See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/231015741

Inversion techniques applied to resistivity inverse problems

Article in Inverse Problems · January 1999

DOI: 10.1088/0266-5611/10/3/011

CITATIONS READS

47 3,573

3 authors, including:

Sheenal Narayan Maurice B. Dusseault

Douglas College University of Waterloo
12 PUBLICATIONS 177 CITATIONS 495 PUBLICATIONS 8,120 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Maurice B. Dusseault on 23 May 2014.

The user has requested enhancement of the downloaded file.

Inverse Problems 10 (1994) 669-686. Printed in the UK

Inversion techniques applied to resistivity inverse problems

Satyendra Narayant, Maurice B Dusseaultt and David C Nobests
T DepartmentofEaithSciences. Universityofwaterloo, Waterloo.Ontario. Canada NZL 3GI
$Department of Geology, University of Canterbury, Christchurch 1, New Zealand

Received 16 February 1993. in final form 24 November 1993

Abstract. This paper reviews various inversion techniques applied to resistivity inverse
problems, and proposes a fairly different approach. The advantages and disadvantages of
various resistivity inversion methods are briefly summarized.
A good inversion method must simultaneously minimize the effects of data error and
model parameter errors. The necessary requirements for inversion of any geophysical data
are a fast forward algorithm for calculating theoretical data for initial model parameters,
and a technique for calculating derivatives of the data with respect to the model parameters.
Unfortunately. the second requirement is not readily available for two-dimensional (ZD)or
three-dimensional (3D) inverse resistivity problems. A few workers have examined the
problem of resistivity inversion, but no claim has yet been made for an optimum inverse
method for ZD or )D resistivity data. Most resistivity inversion methods are also restricted to
surface electrode arrays.
A multi-dimensional resistivity inversion method using perturbation analysis and
reciprocity is proposed. The theoretical basis for this method is briefly developed. This
approach is an efficient way of calculating the derivatives of data with respect to the model
parameters because the inverse formulation thereby developed is a linearized form of the
general non-linear problem. An attempt is also made to provide an update on the state of
knowledge of resistivity inverse problems.

1. Introduction

The interpretation of resistivity survey data for surface and subsurface electrode arrays
was previously carried out by analogue and analytical methods (e.g. Van Nostrand and
Cook 1966, McPhar Geophysics 1966, Scurtu 1972. Bibby and Risk 1973, Telford er al
1976). These methods require a certain degree of symmetry, and are suitable only for
simple geologic situations such as halfspaces, layered earth cases, inhomogeneities
that approximate spherical shapes, and simple faults. Complex resistivity distributions
cannot be dealt with by analytical methods. and must be approached using numerical
techniques. Rapid development of computer technology since 1970 has made interpre-
tation easier for complex geologic structures, and as a result, new techniques of direct
interpretation of data have been developed.
The inverse problem in resistivity interpretation was reported as early as the 1930s
(e.g. Slichter 1933, Stevenson 1934, E j e n 1938, Pekeris 1940, Vozoff 1958 and many
more). Slichter (1933) reported a method of interpretation of resistivity data over a
layered earth using Hankel's Fourier-Bessel inversion formula. It gives a unique
solution if the resistivity is a continuous function of electrode spacings. In practice,

8 Formerly at Department of Earth Sciences. University of Waterloo, Waterloo, Canada.

0266-561 1/94/030669+ 18$19.50 0 1994 IOP Publishing Ltd 669

670 S Narayan et a1

resistivity measurements are limited to a small number of readings taken at discrete

