Geophys. J. R. astr. Soc.

(1975) 42,977-991

Joint Inversion of Geophysical Data

K. Vozoff and D. L. B. Jupp

(Received 1974 September 26)

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

Summary
By jointly inverting several different kinds of geophysical measurements
at a site we avoid some of the ambiguity inherent in the individual methods.
We show how this can be done for the combination of DC resistivity and
magnetotelluric measurements on a layered medium by considering a
simple 3-layer model. The combination resolves the resistivity of the
thin resistive second layer, even though neither of the two methods can do
so alone.
The method is then applied to field data from a shallow sedimentary
basin. A blind zone occurs beneath a thick near-surface conductive shale.
By a study of the eigenvalue structure of the model it can be seen that
resolution in this zone would be slightly enhanced by higher frequency
magnetotelluric data, but additional DC data at larger spacing would
yield no improvement.

Introduction
Inversion of geophysical data consists of operating directly on those data so as to
generate a view of the structure which causes them. It differs from the traditional
forward approach to geophysical interpretation in which a model is assumed, its
response is calculated and compared with the observations, and the model parameters
are then modified in a way which will hopefully improve the comparison. The virtues
of inversion, if it has any, are that it uses the data to the fullest while being more
economical of skilled interpreter time than is the forward approach. Consideration
of the interpretation problem also suggests that it will sometimes be more cost-
effective to acquire a limited number of measurements of several types than many
measurements of a single type, in development and expIoration programs.
The recent geophysical literature includes many works on development and
application of inversion techniques. It is a topic of widespread active research.
Most applied papers have dealt with seismic properties of the Earth, although some
have treated electromagnetic properties (Parker 1970) and localized exploration-
scale problems (Inman, Ryu & Ward 1973).
We treat here the problem of joint inversion of two related kinds of data, DC
resistivity and Ultra Low Frequency electromagnetic (magnetotelliric) measurements
in horizontally-layered conditions. The approach is outlined and examples are
presented in which the combination yields more satisfactory results than either of the
two methods does alone.
977
978 K. Vozoff and D. L. B. Jupp

Pn

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

FIG. 1 The general horizontal layer model.

The two forward problems

Theory for both the D C resistivity (DC) and the Magnetotelluric (MT) methods
in layered media is well known (e.g. see Ward, 1967, pages 93, 117). That is, given a
model containing a number of layers, their thicknesses, and their electromagnetic
properties (Fig. l), it is a nearly trivial exercise to calculate model responses to the
two methods. The analytic expressions for these responses are tabulated in Table 1.
Analysis of the two sets of equations illustrates the characteristics of each which
makes them complementary to some extent It is well known that the MT method
has poor sensitivity to resistive layers: a thin resistive layer is ignored whereas response
of a thick resistive layer depends only on its thickness and not on its resistivity. MT is
highly responsive to conductive layers even if they are very thin. D C on the other
hand is virtually unbiased in the responses (Fig. 2), but for thin layers these depend
only on the product of conductivity-thickness (conductive layers) or resistivity-

Table 1
DC (point source of
current I at r = 0 ) MT (plane wave)

In1 [ZI
+(u) .= arctan -
Re [ZI

Jo(x) - Besscl function Z“ L- W\’P“

 Joint inversion of geophysical data 9 79

( P E R Io D) ‘z
I 10 I00
10

model

In-m Ikm
p2
In-m 2 k m

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

1.0
w _--
Schlumberger D C
Cr MT

t-
z
W
[r:
a
a
a
a
0.1

I 10 100

CUR R E N T ELECTRODE SPACING,

SCH L UM BE R G E R ARRAY
(km)
FIG.2 Responses of the DC resistivity and MT methods to the 3-layer mode1 shown.

thickness (resistive layers). DC has poor resolution in the case of models containing
layers of highly-contrasting properties.
The two methods are complementary in another sense. For practical reasons
it is easy and inexpensive to obtain DC data relating to shallow depths (less than a few
kilometres), but difficult to obtain deeper DC data. Just the reverse is true of MT.
Low frequency MT information which refers to depths beyond a few kilometres is
normally more easily obtained than it is at higher frequencies. Yet MT data are
better interpreted with a knowledge of shallow conditions. In the examples which
follow we demonstrate some advantages of a stereoscopic approach to interpretation.

