used to compute the average bias-corrected
coherence function are discussed in 4.6. The averaged predicted coherence values for
all the station varies with a mean of 0.7 for the measurement duration. They are
generally high in the region of K > 0.3 and low at stations whose period of measurement
coincided with low geomagnetic activity. But stations KG03 and TT08 whose period of
measurement have low geomagnetic indices, show relatively high coherence values and
station VP13 with K indices > 0.3 shows low coherence. These deviation from the general
62
trend, indicate the presence of other factors which might have also contributed to the
coherence distribution.
4.6
Spatial character of coherence
Predicted coherence function defines optimality of the least square solution for MT transfer functions ( 3.3.5). It has widely been used to examine the quality of the measured
MT data (Vozoff [1972]). Though it is possible to compute predicted coherence functions
for all the measured MT fields, telluric coherence functions are more frequently used
for two obvious reasons i.e., 1) The magnetic fields are usually measured at a greater
level of accuracy than the electric fields, 2) Electric fields are more prone to noise from
man-made sources. Due to reasons 1 & 2, the least square estimation of MT transfer
functions usually consider magnetic fields as noise-free inputs and minimize the noise in
electric fields (Sims et al. [1971]). In this section, the distribution of averaged predicted
coherence of E x and E y components for all the stations in the measurement corridor are
described. The averaged predicted coherence is defined as,

1 2
2
ExHxHy + EyHxHy
(4.1)
2
The individual coherence functions within the bracket are defined in 3.3.5. However,
it is relatively well known that the predicted coherence functions get biased due to noise.
Jones [1981] and Jones et al. [1983] describe a simple way to reduce the bias in coherence
functions. For predicted coherence it is defined as,

4
4 2
2
2
2 2
b =
1
(4.2)
1+
v2
v
2 =
Where b is the bias reduced estimate of and v is the degree of freedom for the
estimate. In order to estimate the MT transfer functions, all the time segments available
at the stations were processed and averaged the coherence without using any preferential
stacking algorithms. The idea here is to represent the noise contributions at each site
and time segments with obvious outliers were also kept to estimate impedance tensor and
their associated coherence functions. Considering the bias in coherence estimation in the
presence of noise either correlated between the input and output fields or un-correlated
noise in the input field, the coherence function over respective frequency bandwidth of
measurements were averaged. At each site we have four estimates of predicted coherence
functions in four bands of measurements. The coherence estimates at all the sites are
gridded (with grid surface always going through the data points) and presented as a
contour map for the measurement corridor for each band in Figure 4.7.
A total of 69 sites were used to prepare the map out of 80 stations. The predicted
coherence functions of the telluric fields vary between 0.3 and 0.9 for most of the sites and
over bands of measurements. The spatial distribution of coherence is however not similar
for all the bands. The distribution patterns are similar for band 1 and band 2 (Figure
4.7(a)&(b)) with coherence assuming high values for most of the sites. Except for the NE
apart and a site in the center of the measurement corridor, the coherence values are very
63
Figure 4.7: The band averaged telluric predicted coherence plotted as a contoured map
for all the measured MT sites. (a) Band1 (b) Band2 (c) Band 3 and (d) Band4.
high (>0.8). As band1 and band2 covers the noise contributions from power lines, there
could be two reasons for high coherence for MT data in this range. If the noise source
can be considered as near field ( 2.4) for that measurement site, the MT coherence may
still be very high, but the apparent resistivity will just be a function of frequency, with
least amount of information of subsurface resistivity. On the other hand, if measured
sufficiently far from the noise source, so that the plane wave conditions for MT are met,
then the power line signals can be an ideal source for the high frequency MT. However, it
is impossible to distinguish these two possibilities by analyzing the coherences. The MT
data in the middle frequency range (Band3 0.25 6 Hz) exhibits comparatively low
values for coherence (Figure 4.7(c). As the natural signal strength in this frequency band
are very low compared to other bands, it is not surprising to observe the low coherence
for band3. However, the southern part of the measurement corridor (south of 110 Lat)
has higher coherence values than the northern part. The low coherence for the NE part
observed in Band1 and Band2 (Figure 4.7(a)&(b)) also is repeated in band3. Except for a
site in the north and for an EW (11.50 Latitude) belt in the center, the MT measurements
show fairly high coherence values in the long period measurements (Figure 4.7(d)). On
a closer look at all the four plots it is apparent that a EW belt in the center of the
measurement corridor shows low values throughout the frequency bands.
64
4.7
Geographic relation of noise
As we have seen in 4.5 the geomagnetic indices varied during the measurement period.
However, its direct implication on coherence of longer period data was not clear. One
reason could be the varying effect of man-made noise in the MT data. The spatial
distribution in predicted coherence ( 4.6) consistently showed as EW belt of low values
for all the measurement bands. In order to understand the relation of noises in MT
data measured over South India to the major geographical elements of the region, the
contoured plot for noise/(signal+noise) ratio for all the stations was superimposed on the
geography map for the measurement corridor (Figure 4.8.The noise/ (signal+noise) ratio
is defined as ,
n
= (1 2 )
(4.3)
s+n
Where is the averaged predicted coherence defined in equation 4.1 & 4.2. The
ratio is bounded by values 0 and 1. Major roads (thick black lines) and rail tracks
crisscross the measurement corridor. Major rail/road links between two state capitals viz
Bangalore and Chennai passes through the northern part of the corridor. Cauvery river
flows through the center of the corridor, with major towns like Erode and Sankaridurg on
its banks. In addition to the rail and road clusters, this area hosts a number of cement
factories. Limestone is present all along the fringes of an exposed granite (Sankari)
dome. Availability of limestone and water from Cauvery made this region suitable for
such factories. The measurement corridor is relatively free of anthropogenic disturbance
sources further south (ie south of 110 Latitude). The contour plot for N/(S+N) ratio
shows an interesting relation to the distribution of man made electromagnetic disturbance
sources. The northern region of the corridor shows three isolated highs at stations JN08,
JN03 and JN05. Though it is spatially correlated well with the Chennai Bangalore
rail lines and roads, the other three stations in the area viz. JN01, JN02 and EW02
do not show high N/(S+N) ratios. In the middle of the measurement corridor, 8 MT
stations distributed in and around the towns Erode and Sankaridurg shows very high
N/(S+N) ratio. This cluster of highs in noise ratio clearly shows their affinity towards
the industrial belt mentioned above. Though one expects the effect of the industrial zone
to extend all along the banks of the river Cauvery, except for station TT08, the effect
of noise is not perceptible further West. In line with the observation that the southern
part of the corridor is devoid of major anthropogenic disturbance source, the N/(S+N)
ratios of southern MT sites have relatively low values. In conclusion, in the northern and
central part of the measurement corridor, the noise distribution is related to the known
man-made electromagnetic disturbance sources.
4.8
Discussion
An attempt has been made in this chapter to characterize the signal and noise in the
measured MT data over the Southern Granulite Terrain. Processing of MT data start
with visual inspection of the measured time series. As seen in Figure 4.2 and 4.3, the
time series collected over the SGT shows effects of noise from various sources. Moreover
65
Figure 4.8: Contoured map of telluric Noise/(Signal+Noise) ratio, superimposed on the

major geographical elements of the region. Note N/(S+N) ratio on either sides of Cauvery
river near Sankari.
66
the noise is manifested in the time series with patterns different from the natural signals.
However, visual inspection of time series will not remove all the noise in the measured
data, as many noise processes have temporal characteristics similar to the signals. A more
quantitative definition of signal and noise is required to separate them in MT data. One
straightforward way to do this is to construct a transfer function between the measured
electric and magnetic field. Now the portion of electric and magnetic fields that cannot
be explained by such a relation may be treated as noise. As we have seen in 4.3.3,
criteria for separating signal and noise based on the coherence and errors may not work
always. This necessitated looking at some other properties of the data, specifically, two
independent observations on the processes contributing to signal and noise. In this respect
the indices of geomagnetic activity during the measurement time were compared with
the averaged long period coherence of each site. The agreement of the coherence and the
geomagnetic indices indicate the validity of such an approach (Figure 4.6). However, a few
disagreements are also evident. This indicates the possible noise processes contributing
to the measured data. Spatial relation of the known cultural noise centers and the quality
of MT data was examined. The strong correlation of the high noise/(signal +noise) ratio
to the major industrial belt indicate that the MT signals may get consistently degraded
due to presence of an active noise sources, irrespective of the signal activity. However,
the conclusions drawn from this chapter should be treated rather cautiously in instances
where the underlying noise processes deviate from a Gaussian distribution. The signal
and noise are defined in this chapter as inline and outline components of a least square
solution of MT transfer functions. In the next chapter, we will take a closer look at
the temporal properties of MT data and propose a new signal discrimination method,
without the definition signal and noise from a least square solution.
Chapter 5
The application of the artificial
neural networks to magnetotelluric
time series analysis
67
68
5.1
Introduction.
In the previous chapter the signal and noise characteristics of MT data acquired from
SGT were analyzed. The noises from various sources manifest in MT data in different
forms. A part of noise is obvious in the time series itself. The manual inspection of the
time series is often the first step in MT data processing, which selects a subset of time
segments, followed by statistical transfer function estimation such as robust processing.
However, manual inspection has its own limitations. Firstly, it is a time consuming process, which constitutes 80 % time used in MT time series analysis. Secondly, like other
human decision making processes, MT time series editing is also a subjective process,
wherein the same person may output differently, in a long session of editing. Considering the large volume of MT data collected over SGT, a need to develop a procedure
to automate the manual editing of MT time series was felt. An approach was made to
this problem from pattern recognition angle demonstrating the efficacy of artificial neural networks in discriminating noisy sections of time series against the signals. In first
instance, characteristics of long period (4-128sec, band 4). MT time series were used to
build training / testing data base for artificial neural networks.
5.2
Signal and noise in the magnetotelluric time series
An exact classification of signal and noise characteristics in magnetotelluric time series

is difficult as their sources may have similar spectral content. Unfortunately there are
often more noise sources than signal sources. Still, a generalized classification is possible
according to the origins of both signal and noise. At periods longer than 1 sec, the natural electromagnetic field originates in the upper ionosphere and magnetosphere. Random
bursts of energy originate in charged particles from the Sun and induce sinusoidal electromagnetic waves in the magnetosphere and ionosphere (see 2.2.1 for more details).
The signal amplitude and frequency may vary with the energy and type of activity and
there may be a long gap between the arrival of two trains of signal. Figure 5.1(a) shows
some typical magnetotelluric signal patterns. Magnetotelluric noise sources have been
listed in table 2.1. Any or all the these noise sources may be present together with MT
signal resulting in compounded responses in the recorded time series. Some of the most
common MT noise patterns are presented in Figure 5.1(b).
5.3
Visual Inspection (editing) of magnetotelluric

time series data; why automation ?
Manual editing is often the first step in magnetotelluric data processing. The editor
examines each stack of time series and labels it as either good or bad according to
its signal/noise character. The bad stacks are removed from further processing. The
task involves an intensive amount of pattern recognition. Experience provides a judicial
balancing of signal characteristics such as shape, amplitude, frequency and correlation.
69
Figure 5.1: Common signal and noise patterns in long period MT time series. Samples
are collected from different sites. Note the change in amplitudes. (a) Signal patterns;
Samples of Ex and Hy shows the geomagnetic pulsations. Other channels also show
signals but at longer periods. (b) Noise patterns; Ey & Hx shows different types of
spikes. A step and its decay is shown in Hy . Sample of random noise is shown in Hz .
This process is subjective in nature and the same editor may output differently in long
sequences of editing. A rule of thumb is that if in doubt throw it out. If questioned
about a particular decision, however, the editor may offer a few rules for guidance but
can give no obvious systematic reasoning. This constitutes the major time and human
resource used in MT data processing. The number of stacks of time series recorded
sometimes runs into hundreds. With several such sessions of recordings from a station
and many such MT stations occupied in each survey, there is a pressing need to provide
a more robust alternative, which is less time consuming and more objective.
70
5.4
Magnetotelluric noise characterization
Automation of MT time series editing requires a systematic evaluation of the task performed by the editor. What parameters influence an editing decision? How much importance does the editor give to each parameter? Can it be quantified? As there is
no information available in this regard, an editing evaluation exercise was undertaken.
It involved discussion with different editors and re analysis of previously edited data
and identification of the influence of different parameters in magnetotelluric time series
editing. From the study, it was found that the following factors reasonably represent the
criteria applied in editing:
1. Signal pattern
2. Signal amplitude
3. Correlation of different field components
4. General noise level
5. Quantity of MT data available.
When the general noise level is very high and/or the quantity of data available for
editing is limited, compromises are made in editing. As these are exceptional cases, they
were excluded from the present analysis. The first three factors were analyzed and each
was given a percentage of influence. The percentage of influence for a particular factor
was found by preferential editing using that criterion and comparing results with the
regular mode of editing.
5.4.1
Patterns of signal & noise :
Pattern controls a major part of decision making. The signal pattern is transient overlapped sinusoids. The noise patterns can be classified as follows.
1. Spikes : High amplitude, maximum duration for few data samples. Sometimes a
spike is followed by a transient decay. These are commonly power related.
2. Noise bursts: Plateau or step-like appearance spanning hundreds of data samples.
3. Noisy trace: Random fluctuations due to wind, seismic effects and or low signal.
4. Muted or dead trace: Instrument problem such as broken cable or failure in electronics.
In most cases, the quality of signal can be deduced from the signal waveform. This,
according to our study influences 60% of editing decisions.
71
5.4.2
Amplitude of signals
Although naturally varying electromagnetic fields exhibit a large variation in their

strength, a broad range can be specified. In most cases, the amplitude ratio of orthogonal electric and magnetic signals was found to be a good discriminator. In long
period time series, the electric field fluctuates within +/-100 mV/km and the magnetic
field fluctuates within +/-0.5 nT in a noiseless environment. The channel amplitude may
be increased many times in the presence of certain types of noise. Amplitude criterion
often gives the most reliable information if the contaminated signal has the same pattern
as noise- free signals but enhanced amplitude. This was found to influence 20% of the
decisions.
5.4.3
Correlation between simultaneously measured channels
The electric and magnetic fields are related by a transfer function defined in 3.3.1.2 In
the ideal case the MT signals should be highly correlated and random noise will reduce
the correlation. However, noise can be highly correlated between E and H channels.
Noise from power lines especially near 60/50 Hz is an example. In longer period data,
it was found that the correlation coefficients between orthogonal electric and magnetic
fields (Ex - Hy and Ey - Hx ) could be a signal discriminator. Correlation was found to
influence another 20% of editing decisions.
As the above-described parameters influence the bulk of editing decisions, they were
selected as the basis for automation. An automation scheme was developed using an
artificial neural network for classification/editing of time series data.
5.5
Artificial Neural Networks
5.5.1
Why artificial neural network?
Optimal conventional automation requires statistical characterization of the noise. This

means estimation of certain statistical parameters from the data. In other words the
likelihood ratio of signal/noise is replaced by sufficient statistics of data. The method is
simple and appealing. It works very well when the target signal is known and the noise
has a normal distribution. But in most cases noise does not have a normal distribution
and the likelihood ratio is a complicated nonlinear function of input data. Garth and
Poor [1994] classifies geophysical signals as non-Gaussian, unstructured signals as they
involve least amount of detail. Classification of such signals is best done by labeling
them according to their signal content and then presenting them to a pattern recognition
scheme. Artificial neural networks (ANN) are an emerging tool that have been applied
in many areas of science and engineering where pattern recognition is involved, such as
speech and character recognition. The learning and adaptive capabilities of these models
make them attractive for application to some problems in geophysics (Calderon-Macias
et al. [2000]). The most documented application of ANN in geophysics has been for the
automation of seismic data processing and interpretation (Murat and Rudman [1992],
McCormack et al. [1993], Fish and Kusuma [1994], Dai and MacBeth [1995]). ANN also
72
Figure 5.2: Simple three layer feed forward neural network. The data is processed at
each neuron in the layers. Each neuron performs a summing of inputs multiplied with a
weight parameter and outputs the data through its sigmoid transfer function.
have found other applications in geophysics; such as in interpretation of well log data
(Winer et al. [1991]) , locating subsurface targets from electromagnetic field data (Poulton
et al. [1992]) prediction of upper atmospheric and ionospheric activities (Lundstedt [1996],
Altinay et al. [1997], Koons and Gorney [1991]). More recently, artificial neural networks
have been used in magnetotelluric inversion (Zhang and Paulson [1997], Spichak and
Popova [2000]).
5.5.2
ANN theory
An artificial neural network is an information processing system composed of a large

number of processing elements called neurons, which are modeled on the functions of
neurons in the human brain. ANN differ from conventional pattern recognition techniques
in their ability to adaptively discriminate or learn through repeated exposure to examples
and in their robustness in the presence of high noise levels. ANN do not require a priori
knowledge about the noise distribution of the process under study, as do its statistical
counterparts. Unlike conventional methods, which incorporate a fixed algorithm to solve
a particular problem, ANN perform a mapping, usually non-linear, between the input
and output data, which allows the network to acquire important information on the
problem being solved. It is these characteristics of neural networks that motivated us to
investigate their use in MT data processing.
73
One of the most widely used types of ANN, the feed forward artificial neural network
(FANN) was used in the present study. Its architecture is outlined in Figure 5.2. It
consists of a layer of neurons that accept various inputs (input layer). These inputs are
fed to further layers of neurons (hidden layers) and ultimately to the output layer, which
produces a response. The aim of the technique is to train the network such that its
response to a given set of inputs is as close as possible to a desired output. A number
of algorithms are available for training a neural network. Back propagation is the most
popular training algorithm (Werbos [1990]) and was used in the current study. During
FANN training, each hidden and output neuron process inputs by multiplying each input
by its weights. The products are summed and processed using an activation function,
here, a sigmoid function,
f (x) = 1/(1 ex ),
(5.1)
to produce an output with reasonable discriminating power. The neural network