electrode spacings. arid thus, a unique resistivity response does not exist. Vozoff
(1958) tested Slichter’s method on field and synthetic data generated for three- and
four-layer models. The method becomes unstable if the data are noisy. Zohdy (1972)
proposed a method of direct resistivity-interpretation which is valid for noisy data as
well. However, none of these earlier investigations deal with existence, uniqueness, con-
struction and stability, which are important concerns and must be dealt with within any
inverse problem.
Backus and Gilbert (1967, 1968, 1969) introduced a linear inverse theory for
geophysical problems. They thoroughly discussed model resolution, least-squares fit
of the data. and solution uniqueness. The method is valid even for noisy or insufficient
data, and they quantified the trade-off between resolution and stability for solutions to
inverse problems. The Backus-Gilbert approach, as do many others, suffers from the
difficulty in estimating the degree of smoothness for all admissible models; this is
required to calculate the greatest deviation between the estimated and true models.
Following Backus and Gilbert’s work, generalized linear inverse theory was described
by Wiggins (1972) and Jackson (1972) in terms of linear algebra. They described
classical solutions to inverse problems and resolution of model parameters in terms
of matrix algebra.
The application of generalized inverse methods to resistivity data will be
described. The limitations and refinements of different resistivity inverse methods
proposed so far are summarized in the following sections, followed by a brief presenta-
tion of a multi-dimensional resistivity inversion using an approach different from the
others.

2. One-dimensional (ID) resistivity inversion

The method of generalized linear inverse theory for the resistivity problem was
introduced initially by Inman et a/ (1973). They described the linearization of inverse
resistivity problems and minimization in a least-squares manner to find the best
possible solution. They also discussed important information such as the noise level
(random error) associated with data, and the eigenvalues and eigenvectors of the
system matrix, which are essential to study model parameter resolution. They did not
discuss the singularity of the system m3trix that arises with small but non-zero eigen-
values, which was experienced by Jupp and Vozoff (1975), Vozoff and Jupp (1975),
and Inman (1975) in many resistivity problems. Thus, because of its ill-conditioned
nature, the generalized linear inverse method is unstable for geophysical inverse
problems which have a nearly singular system matrix. Only very simple resistivity struc-
tures do not produce a system matrix which is nearly singular.
Hoer1 and Kennard (1970a, b) showed that linear estimation from non-orthogonal
data (i.e. nearly singular data) could be refined or improved by using biased estimators;
this technique has been named ‘ridge regression’. Marquardt (1970) established the
similarities between the generalized inverse method and ridge regression method, and
proved the suitability of ridge regression methods for problems with small eigen-
values. Because most common resistivity problems involve small eigenvalues, Inman
(1975) introduced ridge regression for inversion of resistivity data. Later, the method
was also used by Petrick et a/(1977) and Glenn and Ward (1976) for one-dimensional
(ID) resistivity inversion.
 Inversion techniques 671

Hoversten e t a1 (1982) studied five different least-squares inversion techniques and

their speeds of convergence over horizontally layered resistivity structures. They
showed that, of the five methods studied, ridge regression requires the fewest number
of iterations to reach a desired minimum. A critical review of least-squares inversion
techniques and their application to layered geophysical problems has been done by
Lines and Treitel(l984). Marcuello and Queralt (1985) proposed an iterative technique
for inverting I D resistivity data, where an error function is minimized. They tried to
resolve equivalence in the resistivity data, and were able to obtain a more realistic
model by applying a regularization process. The drawback of this method is that it
requires that the initial assumed model be close to the true solution.
Parker (1984) addressed the non-uniqueness of ID inverse resistivity problems using
a bilayer expansion method. His models consisted of layers having uniform thickness,
and the solution determined the optimum number of layers and layer thicknesses that
would minimize the deviation in a least-squares manner. Assal and Mahmoud (1987)
developed an algorithm for interpretation of resistivity data over a layered model.
setting the tbickuess of each successive layer equal to the required depth resolution.
They derived spectral reflection coefficients at the earth’s surface for the model.
These coefficients are functions of the resistivity ratios of adjacent layers and are
used in the evaluation of model parameters. Their results indicated that the algorithm
is suitable for continuously varying resistivity with depth.
The basic limitation of all I D resistivity inversion methods is that they only consider
vertical variations in the earth’s resistivity: thus they are generally limited to cases with
simple, planar spatial variations, and cannot be applied to evolving situations such as
arise in in situ monitoring or strongly to three-dimensional stmctures.

