You are on page 1of 9

Geophys. J . Int.

(1993) 112, 115-123

Automatic 1-D inversion of magnetotelluric data by the method of


modelling
J. T. Weaver and A. K. Agarwal
Department of Physics and Astronomy and School of Earth and Ocean Sciences, University of Victoria, Victoria, BC, Canada V8W 3P6

Accepted 1992 July 14. Received 1992 July 14; in original form 1992 April 10

SUMMARY

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


Two popular methods for obtaining a preliminary 1-D geo-electric model from
magnetotelluric data are the simple inversion algorithms due t o Niblett, Bostick or
Schmucker, and the modelling scheme of Fischer & LeQuang. The former provides
resistivity values at different depths, while the latter uses an optimizing program to
obtain the resistivities and thicknesses of a prescribed number of layers which give a
best-fitting response t o the given data. The present scheme combines features of
both methods by using Niblett inversion t o provide an initial resistivity profile from
which, successively, 2 , 3 , 4 , . . . layer models are constructed automatically. The
algorithm provides a response with minimum misfit for each layered model
according t o the method of Fischer & LeQuang, and uses a statistical F-test to
determine when the inclusion of additional layers is no longer justified by the
improvement of fit. The method therefore provides a best-fitting ‘least-layered’
model consistent with the given data without the human intervention required by
the Fischer-LeQuang scheme. Some examples of its application t o both synthetic
and real data are presented.
Key words: conductivity, geo-electric structures, geo-electromagnetic induction,
inversion, magnetotellurics.

model under certain prescribed conditions. The former


1 I N T RODUCT I O N
approach provides directly a series of depths and associated
1-D inversion of magnetotelluric (MT) data is probably resistivities from which an approximate, continuous
regarded as a closed book by many practitioners in the field resistivity profile can be drawn, while the latter yields a
of electromagnetic induction, especially as much attention is best-fitting layered model, the number of layers having to be
now directed towards the more challenging problems arising decided in advance by the user following the general
with the interpretation of 2- and 3-D data. Indeed there are guideline ‘the fewer the better’. It was successfully
many sophisticated 1-D inversion schemes in routine use developed by Fischer & LeQuang (1981, 1982) using a
today, ranging from Parker’s (1980) D + method which minimization routine MINDEF originally written by J. Beiner.
yields a best-fitting but geologically unrealistic model, The second output is more pleasing to the many
comprising a series of infinitesimally thin sheets at different geophysicists who tend to think of the geo-electric structure
depths to OCCAM inversion (Constable et al. 1987) in of the Earth in terms of layered models, but the method is
which the problem is regularized by stipulating a maximum less easy to implement than the first because considerable
smoothness constraint. The first is used mainly as a test on interplay between operator and computer is required. In
the acceptability of data for 1-D inversion, while the second particular, it is necessary to make an inspired guess for the
will generate a smoothly varying conductivity profile which starting model, and later to adjust the number of layers if
is both physically plausible and compatible with the data, that seems appropriate in the light of the results obtained.
Other methods, including those developed by Oldenburg Even if the starting model is generated by the inversion
and co-workers are covered in a review article by Whittall & scheme of Fischer et af. (1981), some judicious intervention
Oldenburg (1992). is usually required before optimization, because the scheme
For a fast, preliminary inversion of MT data, however, it invariably gives a starting model with more layers than are
is customary either to use one of the approximate inversion necessary.
formulae associated with the names of Niblett, Bostick or Despite the fact that Parker (1983) wrote nearly a decade
Schmucker, or to do forward modelling in conjunction with ago ‘It is hard to believe there is a need for any further
some sort of optimization which seeks out a best-fitting effort on iterative inversion in terms of a small number of

115
116 J . T. Weaver and A . K . Agarwal

homogeneous layers’, some new discussions on this theme In this definition R r ) it:)+ denotes the calculated
have appeared recently. Fischer & Weibel (1991) have (dimensionless) impedance at period T, for the k-layer
investigated the inversion of Swiss AMT data using the model, and the weighting functions w, and wf are the
modelling method of Fischer & LeQuang (1981) paying normalized inverse variances uf of the data points whose
particular attention to the trade-off between layer thickness errors of measurement are assumed to be independent and
and resistivity. In a conference presentation of what has normally distributed about the true response. Thus
turned out to be a forerunner of this paper, Weaver &
Fischer (1990) described a new computer program for an
automatic implementation of the Fischer-LeQuang scheme
in which all human involvement in finding a best-fitting
model with the appropriate number of layers was removed.
In this paper we develop the automatic algorithm (called (3)
AUTOMOD) in a more systematic manner by introducing
statistical tests which allow us to determine the best
‘least-layered’ model at some prescribed level of confidence.