learns by modifying the weights of the neurons in response to the errors between the
actual and targeted output values. For a given set or vector of N inputs (x1 , x2 ,....,xn ),
the output of node j is computed as
X

yj = f
Wji xi
(5.2)
where W ji is the weight of the connection between the i th and j th neurons. The
learning rule for the adjustments in the weight between neurons i and j is expressed as
Wji = j oi
(5.3)
where o i is either the output of node i or an input, is a positive constant named

the learning rate and j is the error term of node j. Thus,
j =
E
netj
(5.4)
where
E = 0.5
(yj oj )2
(5.5)
and,
netj =
Wji oi .
(5.6)
Here, yj is the target value for the j th node and o j is the output for the j th node.
The value j is computed as,
X
j = oj (1 oj )
k Wkj ,
(5.7)
if the node is not an output unit. To improve the convergence characteristics, a
momentum gain is added to the weight correction term which stabilizes oscillations
during the learning process ( Boadu [1998]), i.e.,
Wji (n + 1) = j oi + Wji (n),
(5.8)
74
where n is the iteration index. The training of the network is complete if the convergence of weighting coefficients has been achieved. The convergence criterion requires
that the sum square error at the output must be less than a desired tolerable error (Luo
and Unbehauen [1997]).
While training a network, we may start with arbitrary values for the weights Wji . It
is usual to choose the random numbers in the range -1 to 1. Next, the outputs (O) and
errors (E) for that set of weights are calculated, followed by the derivatives of E with
respect to all of the weight (equation 5.7). If increasing a given weight would lead to
more error, the weights are adjusted downwards. If increasing a weight leads to reduced
error, it is adjusted it upwards. After adjusting all the weights up or down, start all over
again and keep going through this process until the error is close to zero. The sequence of
presenting the entire training database, calculating the network response, comparing the
result with the assigned class, propagating the error backwards and adjusting the weights
is called an epoch. Few thousands of such epochs of training are usually required by a
neural network to reach zero error. However, if the number of training patterns exceed
the number of weights in the network it may not be possible for the sum squared error
(SSE) to reach zero.
5.6
5.6.1
Data analysis
Network engineering
As a neural networks massive interconnectivity and inherent non-linearity requires significant computing resources, dedicated mainframe computers and workstations were
traditionally used for neural network training. Most of the applications utilizing ANN
in geophysics deal with single-channel data using a small sliding window moving along
the time series. The current application, where 5 channels of data, each containing 256
points, were to be classified simultaneously, posed a major challenge. A novel approach
was made to accommodate multi-channel data so that all the computing could be done
on a PC. Figure 5.3 gives a schematic representation of the data flow. This method can
be used for any multivariate signal detection scheme.
As the signal shape and pattern controls a major part of the editing decisions, attention was focused on this part of the analysis. The inter-channel parameters, such
as amplitude ratios and correlation coefficients were also included for neural network
training. The signal detection scheme was divided into two steps.
a) Detection of the patterns of individual channels: Patterns of individual channels
were evaluated using a neural network. Within each stack (256 points) of 5 channels,
each channels pattern successively were classified. Thus pattern detection of a single
stack resulted in 5 values corresponding to the 5 channels.
b) Detection of inter-channel parameters: Amplitude ratios (Ex /Hy , Ey /Hx ) and correlation coefficients between channels were computed. These parameters, along with the
pattern quality data from step (a), formed the inputs for another neural network. Output
from this network indicates the overall quality of that data stack.
This separation provided us with the flexibility and ease of operating on a small
computer, without having to compromise on network performance.
75
Figure 5.3: Data flow through the feed forward artificial neural network (FANN) based
editing scheme.
5.6.2
Pattern Training
5.6.2.1
Data used:
Selecting an appropriate training data set is one of the most critical steps in successful
training. How many patterns are required in training? How many patterns should be
dominated by signal and how many by noise? A rule of thumb is that the patterns in
the training set should cover the main categories of signal and noise. The signal pattern
should represent the typical features of a signal with different frequency characteristics.
Obviously there are more noise patterns than signal patterns (Zhao and Takano [1999]).
With an input data space of 256, the number of exemplars required to train a neural
network was large (> 1000 sets). This required us to search for and select a large number
of time series segments. 1200 sets of MT long period time series data stacks (256 s.
length, each stack) were collected from different data sets giving 6000 traces.
5.6.2.2
Pre-processing
Each channel was corrected for trend and bias and normalized between 1 and 0. The
data segments were then manually classified and assigned a value between 0.9 and 0.1,
depending on quality. The label varied between 0.1 (bad) and 0.9 (good) depending on
the pattern quality of the data. There is no fixed generic relationship between the label
and quality ( one can as well have labels in the reverse order).
Out of the classified time series segments, 5000 extreme cases were selected for training. This set contained 2500 very good and 2500 very bad data samples. Care was taken
to include all categories of signal and noise patterns generally found in MT time series.
The database was shuffled to have a random distribution of good and bad data for training. From this data base 3000 exemplars (training vectors) were kept for training and
76
Figure 5.4: Stacked amplitude and phase spectra of the training database. The spectra
of noisy data (class 0.1) clearly different from the signals (class 0.9).
the rest, for testing.
5.6.2.3
FANN Training
The network was presented with the training database of 3000 time series segments.
Training was done with different values of learning rate () and momentum gain () for
different epochs (defined in 5.5.2). Training typically took 60 minutes to complete 1000
epochs on a 500 MHz PC. The longer time for training hampered effective interaction
with the process and limited an exhaustive search for the optimum network configuration
( i.e.; optimize , and the number of hidden neurons). A transform of the input data
was sought which would preserve the essential information about the signal while reducing
the dimensionality of the input space. For this, the time series was Fourier transformed,
after applying a cosine taper to both ends. To test whether the amplitude spectra alone
could be used as a discriminator, the spectra of good and bad signals in the entire training
database were stacked separately. As seen in Figure 5.4, there is a clear difference between
the two types. Noise generally raises the spectral power at higher frequencies, where as
signal has more power at lower frequencies. Amplitude spectra were used for further
training. FFT enabled us to reduce the number of input data from 256 to 128. Training
was attempted with different subsets of spectral harmonics by removing the highest
frequency elements successively. The first 100 harmonics were found to be sufficient
and the training time was considerably reduced (20 minutes). The sum squared error
(SSE ) as a function of epoch for the last training is plotted in Figure 5.5(a). The SSE
reached a minimum of 33.73 for 2000 training samples. In neural network training, more
77
Figure 5.5: Results from the pattern training. (a) The SSE as a function of epoch. The
error reached the minimum floor after 400 epochs. Stability of convergence is demonstrated up to 1000 epochs. (b) Deviation between manually classified and network predicted classes for 500 test time series segments. The scattered points shows the major
deviations. The correct picking constitute 94
importance is given to how the network performs on novel (non training) data than the
SSE of training itself. FANN on test samples (samples which were not used for training)
gave 94 % (472/500) correct classification (Figure 5.5(b) with = 0.09, = 0.1 and 10
hidden neurons. As observed by Dai and MacBeth [1995], although this solution was
considered optimal for the current application, further architecture optimization could
undoubtedly be achieved by a more exhaustive search procedure on a more powerful
computer.
5.6.2.4
Sensitivity analysis
The neural networks sensitivity to the signal to noise ratio was examined using the
following analysis. A time series of 256 points with high signal content was mixed with
a normally distributed random noise series. The trained neura1 network was assigned to
classify the signal. In each run the signal content was changed with a small increment
so that it covered the range 0% to 100%. The network steadily gave values near to 0
(Figure 5.6) until the signal content reached 60%. The output changes to higher values
as signal content exceeds 60% and asymptotes to 1 after it is 80%. It is now evident that
the network is able to detect signals if the signal content is more than 70%. Another
interesting aspect is the narrow region of high variance (when signal content is between
65% and 75%) in the output. A thorough analysis with noise of other distributions was
beyond the scope of the present study.
78
Figure 5.6: Network output versus signal content. The network was simulated by inputs
with varying signal content. A narrow region of high variance exists when signal content
is between 65
5.6.3
Inter channel training
Here, a separate neural network was trained with the inter channel parameters (explained
in 5.6.3). To keep this as the final step in determining the overall quality of the stack,
the 5 pattern quality values predicted by the first step also as input were also included.
The inputs to the network were,
1. 5 pattern quality values from earlier network
2. 2 amplitude ratios (Ex /Hy & Ey /Hx )
3. 2 correlation parameters (Ex - Hy & Ey - Hx )
5.6.3.1
The data
To train the network, 900 stacks were selected from different data sets ( 5 channels,
256 data points each). Care was taken to include an equal number of signal and noise
segments. Each stack was manually inspected to assign a class vector 0.1 or 0.9 depending
on its overall signal quality, as explained previously. A training database was made as
follows.
1. The trained FANN was used to classify the patterns of the 5 simultaneously measured channels within each stack. Output of this processing varied between 0.0
(bad) and 1.0 (good). Figure 5.7(a) shows the pattern classes for Ex and Ey channels versus the stack numbers. The broken bar below the graph indicates the
manually assigned stack class (Black-good; White-bad). An excellent correlation
79
Figure 5.7: Pattern, amplitude ratios and correlation coefficients of 900 stacks which
form the database for inter channel training and testing . The thick line is the running
average over 10 points. The broken bar indicates the overall stack quality - black good
(0.9) and white bad (0.1). (a) Ex (squares) and Ey (triangles) pattern quality predicted
as a function of stack number. (b) Ex - Hy (squares) and Ey Hx (triangles) amplitude
ratio. (c) The correlation coefficients of Ex to Hy (squares) and Ey to Hx (triangles).
between the manually assigned stack class and the pattern class of individual channels computed from FANN processing is evident. This justifies our earlier comment
that signal discrimination largely depends on the pattern of time series signals.
2. As the amplitude ratios between the orthogonal electric and magnetic field were
found to be another signal discriminator, the two ratios were included in the training
database. Further, in order to keep the values within 0 and 1.0 (as necessitated by
the sigmoid function) they were normalized as
AExHy =
Ex
Hy
Ex
Hy
AEyHx =
and
Ey
Hx
Ey
Hx
Ex
Hy
Ey
Hx
(5.9)
(5.10)
80
where, Ex , Ey , Hx and Hy are simple ranges of amplitude (maximum - minimum) of

respective channels for a stack. As plotted in the Figure 5.7(b) the ratios vary considerably and a direct correlation with the signal class is impossible. This parameter
was retained for training, as it adds another dimension to the input data.
3. Correlation coefficients between orthogonal electric and magnetic fields (Ex - Hy
and Ey - Hx ) were calculated for each stack. The correlation coefficient r( ) (
Molyneux and Schmitt [1999]) between two vectors X and Y ( t = 0,1,2.....n) is
given by
t=n
t=n
P
P
X(t)Y (t + )
X(t)
Y (t + )
t=0
t=0
t=0
r( ) = s
t=n
2 s t=n
t=n
2
t=n
P
P
P
P
2
2
n
X(t)
X(t)
n
Y (t + )
Y (t + )
t=0
t=n
P
t=0
t=0
(5.11)
t=0
A value of r =1 indicates perfect positive correlation between two vectors; r = 0