Formulating and solving the joint problem

Our approach to inversion is set out in the accompanying paper ‘ Stable lterative
Methods for the Inversion of Geophysical Data ’ subsequently referenced as SIM
(Jupp & Vozoff, 1975).
For general joint inversion problems, the proper combination of the two data
sets, and weighting of the influence of the same earth model on the different responses,
is very important.
980 K. Vozoff and D. L. B. Jupp
For isotropic, horizontally layered earth models, the MT data and DC data
measure ' earth ' response to two distinct inputs through quite distinct physical
processes. However, the output apparent resistivities have very similar features. For
example, at long (DC) spacings and long (MT) periods, both are asymptotic to
basement resistivity p N . For short spacings and short periods, both are asymptotic
to P1-
When the data are not from an (approximately) layered situation, of course, the
joint process cannot be expected to improve their interpretation. A more complex
model, together with joint inversion, is needed for this case. In SIM we describe how
an iterative method, based on the Gauss iterative method (Kowalik & Osborne 1968),
may be used to solve inverse problems, wherein data are fitted to a layered (or some
other more complex but nevertheless simplified) earth model. The approach is to
evaluate the forward problem to determine how well a current earth model fits the

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

data, and to evaluate the ' parameter influence ' (or Jacobian, or partials) matrix to
determine how the model should be altered to improve the fit.
The ' parameter influence ' matrix

has components

ag. (where w i indicates relative importance of observation i )

J , = wi
axi
which is the variation of the i'th data value, with respect to changes in the j'th para-
meter (layer resistivity, or thickness). That is, it measures the influence of changes in
parameters on the (model predicted) data.
To properly combine the separate DC and MT problems in an inversion method,
the separate influence matrices need to be balanced. In our problems, since the layer
resistivities and thicknesses have implicit constraints
~

pi20 i=l,N+l

hi>O i=l,N
we may force the constraints by letting
~

xj = log pj for j = 1, N + 1

x ~ +=~log+hj ~for j = 1, N .
If, moreover, we use relative errors, the elements of the influence matrix take the form

In this way, the Jacobian is made scale free, and since the relative errors will be
commensurate, the two kinds of data will equally influencethe correction that improves
the current model.
We also show in SIM how the eigenstructure (or singular values) of J may be
used to classify the parameters as Irrelevant, Unimportant and Important.
 Joint inversion of geophysical data 98 1

(i) Irrelevant parameters have no influence on the model data. They correspond
to layers which are out of range of the measurements, or as in the case of thin resistive
layers, to parameter combinations that cannot be resolved from the data.
(ii) Unimportant parameters have only small influence on model data. Inversely,
large changes in these parameters can occur for only marginal improvement in fit
to the actual data. For this reason they must be either neglected, or altered only
marginally during the inversion process.
(iii) Important parameters correspond to the well resolved, and often gross features
that are well represented in the data.
Many inverse problems in geophysics are ill-posed. That is, small changes in the
data can lead to large changes in the model. The ill-posed nature and consequent
numerical instability of the inverse problem is largely contained in its Unimportant

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

and Irrelevant parameters. In SIM we describe an inversion method that filters out
the unstabilizing effect on the joint, or separate, inverse problems.
Joint inversion can increase the numbers of Important parameters of a model to
include some which the methods cannot resolve separately. It is essential to note that
this effect cannot be achieved merely by increasing the number or accuracy of data
values of a single method. Rather, joint inversion is a means of eliminating some
defects inherent in the individual methods.
To analyse the stability of the result, we have used the Damped Error Bounds
described in SIM. These measure the expected variation in the well resolved
(Important) parameters, in response to small variations in the data. The bounds
ascribed to the Unimportant parameters are damped, and only measure variation
due to their correlation with the well resolved parameters. We need to distinguish
these narrow limits from the sharp detail of the Important parameters. The greatest
' expected variation ' occurs for parameters which are near the ' threshold ', or
dividing line between the Important, and Unimportant parameters. Usually, these
are the most interesting parameters, and the ones most strengthened by combining
the two problems.