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


Note that if u?’ at each period T, then
It combines some features of Niblett-Bostick inversion with wR(Tn)= wf(T,). Furthermore, if the variances ui are the
the forward modelling method of Fischer & LeQuang, and same at all data points (e.g. if the same percentage of
although it is more computer intensive than either of the Gaussian noise is present in all the data, or is simply
original procedures on which it is based, this is more than assumed to be the same in the absence of experimental error
offset by the rapid advances in desk-top computer power in bars, or if the data are synthetic and exact) then the weights
the past decade. It can be employed as readily today as were all reduce to unity. Experimental data are usually provided
the other procedures ten years ago. with error bars indicating in some way the standard
The given data are assumed to be in the form of a set of deviation of the Gaussian distribution of error at each point.
values R ( T ) and I ( T ) , the respective real and imaginary In this case
parts of the (dimensionless) complex impedance Z ( T ) / Z , ,
at the periods T = T, (n = 1, 2, . . , , N), and with the error +
uZp) iu?) = A[zmaX(~,)
- Z~’”(T,)] (4)
bars at period Tn defined by a pair of complex values
[Zmi”(T,)/Z,, ZmaX( Tn)/Zo]. The (real) scaling constant Z, where 1 is a factor which takes into account how the error
includes the dimensions of impedance and is introduced to bars are defined; e.g. if they have a length of two standard
ensure that R( T) and I( T) are dimensionless quantities of deviations then 1 = 1/2; if the length is twice the standard
moderate magnitude. A suitable choice for the value of Z, (or probable) error, then A=0.74; if it is the half-width of
in a given investigation might be [max lZ(T,)I x the Gaussian distribution, 1 = 0.43; if it represents a 95 per
min IZ(Tn)I]1’2B where the maximum and minimum values cent confidence interval, 1 = 0.26; etc..
are extracted from the N impedances in the data set or It should be mentioned that Fischer & LeQuang (1981)
alternatively, perhaps, the modulus of the impedance (at preferred to use unconventional weighting factors propor-
some average period) of the best-fitting half-space. MT tional to the linear inverses of the error bar lengths because
responses are usually displayed as graphs of apparent they penalize outlying data points less. Definitions (3) and
resistivity pa = IZ12T/2np, and phase @ = arg Z plotted (4), however, are the natural weights to choose since
against period, with logarithmic scales for apparent minimization of the corresponding s2 yields the maximum
resistivity and period. If the data are supplied as values of likelihood that the model is representative of the true
apparent resistivity and phase, they can be converted to response from which the data were obtained (Bevington
complex impedance values with the formula 1969).

Z(T) = [ R ( T )+ iI(T)]Z, = q2nip,pa(T)/Texp [i$(T)] (1)


2 THE ALGORITHM
before the inversion procedure is begun. However, the
It is assumed that measured values of the complex
option exists in AUTOMOD to treat apparent resistivity and
impedance are available for N-periods and that error
phase data directly as described in section 4.
bounds (variances) are iricluded with the data [if they are
The parameter to be minimized for a least squares fit to
not, the weight functions (3) are all set equal to 11. The
the response of a k-layered model is the variance of fit s’,
strategy adopted in AUTOMOD is described in the 11 steps
(Bevington 1969), where the subscript v signifies the number
listed below.
of degrees of freedom. Since there are 2k - 1 adjustable
(1) A preliminary inversion is undertaken by converting
parameters defining a k-layered earth, k resistivities pi
the complex impedance values to apparent resistivity and
( i = 1, 2 , . . . , k) plus k - 1 layer thicknesses hi ( i =
phase, and applying the Niblett formulae (Niblett &
1, 2, . . . , k - l), the number of degrees of freedom left
Sayn-Wittgenstein 1960) that gives the conductivity
after the model response has been fitted to the 2N measured
K, = K ( z , ) at depth z, as
data is 2N - 2k + 1. The variance of the fit (or simply the
square of the ‘misfit’) is therefore defined by
N

for each period T,. Here m ( T ) = d log,, p , ( T ) / d log,, T is


the slope of the apparent resistivity curve which can be
Inversion of MT data 117

estimated from the phase data by Weidelt’s (1972) result considered, the starting model is defined by
m ( T ) = 1- 4 $ ( T ) / n .
Alternatively, Schmucker’s (1970) formulae
(6) hj= c
JMk

n =n,
d,,
1 1 JMk
-=-
pj hi,=,,
tn (j=1,2, . . . , k-2) (10)

M-1 rG-7

n=l n=l

1 - 1 M-1
could be employed to the same end. -- C rn,
(k-1)
Pk = Pk-1
__ (11)
(2) Unacceptable data points (if any) which do not give Pk-1 h k - l n=nk--I
increasing depths for increasing periods, or yield negative +
where nj = ( j - l)Mk 1, M k = [(M- l)/k], and the best-
resistivities, are temporarily discarded. If there are N - M fitting model with misfit mins2N-2k+, is defined by the
such points, we are left with M pairs of values K , , zn for values hjk), p j k ) ( j = 1, 2, . . . , k - 1) and p p ) .
n = 1, 2, . . . , M. With most data M = N. (7) Before AUTOMOD moves on to construct a (k 1)- +
(3) Based on the inversion in (l), an M-layered structure layer starting model a test is performed which determines