indicates no correlation, meaning the vectors are not similar; r = -1 indicates anticorrelation (meaning two vectors are of same shape but of opposite polarity). The
correlation coefficients for all the 900 windows are plotted in Figure 5.7(c). The
squares represent Ex -Hy correlations and diamonds represent the Ey -Hx correlations. Most of the Ex -Hy coefficients are distributed between 0.2 and 0.6 whereas
the Ey -Hx distribution is between -0.2 and -0.6. It can be clearly seen that the
signal class is = 0.9 (good quality signal) when Ex -Hy and Ey -Hx coefficients are
well separated (when they are closer to 1).
5.6.3.2
FANN training
The training database thus prepared was a 9 x 900 matrix. The rows were shuffled to
produce a random distribution. Around 70% of the database was used for training the
neural network and the remaining 30% was used for testing. As the input vector is only
of length 9, the training procedure was rather simple compared to the pattern training. A
three layer feed forward network was trained with 9 neurons in the input layer, 2 nodes in
the hidden layer and one node in the output layer. The momentum gain was set to 0.1
and the values of learning rate () and number of hidden layer nodes were changed. Each
training consisted of 5000 epochs and took 7-10 minutes to complete. As the training
progressed, the number of hidden layer nodes was reduced to one. The training stopped
when sum squared error (SSE) was 7.9 for 650 samples with = 0.09. Figure 5.8(a)
shows SSE as a function of epoch for the final training. On simulating the network using
the test data, SSE was 3.72 (i.e. 0.1222 per vector ). Figure 5.8((b) shows the deviation
of network-predicted signal class values from the manually classified values. It can be
seen that the network classification was successful.
81
Figure 5.8: Results from inter channel training. (a) The SSE as a function of training
epoch. (b) Deviation between manually classified stack quality and network predicted
for 250 stacks. 235 stacks were classified similar to manual classification.
5.6.3.3
Relative significance of input
The network was fully trained to classify MT signals. To find out the relative importance
of the nine inputs (5 pattern classes, 2 correlations & 2 amplitude ratios) to the FANN, a
validation training was carried out. At each run, the particular input of interest was set to
nil all through the run. This resulted in a greater error compared to the original training
using non-zero inputs. The extent of departure from the previous error is considered
as the relative significance (Altinay et al. [1997]) of that particular point for the neural
network. As can be seen in Figure 5.9, patterns of Ex , Hy , Ey and Hx are dominant
in the neural network response, followed by the correlation parameters. Pattern of Hz
and amplitude ratios have the least influence on the neural network. This roughly agrees
with our earlier classification of the factors influencing editing decisions, but with two
surprises. First is the relative insignificance of Ey as compared to Ex . As the database
for training was equally biased to all the 5 channels, it was not the result of faulty
training. The excellent performance of the neural network on test data also proves this.
Second is the low influence of amplitude parameters. While formulating the problem,
the amplitude and correlation were given equal weight, but the training results disprove
it. As the testing sessions proved the discrimination capability of ANN, the training was
stopped.
82
Figure 5.9: Relative significance of various inputs to the networks, viz amplitude
ratios (A1 and A2), correlation coefficients (C1 and C2) and five pattern qualities
(Ex , Ey , Hx , Hy andHz ). The error deviation against each input is a measure of its significance to the Neural Network.
5.7
Application
The artificial neural network based signal detection scheme was applied to the MT data
collected from SGT. The neural network approach was successful in discriminating bad
time series segments against the good segments in majority of the cases. Here the result
of neural network processing from four sites viz, G12, VP12, JN10 and TT8 are presented.
The selections were made to demonstrate the performance of neural network in the presence of varying degree of noise contamination. Station TT8 situated in the industrial belt
( 4.7) is most affected by noise, followed by JN10 and VP12. G12 is a station occupied
in the Western India, with very high signal-to-noise ratio in the longer period. This was
included to demonstrate neural networks performance on a very good site as well. The
long period data collected in the chief session ( 4.3.1) of each site were subjected to three
type of editing viz 1) Blind editing - selected all the stacks, 2) ANN based editing scheme
and 3) Manual editing by a third person. Once edited, the data were subjected to a
common processing procedure as detailed below. Auto and cross spectra were computed
for each target frequencies (Figure 3.3) from all the available time segments, after the
editing. Transfer functions were estimated ( 3.3.1.3) for each time segments, for all the
target frequencies. Average telluric predicted coherence (equation 4.1) functions were
used to sort the auto and cross spectra. A subset consists of 70% of the highest coherent
spectra sets were the selected and used for estimating the final transfer functions. Note
that this method differs from Coherency Threshold methods, wherein data sets with
predicted coherence below a preset value are rejected for final transfer function estimation. This choice of somewhat old-fashioned algorithm is justified as it allowed to show
the efficacy of neural network based editing. Moreover, the algorithm does not interfere
too much with the data selected. Variance of the MT transfer function was computed
according to equation 3.41 in 3.3.4. General descriptions of results from each site are
given below.
83
Figure 5.10: Comparison of MT apparent resistivity and phase computed from different
mode of editing of data from site G12 . Filled circles represent xy and diamonds represent
yx components. (a) Using all stacks available. (b) By neural network editing. (c) By
manual editing.
1. G12: The station was located over basaltic province of western India. 192 stacks
were available for processing. The time series was generally noise free with long
period geomagnetic pulsations (sinusoids). The occasional spikes were smaller than
the waveforms within which they occur. Figure 5.10(a) shows MT apparent resistivity and phase computed from all available stacks (192). The curve is quite smooth
and without deviation, suggesting that the time series were relatively noise free.
Neither FANN based editing (selected 157 stacks) nor manual editing (selected 150
stacks) improved the curve significantly. Results are given in Figures 5.10(b) and
(c). A small number of noisy segments were easily rejected by the coherency-based
estimator, without any need of editing.
2. VP12: This station was located in the granulite province of South India. A total
of 352 stacks were recorded. Almost half of the stacks carried spikes and step
like features originated from submersible electric pumps and switching of power
supplies. The MT apparent resistivity and phase computed from all the stacks are
given in Figure 5.11(a). The resistivity () values, especially between 1.0 and 0.1 Hz
are scattered and phase () is poorly resolved, with both xy and yx modes being
equally affected. Figure 5.11(b) shows the results of FANN-based editing ( 140
stacks selected ), where, both apparent resistivity and phase are better resolved.
Improvement is clearly evident in the 1.0 to 0.1 Hz range. Manual editing resulted
in selecting 127 stacks and the computed apparent resistivity and phase are similar
to FANN editing (Figure 5.11(c)).
3. JN10: This site was also located in the same region as VP12, but with more noise
in the data. The effect of the noise is evident in the apparent resistivity curves
84
mode of editing of data from site VP12 . Filled circles represent xy and diamonds
represent yx components. (a) Using all stacks available. (b) By neural network editing.
(c) By manual editing.
mode of editing of data from site JN10 . Filled circles represent xy and diamonds represent
yx components. (a) Using all stacks available. (b) By neural network editing. (c) By
manual editing.
85
mode of editing of data from site TT08 . Filled circles represent xy and diamonds
represent yx components. (a) Using all stacks available. (b) By neural network editing.
(c) By manual editing.
computed from all the stacks (Figure 5.12(a)). The phase is also affected in the
range 1- 0.1 Hz. The neural editing picked 80 stacks out of 256 available and gave a
better estimate, as shown in Figure 5.12(b). A less rigorous manual editing picked
107 stacks and gave a similar result (Figure 5.12(c)).
4. TT8: This station was the most affected by noise. Both electric channels, especially
Ex , were affected by noise originating from an industrial belt nearby. Increased
spike activity was observed in all channels. The MT apparent resistivity and phase
computed (Figure 5.13(a)) from all the 256 stacks available gave a very distorted
picture. Both xy and yx components were poorly resolved. Significant improvement
was made by FANN editing (Figure 5.13(b)) which selected 67 stacks out of 256.
The yx component is now smooth and relatively error free. The xy component is
also improved except near 0.1 Hz. Almost the same result is produced by manual
editing (69/256) as shown in Figure 5.13(c). The stacks picked by neural network
and manual editing are compared in the Figure 5.14. Overall the pickings match
quite well. Deviation between the two editing schemes is evident for a few picks
between stack numbers 150 - 180. Between these there are stacks with moderate to
low signal content, which were accepted by manual editing but rejected by neural
editing.
86
Figure 5.14: Comparison of manual and neural signal picking for site TT8. The diamonds
present the neural picking and crosses, the manual.
5.8
Discussion
The application of ANN based editing to magnetotelluric time series brings out some
interesting results. The ANN based noise rejection predicts an overall quality of each
subsection of time series (0 to 1). While this value cannot be treated as errors of the corresponding impedances, the response is a measure of quality of the time series/ impedances
and help improve the process of editing of data more objectively as compared to manual/ visual editing. In a low noise environment such as station G12, the network editing
produces results almost similar to blind editing (using all stacks). On such a data, a
simple coherency-based estimator can do the signal discrimination to a certain level of
satisfaction. However, the neural networks ability to pick out signal from moderate to
high noise environment was evident on the data collected from SGT. In such cases it approximates human intelligence - established from the fact that the neural network based
editing gives a result similar to manual editing. These results satisfy the objective of the
present chapter, ie, to provide a robust alternative to manual editing of magnetotelluric
time series. This experiment, brought out ANNs potential to automate MT time series
editing, efficiently and reliably. The scheme could be adapted into routine processing to
save human time and increase reliability in editing. The current scheme performs about
70,000 floating point operations (flops) per stack for classification. FFT of five channels alone uses 40,000 flops. If the further computation uses the same spectra, some of
the computational redundancy can be removed. It points to the possibility of integrating
ANN based editing into real time processing of MT data. The signal and noise characteristics in magnetotelluric data are very different in different frequency ranges. This is due
to the difference in signal and noise source mechanism in different part of Earths natural
electromagnetic spectrum. Sensor geometry and instrumentation also affect pattern of
signal, which the neural network depends on. This necessitates reformulation of MT signal and noise characteristics for training of ANN, wherever necessary. However, the extra
computational requirement by re-training of ANN may not pose a burden on resources,
taking into consideration the ever-increasing computational power of microprocessors.
The neural network based approach described in this chapter see the time varying
87
processes for a finite duration as a whole. It does not assume a transfer function relation
between the different channels under preview. Noise processes, which will not affect the
pattern, amplitude and correlation parameters of the time series, but have potential to
badly affect the estimation of MT transfer functions, may still escape from the screening
of the neural network based editing. Secondly, once the time series has been selected,
the signal content at all the frequencies from that time segment may not be the same.
This provides for and necessitates a second screening of MT data, now in the domain
of frequency, after an initial data screening by ANN. One advantage of working in the
frequency domain is that estimation is done independently at many frequencies. For
stationary processes, the data at different frequencies are strictly uncorrelated (Chave
and Thomson [2003]). The robust processing techniques effectively search the frequency
space to estimate MT transfer functions and its superiority above other comparable
methods in frequency domain is established (Jones et al. [1989]). However, as stated in
the introduction, in certain cases of noises, robust processing fails. The next chapter
critically analyze and propose improvements to the robust processing techniques for MT
transfer function estimation.
Chapter 6
Estimation of Magnetotelluric
Transfer Functions: Robust
Statistical Methods
88
89
6.1
Introduction
After the raw time series have been inspected by manual or automated noise rejection
methods, the selected segments are used for estimating MT transfer functions. This involves estimating 101 -102 complex frequency domain transfer function elements Z()
from electric and magnetic field time series E(t) and H(t) (approximately 106 real numbers per site) (Egbert and Livelybrooks [1996]). This data reduction, though superficially
simple, can result in useless MT transfer functions, in the presence of noise in measurements. The classical least-square method of computing MT transfer functions, allowing
for noise distributed in the simplest manner was discussed in 3.3.1. However, the drawbacks of least-square (LS) method has been widely recognized and documented in the
past three decades (Sims et al. [1971], Gamble et al. [1979], Egbert and Booker [1986],
Egbert and Livelybrooks [1996]). The failure of the LS method has been attributed to 1)
presence of noise in the input channel and 2) violations of Gaussian noise assumptions.
In the first instance, the linear statistical model, (equation 3.26) through which the
natural electromagnetic fields are related to each other, considers the input as noise free
and the noise is restricted to the output or predicted data. In MT it is usual to assume
the magnetic fields as input and electrical fields as output. It follows that the noise in
magnetic fields can down bias the transfer function estimates ( see 3.3.2). To avoid these
bias errors, Gamble et al. [1979] proposed the measurement of two remote magnetic field
components as references and remote reference processing substantially improved (Jones
et al. [1989], Shalivahan and Bhattacharya [2002]) MT transfer functions, over the single
station least squares approach. In the second case, Gaussian distribution of errors is
assumed by the LS estimator. The observed errors in magnetotelluric data often have a
Gaussian distribution, but with heavier tails due to the presence of outliers (abnormal
data). These non-Gaussian noise produces scatter (or low point to point continuity) in
the processed data. If ignored, these outliers can corrupt the estimated magnetotelluric
transfer functions, making them useless for geologic interpretation. A number of processing methods have been proposed which adaptively weights or screen the data. Stodt
[1983] showed the usefulness of weighting the subsections of MT data according to their
predicated coherence (see 3.3.4). However, such approaches may fail in the presence of
correlated noise in the data. Substantial improvements were bought out by the application of robust M (Hampel et al. [1986]) statistical procedures to MT and geomagnetic
time series analysis (Egbert and Booker [1986], Chave and Thomson [1989], Chave et al.
[1987], Larsen [1989], Sutarno and Vozoff [1991], Egbert and Livelybrooks [1996], Egbert [1997], Ritter et al. [1998], Nagarajan [1998], Smirnov [2003], Chave and Thomson
[2003]). The success of robust procedures may be attributed to three factors. First, its
superiority to other data processing techniques is established (Jones et al. [1989]). Second, these procedures can be justified rigorously (Egbert and Livelybrooks [1996]) and
third it can be easily implemented using iterative-weighted LS procedures and extended
to remote reference processing (Chave and Thomson [1989]).
In this chapter, two new approaches are proposed to improve the performance of
robust statistical procedures on MT time series. The basics of the robust M procedures are
discussed in 6.2 and more stress is given to the computational aspects. Non-parametric
estimators such as Jackknife (Efron [1982]) were used to robustly compute the variance
90
of MT transfer functions (Chave and Thomson [1989]). Its use as an effective initial guess
for robust procedures is discussed in 6.2. It is shown that in majority of the cases, the
use of Jackknife for initial guess resulted in better estimation of MT transfer function as
compared to LS estimations. It is usual in MT to sub divide the time series, estimate
the spectral density matrices for each segment individually and then robustly average
the spectra or transfer functions between the sub segments (section averaging). Within a
segment, it is common to use a limited number of target frequencies and obtain smooth
spectra by averaging several adjacent Fourier harmonics (frequency band averaging).
The documented researches on robust estimation of MT spectral densities and transfer
functions concentrate on section averaging. This arises from the assumption that, within a
narrow frequency band, the distribution of Fourier coefficients are of Gaussian nature and
a simple average (LS) gives the best estimate. In 6.3 it is shown that this argument often
fails and the problem of contamination is applicable to band averaging as well. A robust
weighting approach is proposed for estimation of cross and auto spectral estimation within
a band, without making specific model assumptions concerning signal or noise. Both these
proposed procedures, while applied on a large volume of MT data collected over SGT,
South India, met with moderate to good improvement of MT transfer functions. Majority
of these data were collected in single station mode, except for few stations, where remote
reference data were available. A synchronization problem of MT equipment and GPS was
the reason for single station recording during the study ( 4.3.1). However, the robust
processing procedures outlined in this thesis can easily be extended to remote reference
processing and such an example is shown for station VP10 ( 6.3.5). The application of
the proposed robust processing methods are discussed in 6.2.5 and 6.4.
6.2
Robust estimation of MT transfer functions
The term robust was coined in statistics by G.E.P. Box in 1953 (Hampel et al. [1986],
Flannery et al. [1992]). Various definitions are possible for a robust procedure. But
in general referring to a statistical estimation like MT transfer function, it means one
which is relatively insensitive to the presence of a moderate amount of bad data or to
inadequacies in the statistical model and that reacts gradually rather than abruptly to
perturbations of either (Jones et al. [1989]).
6.2.1
Why robust methods?
LS estimators are best on a data with Gaussian (normal) distribution of errors. It is well
known that the break down point of least square (LS, or L2) estimates is zero. Break down
point describes the smallest percentage of bad data that can corrupt an estimate. This in
turn means that the presence of few outliers can corrupt an LS estimate. The resistance
of median and other L1 estimates (L1 minimizing the first power of residuals) to the
presence of outliers is well - documented in the geophysical context (for e.g. Claerbout
and Muir [1973]). Simple median has a break down point of 50 %. This led to the
suggestion that L1 norm can replace L2 norm in many geophysical estimation problems.
However, Chave et al. [1987] pointed out major drawbacks with L1 estimator on practical
91
data sets. It was shown that L1 estimator requires about 60% more data to achieve the
same parameter uncertainties as L2 estimator. Also, the natural probability distribution
function for L1 is double exponential (Laplace) which make statistical inferences difficult.
This suggests that it is desirable to treat the outliers within the framework of a Gaussian
model, rather than outright abandonment of that model (Chave et al. [1987]). One
way to achieve this is to identify the outliers and process the remaining segments of
data with usual LS procedures, one such example is the ANN editing discussed in 5.
However, in many situations like time series analysis, detection of outliers itself becomes
extremely difficult and it may result in rejection of all the data as well. As a consequence,
robust methods are developed, which can accommodate outliers and still minimize their
influence. Without going into detail, the situation in which robust methods are desirable
is shown in Figure 6.1. The probability distribution function shown in Figure 6.1(a) has
heavier tails than expected for Gaussian distribution. Any fluctuation in these tails may
lead to inaccurate estimate of the location of central peak (Flannery et al. [1992]). The
simple line-fitting problem, given in Figure 6.1(b) shows the influence of few outliers on
the estimation of the slope of the line, constrained to go through the origin. An estimate
that is robust to the presence of the outliers should instead produce a fit, which satisfies
majority of observations. Thus the need for robust estimates is seen.
6.2.2
Robust M estimators
Statisticians have developed various robust statistical estimators. For MT transfer function estimation, M estimators are more relevant (M stands for maximum likelihood)
and are discussed in detail here. A brief discussion of development of robust estimation
in MT follows. Let us reproduce the MT transfer function equation 3.26,
E=ZH+r
(6.1)
where there are N observations so that E and r are N vectors, H is an N X 2 matrix

and Z is a rank two vector. The last variable in equation 6.1, r is the difference between
measured and predicted output field (here, Electric) and is the residual parameter to
be minimized. The classical way of solving MT equation is the least squares technique
(Swift [1986], Sims et al. [1971]), where the square of the residual power in equation
6.1 is minimized to yield a solution for Z. Let r i be the residual of the i th observation,
the difference between observation and prediction. The standard L2 method tries to
minimize r i 2 . The M estimators try to reduce the effect of outliers by replacing the
squared residual r i 2 by another function of residuals, yielding, min (r i ), where is a
symmetric, positive defined function, called loss function with a unique minimum at zero.
For standard L2, ( r i ) = r i 2 /2, while for the L1 estimator ( r i ) = | r i |. In general, if
(r ) is chosen to be log f(r), where f(r) is the true probability density function (pdf), of
the residuals, then the M estimator is maximum likelihood (Chave and Thomson [1989]).
However, as it is difficult to obtain a pdf from finite observations, the loss function is
chosen in theoretical ground. Performing minimization,
X
i
(ri )zij = 0,
f or
j = 1, 2.
(6.2)
92
Figure 6.1: Examples where robust statistical methods are desirable: (a) A one dimensional distribution with heavy tails (b) A distribution in two dimensions fitted to straight
lines. Adapted from Flannery et al. [1992]
93
Type
L2
L1
Huber 1) if |r| < k
Huber 2) if |r| k
Tukey 1) if |r| < k
Tukey 2) if |r| k
Loss- (r)
r2
2
|r|
x2
2
k
2
k = (|r|
k2
(1 [1 ( kr )2 ]3 )
6
k2
6
Influence (r)
R
sign(r)
R
sign(r)
r[1 ( kr )2 ]2
0
Weight w(r)
1
1
|r|
1
k
|r|
[1 ( kr )2 ]2
0
Table 6.1: Few commonly used influence functions. Adapted from Zhang [1996].
Where (r )= ( r)/ (r), is called the influence function and z ij is one component
of the 2 X 2 tensor Z. The equations will reduce to least squares, when ( r) = r2 /2,
(r )=r or to least absolute deviations when ( r) =|r|, (r) = sign(r). To normalize
the equations such as 6.2, it is common to divide the equation with a robust estimate of
scale for which median absolute deviation (MAD) is a good choice (Sutarno and Vozoff
[1991]).
med|ri med(ri )|
(6.3)
M AD
is the theoretical counter part of an appropriate pdf. Now equation 6.2
d=
Where MAD
becomes,
X
i
ri
( )zij = 0.
d
(6.4)
To solve this equation, it is easiest to write it as a weighted least square, by defining