Example using model data

To see whether joint inversion treats thin layers as anticipated we first studied a
three-layer model. The model chosen (Fig. 3) has a thin resistive second layer which
is extremely difficult for either method to resolve by itself. The apparent resistivity
curves are shown on an expanded linear vertical scale in Fig. 4.

Table 2
Three-layer inversion of three-layer model responses
Inputs: (i) Three-layer model responses (Fig. 4) consisting of MT apparent resistivities at
16 frequencies and Schlumberger (DC) apparent resistivies at 10 electrode
spacings, truncated to three figures.
(ii)
Models p1 p2 p3 h1 h2 RMS No. ofsignif.
error singular values
Actual 1 10 1 100 10 - 5
Starting 1.0 1.0 1.0 50 100

Outputs: DConly 1.0 1.9 0.94 89 71 0.2% 4

MTonlj 1.0 1.1 1.0 68 82 0.3% 4
COIIDb? 1.0 9.7 1.0 100 10 0.2% 5
982 K. Vozoff and D. L. B. Jupp

Table 3
Five-layer inversions of three-layer model responses
Inputs:
To 3 (a) Data as to Table 2
To 3 (b) Same data mixed with 3 per cent gaussian error
To 3 (c) Same data mixed with 6 per cent gaussian error.

Models p1 p2 p3 p4 p5 hl h2 h3 h4 RMS No. of

error SSV

Actual 1 1 0 1 - - 100 10 co -
Starting 1 1 1 1 1 20 50 50 50
outputs

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

3 (a)DC 1.0 0.99 1.8 1.6 0.94 23 60 57 50 0.2% 5
MT 1.0 0.97 1.2 1.0 1.0 20 49 51 50 0.3% 5
Coupled 1.0 0.96 5.4 0.78 1.0 40 55 23 28 0.2% 6

3 (b)DC 0.82 1.1 1.9 1.9 0.73 15 86 60 56 1.5% 5

MT 1.6 0.66 2.5 0.75 1.0 19 41 61 47 2.3% 5
Coupled 0.96 0.99 6.0 0.87 1.0 19 70 16 26 2.5% 7

3 (c)DC 1.0 0.97 2.8 2 . 7 0.47 31 80 60 60 4.3% 4

MT 2.1 0.51 4.3 0.48 0.99 21 33 96 40 4.7% 5
Coupled 1.0 0.89 2.7 0.57 1.0 34 53 83 60 5.0% 7
1-1 = 0.01 (defined in SIM)

DC and MT forward models were calculated. These were variously truncated and
mixed with noise, and used as inputs for inversions. Table 2 lists results of the individual
inversions and of the coupled inversion, to a 3-layer model, of noise-free data truncated
to three figures. These are illustrated in Fig. 5. Five parameters (pl, p,, p3, t i , , k,)
are involved so there are five singular values. Only the four largest are resolved in the
individual inversions, as indicated by normalized singular values +
0.01 and p,
remains unresolved. The solutions attained show this.
Joint inversion brings all five singular values into the significant range, resolving
both pz and h,.
In reality, the number of layers present will seldom equal the number in the model
chosen for inversion. To study the consequences of an excess of layers in the starting
model, the same 3-layer data were inverted to a 5-layer model (9 singular values) with
results shown in Table 3(a) and Fig. 6. Here the joint model again did better than the
other two, but the resistivity-thickness product is in error by 24 per cent (within
the predicted Damped Error Bounds).