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


is assembled with boundaries at depths z = 5, where whether the k-layer model is so highly unacceptable that it
should be rejected. For this purpose the program computes
(7)

and with the layers ~ n < z z ( 5 ; 1 + l ( n = 1 , 2 , . . . , M - 1 )


having the conductivities K, given by equation (5). The
which is a standard x2 distribution with 2N - 2k + 1 degrees
basement half-space z > cM
is assigned the final conductivity
of freedom. The hypothesis that the k-layered model is
K~ corresponding to the period T, in equation (5). The
acceptable in the sense that the computed value of x$-,~+,
thicknesses d , and conductances rn (n = 1, 2, . . . , M - 1) of
would be attained or exceeded by virtue of the randomness
the M - 1 layers above this half-space are given by the
in the data alone is now tested. The probability of x2
formulae
exceeding a given value can be found, of course, in standard
dn = fnt rn =Kndn. (8) statistical tables, but in AUTOMOD it is computed directly
using published subroutines (Press et al. 1989). If the
(4) The search begins with a two-layer model comprising probability is found to be less than the pre-assigned value of
a top layer of thickness h , and resistivity p 1 defined by 0.1 per cent, then the hypothesis (and with it the model) is
M-1
1 1M-l rejected at the 99.9 per cent confidence level. If no
h,= C
n=l
dn, -=-
p1 h1
C
n=l
rn (9) variances are supplied with the data and cannot be
estimated, or if the data are synthetic and exact, or even if
overlying the half-space z > h , of resisitivity pz = 1 / ~The ~ . the error estimates are believed to be highly unreliable, then
three model parameters h , , p 1 and p2 are adjusted by the this test has no meaning, and it is therefore by-passed. In its
minimization routine (with a prescribed maximum number place we simply require that the minimum misfit s, defined
of calls and step size for the parameter variations, with layer with unit weighting functions be less then 0.1.
thicknesses constrained to be greater than 10 m, and with (8) When k 2 3 and the model has not been rejected at
resistivity variations limited to the range 0. 1-lo6 P m ) until a step (7), a further test is made by comparing the current
minimum of the misfit, mins,,_,, defined in equation (2), is value of x2 with the previous one which had 2N - 2k + 3
found. The ‘best’ two-layer model is then defined by the degrees of freedom. This is done by taking their difference
corresponding parameter values hi’), p p , p g ) . Note that s2 and forming the ratio
is calculated using all N data points including any that had to
be removed from the Niblett inversion. In other words the F(2,2N - 2k + 1) = (xfN-2kt3 - x f N - 2 k + L)l2
response is fitted to all the data even though only a subset of X2N-2k+1/(2N - 2k + 1>
the data may have been used to construct the starting which represents a measure of the improvement in fit as a
model. result of adding the additional layer. It behaves as an F
( 5 ) The next step is to construct a three-layer model distribution (Bevington 1969) being the ratio of two reduced
defined by the resistivities p l , p 2 , p3 and thicknesses h , , h, x2 distributions with 2 and 2N - 2k + 1 degrees of freedom
where respectively. If ~7:~)= uy’ at each period Tn (even though
M2 M-l the variance may change from period to period) then
h,= c dn,
n=l h 2 = ( z l dn)-hl. equation (13) reduces with the aid of equation (12) and
equation (2) to
F(2, 2N - 2k + 1)
- (2N - 2k + 3)S&.-zk+3 - (2N - 2k + 1)SfNP2k+1
Here M2= [(M- 1)/2], the square brackets meaning (14)
‘integral part o f . The minimization routine is again called to %;N--2k+l
find a best-fitting three-layer model, characterized by the which depends only on the normalized weighting functions
parameters h?, h y ) , p?, p p l , p$3)and misfit min s ~ ~ - ~ . rather than the variances, and therefore continues to be
(6) The procedure continues adding one more layer each defined even when the variances are unknown but are
time. At the stage where a k-layer model is being assumed equal for all data points (or if the data are synthetic
118 J . T. Weaver and A . K . Agarwal

and exact). The hypothesis that the value of F obtained (11) A final model selected by the criteria described in
could have arisen by chance is now tested by computing, @)-(lo), and not rejected at the 99.9 per cent level by the
with the aid of published subroutines for the probability x2 test in step (7), is now resubmitted to the minimization
distribution of F (Press et al. 1989), the probability of the routine for further refinement with a smaller step size in the
given value being reached or exceeded. If it is found to be parameter variations, and a greater number of calls to the
less than 5 per cent, the hypothesis is rejected and it is routine permitted. This allows the search to continue
concluded with 95 per cent confidence that the new layer onwards, if necessary, towards the local minimum which
contributed significantly to the improvement of the fit. In may be on the floor of a rather long, flat, curved valley in
this case AUTOMOD advances without interruption to an parameter space (Fischer & LeQuang 1982). When the true
+
investigation of models with k 1 layers. Should the local minimum has been located, the model parameters
hypothesis not be rejected, however, then a flag is set to piK), hiKK’(k = 1, 2, . . . , K - 1) and p F ) are displayed
prepare the search for possible termination at the (next) together with the misfit parameter mins,,-,,+,, and the x2
+
(k 1)th step in the inversion procedure. If the flag is level of acceptance (if available).
already in the ‘on’ position, having been set during the
testing of the (k - 1)-layer model, then the search stops here 3 COMMENTARY