a weighting function, w i = (r i /d )/r i and rewriting, equation 6.4,
X
wi ri zij = 0, f or j = 1, 2.
(6.5)
i
The weights are computed based on the residual r and scale parameter d, from the previous iteration and they are initialized using a least square solution (Chave and Thomson
[1989]). The major difference between the solution given above and a normal weighted
LS procedure is that, in the first case the weights are computed based on the residuals
and scale estimate from previous iteration, where as in later case, weights are computed
based on the data itself.
6.2.3
Choice of influence functions
The influence function (r) measures the influence of a datum on the value of the parameter estimate. We have already seen the influence function for L1 and L2 estimates. Given
in the table 6.1 are four different influence functions. (Zhang [1996]). Their graphical
representation is given in Figure 6.2.
From the table 6.1 and plot (6.2) it may be deduced that,
94
Figure 6.2: Schematic diagram showing loss, influence and weight functions for Least
Square (LS or L2) , Least Absolute (L1) Huber and Tukey estimators. Values shown in
y axis are arbitrary See text for discussion. Adapted from Zhang [1996]
1. L2 (least-squares) estimators are not robust because their influence function is not
bounded. The larger the residual, the heavier weight it gets.
2. L1 (absolute values) estimators are not stable and their weight function at r=0 is
unbounded and solution may become undetermined.
3. Huber function is a parabola in the vicinity of zero (like L2) and increases linearly
at a given level |r| >k (like L1). Where k = 1.5d gives 95% efficiency with Gaussian
data. This is most widely used and suitable to residuals drawn from a probability
distribution that is Gaussian in the center and Laplacian in the tails. This blend
of L2 and L1, is more suitable for the distribution presented in Figure 6.1.
4. Tukeys function is very severe for outliers (Figure 6.2)
However, use of either Huber or Tukey function alone is not advisable. The Huber
weights fall off slowly for large residuals and never descend to zero and thus do not
provide adequate protection against large outliers (Chave and Thomson [1989]). Tukeys
influence function outputs near zero values for slightly higher residuals. Used alone, it
may result in rejection of all data, or acceptance of very few. Hence it is advisable to
use a Huber function for the first few iterations and then use Tukeys function as a final
weighting to protect against large outliers (Egbert and Booker [1986]).
The response of influence functions to a set of MT data is presented in Figure 6.3.
The data (VP13) were collected in the period range 4 sec to 128 sec at a sampling rate
1Hz. The time series were sub segmented into sections of 2048 data points each and 105
such overlapped sections were available for processing. Figure 6.3(a) shows the E x -E xp
(defined in 3.3) residuals at frequency 0.1875 Hz for each stack. Majority of the residuals
assumes small and similar values with a small amount of scatter. However, there are
large outliers present (16), especially near stacks 20, 40 and 100. The weight functions
computed according to Huber, L1 and Tukey functions for these residuals are presented
in Figure 6.3(b). L2 weights are always 1. L1 weights (inverted triangles) show low values
for the outliers. However, their weights become very large for very small residuals. Huber
weights (stars) are combinations of L1 and L2 weights, with a cut off at k = 1.5d. They
95
Figure 6.3: Responses of different influence functions to a set of residuals from MT data
processing. Station VP13 Shows the Ex residuals for 0.1875 Hz for all the 105 stacks. (b)
Shows the response of three influence functions to the residuals. See text for discussion.
are less influenced by the large outliers (as L1) and can accommodate the fluctuations
within small residuals (near the Gaussian peak, like L2). As seen for values between
stacks 50 and 60, Huber function down weights the large residuals, where the influence is
same as L1. The steady response of Huber weights is more clearly demonstrated between
stacks 60 and 80. Tukey weights show very low weights (< 104 ) for large outliers. If
used alone, it may result in deselecting most of the data sets. Severity of Tukeys weight
dominate everything else, as demonstrated at stacks 50-60, 100 to 105. However, it also
rejects usable data especially at stacks 10, 40-50 and 70.
6.2.4
Implementation for MT
The application of robust processing developed for the estimation of magnetotelluric

transfer functions is presented in this section. A MATLAB [2001] code robspm.m was
realized by integrating the algorithms discussed by Egbert and Booker [1986], Chave
and Thomson [1989], Ritter et al. [1998], with the two new approaches introduced in this
chapter. All the computations are performed in frequency domain. It is assumed that the
time series were collected for sufficient duration of time. It is then divided into segments
(subsets / stacks) of fixed length. Size of subset is chosen based on the lowest frequency of
interest and a target value for degrees of freedom. Each segment was tapered by a Hanning
window. The segments may be overlapped, provided corrections to dof are made. After
Fourier transform, each channel was divided by calibration transfer function ( 1.3.4)
to remove the effect of instrument and sensors. A matrix of auto and cross spectra
between channels (5 x 5 for single station set up), was made for a particular frequency
by band averaging the adjacent frequency points respecting a narrow frequency window
96
(this averaging is critically analyzed in 6.3). The preprocessing results in generation of

a 4 dimensional matrix for a site, with size L segments, N frequency and 5 x 5 channels.
This forms the basic data for robust estimation. A flow chart for the robust processing
implemented in this thesis is shown in Figure 6.8(a) and discussed towards end of this
section
6.2.4.1
Initial guess of transfer function
One component of equation 6.1 may be written as,

Exl = Zx xl Hxl + Zx y l Hyl + rxl
(6.6)
Where E and H are frequency domain components of electromagnetic field recorded

in l = 1,2,. . . .L time segments. Zxx and Zxy are components of Z (equation 6.1). The
algorithm described here will use the above equations. However, the same steps can be
used to estimate the other components of the tensor and also magnetic transfer functions
(equation 3.24). Estimations are done independently for all the N frequencies, so that
the frequency terms have been omitted. To get an initial guess for transfer functions,
global spectral matrix is obtained by averaging all the L cross and auto spectra sets for
a particular frequency, in the least square sense. For example one element of the global
5 x 5 matrix is obtained by,
L
Ex Hy

1 X
Ex Hy l
=
L l=1
(6.7)
Conventional way of getting an initial estimate for Z is by solving the equation 6.1
in least square sense (see 3.3.1.3), with the global cross and auto spectra as obtained
above.
6.2.4.2
Jackknife estimate as initial guess
The idea of using LS as an initial guess is to keep the solution somewhere in the convergence path and then to iterate to a solution with robust re-weighting. If one can keep
the initial solution nearer to the actual/desired one, the convergence can be fast. Here
the use of a simple nonparametric estimator called Jackknife (Efron [1982], Chave and
Thomson [1989], Eisel and Egbert [2001]) as a better replacement for LS as initial guess is
demonstrated. Chave and Thomson [1989] demonstrated its ability to estimate variances
of the MT transfer functions robustly, alleviating the need to accurately compute the dof
as needed by conventional variance estimate (such as given in 6.16). However, Eisel and
Egbert [2001], while discussing the results from processing of 2 years of continuous MT
data, commented that they are systematically too large for most of the periods. The
jackknife estimator was used to compute an initial MT transfer function (not variance)
in the robust processing routine. This is shown to be resistant to outliers and thus may
produce an initial guess, nearer to the true one. One advantage of Jackknife method is
its computational simplicity. This is important, as these routines will be regularly called
in robust processing. Consider Z as any one component of the transfer function derived
97
by locally solving the equation (6.1). From all the time segments we have L number of
estimates of Z. Let Z mean be the mean based on all the data. The data are then divided
into L groups of size L-1 each by deleting an entry in turn from the whole set. Let the
estimate of Z based on ith subset, when the ith datum has been removed be Z i (Chave
and Thomson [1989]). The jackknife mean is given as,
L
Zjackknif e = LZ mean
L1X
Zi
L i=1
(6.8)
The quantity in the above equation was originally introduced as a low bias replacement
for regular mean (Chave & Thomson, 1989, Effron, 1982). We may construct a difference
vector,

L

X

dif f
Zi
= Zi
Zi .
(6.9)

i=1
The Jackknife variance is then given by,

L
Z var =
L 1 X dif f 2
Zi
.
L i=1
(6.10)
Note that jackknife variance is entirely different from the conventional variance which
are parametric estimations. (In the sense they depend on number of dof, value of
coherence etc). Advantages and disadvantages of this variance are thoroughly discussed
by Chave and Thomson [1989], Eisel and Egbert [2001]. The aim here is to investigate
the use of jackknife estimate as an initial guess for robust processing. Iterations can be
performed by successively deleting data rows (stacks) which gives maximum difference
as in equation 6.9. After each iteration, the variance of current iteration is compared
with the previous one. Iterations are stopped when there is no more improvement to the
variance, or the data get exhausted.
An example of the application of Jackknife estimation of MT transfer function is
plotted in the Figure 6.4. From the station VP10, a total of 103 overlapped sections
(length 1024) of time series data were available. An equal number of transfer functions (Z)
were generated from these data sets. The jackknife difference as calculated by equation
6.9 for the first iteration is shown in Figure 6.4(a). The plot shows large differences
especially at the beginning and end of the stacks. To proceed data row (stack) with
highest difference will be deleted. Figure 6.4(b) shows the progressive decrease of variance
as the iteration progresses. In this case, iteration was stopped at 28. Figure 6.4(c) and
(d) compares the output of a least square and jackknife estimations of MT apparent
resistivity and phase values (YX). Figures 6.5(a) to (d) depict the jackknife estimation
of MT transfer functions for site VP13. As can be seen from the plots, the jackknife
gives better estimate of the parameters as compared to the LS. The LS estimations were
distorted by few and large unusual data elements. However, the jackknife estimations
are not advised as the final ones, as the variability of jackknife estimate can be large
for some statistical distributions (Chave and Thomson [1989]). It may be concluded
that jackknife can be good start for robust processing, considering its computational
98
Figure 6.4: Comparison of Least Square and Jackknife estimation of MT transfer functions for station VP10 (a) Jackknife difference for the first iteration. (b) Variance as a
function of iteration number. (c) and (d) comparison of and values from LS & JK
processing
99
Figure 6.5: Comparison of Least Square and Jackknife estimation of MT transfer functions for station VP13 (a) Jackknife difference for the first iteration. (b) Variance as a
function of iteration number. (c) and (d) comparison of and values from LS & JK
processing
simplicity, non-parametric nature and superior performance over LS.
100
6.2.4.3
Scale estimate
Once the initial guess for transfer function is made, we derive the residuals r l , for all the
time segments for a particular frequency as,
2

2
rxl = Exl Zxx Hxl Zxy Hyl
(6.11)
Equation 6.11 describes one way of assessing the quality of a least square solution
to the observations in E and H. There are also other possibilities like variance Z,
coherence 2 etc to be used as quality parameters (Ernst et al. [2001]). However, the
residuals between predicted and observed fields are popularly used to assess the quality
of LS solution of MT transfer function and are used in this thesis as well.
An initial MAD scale estimate for the robust weighting procedures is obtained as,
dM = 1.483med.(|rl med.(rl )|).
(6.12)
For MT data the residuals in the above equations are complex in nature. One way
of measuring the residual size is by its magnitude and it is preferred, as it is rotationally
invariant (Chave and Thomson [1989]). In a simple way, the weight assigned to a complex
value changes both real and imaginary parts in the same way such that its phase remains
the same.
An upper limit to the scale estimate may be given as k M = 1.5d M (see 6.2.3 case 3).
6.2.4.4
Robust transfer function estimation
The Huber weights wl for each stack l is calculated according to kM .

1 f or rl kM
l
w =
kM
f or rl > kM
rl
(6.13)
Now using the weights, new auto and cross-spectral estimates are given by,
Ex Hy
L
l
1 X l
= L
w Ex Hy
P l
w l=1
(6.14)
l=1
From the modified spectra sets, new estimates (robust) of transfer function Z are
obtained. To estimate the new MAD scale estimate, we must take into account the
weighting process.
d2H
L
L X l l 2
w (r )
= 2
Lc l=1
(6.15)
where, Lc is the number of weighted events with w l = 1 in step 6.13. Now by using
the new upper limit k H = 1.5 d H , the steps 6.11 and 6.15 are repeated, resulting another
estimation of Z xx & Z xy . Though, it is advised in the literature to iterate equations 6.11
to 6.15, to our experience the transfer functions hardly changes after second iteration.
101
6.2.4.5
Tukey weights
The extreme outliers in the data are removed by again weighting the spectral matrices
with Tukeys biweight criterion. In order to compute a new scale estimate. Following
Ritter et al. [1998],
1
L
d2T =
1
L
L
P

1
L
P
(wl rl )2
l=1
2
l
r
kH
15
rl
kH
2
(6.16)
With an upper limit k T = 6d T , we obtain Tukey weights w l as (reproducing from

Table 6.1),
l
(
r
1
f or rl kT
kT
wl =
(6.17)
0
f or rl > kT
Spectral matrices are weighted again with this new weight. This forms the final step
in robust processing. The transfer functions are computed from the weighted spectral
matrices.
6.2.4.6
Computing the variance
The robust processing procedure modifies the degree of freedom of individual estimates.
While computing the variance, this also must be taken into consideration. In the processing code, it was realized for each weighting step as,
dofold X l
w .
L
Now the variance for one element of the transfer function is given by,
dofnew =
F (k, 2dof 4, = 0.05)[1 Coh2 (Ex Exp )]Ex Ex

(Zxy ) = k
2dof 4
[1 Coh2 (Hx Hy )]Hx Hy
2
(6.18)
(6.19)
M
uller [2000]. Where F is the Fischer distribution, with k = 4 (See Bendat and
Piersol [1971]). An indirect estimate of the degree of noise level can be obtained from
the misfit of the final model (impedances) with the weighted data sets (cross and auto
powers) of the final iteration. However this assumes that, once the outliers are removed,
the residuals (model predictions observations) gets a Gaussian distribution
6.2.4.7
Quantile Quantile plots
In Figure 6.6, comparison of Least Square and the robust processing method (described
above) is demonstrated for station TT08. The Ex -Exp residuals from LS estimates of
transfer functions for all the stacks of frequency 0.0791 Hz are plotted in Figure 6.6(a) and
b (zoomed version). Inverted triangle represents the un-weighted (LS) residuals and stars
represent the weighted (Robust) residuals. In general, the residuals follow a Gaussian
distribution, but superimposed by few outliers. The Quantile Quantile (QQ) plot in
102
Figure 6.6: Comparison of Least Square and Robust processing of magnetotelluric data
station TT08 for frequency 0.0791Hz . Triangles represent LS processing, and stars
represent robust (RB) processing (a) and (b) time series of Ex residuals. C) Quantile
Quantile plot of Ex residuals d) MT apparent resistivity and phase values from LS and
robust processing. See text for discussion.
Figure 6.6(c) better explains this statement. QQ plot describes the departure of observed
data set from a particular distribution, here Gaussian distribution ( 3.2.1.1). If the
residuals are drawn from a Gaussian distribution, the QQ plot will be an approximately
straight line (Chave et al. [1987]). As can be seen from the Figure 6.6(c), up to 1.5
(standard normal quantile) the residuals (LS inverted triangles) follow a straight line.
However, the departure from Gaussian distribution is evident above 1.5. The weighted
residuals (Robust stars) closely follow the straight line, indicating the effectiveness of
robust processing methods to eliminate non-Gaussian errors. The effect of these outliers
on computed MT apparent resistivity (xy ) and phases (xy ) are shown in Figure 6.6(d).
The highly irregular nature of the LS estimate (triangles) are the results of a few outliers
in the data influencing the transfer function estimate. The robust processing resulted in
better estimation of both xy and xy for all the frequencies.
6.2.5
Application
6.2.5.1
Flowchart
A flow chart for the robust processing scheme developed in line with the discussions in
the above sections is given in Figure 6.7(a). Robust processing scheme receives a spectral
103
Figure 6.7: Flow chart for robust processing scheme. The gray area represents the proposed initialization of the transfer functions using Jackknife. This routine, concentrates
on the section averaging of MT data. See text for discussion.
104
Figure 6.8: Data flow through the flow chart representing robust processing of MT data.
The gray area represents the proposed Jackknife initialization. (a) Process A uses Jackknife (JK) as initial guess and (b) process B uses Least Square (LS) as initial guess. See
text for discussion.
matrix (4 dimensional) as input. The process starts with an option of least square (LS)
or Jackknife (JK) solution as initial guess. Then the data successively go through the
weighting procedures. In majority of cases, more than two iterations may not be needed
and one can apply the Tukeys weights for final estimation of spectra sets. Finally the 4D
(L stacks, N frequencies n x n channel) spectral matrix gets reduced to 3D (N frequencies
n x n channel), which is the final output of the robust algorithm. This routine, written
in MATLAB [2001] takes 10-20 seconds on a 500 MHz PC, with a typical 5 channel MT
time series with a total of 106 values. Two processes are defined to demonstrate the
superiority of robust processing with Jackknife as initial guess. The data flow for the two
processes are plotted on the flow chart in Figure 6.8(a) & (b). Process A uses robust
processing with Jackknife (JK) as initial guess, where as Process B uses least square
(LS) for the same. The two processing schemes are compared in the following section,
according to their performance on application to 8 MT stations from SGT.
6.2.5.2
Comparison of robust processing schemes
The eight stations chosen for the demonstration are evenly distributed in the measurement corridor (Figure 4.1) and the results are presented in Figures 6.9 and 6.10. In all
the plots, the solid symbols represent Process A and open symbols represent Process B.
The station JN12 (Figure 6.9(a)) located in the northern half of the corridor was affected
by a large number of spikes in the longer periods and by power line harmonics in the
range 10 to 100 Hz. This is evident as the xy in this frequency range exhibits grossly
different values from Process B. Robust processing with Jackknife (Process A) improved
105
the estimates of xy in the frequency range 10 to 100 Hz. In the longer period, Process A
resulted in smoother and values. In station VP16 (Figure 6.9(b)), no major change
is observed between results of processes A and B. Perhaps as the station is relatively less
affected with noise, the initial LS solution itself is comparable to that of Jackknife (A).
Even then Process A resulted in better estimation of values in the frequency range 10
Hz to 1Hz. The results from OK16 (Figure 6.9(c)) show that process A (JK) initialization created a down bias in xy values as compared to process B. However, the process
A resulted in smoother values as compared to process B. In a later discussion it will
be shown that how this bias may be removed from Process A. The Process A resulted
in smoother estimates of values for station VP12 (Figure 6.9(d)) as compared to the
output of process B. However, the sharp drop of values computed through Process A
near 6 seconds seems to be unreal.
The dramatic improvement of values (especially for periods > 1 Sec) at station
VP14 (Figure 6.10(a)) by Process A as compared to Process B gives another evidence
that robust processing results in better estimation, if started with a good initialisation.
It is also seen from the Figure 6.10(c) that the values from process A are less scattered
in the same range of period. Station TT08 is located near the industrial zone, along
the Cauvery river (Figure 4.1, see 4 for discussion on this). It was written earlier
that the long period time series from this station is affected with high magnitude spike
activity. The Jackknife estimates (A) resulted in better estimates of yx values especially
by defining a smooth curve between 1 and 10 sec (Figure 6.10(b)). Though Process A,
could not produce better results for xy , it failed gracefully by predicting a trend for
the xy , in the period range 1 to 10 sec, which was absent from the Process B. The Ex
channel of station TT04 was severely affected by near source noise ( 2.4) in the frequency
range 0.1 Hz to 10Hz, as manifested in a 450 raise in xy values and small values for xy
values. The conventional robust processing, which was biased by the LS initialisation,
failed in the presence of majority of noisy data. Jackknife produced a better initial guess
and when followed by the main robust scheme, dramatically improved the xy and xy
values as shown in Figure 6.10(c). However, in the longer period the yx values seem to
be biased down compared to conventional processing (Process B). At station OK18, the
values from both processing do not differ much. Still, the xy values from Process B,
seems to be scattered for periods > 10 sec. Also the steep decrease in xy values with
increase in frequency from1 Hz to 10 Hz seem to be unreal and inconsistent with xy
values. In both the frequency ranges Process A produced a better estimates as shown in
Figure 6.10(d).
6.3
6.3.1
Robust Band averaging