P, =Ion-m
T
1
P, = In-m h,=100m

he= 10rn
P,= In-m
FIG.3 The model used in the 3-layer inversion study.
 Joint inversion of geophysical data 983
CURRENT ELECTRODE SPACING, SCHLtJMBERGER ARRAY (krn)
10 10 0
I60 I ,

I50

-E
I
c
Y
140

>
t
2 130
I-

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

K? +
v,
w +
120 +
+ DC
I-
z +
W
cL 110
3
a
a
100

90 I I I I I I I I I I I I I I I l l 1 I I I I l l 1

(PERIOD@ (se&
FIG.4 Input data (truncated) to the 3-layer inversion study.

As a further concession to reality, noise was added to the truncated data before
inversion. Gaussian-distributed random noise was added, at levels of 3 and 6 per cent,
with results given in Tables 3(b) and 3(c) and Fig. 7. Results deteriorate gradually,
but considering the nature of the model it seems remarkable that the results remain as
stable as they are.
Additional experiments with changing the lower threshold limit ji (see SIM) made
no clear differences.

Field example
DC and MT measurements were made in conjunction with the BMR* at Pirlta,
Vic., approximately 50 km WSW of Mildura. The area is in the Murray Basin of New
South Wales, South Australia and Victoria. Sparse geophysical and well data
indicate that it is comprised of a thin cover of near-horizontal Permian and younger
sediments overlying older granites and folded metasediments. The cover may be no
more than 1-13 km thick. Gravity and magnetic coverage is not complete but indicates
some lateral contrasts at depths of 2-3 km with a generally north-easterly strike.
Refraction seismic measurements by the Bureau of Mineral Resources (Watson
1962) show a strong refractor at about 0.6 km depth in the Pirlta area. This may be a
conglomerate encountered in the only well in the area, AOG Wentworth No. 1,
60 km to the north-east. A later BMR reflection survey obtained coherent primary
reflections at nearly 15 s from a 100 kg explosive charge (Branson, Moss & Taylor,
* Bureau of Mineral Resources, Geology and Geophysics, Canberra.
984 K. Vozoff and D. L. B. Jupp

THREE L A Y E R INVERSION OF THREE LAYER MODEL

10
8

4
E

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

I
E
c
0
v

2
>
k
2
I-
K2
v)
1.0 c

.8

.6

4
10 100 I000
DEPTH (m)

FIG.5 Result of inverting 3-layer truncated data to a 3-layer model.

1972). Such a remarkable result would seem to indicate minimal scattering and hence
simple structural conditions. This is in accordance with the Geomagnetic Depth
Sounding results of Lilley & Bennett (1973) which show negligible apparent lateral
variation of electrical conductivity at periods of from one minute to one day.
The resulting impression is that the Pirlta area is basically horizontally layered,
but with minor lateral variation in magnetic susceptibilities and densities in the
2-5 km depth range.
DC measurements were made along a pair of perpendicular lines crossing at their
centers. Schlumberger arrays were used with current electrode spacings from
20 m to 6 km (NS) and 200 m to 6 km (EW). These data together with the inversion
results are shown in Fig. 8. A shallow, conductive (< 1Q-m) shale is widespread,
encountered in many water wells (Polak, private communication).
MT measurements were carried out at a site 5 km to the north-west using techniques
virtually identical to those described earlier (Vozoff, 1972). Resulting apparent
resistivities and principal axis directions are shown in Fig. 9. (For reasons not yet
ascertained, data at frequencies above 0.1 hz were unusable.) The major principal
axis direction agrees well with the strike direction of the gravity and magnetic data.
When the DC and MT data were jointly inverted the results of Fig. 10 were
obtained. Table 4 lists the normalized singular values and their eigenvectors for the
final models. These give the interesting result that there is a gap between the second
 Joint inversion of geophysical data 985

F I V E L A Y E R INVERSION OF THREE LAYER MODEL

Truncated Data

10
8

6
I1 -

!
-
E
4 I
I

!
I I
E I

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

I
c
0 1
v I

> 2 I

.4
10 100 1000
D E P T H (m)

FIG.6 Result of inverting 3-layer truncated data to a 5-layer model.