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


because there is no strong evidence to suggest that inclusion
In step (4) it is possible to tighten the bounds on the
of the last two layers has contributed significantly to an
resistivity and thickness parameters before calling the
improvement in the fitting of the response curves to the
minimization routine. This could be justified if, for example,
data. The program therefore returns to the best-fitting
prior geological information suggested that resistivities
(k - 2)-layer model (i.e. the last one to show improvement
outside a certain range would definitely not be expected in
over those with fewer layers) as the ‘least-layered’ model of
the region under investigation. The minimization program
choice. Occasionally the hypothesis will be rejected even
that is actually called at present is the original routine
though the flag had been previously set one stage back. In
MINDEF used by Fischer & LeQuang. Obviously this could
this event it is assumed that the lack of improvement in the
be easily replaced by any of the more modern routines now
fit when an extra layer was added to the (k - 2)model was a
freely available.
temporary aberration and that the inclusion of further layers
The test in step (7) is essentially what Parker (1983) called
is still justified. The flag is therefore cleared, and the
an ‘old fashioned approach’ which is ‘quite reasonable’. It is
program proceeds as before. If F should turn out to be
used here with a very high confidence limit (99.9 per cent) in
negative (meaning that the fit is no longer improving) then
order to provide a preliminary screening of possible models
the search halts and the (k - 2)- or (k - 1)-layer model is
and to eliminate only those that are clearly unacceptable. If
selected depending on whether the flag has already been set
a final model is rejected at this level of confidence it is
or not. probably because the data are not 1-D, or because the
(9) In the unlikely event that the search for a
errors have been seriously underestimated, or possibly
‘least-layered’ model continued in this manner without
because the errors are not normally distributed. It is of
success until all M layers in the starting model were used up
interest to note that according to Parker (1983) the
(or equivalently, if M = N, until the available degrees of
best-fitting model to the full COPROD data set was rejected
freedom were exhausted) the program could be led,
with more than 99.9 per cent certainty, but when the set was
especially if the data set were large, into an unreasonable
limited to the 15 well-estimated responses it passed the x2
number of calculations entailing lengthy and pointless
test at a confidence level as low as about 70 per cent.
searches for minimum misfits in a series of multi-layered
Step (8) provides some sort of quantitative backing to the
models. As a safeguard against this happening, AUTOMOD is
philosophy expressed by Fischer & LeQuang (1981)
programmed to select the model with the smallest misfit
regarding the selection of the right number of layers, and
after it has reached a 10-layer model without the test criteria
removes all dependency on human judgement in this task by
described in (7) and (8) being satisfied.
basing the procedure on statistical tests that can be
(10) Once the ‘least-layered’ model, or failing that the
implemented automatically.
one with minimum misfit, has been selected with (say)
A non-automatic option is retained under which the
L ( L 5 10) layers, it is subjected to a further test to check
program will compute a best-fitting model from a given
that resistivities in adjacent layers are sufficiently different in
starting model with a fixed number of layers. This may be
magnitude. For this the parameters
required if, despite the imposition of various criteria for
model selection described above, the final model generated
is not fully satisfactory in some aspect-perhaps two
adjacent layers whose resistivities just fail to pass the cut-off
are required to satisfy Pk 2 6 where 6 is specified, e.g. if
test in (10) need to be combined. In addition, the
6 = 0.2 then neighbouring resistivities must differ by more
non-automatic version of the program can be used to
than 20 per cent. Each time this condition is not met, the
investigate ‘parameter trade-off (Fischer & LeQuang 1982;
adjacent layers are combined into a single layer whose
Fischer & Weibel 1991).
resistivity is an average of the two previous values and the
total number of layers is reduced by 1. Thus if Pk < 0.2 for j
4 A P P A R E N T RESISTIVITY A N D PHASE
values of k, then the ‘least-layered’ model selected will be
DATA
changed to one with K = L - j layers. This normally
happens only when AUTOMOD has failed to select a clear It was mentioned in section 1 that data are often supplied as
‘least-layered’ model in the early stages of the search. apparent resistivity and phase values with associated error
Inversion of M T data 119
bars. AUTOMOD can deal with such data directly, the derivative, while m(Tl) is found by fitting a parabola
modifications required being quite straightforward. through the three points (log,, pa(T,), log,,, TI),
It is usual to take the logarithm of (1) and to regard the (log,, P,(T2), log10 T2) and (log,, P a ( m log,,, TA and
errors in pa as having a log-normal distribution. Thus in differentiating at the end-point T,. An analogous procedure
place of (1) we have at some fixed period is used for calculating m(TN). Similar modifications are
required if only phase data are inverted,
In Z ( T ) = (1/2) In [p,(T)/p,] + i#(T)+ const
where po is some constant representing a scale of 5 APPLICATIONS
measurement for pa. The variables to be fitted when
comparing the computed model response with the data are We consider first the five-layer structure chosen as a test for
(1/2)1n(pa/po) and #. The square of the misfit (2) is the original inversion scheme of Fischer et af. (1981). The
therefore replaced by given model, and the model returned automatically, are
shown by tabulated values in Table 1, together with the
2
1 corresponding values of the misfit min s, (the square root of
'ZN-Zk+l =
2N - 2k +1 Pan the variance of fit). The synthetic data supplied to the