Introduction
Role of section and band averaging of auto and cross spectra in the context of magnetotelluric signals was discussed in 3.2.3.4. It was shown by Jenkins and Watts [1968]
that subdividing a time series into sections of length M and forming smoothed spectra
is equivalent to smoothing the spectrum of undivided time series by a sinc (sin(x)/x)
106
Figure 6.9: Comparison of robust processing results using Least Square (LS) and Jackknife (JK) initialization for stations JN12, VP16, OK16 and VP12. Process A refers to
robust processing with JK initialization, where as Process B refers robust processing with
LS initialization See legend for symbol identification
107
Figure 6.10: Comparison of robust processing results using Least Square (LS) and Jackknife (JK) initialization for stations VP14,TT08,TT04 and OK18. Process A refers to
robust processing with JK initialization, where as Process B refers robust processing with
LS initialization. See legend for symbol identification
108
function. In magnetotellurics, it is a common practice to use the combination of section

and band smoothing, due to various reasons.
1. With magnetotelluric data, one seeks to estimate a time stationary quantity (transfer functions), which varies smoothly over frequency Band averaging over a spectral window facilitates this.
2. The measured time series may not be a contiguous sequence as there may be gaps
in time series recording due to instrumental reasons This necessitates the use of
section averaging in addition to band averaging
3. A short duration noise with high power, may corrupt larger number of FCs if band
averaging only used, than the case where spectra is estimated from sub-segmented
time series sections (section averaging) (Egbert and Booker [1986]) - This make it
necessary to reject parts of time series affected by noise.
The robust statistical procedures, as the one presented in 6.2, concentrates only
the section averaging. Sims et al. [1971] commented that the average over adjacent
frequencies would facilitate the estimation of spectra and might give sufficient dof to
reduce noise powers in cross spectra. The mention of band averaging can also be found
elsewhere in context of magnetotelluric time series processing (Swift [1986], Gamble et al.
[1979]). While applying the robust procedures on MT transfer function estimates, Chave
et al. [1987], Chave and Thomson [1989] opined that the same approach might not be
amenable to band averaging. Egbert [1997] discussed robust estimation of spectral density
matrices (SDM) for multi-station MT processing. However, he limited himself to the
problem of estimating SDMs from a set of sections, which are individually averaged
over a frequency band. Recognizing the need for better estimation of cross and auto
spectra before the actual estimation of transfer functions, Ritter et al. [1998] suggested
procedures to predict single event spectra by comparing with a global (estimate from all
the stacks/sections) average. While comparing two processing techniques for synthetic
MT time series processing, Ernst et al. [2001], commented that the use of lesser number
of frequency points in the vicinity of target frequency to average the spectra might help
to prevent over smoothing of long period estimates. Recently Smirnov [2003] computed
MT transfer functions for all the Fourier coefficient pairs and used a repeated median
estimator to arrive at a better estimate of MT transfer function. However, such an
approach may fail at the instance of strongly polarized fields (either due to the Geology or
to coherent noise), as it may not give enough degree of freedom for a practical estimation.
The effect of few outliers on MT data processing is well recognized and it is demonstrated
in the previous section. However, effect of outliers in band averaging in the vicinity of
a target frequency has not been estimated. It is usual in MT, to apply a small window
over the target frequency (f ) (Figure 3.3) and average the cross and auto spectra with
respect to this window. For example the cross spectra between E x and H y at the target
frequency f from a single event spectra is obtained by,
Ex Hy (f )
1
=
M
M/2
X
k=M/2
Ex (f + k)Hy (f + k).win(k)
(6.20)
109
To give lesser weights to the harmonics farther from the central frequency, it is usual
to average over a window function. The number of harmonics to be averaged depends
upon the position of target frequency within the spectrum and the length of the time
series segment. If the time series were segmented into smaller sections, the number of
harmonics to be averaged over a particular frequency reduces. However, this will be
compensated by the increased amount of data due to larger number of sections. The
robust processing methods thus advocated the need for shorter time windows (Egbert
and Booker [1986]) . Even if the time series does not arise from a Gaussian process, it
is assumed that the Fourier coefficients (FC) that result from the summarization of time
series, follow a Gaussian distribution, (from Central Limit Theorem) (Chave et al. [1987],
Menke [1984]). So within a band it was felt reasonable to assume a complex Gaussian
distribution for the FCs and LS averages were used to estimate auto and cross spectra.
However, it will be shown in the next section that this assumption often fails due to the
presence of some unusual data. A weighted robust averaging is essential to remove the
outliers within a frequency band well.
6.3.2
Effect of frequency band width
In Figure 6.11 the concept of frequency band averaging is presented for varying window
lengths, effectively using more FCs in each window in Figure 6.11 (a) to (d). The term
frequency band refers to the narrow band around a target frequency, whereas the term
band 1, band 2, etc represents the measurement band. Parzen window (3.2.3.5; also see
Jenkins and Watts [1968]) is applied in frequency domain, whereas Hanning window is
applied in time domain (Bendat and Piersol [1971]). The upper four plots (a to d) show
the Parzen window with different sizes. Parzen window is preferable over a boxcar, as 1)
it gives less and less importance to farther harmonics from the target frequency and 2)
Boxcar ( 3.2.3.3) in frequency domain has a sinc (sin(x)/x) impulse response, which
is not desirable in spectrum estimation. The vertical broken lines are drawn through
each target frequency at which MT transfer functions are to be computed. The bottom
plot show a sample E x E x spectra for data collected in the frequency range 8 Hz to 0.25
Hz, with a sampling interval 32 Hz. The spectrum at each target frequency is usually
estimated by averaging the adjacent harmonics with respect to the window. Though
the window width seems to increase as the frequency decreases, the effective bandwidth
remains the same, considering the distribution of FC on a logscale (Wight and Bostick
[1980]).
In order to study the effect of frequency band averaging on MT transfer functions, the
xy and yx for a particular frequency were computed using different frequency window
lengths. Figure 6.12 show xy and yx values at 6Hz plotted as a function of number of harmonics used for the station VP13. The time series were subdivided into sections of 2048
length and robust procedure described in 6.2 is used to estimate the transfer functions.
Variance of each was estimated according to equation 6.17, taking into consideration the
modification to dof by the robust procedures. Two results from this computation are: 1)
As the estimate is made over longer window sizes, the expected variance decreases. This
can be seen in the form of reduced error bars for the estimates with larger number of
harmonics. 2) However, the reduction in variance comes at the cost of bias error. As the
110
Figure 6.11: Spectrum estimation in MT using band averaging. (a) to (d) windows in
frequency with different radii. Though the window radius seems to increase as period
increases, effective bandwidth remains the same, as spectrum in long period contains fewer
Fourier harmonics. (e) Sample Ex Ex spectrum in the band 4, with target frequencies
projected as dotted lines. It is common to have 10 12 target frequencies per band.
111
Figure 6.12: Apparent resistivities as a function of frequency window length. Triangles

represent xy and circles yx . Error bars represents 95% confidence interval. See text for
discussion.
window size increases the computed xy and yx values get biased down. The xy values,
which are around 7000 Ohm m at a frequency window with less than 100 harmonics, get
biased to 5000 Ohm m with number of harmonics > 250. This dependence of resistivity
values on frequency bandwidth has nothing to do with the geology and are clearly due to
noise. To be more specific, the noise power in the magnetic field auto spectra may not get
averaged out as the number of data point increase as they are squared and positive and
any outliers may worsen the situation. It is worthwhile to examine the distribution of
these residuals to check the assumption of Gaussian distribution for the spectra within a
band. Figure 6.13(a). shows the magnitude of a cross spectrum obtained by multiplying
H x with H y respecting the Parzen window in frequency domain, at frequency 6 Hz for
station VP13. Few frequency points with grossly different spectral amplitude than the
majority are seen. A Quantile-Quantile plot of spectra of real part of the cross spectrum
(Figure 6.13(b)) shows (inverted triangles - unweighted) large deviation from linear function at the higher quintiles, indicating the departure from a Gaussian distribution. This
shows that the individual spectral coefficients within a frequency band might not follow
a Gaussian distribution and an LS average is not desirable in such cases.
6.3.3
Robust estimation of spectra
Here, a robust weighting procedure is proposed to minimize the effect of outliers in the
estimation of cross and auto-spectra from a frequency band. Once spectra are estimated
in a robust sense, any type of section averaging procedures may be adopted for the estimation of MT transfer functions. Though theoretically possible, computation of transfer
functions using a pair of spectra sets from two adjacent FCs (Smirnov [2003]) is unstable
112
Figure 6.13: Concepts of robust band averaging. (a) Magnitude of cross spectrum between Hx and Hy for station VP13, around a target frequency 6 Hz. The spectra are
multiplied by a Parzen window. LS least square, RB Robust. (b) Quantile Quantile
plot for real part of the same cross spectrum. Inverted triangles unweighted, Circles
robust weighted. See text for discussion.
113
because of the chances of having same polarization. So minimization of residual noise

from an LS transfer function estimate is undesirable within a frequency band. Without
assuming any signal & noise components, a Huber weighting procedure was used, assuming that the spectra are random variables. For complex data, we used the magnitude to
define the weights (a discussion in this regard is given by Chave and Thomson [1989]. The
algorithm devised for robust spectrum estimation is as follows. At each target frequency
we have M number of FCs from two channels, in the vicinity of the target frequency, f.
We would like to compute the cross spectra between two channels, say Ex and H y from
M realizations of E x H y *, equally distributed around the target frequency, f.
M
M
...f...
(6.21)
2
2
Where function win is a Parzen window ( 3.2.3.5) and equation 3.22) with radius
M/2. Residuals r k were calculated by subtracting median (|S|) from each value in |S|. A
MAD scale estimate is obtained as in equation 6.12 as,

f
SExHy
= Ex Hyk win(k),
where k =

= med. rk med. rk
Then the Huber weights w k are used to down weight the data that exceeds

1 if rk
wk =
if rk >
rk
(6.22)
ri
.
(6.23)
If the outliers have been eliminated the spectra are the sum of almost normally distributed cross products. The robust estimate of spectra (here the cross spectra between
E x and H y ) are obtained as,
Ex Hy (f )
1
=
M
M/2
X
Ex (f + k)Hy (f + k).wk .W in(k)
(6.24)
k=M/2
It may be followed by iteration with Tukeys weighting ( 6.2.4.4) to reduce the