and fifth layers, beyond and reach of the DC data but too resistive and shallow for
access by most of the MT spectrum. In fact a study of the Jacobian shows that the
highest MT frequencies also respond to the shallow conductor but not to layers 3
and 4 beneath it.
That such a situation could arise beneath a thick conductive bed seems reasonable.
However, if that depth range happened to be of particuIar interest, what measurements
would help to define its conductivities? Would it help to take DC measurements at
larger spacings, or to acquire higher frequency MT data? One way to answer these
questions is to examine changes in the list of well-resolved parameters as higher
frequencies or larger resistivity electrode spacings are added to the list of measure-
ments. This was done with the major axis MT data set and the EW DC set.
The original data set included pa at 33 frequencies, from 2.45 x to
7-7 x hz, and 12 sets of electrode spacings from 200 m to 6 km. We found that
extending DC measurements to 26km had no significant effect on the Damped
Error Multipliers. Extending the frequency range upward to 1.0 hz reduced the
Damped Error Multipliers for p 3 , p4, h3, and h,, but they remain Unimportant. The
conclusion is that, in this configuration of parameters, layers 3 and 4 represent a
zone of ' near-blindness ' to this combination of measurements.
A number of interesting problems arise in attempting this application. Both DC
and MT measurements were made with wires in the N-S and E-W directions. In the
normal course of MT analysis it was found that the principal impedance axes at low
frequencies were approximately N45E and N45W. Apparent resistivities are computed
986 K. Vozoff and D. L. B. Jupp

FIVE LAYER INVERSION OF T H R E E L A Y E R M O D E L 6% GAUSSIAN NOISE ADDED

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

.4
I
10 100 100
DEPTH (m)

FIG.7 Result of inverting 3-layer data with 6 per cent error to a Slayer model.

along principal axis directions. Which DC curves should each be coupled with for
joint inversion ?
In this particular instance an effort was made to observe possible non-isotropy
while making the DC measurements. Potentials were measured perpendicular to the
current line at several spacings, but no detectable field was measured in that direction.
It was concluded that conditions were isotropic within the depth range of importance
to the DC work. There is very little difference between the two DC curves, so that
the conclusion could be accepted in this case. Generally it will be necessary to ascertain
strike direction, and to make DC measurements in those co-ordinates, if measurements
are to be coupled. This will have to be done cautiously, in view of the ' anisotropy
paradox ' (Maillet 1947) applying to anisotropic media.

Discussion
The examples presented constitute only a preliminary test of the concept. In
the field example the data barely overlap in their coverage (because of equipment
malfunction), while there is nearly complete coverage by both in the three-layer
model studies: one would normally expect a situation somewhere between.
Nevertheless they do illustrate the potential of joint inversion to extract the utmost
from a method and to define its limits.
 Joint inversion of geophysical data 98 7
10

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

(r

t-
z
W
LL
U
a
a
a

0.1
10 100 1000 lop00
CURRENT E L E C T R O D E SPACING - SCHLUMBERGER ARRAY (rn)

100 100
---- NS

E
I
-- EW
cf NS 10
1
10
t
t
>
-
t
!vc)
w -
n
w 10. - 10
2
a EW 1
c

01 t I .01

Other tests were run on the three layer model to determine the minimum amount
of data necessary for a second method to exert a significant effect on the result. It
was found that three or four point of DC data were sufficient provided they were
well distributed. Ten or twenty extra points had little further effect. This appears to
bear out one's intuitive ideas that with noise-free data it is only necessary that the
response function be adequately sampled. Field data will always include noise, so
there i s a good statistical reason to make more than a bare minimum of measurements.