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


program consisted of 52 points, the 26 pairs of R(Tn) and
I( )
'
7 values obtained from forward calculations for
logarithmically equi-spaced periods Tn over the five decades
where the new weighting functions are in the AMT range from 10-3s to 100s with Gaussian
random noise superimposed at all periods.
There are frequent references in the literature to synthetic
data to which a certain percentage of Gaussian noise has
been added, but it is not always made clear what the stated
percentage refers to. Here we use a Box-Muller method to
generate 52 Gaussian random deviates ul, u2, . . . with zero
mean and unit variance (Press et al. 1989; Constable 1991)
n=l and, if we are adding p per cent noise, the data points
actually used for the inversion are R(Tn)+ p ~ , ~ - , / 1 0 0and
Here (uf))' is the variance of (1/2)In[pa(T,)/po] while
0 n = 1, 2, . . . , 26. Thus the percentage
I(Tn)+ p ~ , ~ / 1 0 for
(u:@))' is the variance of #(T',). The relationship between
p refers to the standard deviation of the Gaussian
the variances and the error bars supplied with the data
distribution. Put in another way, for A = 1/2 in (4), the error
needs to be clarified. Corresponding to equation (4) in
bar Zmax(T,)- Zmin(T,) is pZ0/50. [With apparent
section 1 we have
resistivity and phase the corresponding noisy data would be
o l #('GI + P V 2 n 1 1 0 0
(1/2) In [ ~ a ( T n~) X (P~ ~ 2 n - 1 1 1 0 0 ) l ~and
in radians; e.g. if 2 per cent noise is added, the 95 per cent
Note that this implies how the scales of logl,pa and #
interval, corresponding to A =0.26 in (18), defines error
should be drawn if the error bars on apparent resistivity and
bars with p y = 1.08 x pa and p:'" = 0.93 X pa, while
phase are to have equal lengths when the standard deviation - #""" = 4.41O.I
of their errors of measurement are the same at any given
period. Since one unit on the loglo pa scale corresponds to a Table 1. Five-layer test model. The top table of values were
decade in pa or a unit of standard deviation proportional to obtained from synthetic impedance data; in the bottom table
(1/2)ln 10, an equal standard deviation in # is similarly synthetic apparent resistivity and phase values were used.
proportional to (1/2) In 10 radians, or approximately 66". Layer 0% Noise 2% Noise 10% Noise
Thus one unit on the log,,p, scale should have the same pPm) h (km) PWm) h (km) P W ~ ) h (km)
length as 66" on the # scale. 1 450 0.10 442 0.10 415 0.10
2 50.0 0.40 49.4 0.40 46.9 0.39
The procedure followed by AUTOMOD is exactly the same 3 28.0 3.00 28.1 1.58 28.5 1.41
as before except that it is now the misfit (16) that is 4 45.0 7.00 48.0 8.64 140 15.6
minimized. The new x2 function is formed by replacing R in 5 1.00 x 104 6.72 x 103 3.78 103
(12) with In pa and I with #. The F test function is still Misfit rnins13 = 6.33 x 1O-I rnins43 = 1.88 x lo-' rnins13 = 9.55 x lo-?

defined by (13) and (14) in terms of the new x2 and s 2 Layei Actual Model
functions. P W ) h(km)
Occasionally a data set may only contain apparent 450 0.10
resistivity values, in which case the number of degrees of 50.0 0.40
28.0 3.00
freedom is reduced by N and the misfit squared becomes 45.0 7.00
1 2 w,(Tn)((1/2)1n~)'.
N
(19) 1.00 x 10' 05

S L k + l =
Layer 0% Noise 2% Noise 10% Noise
N-2k+ln=1 Pan
P W ) h (km) P(W h(km) p ( W h (km)
The definitions of x' and F are revised accordingly. In the 1 450 0.10 424 0.10 379 0.11
2 50.0 0.40 48.6 0.40 44.3 0.34
absence of phase data, it is also necessary to modify the
3 28.0 3.01 27.3 3.09 26.1 5.60
preliminary inversion procedure described under (1) in 4 45.1 7.00 47.6 6.94 998 63.6
section 2. The slope m(Tn) (n = 2,3, . . . , N - 1) is 5 LOO x 104 7.89 x 103 1.09 x 10' m
estimated from the central difference representation of the Misfit rnin = 2.52 x lo-' rninsrs = 1.78 x lo-' rninq3 = 8.91 x
120 J . T. Weaver and A . K . Agarwal