influence of large outliers. However, the experience shows that this may not be required
in majority of the cases. Figure 6.13 shows results of robust band averaging. Figure
6.13(b) shows Q-Q plot of robust weighted H x H y cross spectrum. Represented with
circles, the robust weighting clearly reduced the outliers at higher quantiles compared
to un-weighted (or LS) spectra (triangles). The LS estimate gave a value of 2.1563e-010
-3.9535e-011i nT2 /Hz for the cross spectrum, while the robust estimate gave 1.4426e-010
-2.4245e-011i nT2 /Hz, both at 6 Hz. This means the LS estimate gave a value 60%
higher than that of robust weighted estimate of spectrum at 6Hz. This reduction in
spectral amplitude (solid horizontal line, Figure 6.13(a)) by robust weighting could be
attributed to the elimination of few large outliers, which biased the LS estimate (broken
line, Figure 6.13(a)) up. Figure 6.14 shows a series of band averaging instances from the
MT data at station VP13. All the eight plots show cross spectrum estimation of between
Hx and Hy , for different target frequencies and different Parzen (Figure 3.3) radius sizes.
Note that X axis do not represent the frequency itself, but the wave number from FFT.
Y axis shows the magnitude of the cross spectrum in nT2 /(Hz). The two horizontal bars
114
show the averages from LS (broken) and robust (solid) weighting scheme. Column (a)
band averaging for four target frequencies and column (b) shows the band averaging at
single target frequency (6Hz) with different radius sizes. From these results it can be
seen that there exists a consistent bias error in spectrum estimation, which in turn may
corrupt the transfer function computed out of it as well. As the magnitude of the spectra
is a positive function, outlier contamination can only bias it upwards and that is exactly
seen in these plots, where robust weighting procedures always resulted in smaller values
as compared to LS. The robust band average always resulted in smaller magnitudes for
spectra. Therefore the need for robust band averaging is seen.
6.3.4
Flow chart
Flow chart for the proposed robust band averaging to estimate cross and auto spectral
matrices from time series is presented in Figure 6.15(a). The entire codes were written
in MATLAB [2001]. The input to the routine can be either standard 5 Channel single
station MT time series or 7 channel with remote reference channels. Pre-processing of
data involve removal of trend and bias ( 3.2.3.1) , sub-segmentation, Fourier transform
( 3.2.1.5) and calibration ( 1.3.4. There are two options for band averaging, viz. least
square (LS) and robust (RB). While LS perform conventional frequency band averaging
respecting a window around target frequency ( 3.2.3.5), RB will perform a weighted
averaging approach as presented in the above section (equation 6.24). The averaging
routines are called in confined loops (L, N and n x n). With a 5 channel MT time series
of 80 stacks of 1024 data points each, the averaging routine will be called 24,000 times
to estimate spectra at 12 target frequencies. This necessitates the optimization of the
processes within the averaging routine. For this reason, the robust weighting procedure
was designed as simple as possible and the final iteration by Tukeys weights was avoided.
As a result, the processing of a time series of comparable dimension as earlier stated, takes
1 minute on a 500 MHz Pentium PC. Further reduction in processing speed is possible
by rewriting the codes in C or C++. The routine will deliver a 4 dimensional spectral
matrix as output, which will form input to any section average processing, including
robust processing.
6.3.5
Validation
Four more processing schemes are defined based on this flow chart to validate the results
of robust band averaging. These are 1) single station LS band averaging (Process C),
2) single station robust band averaging, (Process D) 3) remote reference and LS band
averaging (Process E) and 4) remote reference and robust band averaging (Process F).
They are represented in the Figure 6.15(b) to (e) respectively. These processes deliver
the 4D spectral matrix, from the MT time series data. The process A, defined in 6.2.5,
which is a robust section averaging method employing a Jackknife initial guess will be
used to estimate MT transfer functions for all the four processes defined above. In order
to demonstrate the efficacy of the robust band averaging methods, station VP10 was
chosen, where, in addition to the single station data for all the bands, one session of
remote reference data were also available from station VP13. This enabled us to exam-
115
ine the performance of the proposed robust processing procedure with remote reference
processing as well. The results from four processing methods on the data set are presented in Figure 6.16. Figure 6.16(a) shows the xy component, where as yx component
is plotted in Figure 6.16(b). The values are not shown, as they are not much varied
by the processes described above. However, the values are included at a later stage
when the applications of the procedure is discussed. The quantity of remote reference
data for period range 4 sec to 128 sec (band 4) was insufficient for a through estimation.
However, this was included to get a complete evaluation. It can be seen from the figures
that the best and unbiased estimate for values are derived from process F (stars). The
values are clearly higher and less smoother than the other estimates, except for a few
data point between 10 and 100 Hz. The conventional robust remote reference processing
(E -solid line) , are similar to the results from process F, except for few data points
in dead band, where they are biased down compared to process F. The usefulness of
robust band averaging results are more evident in single station estimates (Process C &
D). Clearly both these process output values that are down biased compared to remote
reference data. However, the robust band averaging (Process D) improves the situation,
with the values that are nearer the their remote reference counter parts. This clearly
shows the effect of robustly averaging the spectra sets within a band. On single station
estimates, the robust band averaging reduces the down bias as compared to LS band
average. Even remote reference processing can benefit from robust band averaging, as
evidenced from the improved RR estimate of xy and yx values by robust band average
in the dead band range (1 to 10 seconds). This analysis could not be extended to all
other data sets, as reference data sets were not recorded at majority of stations due to
instrument problems.
In the second approach to validate the results of robust band averaging, the estimates
of values from robust band averaging were compared with E-referenced and H-referenced
estimators on spectra generated from conventional robust processing. Standard LS estimators have the property that E-reference (transfer function from admittance analysis)
are biased upward by un-correlated noise on the electric field components and the Hreference estimates are biased downwards by the noise in un-correlated noise in magnetic
field. (See 3 for more discussion). Obviously downward and upward biased estimates
give an envelope within which the true transfer function should lie (Jones et al. [1989]).
The up and down ( 3.3.1.3) biased estimates were computed from process C and
compared it with the result of Process D. Both these processes are done in single station
mode and the results from four stations are plotted in the figures 6.17 to 6.20. Figure
6.17 shows the data from station JN12. The up biased estimates from Process C shows
larger deviation near the dead band. This is common, as the s/n ratio at this frequency
range is low compared to other ranges. As clearly seen from Figure 6.17, the xy and
yx components are less biased from Process D. This is more pronounce near the dead
band (between 1 and 10 seconds). Near the period of 100 seconds, the xy (top panel)
values from Process C seems to be heavily biased, which to an extend restored by Process
D. Station VP16 is relatively noise free and xy and yx values are well defined over the
full bandwidth by conventional processing. Processing D results in an inner envelop of
up and down biased estimate within the same estimates by Process C (Figure 6.18).
The results from station KG02 (Figure 6.19) shows large deviations between the biased
116
estimates for periods > 1 seconds. Here also the robust band average results in reduced
split between both estimations. Perhaps the use of robust band averaging (Process D) is
well explained by the results from the processing of long period data from station OK18
(Figure 6.20). The xy values at the periods > 4 seconds are obviously affected by poor
s/n ratio as evidenced by the large split between the up and down biased estimates
of Process C, especially for periods > 10 seconds. The yx vales in the same period
range are not that affected with noise, as the up and down biased estimates follow same
curve. However, is evident in yx values between 0.1 to 10 Hz. Here again, the robust
band averaging (Process D) resulted in reduced split between the up and down biased
estimate.
6.4
Results and discussion
From the validation experiments carried out in the earlier section, it is established that
the robust band averaging reduces bias in the estimation of values and may even improve the remote referenced solution for transfer functions. In this section a combination
of robust band averaging (to produce robust spectral matrix) and robust section averaging (to estimate transfer functions) (Process D) were used to the same set of sites
used for validation. For discussion the and values and their associated telluric predicted coherencies (positive square root of coherence function, 2 ) derived from the above
combination were compared with that of conventional spectrum estimation with robust
transfer function estimation (Process C). The long period MT data were subjected to
ANN editing as detailed in 5. The results are shown in Figures 6.21 to 6.24. The top
panel in each Figure compares the coherencies, the middle panel compares the apparent
resistivities and the bottom panel compares the phase values.
6.4.1
VP14
This site was occupied in the southern part of the measurement area where the main
rock type exposed is gneiss ( see 4.2). The predicated coherencies from process D are
consistently higher than that process C. The increase in predicted coherencies are more
clear in the frequency range 0.1 Hz to 10 Hz. This is also attested in the form of improved
estimates of and values in the bottom panels. However, in the xy coherencies from
process D for periods > 1 seconds are less than compared to Process C. (Figure 6.21(a))
6.4.2
TT08
The noise environment for the station TT08 was described in 5, in the context of application of ANN. The improvement in the estimation of and values by Process D
are evident in the middle and bottom panel of Figure 6.21(b). Frequency to frequency
variability of values were remarkably reduced for the periods > 1 sec. The Process
D also increased the values of predicted coherencies as shown in the top panel. The
improvement in values indicate a better estimate of cross spectra sets, where as the
reduction in bias may be attributed to the better estimation of auto spectra. This is
117
stated to emphasize that robust band average give cleaner spectral matrices, from which
the robust section average can further, improve upon.
6.4.3
TT04
This station, also falls in the industrial belt near the Cauvery river. The xy components
( and ) between 0.1 second and 10 seconds were seriously affected by a coherent noise
source, which distorted these curves as shown in Figure 6.22(a). While discussing the
results of robust processing with jackknife initialization (Process A) ( 6.7), it was seen
that Process A resulted in recovering the heavily distorted values of this station. However, Process A induced a bias in long period yx values. The robust processing with
robust band averaging and jackknife initial guess substantially reduced this bias as seen
in the middle panel of Figure 6.22(a). The improvement in predicted coherencies are
evident, perhaps the difference is less compared to TT08 or VP14
6.4.4
OK18
The station lies in the south end of the measurement corridor (see Figure 4.1). Electric
water pumps for domestic and small scale agricultural purposes were operational in the
area, during the measurements. This resulted in increased noise activity in the electric
fields, mainly for long period data. The high frequency data were influenced by power
line related noise. It can be seen from the plot (Figure 6.22(b)) that the xy component of
the data are more affected than yx components. The heavy bias for xy values at periods
> 1 second are associated by a low in predicted coherency values confirm this. Process
D gave reduced bias in values and smoother values, which are consistent with their
orthogonal counterparts. The increase in predicted coherencies brought by Process D is
also evident in Figure 6.22(b).
6.4.5
JN12
The MT station JN12, lies in the northern part of the measurement corridor, near the
town Dharmapuri. The area has as average long period noise/(signal + noise) ratio of
0.35 (Figure 4.8) and relatively less noisy. The resistivity values monotonously reduce
with increase in period (Figure 6.23(a)). The Process C poorly estimated the xy values
especially in the frequency range 100 Hz and 10Hz. This resulted from power related
noise that biased up the xy values near 50 Hz. The robust band and section averaging
with jackknife initial guess (process D) gave a better estimation of all but two xy values
in the band. However, there was not much improvement obtained for long period values
from robust processing. For xy values at periods > 4 seconds, there small but perceptible
reduction in bias by Process D. This is also attested by the increased predicted coherencies
values as shown in the top panel of Figure 6.23(a).
118
6.4.6
VP16
The site falls in the southern portion of the measurement corridor, over exposed gneisses
(Figure 4.1). The moderate geomagnetic indices (Figure 4.6) for the period of survey
were favorable for the occurance of MT signals in the long period. Average noise ratio
(Figure 4.8) of 0.4 for the long period data also indicate that the site is relatively noise
free. The Process C resulted in almost same estimate as Process D as shown in the middle
and bottom panel of Figure 6.23(b). The predicated coherencies for the estimates from
process D do have higher values as compared to the that of process D. However, near
10 seconds, the robust band averaging resulted in a small kink (upward) which was not
there in the estimates from Process C.
6.4.7
OK16
This station lies in the southern end of the measurement corridor. The site is away from
the major cultural noise sources as shown in Figure 4.8. However, the effect of noise
from other sources are evident in the and values (Figure 6.24(a)) from Process C.
Remarkable improvement in the smoothness of values resulted from the application
of process D. The improvement is also evident in xy values , as Process D resulted in
smoother and less biased estimates especially in the frequency range 1 Hz to 10 Hz. This
is also attested by the increased telluric predicted coherencies as shown in the top panel
of Figure 6.24(a).
6.4.8
VP12
Situated far south of the major industrial zone along the Cauvery river, the station VP12
is relatively free from cultural noise sources. The and values from Process C and
D are almost same, except for four frequencies in dead band (period 1 second and 10
seconds) as shown in Figure 6.24(b). In this frequency range, the Process D produced
values that are higher than that of Process C. However, as there are no appreciable
increase in predicted coherencies from Process D in this period range, increased estimates
in values cannot be confirmed as an improvement. Between 10 and 100 Hz, the Process
D resulted in better estimation of xy values as compared to that of Process C. The
predicted coherencies from Process D are slightly higher than that of Process D in the
frequency range 10 Hz to 0.1 Hz.
6.4.9
Comparison of results from the Vellar - Palani profile in

SGT
The results presented in Figure 6.25 compare the MT data sections along the Vellar Palani profile (the broken line in the Figure 4.1) over the Southern Granulite Terrain.
The left panels show the xy and xy sections from ProcMT, the MT data processing
software of Metronix (Ellinghaus [1997]) and the right panels show the results from the
proposed approach. The data are not smoothed or contoured to better represent the
reality. The data that are missing/ out of range are represented as white patches. The
119
length of the FFT and parzen windows were kept identical for both schemes. For ProcMT,
the available time series were manually edited and the MT impedances were estimated
by Robust-M processing. While using the scheme proposed here, the time series were
not manually edited. The long period data were subjected to ANN editing. The cross
and auto spectra were computed by robust band averaging. The MT impedance were
estimated by robust processing with jackknife initialization.
The resistivity data ranges from 101 to 105 Ohm-m, typical of the highly resistive crust.
The high frequency data (>1 Hz) shows a relatively resistive crust in the northern part
of the profile (especially north of station VP10). However, point to point discontinuities
are evident along majority of the stations for the data processed by ProcMT (Figure
6.25a). Figure 6.25b shows the xy section derived from the scheme proposed. A greater
smoothness for data are now evident between stations, and the two major resistivity
blocks are clearly visible. In addition, the smaller resistive structures within the high
frequency data below the southern part of the profile are more evident. The improvements
are not so evident for the low frequency data (<1 Hz). The apparently conductive patches
appearing in both the plots along 0.1 Hz seem to be effect of unresolved noise (VP19,
VP11, VP06,VP03,VP02 and VP01). However, reasonable improvement in the point to
point continuity are evident for data from other stations (VP1,VP10,12-17). The bottom
panels show the corresponding phase data. Phase data has low values in high frequency
ranges, systematically increases to higher values below 1Hz. The phase data processed
by ProcMT (Figure 6.25c) shows a near random distribution of values among most of the
frequency/phase domain. Eventhough the proposed MT processing scheme does not use
manual editing of the time seires, the results (Figure 6.25d) are more smooth, and with
a better point to point continuity. These results show the efficacy of the MT processing
scheme proposed in the thesis.
6.4.10
Discussion
The results presented above clearly show the efficacy of the robust band averaging in reducing the uncertainties in transfer function estimation. Robust band averaging attached
to the robust section averaging using Jackknifed initial guess (Process D) is superior to
the other combinations. Thus the chapter suggest two important improvements to the
robsut estimation of MT transfer functions. They are: (1) The role of initial guess is
very important in robust processing and the conventional way of assigning a LS estimator to this task often fails in the presence of large outliers. In such cases, the initial LS
estimator locates the solution too far away from the true solution such that the iterative
robust scheme may not reach the true estimate starting from it. Where as Jackknife
estimates offer simple and efficient way to initialize the transfer functions. This makes
the job of iterative weighting procedures easy in converging to the true estimate. (2)
With the assumption that as the Fourier transform is summarization of a large number
of time variables, the individual Fourier Coefficients (FC) are often assigned a complex
Gaussian distribution according to Central Limit Theorem (Menke [1984]). Following
this assumption, MT spectral estimates were made by Least Square averaging of a band
of frequencies in the vicinity of the target frequency. It was shown in this chapter that
the effect of outlier, which was well recognized in the section averaging of MT data, need
120
to be accounted in frequency band averaging as well. The bias effect in values caused
by ignoring this factor is demonstrated on MT data collected from SGT. The robust
weighting approach proposed in this chapter can alleviate this problem to an extent as
demonstrated by its application to the field data. The robust weighting procedures for
spectral estimation may easily extended to robust processing as well.
In summary, the Robust band averaging deals with the averaging of the adjacent
Fourier coefficients around a target frequency while estimating the auto and cross powers
(or power spectral densities PSD). As shown in Figure 6.12, the apparent resistivity
value has a systematic dependency on the number of harmonics averaged. This means
the conventional way of computing PSD by band averaging leads to biased estimates
of apparent resistivity values. Robust band averaging, proposed , here accounts for the
abnormal Fourier coefficients while averaging around a target frequency. This clearly
results in
1. Minimizing the effect due to non Gaussian distribution of Fourier coefficients. Once
the spectral coefficients are robustly weighted, they are more appropriate for a LS
averaging (Figure 6.13)
2. The robust band averaging results in improved (bias less) estimates of power spectra
(Figure 6.14)
3. Finally the MT transfer functions computed from the above PSDs have low bias
compared to the conventional processing (Figure 6.16 and 6.3.5)
It may be concluded that the two new approaches made in the robust processing of
magnetotelluric data clearly showed it effectiveness while applied to a large volume of
data collected from SGT and are easily adaptable to the conventional MT time series
processing routine.
121
Figure 6.14: Comparison of robust and least square band averaging for different frequencies and band (Parzen) radius for station VP13. Magnitude of Hx Hy (nT 2 /Hz) cross
spectra are plotted for all the cases. X-axis show the frequency bins. Solid line represents
robust average and broken line represents Least Square average.The left column represent the the band averaging for different target frequencies, where as the right column
represents the band averaging with different radius length for the same target frequency.
See text for discussion
122
Figure 6.15: Flow chart representing robust band averaging. Gray area represents the
proposed processing routine (a) Shows the different steps in robustly estimating the
spectral matrix from raw time series Schematic diagrams (b) to (e) shows four processing
schemes defined. SS single station, RR remote reference, RB robust band averaging,
LS Least square
123
Figure 6.16: Results of single station (SS) and remote reference (RR) processing for
station VP10. Process C (SS) and E (RR) use least squares band averaging, whereas
process D (SS) and F (RR) use robust band averaging. The MT transfer functions were
derived from a robust section averaging 6.2.4 from the spectra sets produced by Process
C to F. See text for discussion.
Figure 6.17: Comparison of upward and downward biased estimates of apparent resistivities for JN12 for processes C and D (a) XY component and (b) YX component. UP - up
biased (E- reference) and DN down biased (H-references). Symbols are same for both
plots. See text for discussion.
124
Figure 6.18: Comparison of upward and downward biased estimates of apparent resistivities for VP16 for processes C and D (a) XY component and (b) YX component. UP
- up biased (E- reference) and DN down biased (H-references). Symbols are same for
both plots. See text for discussion.
Figure 6.19: Comparison of upward and downward biased estimates of apparent resistivities for KG02 for processes C and D (a) XY component and (b) YX component. UP
125
Figure 6.20: Comparison of upward and downward biased estimates of apparent resistivities for OK18 for processes C and D (a) XY component and (b) YX component. UP
126
Figure 6.21: Comparison of telluric predicted coherency, apparent resistivity and phase
values from robust processing with and without robust band averaging for stations VP14
and TT08. The explanation for symbols are given in legend. See text for discussion.
127
values from robust processing with and without robust band averaging for stations TT04
and OK18. The explanation for symbols are given in legend. See text for discussion.
128
values from robust processing with and without robust band averaging for stations JN12
and VP16. The explanation for symbols are given in legend. See text for discussion.
129
values from robust processing with and without robust band averaging for stations OK16
and VP12. The explanation for symbols are given in legend. See text for discussion.
130
a)
b)
c)
d)
Figure 6.25: Comparison of results from two different processing scheme for MT data:
Left panel shows the results from ProcMT and the right panel shows the results from the
robust processing and ANN editing.
Chapter 7
Discussion and conclusions
131
132
7.1
Introduction
Magnetotellurics has been established as a versatile exploration tool, serving the needs
of shallow resource exploration as well as deeper crustal and lithospheric studies (Vozoff [1991], Jones [1992]) Continuing improvements in instrumentation, data acquisition
procedures, data processing and modelling of MT measurements provide a fertile environment for experimentation and application. In order to determine the sub-surface
electrical conductivity, it would be necessary to obtain the best possible transfer functions from the data collected. Developments in this aspect of MT have been outlined in
Chapters 3, 4, 5 and 6. Just as in the preparations for a MT experiment it is possible
to consider several scenarios, models and contingencies and attempt to plan the optimal
survey. Similarly it is observed in this work that while processing the time series, it is
necessary to study all the characteristics possible, in space, time and frequency domain
and attempt through a suite of processing routines to obtain the best or optimal curves.
It is concluded here that the study of data characteristics after the survey also shed light
on the most suitable methods of processing. It is clearly demonstrated in Chapters 5 and
6 that the same routine may not be optimal for different data sets with different noise
characterization (Egbert and Livelybrooks [1996]). Further, for the same measurement,
one process would eliminate some types of noise and another process the other.
Significant developments for MT transfer function estimation in magnetotellurics, in
the past three decades were the introduction of remote reference processing (Gamble et al.
[1979]) and robust processing (Egbert and Booker [1986]). This was complementary to the
requirement of more precise MT estimates by the new and sophisticated data modeling
algorithms developed simultaneously. Advances in both these fields accelerated in 1980s
with the availability of cheap and fast microprocessors. Thus the tremendous development
in MT methodologies in the past three decades, greatly owe to the advances in digital
computing. The effective depth of MT probing ranges from few hundred meters typical
for geo-technical investigation (Garcia and Jones [2002]) to more that hundred kilometers
for deep crustal studies (Chen et al. [1996]). This greater range of depths demands wider
bandwidth of measurements, which in turn demands better instruments/sensors. To
estimate the transfer function for full bandwidth, as precisely as possible, data processing
algorithms need to be improved as well.
7.2
Summary
The time series processing steps introduced in this thesis were applied to the following:
1. Identifying the noise sources in space domain ( 4),
2. Discrimination of bad segments of time series data against good segments in time
domain ( 5) and
3. Effectively down weighting the unusual data in a robust sense in time - frequency
domain ( 6).
133
In the first instance, an attempt was made to characterize the signal and noise components in the MT data collected from SGT. The MT data were analyzed in time, frequency
and space domain in this regard. The comparison of indices of geomagnetic activity with
the long period coherence showed their general agreement. The effect of noise from an industrial zone was evident in the long period data, while analyzing the noise/(signal+noise)
for all the sites in relation with the known cultural noise sources.
The second approach was made, in response to another requirement of MT exploration
- an automated method of processing. For example, the data acquired in SGT, amounted
to 106 real numbers of data per site, and with 80 such stations, the magnitude of data
reduction in MT is a major task. In such scenario, the editing of MT time series becomes
challenging and may consume more that 80 % of the time spent on time series analysis.
The artificial neural network approach introduced in 5, is a very promising solution to
this. It not only proved to be a good replacement to manual inspection of time series,
but turns out to be a stable and objective decision-maker as well.
In the third approach made in this thesis, the robust processing algorithm was critically reviewed and two steps were introduced to further improve this type of estimators
in the presence of severe noise contamination as found in SGT. The two approaches
namely Jackknife and Robust band averaging were used to replace simple Least Square
procedures, that were still in use in two steps of robust processing. The effects of a few
outliers on MT data are well recognized and robust processing was essentially introduced
to tackle this (Eiesel & Egbert, 2001, Chave et al 1987, Chave & Thomson, 1989). Some
of the computational steps retained LS averaging, overlooking its effect. The effect of
unusual data at the frequency band averaging level has been evaluated in this thesis and
a robust band averaging technique suggested to alleviate this problem. In the same manner, Jackknife estimates were used to find an initial guess for robust processing, replacing
the commonly employed Least Squares.
Holistically, the last two approaches may be considered as a natural outgrowth of
MT time series processing procedures, against the back drop of increasing computational
power and improving instrumentation. Neural networks, with their computational complexity, were rarely used in geophysics in 1980s. With the availability of fast and cheap
microprocessors, neural networks are widely used in a variety of applications in geophysics
today. When introduced, robust processing (in MT) concentrated on the averaging of
transfer functions. Later, by going back one step in the processing algorithm, Egbert
[1997] proposed robust estimation of global spectral density matrix (SDM) from all the
available SDMs. The approach made in this chapter may treated as one more step back
in the processing level, wherein, the individual SDMs from a single time section is also
estimated in robust sense.
7.3
Neural Network and Robust processing
Neural network approach introduced in Chapter 5 and the robust processing techniques
(Chapter 6) discriminate signal against noise. The fundamental differences between these
two methods are
1. ANN based editing is performed in time domain alone, where as robust processing
134
is done on spectra sets frequency domain, generated from short time segments.
2. ANN based editing is used to discriminate signal patterns, correlation structure
and amplitude ratio to search for and select signal elements from the subdivided
time series. Robust processing attempts to down weight data that deviate from the
fit with the transfer function (large residuals).
3. Decision of ANN editing is binary, i.e. it either accepts or rejects the time segment.
Robust processing algorithm tries to minimize the effect of larger residuals, with
proportional weights and the decision is analog. Tukey weighs at the final iteration
in robust processing may be considered binary. However, it affects only a part of
the data.
4. ANN requires previously saved intelligence (weights) to perform processing, where
as robust processing requires none.
In the figure 7.1 & 7.2, the results of processing MT data from two sites are given. The
MT time series collected were subjected to the robust processing algorithm described in
6 (Process D), with and without prior ANN editing. The derived apparent resistivity &
phase for stations JN10 and OK18 are plotted in Figure 7.1(a) &7.2(a). In Figure 7.1(b)
& 7.2(b), the comparison of neural network picking and robust weights are compared
for all the stacks. The upper panel in Fig Figure 7.1(b) & 7.2(b) shows the neural
network decision as a function of stacks. In the lower panels, the robust weights from
the final step in robust processing ( 6) are presented as a function of frequency and
stacks. Dark areas indicates good data, white indicates bad data and the gray levels
in between indicate intermediate data. As can be deduced from these plots, the robust
processing with and without NN editing at station JN10 seems to yield the same results.
On closer inspection of the network picking and robust weights in same figure(Figure
7.1(b) & 7.2(b)), it is seen that from stacks 1 to 15, the neural network consistently
gave zero and the corresponding robust weights also gave smallest values, throughout
time and frequency scale with the exception of stack 3 & 4. In the same way stacks
around 30 also gave similar results for robust and NN processing. It may be concluded
that the similar results from robust processing with and without neural network editing
is due to fact that the signal discrimination were similar, though the approaches were
made in different domains. In the second example the processing results from stations
OK18 is presented. The robust processing alone resulted in biased estimate of xy values
and scattered xy values (open circles) with large error bars. However robust processing
after neural network editing resulted in better estimation of xy and xy (stars). (Figure
7.2(a)). On verifying the neural network output and robust weights as a function of stacks
(Fig 7.2(b)), it is observed that the two considerably vary especially between stacks 20
and 60. The neural network rejected around 12 stacks in this interval, whereas the robust
weights are near the value 1 for all frequencies. However between stacks 60 and 80, the
robust procedures down weight around 5 stacks, where as the neural network did not
reject any. The reason is that the outliers and/or coherent noises which are obvious in
time domain between stacks 20 and 60, failed to produce larger outliers and thus were
not down weighted by robust procedures.
135
a)
b)
Figure 7.1: Comparison of robust processing (RB) with and without neural network (NN)
editing for Station JN12. (a) Apparent resistivity and phase values from both processing
schemes. (b) Comparison of robust and neural network weights.
Those data, which the neural network passed on to robust processing between stacks
60 an 80 with noise that was not evident in time domain, were down weighted by robust
procedures. This clearly shows the need to combine the two techniques in order to
discriminate / down weight a majority of noisy data.
136
a)
b)
Figure 7.2: Comparison of robust processing (RB) with and without neural network (NN)
editing for Station OK18. (a) Apparent resistivity and phase values from both processing
schemes. (b) Comparison of robust and neural network weights.
7.4
Conclusions
Major conclusions that can be drawn from the thesis are