K
 Table 4
Normalized singular values of Jacobian MT major axis/DC E-W
0.100+01 0.105+00 0.995-01 0.335-01 0.286-02 0.732-03 0.677-03 0.214-03 0.275-04 0.657-05 0.715-17
Parameter space eigenvectors (V matrix)
P1 1 0.006 -0.022 0.025 0.150 -0.084 0.914 0.214 -0.295 0.024 -0.006 0.006
P2 2 0'716 -0.474 0.458 -0.229 0.008 0.008 -0.002 -0.005 0.OOO -0.OoO 0.000
P3 3 0.060 0.039 -0.043 0.022 0.011 -0.056 -0.009 -0,153 0.060 -0.033 0.980
P4 4 0.004 0.003 -0.003 0.002 0.015 -0.046 -0.005 -0.165 0.065 0.983 0.000
Ps 5 0.013 0.017 -0.010 0.023 0.990 0.080 0.102 0.051 0.004 -0.002 -0.OoO
PS 6 -0.001 -0.004 -0.003 -0.001 -0.082 -0.221 0.970 0.021 0.003 -0.001 0.000 P
hi 7 -0.002 -0.222 0.243 0,932 -0,003 -0.141 -0.033 0,050 -0'004 0.001 -0.001 c
hz 8 -0.691 -0'447 0.501 -0.235 0.034 -0.01 1 0.001 -0'096 0.032 -0.013 0.069 R
h3 9 -0.059 -0-040 0.045 -0.026 0.065 0.270 0.041 0.848 -0.358 0.180 0.184 %
h4 10 -0.004 -0.001 0.005 -0.003 0.030 0.087 0.008 0.359 0.929 0.004 0.005 w
h5 11 0.035 0.723 0.690 -0.007 -0.006 -0.002 0.004 -0.002 -0.002 -0.OOO -0.001 a
P
Normalized singular values of Jacobian MT minor axis /DC N-S F
0*100+01 0.419-00 0.307-00 0.149-00 0.108+00 0.291-02 0'173-03 0.104-03 0,725-05 0 569-05 0.148-05
Parameter space eigenvector (V matrix)
P1 1 0.013 -0.017 0.660 0.642 0.390 -0.OoO O.Oo0 -0.OoO 0.OoO
P2 2 0.740 -0.110 0.115 -0.434 0.489 -0.001 0.001 -0.OoO 0.OoO 0.m -0.OoO
P3 3 0.040 -0.004 -0.010 0.029 -0.033 -0.001 0.080 -0.090 0.991 -0.001 -0.019
P4 4 0.002 -0.m -0-001 0.002 -0.002 -0.OoO 0.062 -0.147 O.Oo0 0.OOO 0.987
P5 5 0.001 -0.m -O*OOO 0.001 -0.002 -0.01 1 0.941 0.335 -0.046 0.017 -0.009
PS 6 -0.OoO 0.012 0.001 -0.002 0.003 1.OOO 0.011 0,003 0.OoO 0.001 -0.m
hi 7 0.016 -0.023 0.723 -0.405 -0.559 0.0oO -0.OoO O.Oo0 -0.OoO -0.OoO 0.000
hz 8 -0.651 0.116 0.166 -0'484 0.542 -0.004 0.025 -0.058 0.053 0.019 -0.006
h3 9 -0.038 0.012 0.010 -0.029 0.033 0.001 -0.291 0.857 0.107 -0.380 0.146
h4 10 0.001 0.020 0.001 -0.002 0.002 -0.OoO -0.138 0.347 0.045 0.924 0.060
h, 11 0.161 0.986 0.022 0.010 -0.016 -0.011 0.004 -0.01 1 -0.005 -0.017 -0.002
Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2
 MINOR AXIS PRINCIPAL AXIS
DIRECTION A P P A R E N T RESISTIVITIES (n-m)
(Deg E of Mag N ) -
- 0
N P a a, - - 0 0
I 0 I
I 1 I 1 0 I 1 I I I 1 1 1 1 I I I I I I l l
0

U
. .
m
d . ;D
.e
-
0
0 8 0
.@

-
- 0
0

ul 0
(D X
0 -I
u,
v D

e -
I I . I I I I
I I I I I I I 0
Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2
990 K. Vozoff and D. L. B. Jupp