The models obtained for three different levels of noise, 0 equi-spaced periods between 2.5 X lop3 and 250 s on a
per cent, 2 per cent and 10 per cent, are shown in Table 1. logarithmic time scale. We have actually used 26
Note that the program correctly selects a five-layer model in equi-spaced periods again (so as to retain the five periods
each case whereas the original inversion scheme of Fischer per decade) in the same range and have included 2 per cent
et al. returned twelve layers which, after inspection, were and 10 per cent random Gaussian noise. The models
reduced manually to a five-layer starting model for returned by AUTOMOD are also shown in Table 2-the
submission to their minimization routine. The agreement agreement with the original model is quite good for 2 per
with the original model is excellent-perfect, in fact, to cent noisy data (perfect agreement was again achieved with
three significant figures or two decimal places when the exact data) but when 10 per cent noise was added the signal-
exact data are used-but the noisier the data, the greater to-noise ratio at the longer periods became very small as a
the misfit s, as expected. It is of interest to mention that the result of the impedance values being relatively small there.
F-test correctly selected five-layer models by the standard The deterioration in the model obtained from these data is
tests in section 2 (8) when noisy data were used, but halted obvious from the values given for the deeper layers in Table
the search procedure at six layers with the exact data as a 2. The models obtained are compared graphically in Fig.
result of F becoming negative. Since the flag had not 2(b). In their 1991 paper Dosso & Oldenburg inverted data

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


previously been set, however, the program returned to the produced by the same model but for only 12 periods in the
five-layer model according to the criteria listed at the end of same range. We repeated this exercise and found very
section 2 (8) (this latter behaviour was the result of a very similar results to those in Table 2.
pronounced minimum of s’, being reached when Y took the Further examples of the inversion of synthetic data to
value 43 corresponding to 5 layers). which 2 per cent noise has been added are shown in Table 3.
Figure 1 shows the impedance data points containing 10 The models all have four layers which are generally deeper
per cent noise (with their error bars two standard deviations than those in Table 1, and are therefore more relevant to
in length) that were used to generate the model in column 3 MT rather than AMT investigations. They have been
of the topmost Table 1. The response curves for the final designed to illustrate the performance of AUTOMOD on four
model obtained, together with the responses of the actual distinct types of structures, which can be roughly described
model, are also depicted. The random scatter of the points as a ‘sandwiched conductor’ (model I), a ‘sandwiched
around the true response curve is clearly evident in the resistor’ (model II), a ‘staircase of decreasing resistivity’
figure but the deviation of the response curve computed (model III), and a ‘staircase of increasing resistivity’ (model
from the model generated by AUTOMOD is barely visible. In IV). Resistivities and layer thicknesses have been chosen so
Fig. 2(a) the models tablulated in Table 1 are reproduced as as to make the models at least fairly plausible from a
diagrams showing the variation of resistivity with depth. geological point of view. Thus the basement half-space in
As a second example we use the test model adopted by model 111 is placed at a greater depth than in the other
Dosso & Oldenburg (1989, 1991) defined in Table 2 and models in order to be more suggestive of the conductive
illustrated in Fig. 2(b). In the 1989 paper they used 25 mantle; in model IV the corresponding depth is much
smaller because the now highly resistive basement must
represent the middle crust and not the mantle. Given the
depths of the layers, only 16 logarithmically equi-spaced
points in the longer period MT range of 1 - lo3 s were used
to compute the synthetic data, leaving just 25 degrees of
14- freedom when calculating the misfit for the responses of the
four layer models found in each case. The results provided
12-
by AUTOMOD again accurately reflect the structures of the
10. original models from which they were generated. Only the
[(T) depth and resistivity of the basement in model 111 deviate
8- substantially from their true values (albeit preserving the
general trend), but this is hardly surprising because the
6-
structure in this model extends to a greater depth than in the
4-
others. Indeed, when the period range was extended by
another decade, these slight deviations disappeared. With
2- different variations of random noise AUTOMOD occasionally
generated five- or three-layer models which nevertheless still
0- closely resembled the original model, the tendency being
either to split one of the original layers into two separate
- 2 1 -2 I
-3 -2 -1 0 1 -3 -2 -1 0 1 layers whose resistivities just failed to meet the criterion in
LOG T LOG T test (lo), or alternatively to combine adjacent layers into
one.
Figure 1. Impedance data points with 10 per cent noise used as the
Turning to real data with error bars, we re-examine the
input to AUTOMOD to generate the five-layer model given in column
3 of Table 1 . The response curves depicted by broken lines are the
data set analysed by Fischer & LeQuang (1982) of p, and 4
real (left diagram) and imaginary (right diagram) parts of the values recorded at station Ahausen in Germany for 23
dimensionless impedance for the actual test model; those drawn periods between 8 s and 1002s. After a series of trials
with solid lines are the corresponding responses for the five-layer Fischer & LeQuang decided that a four-layer model was
model yielded by AUTOMOD. appropriate for the Ahausen data (see their Fig. l l ) , and
Inversion of M T data 121
-2
(a)

-1

LOG 2

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


2

LOG P o

0
(b)

2 (km)

10

15

20
-2 0 1 2 3
LOG Pa

Figure 2. (a) The five-layer test model in Table 1 (solid line) and the corresponding models returned by AUTOMOD from data to which 2 per
cent (broken line) and 10 per cent (dotted line) noise has been added. (b) The same diagram for the resistivity model used by Dosso &
Oldenburg (1989). Note that the depth scale in km is logarithmic in (a) and linear in (b).