1. The magnetotelluric data collected from SGT encountered severe interference from
cultural noise sources, and their effects are fully understood only if analyzed in a
137
multi- domain approach. Routine application of any one of the data processing
methods is bound to give inferior results in such cases.
2. In time domain, the application of artificial neural networks clearly demonstrated
its usefulness as an effective signal discriminator, even in the presence of large noise
contamination.
3. The prominent weak points in robust processing of MT data are the uses of least
square estimators in two of the steps, namely estimation of spectra and transfer
function initialization.
4. The application of a non-parametric estimator called Jackknife to initialize MT
transfer functions, resulted in improved robust estimation of MT transfer functions. Robust band averaging reduces the uncertainties in estimating the spectra of
magnetic and electric field components. Robustly estimated spectra always result
in better MT transfer functions.
5. On applying these methods to a large numbers MT stations in SGT, the effectiveness of a combined use of neural network and robust processing, in eliminating
severe noise was established and demonstrated.
7.5
Suggestions for future work
The natural signal sources for MT data >1Hz (AMT) are the world wide thunderstorms
( 2). These signals are manifested in high frequency MT time series (band1 of table
2.1) as transient bursts. It is well known that the MT apparent resistivity values are
systematically biased down at frequencies > 103 Hz. The frequency range 103 to 104 Hz
contains the dead band of AMT (Garcia and Jones [2002]), where the natural signal
energy is very low compared to the higher and lower frequencies (Figure 2.1). The
transient signal such as we found in Fig 2.4, has enough spectral energy in the AMT
dead band. However, there are two reasons, why they fail to improve the MT estimates
in the dead band: 1) The MT signals appear as transient envelopes of signals, on the
background of power transmission harmonic noise (and 50 and 150Hz, in this case) and
their occurrences are random. 2) It is widely recognized that FFT fail to perform spectral
analysis on transient signals, as the amplitude spectra thus derived do not give any
temporal information. Due to these reasons, the sparse signal activity in this frequency
range gets down weighted in spectral estimation as well as in robust processing. It would
be worthwhile to explore the use of the wavelet transform (Kumar and Foufoula-Georgiou
[1997] to discriminate transient signals in the high frequency MT time series against the
majority of background noise. Wavelet transforms have the unique ability to retain
temporal characteristics, while giving the spectral information. It is thus important to
estimate MT transfer function in dilatation translation domain of wavelet transform
(somewhat equivalent to frequency-time domain), as it allows us to effectively isolate
energetic signal events both in frequency and time domain.
Thus it may be seen that continuing improvement in MT data acquisition, processing
and modeling complement each other with the need for ever-more accurate imaging of the
138
Earths sub-surface structure. In the case of MT data processing, novel applications of

computational and statistical advances in signal processing, continue to be implemented
as in the work outlined above. Multi-domain analysis of signal discrimination would be
favored as the benefits of such approaches are validated by more data analysis.
Bibliography
D. Allsopp, M. Burke, D. Rankin, and I. Reddy. A wide band magnetotelluric recording
system. Geophys. Prosp., 22:272278, 1973.
O. Altinay, E. Tulunay, and Y. Tulunay. Forecasting ionospheric critical frequency using
neural networks. Geophys. Res. Lett., 24:1467 1470, 1997.
R. Bailey. Inversion of the geomagnetic induction problem. Proc. Roy. Soc. Lond., 315:
185194, 1970.
M. Balser and C. Wagner. Observations of earth-ionosphere cavity resonance. Nature,
188:638 641, 1960.
R. J. Banks. The effects of non-stationary noise on electromagnetic response estimates.
Geophys. J. Int., 135:553563., 1998.
J. S. Bendat and A. G. Piersol. Random data: Analysis and measurement procedures.
John Wiley & Sons, New York., 1971.
F. K. Boadu. Inversion of fracture density from field seismic velocities using artificial
neural networks. Geophysics, 63:534545., 1998.
L. Cagniard. Basic theory of the magneto-telluric methods of geophysical prospecting.
Geophysics, 18:605635., 1953.
C. Calderon-Macias, M. K. Sen, and P. Stoffa. Artificial neural networks for parameter
estimation in geophysics. Geophys. Prosp., 48:2147., 2000.
L. Chaize and M. Lavergne. Signal et bruit en magnetotellurique. Geophys. Prosp, 18:
6487., 1970.
A. D. Chave and D. J. Thomson. Some comments on magnetotelluric response function
estimation. J. Geophys. Res., 94:1421514225., 1989.
A. D. Chave and D. J. Thomson. A bounded influence regression estimator based on the
statistics of the hat matrix. Appl. Statist., 52(3):307 322., 2003.
A. D. Chave, D. J. Thomson, and M. Ander. On the robust estimation of power spectra,
coherences and transfer functions. J. Geophys. Res., 92:633648, 1987.
139
140
L. Chen, J. Booker, A. Jones, N. Wu, M. Unsworth, W. Wei, and H. Tan. Electrically

conductive crust in southern tibet from indepth magnetotelluric surveying. Science,
274:16941696., 1996.
J. F. Claerbout and F. Muir. Robust modeling with erratic data. Geophysics, 38(5):
826844., 1973.
J. Clarke, G. T. D., W. M. Goubau, R. Koch, and R. Miracky. Remote-reference magnetotellurics equipment and procedures. Geophys. Prosp., 31:149170., 1983.
S. C. Constable, R. L. Parker, and C. G. Constable. Occams inversion: A practical algorithm for generating smooth models from electromagnetic sounding data. Geophysics,
52(3):289300., 1987.
H. Dai and C. MacBeth. Automatic picking of seismic arrivals in local earthquake data
using an artificial neural network. Geophys. J. Int., 120:758 774., 1995.
D. L. Dekker and L. M. Hastie. Sources of error and bias in magnetotelluric depth
sounding of the Bown Basin. PEPI, 25:219225., 1981.
S. A. Drury, N. B. W. Harris, R. W. Holt, G. F. Reeves-Smith, and R. T. Wightman.
Precambrian tectonics and the evolution of South India. J. Geol., 92:320., 1984.
J. Dungey. Electrodynamics of the outer atmosphere in the physics of the ionosphere.
The physics of the Ionosphere, The Physical Society (London), page 229, 1955.
W. D. E. Mt/emap Data Interchange Standards, Revision. Society of Exploration Geophysicists, USA., Revision 1.0, 1988.
B. Efron. The jackknife, the bootstrap and other resampling plans. Society for Industrial
and Applied Mathematics, Philadelphia, 1982.
G. D. Egbert. Noncausality of the discrete-time magnetotelluric impulse response. Geophysics, 57:13541358., 1992.
G. D. Egbert. Robust multiple station magnetotelluric data processing. Geophys. J. Int.,
130:475496., 1997.
G. D. Egbert and J. R. Booker. Robust estimation of geomagnetic transfer function.
Geophys. J. Roy. astr. Soc., 87:173194., 1986.
G. D. Egbert and D. W. Livelybrooks. Single station magnetotelluric impedance estimation: Coherence weighting and the regression M-estimate. Geophysics, 61:964970,
1996.
M. Eisel and G. D. Egbert. On the stability of magnetotelluric transfer function estimates
and the reliability of their variances. Geophys. J. Int., 144(1):6582., 2001.
A. Ellinghaus. PROCMT - Users guide, Revision 2. Metronix GmbH, Braunschweig,
Germany, 1997.
141
T. Ernst, E. Y. Sokolva, I. M. Varentsov, and N. G. Golubev. Comparison of two techniques for magnetotelluric data processing using synthetic data sets. Acta Geophysica
Polonica, XLIX (2):213243., 2001.
B. C. Fish and T. Kusuma. A neural network approach to automate velocity picking.
in Society of Exploration Geophysicists 64th annual international meeting; Technical
program, expanded abstracts with authors biographies, 64:185188., 1994.
B. P. Flannery, W. H. Press, S. A. Teukolsky, and W. T. Vetterling. Numerical recipes in
c. The art of scientific computing, Cambridge University Press, UK, page 1020, 1992.
T. D. Gamble, W. M. Goubau, and J. Clarke. Magnetotellurics with a remote reference.
Geophysics, 44:5368, 1979.
X. Garcia, A. Chave, and A. Jones. Robust processing of magnetotelluric data from the
auroral zone. J. Geomag. Geoelectr, 49:14511468., 1997.
X. Garcia and A. Jones. Atmospheric sources for audio-magetotelluric sounding. Geophysics, 67 (2):448 458., 2002.
L. Garth and H. Poor. Detection of non - gaussian signals: a paradigm for modern
statistical signal processing. Proc. IEEE, 82:1060 1095., 1994.
M. Ghil, M. R. Allen, M. D. Dettinger, K. Ide, D. Kondrashov, M. E. Mann, A. W.
Robertson, A. Saunders, Y. Iian, F. Varadi, and P. Yiou. Advanced spectral methods
for climatic time series. Rev. in Geophysics, 40(1):141, 2002.
W. M. Goubau, T. D. Gamble, and J. Clarke. Magnetotelluric data analysis: Removal
of bias. Geophysics, 43:11571166., 1978.
R. W. Groom and R. C. Bailey. Decomposition of magnetotelluric impedance tensors in
the presence of local three-dimensional galvanic distortion. J. Geophys. Res., 94(B2):
19131925., 1989.
GSI. Geological and Mineral map of Tamil Nadu and Pondicherry (scale 1:0.5 million).
Geological Survey of India, Calcutta, India, 1995.
F. Hampel, E. M. Ronchetti, P. J. Rousseeuw, and W. Stahel. Robust statistics: The
approach based on influence functions. John Wiley & Sons Inc. , New York, USA.,
1986.
T. Harinarayana, D. N. Murty, S. P. E. Rao, C. Manoj, K. Veeraswamy, K. Naganjaneyulu, K. K. A. Azeez, R. S. Sastry, and G. Virupakshi. Magnetotelluric Field
Investigations in Puga Geothermal Region, Jammu and Kashmir, India: 1-D Modeling. Geothermal Resources Council Transactions, 122:393397, 2002.
T. Harinarayana, K. Naganjaneyulu, C. Manoj, B. P. K. Patro, S. K. Begum, D. N.
Murthy, Madhusudana-Rao, V. T. C. Kumaraswamy, and G. Virupakshi. Magnetotelluric investigation along Kuppam - Palani geotransect, South India - 2-D modeling
results. Memoir Geological Society of India, 50:107124, 2003.
142
M. Hatting. The use of data-adaptive filtering for noise removal on magnetotelluric data.
PEPI, 53:239254., 1989.
W. Hernandez and J. Jacobs. Discussion on enhancement of signal-to-noise ratios in
magnetotelluric data, by kao dw and rankin d. Geophysics, 44:1594 1596, 1979.
D. B. Hoover, F. Frischknecht, and C. L. Tippens. Audiomagnetotelluric sounding as a
reconnaissance exploration technique. J. of Geophys. Res., 81:801809., 1976.
G. M. Jenkins and D. G. Watts. Spectral analysis and its applications. Holden-Day, San
Francisco, California., 1968.
A. Jones. Electrical conductivity of the continental lower crust. in D.M Fountain et al.
(eds). The continental lower crust, Elsevier:81143., 1992.
A. Jones, A. D. Chave, G. D. Egbert, D. Auld, and K. Bahr. Comparison of techniques for
magnetotelluric response function estimation. J. Geophys. Res., 94(B10):1420114213.,
1989.
A. G. Jones. Transformed coherence functions for mutivariate studies. IEEE trans. ASSP,
29 (2):317319., 1981.
A. G. Jones, B. Olafsdottir, and J. Tikkainen. Geomagnetic induction studies in Scandinavia III : Magnetotelluric observations. Jour. Geophys., 54:3550., 1983.
A. Junge. Characterization of and correction for cultural noise. Surv. Geophys., 17:361
391., 1996.
D. W. Kao and D. Rankin. Enhancement of signal to noise ratio in magnetotelluric data.
Geophysics, 42:103110., 1977.
R. Karmann. Search coil magnetometers with optimum signal-to-noise ratio. in Vozoff,
K. (ed) Magnetotelluric methods: SEG Geophysics Reprint Series 5, pages 215218,
1977.
A. A. Kaufman and G. V. Keller. The magnetotelluric sounding method. Elsevier Scientific Publishing Company, Amsterdam, 1981.
H. C. Koons and D. J. Gorney. A neural network model of the relativistic electron flux
at geosynchronous orbit. J. Geophys. Res., 96(A4):55495556., 1991.
P. Kumar and E. Foufoula-Georgiou. Wavelet analysis for geophysical applications. Reviews in Geophysics, 35(4):385412., 1997.
G. Kunetz. Processing and interpretation of magnetotelluric soundings. Geophysics, 37:
10051021., 1972.
J. C. Larsen. Transfer functions: Smooth robust estimates by least squares and remote
reference methods. Geophys. J. Int., 99:645663, 1989.
143
J. C. Larsen, R. L. Mackie, A. Manzella, A. Fiordelisi, and S. Rieven. Robust smooth