MINOR M MAJOR MT
AXIS AXIS

25
48
c47-491
+
19

i (4.6)

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

.74
[7-.8]
81
[6-11]

w
-1
Q
427
(5 24) &
V
0
I
I-
E 730
Y
- (830)-
- (10801 & (772)
(78) error bars a n d brackets
show D a m p e d E r r o r Bounds
f o r resolved parameters
assuming d a t a a c c u r a t e
a to f 5 %
w curved brackets ( ) show
0
poorly resolved parameters
(2000)

2500
E
Y

0
13.3 krn -I-
(017)

E
Y
0
0 8 5 krn

FIG. 10 Results of joint inversions to layered models, Pirlta, Victoria.

 Joint inversion of geophysical data 99 1
Joint inversion has some implications for the design of measurement programs.
The location of new measurements for greatest effectiveness depends on a very
non-linear way on the true conductivity distribution. If each measurement is expensive
then one would make a few, invert, and (from the result) find the best parameters
for the next set of measurements, etc. In geophysics it is often more expensive to
return to a site than to take extra measurements, so that an excess of measurements
is made to start. However there are clearly some situations when the reverse is true.
There exists an unavoidable influence of the starting model on the final model. It
appears to be less with the present strategy than with many earlier approaches, and
the Damped Error Bounds appear to accurately define the tolerance ranges for signifi-
cant parameters.
Extending joint inversion to include other kinds of electrical or electromagnetic
measurements can be done in the same way as we have done here. It would be equally

Downloaded from http://gji.oxfordjournals.org/ at Pennsylvania State University on May 16, 2016

interesting to include indicators of other physical properties-seismic surface wave
characteristics for example-but their incorporation within a common framework
will require further study.

Acknowledgments
The Australian Research Grants Committee and Macquarie University supported
this research. The Bureau of Mineral Resources, Geology and Geophysics made it
possible to acquire the field data. In particular Mr D. Kerr and Dr E. Polak of the
BMR and Dr R. J. G. Lewis of Macquarie University (now at the University of
Tasmania) made major contributions to the collection and analysis of the data. We
are also indebted to Dr Jack Cribb of the CSR Research Department and Dr David
Johnson of Macquarie University for their critical comments.

References
Branson, J. C., Moss, F. J. & Taylor, F. J., 1972: Deep Crustal Reflection Seismic
Test Survey, Mildura, Victoria and Broken Hill, N.S.W. Bureau of Mineral
Resources Record 1972/127, Canberra.
Inman, J. R., Ryu, J. & Ward, S. H., 1973: Resistivity inversion, Geophysics, 38,
(6), 1088-1 108.
Jupp, D. L. B. & Vozoff, K., 1975. Stable iterative methods for the inversion of
geophysical data, Geophys. J. R. astr. SOC.
42, 957-976
Kowalik, J. & Osborne, M. R., 1968. Methodsfor unconstrainedoptimizationproblerns,
American Elsevier.
Lilley, F. E. M. & Bennett, D. J., 1973. Micropulsations recorded by an array of
magnetic variometers, J. geophys. Res., 78, (lo), 1603-1607.
Madden, T. R., 1972. Transmission systems and network analogies to geophysical
forward and inverse problems. Tech. Report No. 72-3, Department of Earth
Sciences, MIT, 45 pp.
Maillett, R., 1947. The fundamental equations of electrical prospecting, Geophysics,
12, (4), 529-556.
Parker, R. L., 1970. The inverse problem of electrical conductivity in the mantle,
Geophys. J. R. astr. Soc., 22, 121-138.
Vozoff, K., 1972. The magnetotelluric method in the exploration of sedimentary
basins, Geophysics, 37, (l), 98-141.
Ward, S. H., 1967. Electromagnetic theory for geophysical applications, in Mining
Geophysics, v.2, Society Exploration Geophysics, Tulsa.
Watson, S. J., 1962. Murray Basin Seismic Survey, 1960, Bureau of Mineral Resources
Record No. 1962/164, Canberra, 19 pp.