Table 3. Synthetic models I-IV; the data processed by AUTOMOD


included 2 per cent Gaussian noise.
Layer Model I AUTOMOD Model II AUTOMOD
P(Wh (km) ~ ( n m )
h(km) ~ ( n m )h ( k m ) ~ ( n m ) h (km)
Table 2. The test model used by Dosso & Oldenburg (1989). The 1 1000 1 2.02 103 0.99 100 5 99.5 4.81
results are presented in the same format as in Table 1. 2 10 4 10.1 4.10 1000 10 713 105
3 1000 25 I 42 x 103 24.5 100 15 98.5 14.7
Layer Actual Model 2% Noise 10% Noise 4 10 m 9.24 m 10 c4 8.98 30
Misfit minS25 = 1.5i x m i n s Z 5= 1 5 7 x 10.'
~ ( f l m ) h(km) ~ ( n m ) h(km) p(flm) h(km)
1 250 0.6 250 0.60 252 0.60 Layer Model 111 AUTOMOD Model IV AUTOMOD
2 25 1.4 24.8 1.58 25.3 0.33 ~ ( f l m ) h ( k m J ~ ( n m ) h (km) p(llm) h (km) p ( R m ) h (km)
3 100 4.0 244 2.80 14.32 0.37 1 1000 10 998 10.0 2 1 2.00 0.96
2 100 20 99.6 20.1 10 2 8.28 1.82
4 10 4.0 11.5 5.73 57.0 14.3
3 10 20 8.23 22.5 100 17 92.6 li0
5 25 00 52.5 00 0.10 00 4 2 m 0.10 m 1000 m 902 ?j

Misfit mins13 = 1.78 x minsla = 1.26 x lo-' Misfit min325= 1.46 x 10.' minsz. = 1 5 3 x
122 J . T. Weaver and A . K . Agarwal

Table 4. Models obtained from Ahausen data. Table 5. Models obtained from data at station MUS.
Layer AUTOMOD Fischer & LeQuang Layer AUTOMOD Fischer & Weibel
P(Rm) h (km) PPm) h (km) P(flm) h (km) Pwm) h (km)
1 2.71 2.95 2.69 2.83 1 216 0.10 200 0.09
2 75.9 4.08 34.6 4.11 2 5.68 x lo4 0.75 1.14 x lo4 0.78
3 1.18 4.02 1.39 5.18 3 751 2.22 750 2.20
4 11.3 00 16.5 00 4 2.24 x 103 oo 2.23 x 103 00

Misfit mins39 = 2.04 x lo-' mins39 = 2.11 x lo-' Misfit mins2.I = 2.81 x mins2.r = 2.94 x

their best model is shown in Table 4 (see Tables 3 and 4 in recomputed according to definitions (16) and (17). The
their original paper). AUTOMOD also returned a four-layer models agree very well except for the resistivity p2 of the
model whose parameters are tabulated alongside the second layer. This is not surprising, however, because in
original ones in Table 4. The misfit value quoted by Fischer discussing 'parameter trade-off with these data, Fischer &

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


& LeQuang is not directly comparable with ours (see the LeQuang (1982) showed that p2 could increase without
final sentence in section 1) and has therefore been bound with little effect on the value of the misfit. In other

-3 -2 -1 0
LOG T

75'-

+ 6oo-

15'- €
Inversion of M T data 123

words, above a certain lower limit the value of p2 has very period when one of us (JTW) spent 6 months at the
little significance. Observatoire Cantonal in Neuchbtel, are gratefully acknow-
We conclude with an example of a fairly recent Swiss ledged. Thanks are also due to Dr Pierre Schnegg for his
AMT data set of apparent resistivity and phase values assistance in preparing the two real data sets which were
recorded at a station (code-named MUS)which has been the used in this study and for explaining how the optimization
subject of a detailed investigation by Fischer & Weibel routine MINDEF is implemented. This work was supported
(1991). The model constructed by AUTOMOD has four layers by the Natural Sciences and Engineering Research Council
in agreement with the best model obtained by Fischer & of Canada and, during the period spent in Neuchbtel, by the
Weibel; the parameters of both models are compared in Swiss National Science Foundation.
Table 5. The response curves for the model in column 2 are
shown in Fig. 3. The agreement with the Fischer & Weibel
REFERENCES
best model is so remarkably close (even though the
weighting factors (3) and (4) were different in the two Bevington, P. R., 1969. Data Reduction and Error Analysis for the
schemes) that the curves are almost identical to those in Fig. Physical Sciences, McGraw-Hill, New York.
1 of the paper by Fischer & Weibel (1991). Constable, S. C. 1991. Comment on ‘Magnetic appraisal using