MT transfer functions. Geophys. J. Int., 124:801819., 1996.
H. Lundstedt. Solar origin of geomagnetic storms and prediction of storms with the use
of neural networks. Surv. Geophys., 17:561573., 1996.
F. Luo and R. Unbehauen. Applied neural networks for signal processing. Cambridge
University press, Cambridge, UK., page 123, 1997.
J. Macnae, Y. Lamontagne, and W. G. F. Noise processing technique for time domain
electromagnetic systems. Geophysics, 49:934948., 1984.
T. Madden. Spectral, cross spectral and bispectral analysis of low frequency electromagnetic data. ed. Bleil D.F., Pentium Press, New York., 1964.
T. Mahadevan. Deep Continental Structure of India: A Review. Geol. Soc. India Memoir,
28:569, 1994.
C. Manoj and N. Nagarajan. The application of artificial neural networks to magnetotelluric time series analysis. Geophys. J. Int., 153:409423., 2003.
M. Mareschal, R. D. Kurtz, and R. C. Bailey. A review of electromagnetic investigations
in the Kapuskasing Uplift and surrounding regions; electrical properties of key rocks.
Canadian Journal of Earth Sciences, 31(7):10421051., 1994.
MATLAB. Users manual, version 6, mathworks inc. MathWorks Inc., Natick, USA.,
2001.
M. D. McCormack, Z. D. E., and D. D. W. First-break refraction event picking and
seismic data trace editing using neural networks. Geophysics, 58:6778., 1993.
M. L. McCraken, K. G. Oristaglio and G. W. Hohman. Minimization of noise in electromagnetic exploration systems. Geophysics, 51:819832., 1986.
G. W. McNeice and A. G. Jones. Multi-site, multi-frequency tensor decomposition of
magnetotelluric data. Geophysics, 66:158173, 2001.
W. Menke. Geophysical data analysis, discrete inverse theory. Academic Press Inc,
Orlando, Florida, USA., 58:6778, 1984.
Metronix. Geophysical Measurement System (GMS05). Operating Manual, Revision 3.1,
Metronix GmbH, Germany, page 112, 1997.
D. C. Mishra. Geophysical evidence for a thick crust south of Palghat-Tiruchi gap in the
high grade terrains of South India. Jour. Geol. Soc. India, 33:7981., 1988.
D. C. Mishra and M. Venkatarayudu. Magsat scalar anomaly map of India and part
of Indian ocean - Magnetic crust and tectonic correlation. Geophys. Res. Lett., 12:
781784., 1985.
144
J. B. Molyneux and D. R. Schmitt. First-break timing: Arrival onset times by direct

correlation. Geophysics, 64:1492 1501., 1999.
A. M
uller. A new method to compensate for bias in magnetotellurics. Geophy. J. Int.,
142:257269., 2000.
M. E. Murat and A. J. Rudman. Automated first arrival picking; a neural network
approach. Geophys. Prosp., 40:587604., 1992.
N. Nagarajan. Application of robust estimation of transfer function for a magnetovariational array in Eastern India. in Deep Electromagnetic Exploration ed. KK. Roy et al,
Narosa Publishing House, India, page New Delhi, 1998.
S. M. Naqvi and J. J. W. Rogers. Precambrian Geology of India. Oxford Monographs in
Geology and Geophysics No 6, Oxford University press, Oxford., page 223, 1987.
N. Nityananda, A. K. Agarwal, and B. P. Singh. Induction at short period on the
horizontal field variation in the Indian Peninsula. PEPI, 15:59., 1977.
N. Nityananda and D. Jayakumar. Proposed relation between anomalous geomagnetic
variations and tectonic history of South India. PEPI, 27:223228., 1981.
G. Oettinger, V. Haak, , and J. C. Larsen. Noise reduction in magnetotelluric time-series
with a new siganl-noise separation method and its application to a field experiment in
the Saxonian Granulite Massif. Geophys. J. Int., 146:659 669, 2001.
A. L. Padilha, I. Vitorello, , and L. Rijo. Effects of the Equatorial Electrojet on magnetotelluric surveys: Field results from Northwest Brazil. Geophys. Res. Lett., 24(1):
8992., 1997.
J. Park and A. D. Chave. On the estimation of magnetotelluric response functions using
the singular value decomposition. Geophys. Jour. of Roy. astr. Soc., 77:683709., 1984.
L. Pedersen. Some aspects of magnetotelluric field procedures. Surveys in Geophysics, 9:
245257., 1988.
G. Petiau and A. Dupis. Noise, temperature coefficient and long time stability of electrodes for telluric observations. Geophys. Prosp., 28:792804, 1980.
M. Poulton, B. Stenberg, and C. Glass. Location of subsurface targets in geophysical
data using neural networks. Geophysics, 57:1534 1544., 1992.
E. Pulz and O. Ritter. Entwicklung einer kalibriereinrichtung fuer induktionsspulenmagnetometer (search coils) am GeoForschungsZentrum Potsdam. Scientific Technical
Report, GeoForschungsZentrum - Potsdam 10/01., 2001.
B. Radhakrishna. Suspect tectono-stratigraphic terrain elements in the Indian subcontinent. Jour. Geo. Soc.India, 34:124., 1989.
145
S. S. Rai, D. Srinagesh, and V. K. Gaur. Granulite evolution in south India - a seismic

tomographic perspective. Mem. Geol.Soc. of India, 25:235 265., 1993.
M. Rao. Study of geomagnetic pulsation characteristics and their solar cycle dependence
in the equatorial region India? Ph.D. thesis, Osmania University., 2000.
M. Rao, S. V. S. Sarma, N. Nagarajan, T. Harinarayana, C. Manoj, K. Naganjaneyulu,
and V. T. C. Kumaraswamy. On the magnetotelluric source field effects in the Indian
equatorial region - an experimental study. Presented at 16th EM induction workshop
held at Santa Fe, USA:1222nd June, 2002.
A. B. Reddi, M. P. Mathew, B. Singh, and P. S. Naidu. Aeromagnetic evidence of crustal
structure in the granulite terrane of Tamilnadu - Kerala. Jour. Geol. Soc. India, 32:
368 381., 1988.
O. A. Ritter, , and G. J. K. Dawes. New equipment and processing for magnetotelluric
remote refernce observations. Geophys. Jour. Int., 132:535548., 1998.
W. Rodi and R. L. Mackie. Nonlinear conjugate gradients algorithm for 2-d magnetotelluric inversions. Geophysics, 66:174187., 2001.
Y. Sarma, T. Sastry, and S. Sarma. Equatorial effects on nighttime Pc3 puslation. Ind.
J. Radio & Space. Phys., 11:2932, 1982.
T. S. Sastry, Y. S. Sarma, S. V. S. Sarma, and P. V. S. Narayan. Daytime Pi pulsations
at equatorial latitudes. J. Atmos. Terr. Phys., 45:733741, 1983.
U. Schmucker. Auswerteverfahren gottingen, in protokoll 7 kolloquium elekromagnetische tiefenforschung. Niedersachsisches Landesamt, Hannover., pages 163188, 1978.
Shalivahan and B. B. Bhattacharya. How remote can the far remote reference site for
magnetotelluric measurements be? Jour. Geophys. Res, 107(B6):ETG 11, 2002.
W. E. Sims, F. Bostick, and H. W. Smith. The estimation of the magnetotelluric
impedance tensor elements from the measured data. Geophysics, 36:938942., 1971.
W. Siripunvaraporn and G. Egbert. An efficient data-subspace inversion method for
two-dimensional magnetotelluric data. Geophysics, 65(3):791803., 2000.
M. Y. Smirnov. Magnetotelluric data processing with a robust statistical procedure
having a high breakdown point. Geophys. J. Int., 152:17., 2003.
J. T. Smith and J. R. Booker. Rapid Inversion of Two - and Three-Dimensional Magnetotelluric data. J. Geophys. Res., 96(B3):39053922., 1991.
V. Spichak and I. Popova. Artificial neural network inversion for magnetotelluric data in
terms of three-dimensional earth macroparameters. Geophys. J. Int., 142:1526., 2000.
J. A. Stodt. Noise analysis for conventional and remote reference magnetotellurics. PhD.
dissertation, Univ of Utah, Logan, 1983.
146
D. Sutarno and K. Vozoff. Phase-smoothed robust m-estimation of magnetotelluric

impedance functions. Geophysics, 36:938942., 1991.
C. M. J. Swift. A magnetotelluric investigation of an electrical conductivity anomaly in
Southwestern US. in Magnetotelluric methods, Geophys reprint series vol 5 ed. Vozoff,
K., Soc. Explor. Geophys. Tulsa, Okla, pages 156166, 1986.
L. Szarka. Geophysical aspects of man-made magnetic noise in the earth - a review.
Surveys in Geophysics, 9:287 318., 1988.
N. K. Thakur, M. V. Mahashabde, B. R. Arora, B. P. Singh, B. J. Srivastava, and S. N.
Prasad. Geomagnetic variation anomalies in Peninsular India. Geophys. Jour. Res.
astr. Soc., 86:838854., 1986.
A. Tikhonov. On determining electrical characteristics of the deep layers of the Earths
crust. Dokl. Akad. Nauk S.S.S.R., 73:295297., 1950.
J. M. Travassos and D. Beamish. Magnetotelluric data processing: A case study. Geophys.
J. Int., 93:377391., 1988.
K. Vozoff. The magnetotelluric method in electromagnetic methods in applied geophysics.
ed by M.N. Nabighian, Society of Exploration Geophysics, Tulsa, USA., 1991.
K. A. Vozoff. Magnetotelluric method in the exploration of sedimentary basins. Geophysics, 37:98141, 1972.
J. R. Wait. On the relation between telluric currents and the earths magnetic field.
Geophysics, 19:281289, 1954.
S. H. Ward. The electromagnetic method. in SEG Mining geophysics volume II, Mining
geophysics, Tulsa, pages 236 372., 1967.
J. T. Weaver. Mathematical Methods for Geo-Electromagnetic Induction. Wiley, New
York, 1994.
P. Weidelt. The inverse problem of geomagnetic induction. Zeitschrift f
ur Geophysik, 38:
257289., 1972.
P. J. Werbos. Backpropagation through time: What it does and how to do it. Proc.
IEEE, 78:1550 1560., 1990.
D. E. Wight and F. X. Bostick. Cascade decimation: A technique for real time estimation
of power spectra. in Vozoff, K. (ed) Magnetotelluric methods: SEG Geophysics Reprint
Series 5, pages 215218., 1980.
J. Winer, R. J. A., R. J. R., and R. Moll. Predicting carbonate permeabilities from
wireline logs using back propagation networks. 61st SEG meeting, Houston, USA.(
Expanded abstract), pages 285288., 1991.
147
E. Yee, P. R. Kosteniuk, and K. V. Paulson. The reconstruction of the magnetotelluric

impedance tensor: An adaptive parametric time-domain approach. Geophysics, 53:
10801087., 1988.
Y. Zhang and K. V. Paulson. Magnetotelluric inversion using regularized Hopfield neural
networks. Geophys. Prosp., 45:725743., 1997.
Z. Zhang.
Robust M estimators.
Appeared at Internet URL http://wwwsop.inria.fr/robotvis/personnel/zzhzng/Publis/Tutorial-Estim., 1996.
Y. Zhao and K. Takano. An artificial neural network approach for broadband seismic
picking. Bull. Seis. Soc. Am., 89:670680, 1999.
K. L. Zonge and L. J. Hughes. Controlled source audio frequency magnetotellurics. in EM
methods in Applied Geophysics, ed M.N. Nabighian, Society of Exploration Geophysics.
Tulsa, USA, 1991.
Errata
The Following changes/ additions have been incorporated into the thesis based on the
examiners comments.
1. Case in which noise is difficult to be removed is given in 2.1, page 19. A description
of noises that cause the scatter in the processed MT impedances is given in the
second paragraph of 6.1, page 89. Error bars are defined in 3.3.4, Page 46.
2. Two sentences on quantifying the noise in MT data after processing are inserted in
5.8, page 86 (for ANN) and 6.2.4.6, page 101 (for robust processing).
3. Coherent and incoherent noises are defined in the first paragraph of 3.3.3, page
45.
4. Use of coherency to screen MT data in dead band is given in the second paragraph
of 3.3.3, page 45.
5. Choices of data rejection gates other than coherency are discussed in the last paragraph of 3.3.3, page 45.
6. Shortcomings of Kao and Rankin [1977]s iterative MT processing scheme is given
in the last paragraph of 3.3.5, page 46.
7. Specific improvements obtained by band averaging are inserted in the second paragraph of 6.4.10, page 119.
8. The thesis does not claim or deals with the averaging of up and down biased
MT impedances.
9. MT processing results along the Vellar-Palani profile in the SGT are given in 6.4.9,
page 118.

Magnetotelluric Data Analysis Using Advances in Signal Processing Techniques

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Magnetotelluric Data Analysis Using Advances in Signal Processing Techniques

Uploaded by

Copyright:

Available Formats

.

Magnetotelluric data analysis using

National Geophysical Research Institute, Hyderabad 500 007,

THESIS SUBMITTED TO THE OSMANIA UNIVERSITY FOR THE

Magnetotelluric data analysis using

resulted in better estimation of MT transfer function as compared to LS estimations. It

3 Magnetotelluric time series analysis

5 The application of the artificial neural networks to magnetotelluric time

Robust Band averaging . . . . . . . . . . . .

7 Discussion and conclusions

Quasi planar nature of electromagnteic wave front at great distance from

Current circuit electric railway systems. I is the current in the overhead

MT stations superimposed over the geology of the measurement corridor

The band averaged telluric predicted coherence plotted as a contoured map

5.10 Comparison of MT apparent resistivity and phase computed from different

Comparison of robust processing (RB) with and without neural network

Frequency bands for measurements, GMS05 (after Metronix [1997]). . . .

Classification of noise sources that affect MT measurements. . . . . . . .

The locations and measurement details for the MT stations

Few commonly used influence functions. Adapted from Zhang [1996]. . .

Magnetotelluric signals originate in Earths magnetosphere and atmosphere from different

Magnetotelluric methods evolved out of the observation of similar variation in Earth

Figure 1.2: Behaviour of plane EM wave at the surface of a conducting Earth

where I = N-1,N-2,. . . .1, and

Figure 1.3: Diagrammatic representation of 1D, 2D and 3D situations.

However, earth cannot be approximated to perfect homogeneous media and it is usual

Thus each tensor element is Z ij = E i /H j (i,j = x,y).

Once rotated optimally, the tensor may be written as,

Amplitude and phase of impedance:

Sensors and Field Procedures

Electric field sensors

Figure 1.5: Implanting of Electrical Field sensor

Figure 1.6: Installation of induction coil magnetometer

Figure 1.7: Simplified equivalent circuit for induction coil magnetometer

Lower freq. Hz Upper freq. Hz Sampling rate Hz

Magnetotelluric method uses naturally occurring electromagnetic field as source signal.

Thunder storm activity

fn = 7.49 [n(n + 1)] 2

Table 2.1: Classification of noise sources that affect MT measurements.

Noises from power line signals

Noise from electric traction

Noises from instrument & sensors

Noises from failure in electronics in equipment, quantization errors at A/D converter

Noises from the other sources

Switching of submersible pumps causes disturbances in measured fields and is observed in

Effect of active electrical noise in magnetotelluric

Random data analysis

Properties of random data

Probability density function

Where is the standard deviation of x ( 3.2.1.2). Gaussian or normal density function

Mean square and variance

The positive square root of the variance is called standard deviation.

Median and average absolute deviation.

Power spectral density function

Joint signal properties

In many applications such as magnetotellurics, it is desirable to study the joint signal

Cross correlation functions

Figure 3.2: Crosscorrelogram of two magnetotelluric time series components.

Cross-spectral density functions

When considering physical measurement of two random processes, it is often desirable

Before considering the processing of random data, magnetotellurics in particular, it is

Trend and bias removal

Where the time series H(t) is measured for a period of time T.