Downloaded from https://academic.oup.com/gji/article-abstract/112/1/115/652969 by guest on 09 April 2020


simulated annealing’ by S. E. Dosso and D. W. Oldenburg,
Geophys. J. Int., 106, 387-388.
6 CONCLUSIONS Constable, S. C., Parker, R. L. & Constable, C. G. 1987. Occam’s
Inversion: a practical algorithm for generating smooth models
Given a 1-D magnetotelluric data set, the program from EM sounding data, Geophysics, 52, 289-300.
AUTOMOD will always produce, automatically, a ‘sensible’ Dosso, S . E. & Oldenburg, D. W., 1989. Linear and non-linear
layered model with as few layers (maximum 10) as is appraisal using extremal models of bounded variation,
‘reasonable’ to give a response that best fits the data, in the Geophys. J . Int., 99, 483-495.
sense that the misfit is minimized in the neighbourhood of Dosso, S. E. & Oldenburg, D.W., 1Y91. Magnetotelluric appraisal
an initial layered model which is derived from Niblett using simulated annealing, Geophys. J. Int., 106, 379-385.
Fischer, G. & LeQuang, B. V., 1981. Topography and
inversion. By ‘reasonable’ we mean that the number of minimization of the standard deviation in one-dimensional
dimensions in the model space selected is not so large (for magnetotelluric modelling, Geophys. J. R. astr. Soc., 67,
the given data) that extraneous minima are introduced as a 279-292.
result of the multiplicity of layers, or so small that the Fischer, G. & LeQuang, B. V., 1982. Parameter trade-off in
minimum misfit obtained is unacceptably large; and by one-dimensional magnetotelluric modelling, J. Geophys., 51,
‘sensible’ we mean that the program will generally steer 206-2 15.
clear of minima associated with very thin layers (even Fischer, G . & Weibel, P., 1991. A new look at an old problem:
though such minima may be smaller than the local minimum magnetotelluric modelling of 1-D structures, Geophys. 1. Int.,
found), because the search for a local minimum always 106, 161-167.
starts from a ‘sensible’ model. Fischer, G., Schnegg, P.-A,, Peguiron, M. & LeQuang, B. V.,
1981. An analytic one-dimensional magnetotelluric inversion
Minimization of the response function c ( T ) =
scheme, Geophys. J. R . astr. Soc., 67, 257-278.
-iTZ(T ) / 2 n p 0 instead of the impedance or apparent Niblett, E. R. & Sayn-Wittgenstein, C., 1960. Variations of
resistivity and phase was also tried with AUTOMOD,but we electrical conductivity with depth by the magnetotelluric
found the program to perform slightly less robustly in this method, Geophysics, 25, 998-1008.
case. Another available option that may be useful when Parker, R. L., 1980. The inverse problem of electomagnetic
treating real data sets for which prior geological information induction: existence and construction of solutions based on
is available, is to fix the resistivity and depth of the bottom incomplete data. J . geophys. Res., 85B,421-4428,
layer. Parker, R. L., 1983. The magnetotelluric inverse problem.
While no new theoretical insights into 1-D inversion have Geophys. Surv., 6 , 5-25.
been presented here, the rather vague criteria (requiring Press, W. H., Flannery, B. P., Teukolsky, S. A. & Vetterling, W.
T., 1989. Numerical Recipes, Cambridge University Press,
human judgement) for selecting the right number of layers
Cambridge.
in a modelling scheme have been replaced by recognized Schmucker, U., 1983. Anomalies of geomagnetic variations in the
statistical tests which can be implemented automatically at a southwestern United States. Bull. Scripps Inst. Oceanography,
stated level of confidence. Automation of the method and 13, University of California Press, California.
its speedy execution have become possible primarily because Weaver, J. T. & Fischer, G. 1990. Finding a one-dimensional
of the emergence of relatively inexpensive and powerful layered model: a combination of the Niblett-Bostick and
desk-top computers over the last few years. Fischer-LeQuang schemes. Paper presented at: lcXh Workshop
on Electromagnetic Induction in the Earth, Ensenada, Mexico.
Weidelt, P., 1972. The inverse problem of geomagnetic induction,
ACKNOWLEDGMENTS Z. Geophys., 38, 257-289.
Whittall, K. P. & Oldenburg, D. W., 1992. Inversion of
The indirect contributions to this work by Dr Gaston magnetotelluric data for a one-dimensional conductivity.
Fischer through his informative discussions on the modelling Geophysics Monograph Series No. 5, ed. Fitterman, D. V.,
approach t o inverting magnetotelluric data, during the Society of Exploration Geophysicists, Tulsa, Oklahoma.

You might also like