3. Two-dimensional (ZD) resistivity inversion

Until recently, direct interpretation of resistivity data using inversion methods was
common only for horizontally layered structures. However, layered models are
inadequate in applications such as mineral exploration; study of dikes, valleys, contact
zones. and geothermal fields; monitoring of steam, water or chemical flooding for
enhanced oil recovery; mapping of groundwater contamination: and in situ monitoring
of mining methods. Numerical modelling techniques for surface electrode arrays (e.g.
Madden 1967, Coggon 1971, Lee 1975, Jepsen 1969, Mufti 1976, Dey and Morrison
1976) have been extensively used on a trial-and-error basis to interpret resistivity
data in terms of two-dimensional (?D)geologic structures. Numerical modelling of resis-
tivity mapping using subsurface electrode(s) permitting better accuracy and resolution
than surface electrode use only is described by Daniels (1977, 1978. 1981, 1983). Trial-
and-error modelling (i.e. optimization of a model based on a forward solution) for
interpreting the resistivity field data is rather difficult and time consuming. At the
same time, forward modelling does not yield much information on resolution. Thus,
there is a need to develop an inverse technique to interpret resistivity data in terms
of ?D or 3~ subsurface structures.
Pelton e t a1 (1978) developed a computer algorithm for inversion of :D resistivity
and induced polarization (IP)data. This method calculates a data array containing solu-
tions of the forward problem for the range of anticipated model parameters (i.e.. depth
of overburden, thickness and width of anomalous body, resistivity of anomalous body,
resistivity of the host rock, and centre of the body). Then, it uses spline interpolation
672 S Narayan et nl

between the various stored forward models. Spline interpolation makes the forward
problem continuous with respect to the model parameters, so that the derivative with
respect to the parameters may be obtained. This algorithm is not well suited to
complex cases because the calculation of partial derivatives with respect to the data
and interpolation between the models is extremely difficult.
Smith and Vozoff (1984) and Tnpp et a1 (1984) proposed a 2D resistivity inversion
using a finite difference technique, and a transmission surface analogy with Cohn’s
sensitivity theorem. respectively. Their schemes are similar and suitable for complex
ZD models, but do not incorporate the effects of topography on resistivity data in the
inversion scheme. Although a terrain correction has been advocated by Fox et a1
(1980) and Holcombe and Jiracek (1984), it may not be possible for a complex ZD
model to completely separate the resistivity anomalies because of topography from
those associated with subsurface features by using a simple correction. Tong and
Yang (1990) developed an algorithm for ?U resistivity inversion where topography is
considered in the model, allowing for the direct inversion of resistivity data obtained
from a rough terrain, without applying external corrections in advance.
Automatic ZD resistivity~inversionusing alpha centres has also been reported (Shima
1990). The alpha centres method is a non-linear transformation, and has been applied
to many simple resistivity problems involving direct current conduction equations.
Shima (1990) developed an algorithm to invert resistivity data gathered over complex
ZD structures, formulating the forward problem by the alpha centres method, and the
inverse problem by steepest descent and Gauss-Newton methods. His field investi-
gations show that the method has good application for resistivity surveys in steep,
mountainous areas.
LaBrecque and Ward (1988) outlined the inversion of borehole-to-borehole resistiv-
ity data. However, their work was confined to apnsmatic anomalies in a resistive back-
ground. Most of the inversion techniques based on non-linear inverse formulation
require that the resistivity of a set of blocks each of fixed size, shape and position be
refined/adjusted to fit the field data. The geometry of the model also needs to be
defined in advance.
Recently, several papers describing imaging of subsurface resistivity structures have
appeared in the literature. Fry and Neuman (1985) introduced a technique to image
subsurface features for resistivity problems using an impedance-computed tomography
algorithm. The algorithm employs the solution of the Poisson equation without
assuming straight line behaviour of the current flow paths. This approach usually
requires extensive computation and a large number of iterations based on the
Newton-Raphson scheme to reach the h a 1 solution. Daily and Owen (1991) describe
a tomographic inversion scheme to image the resistivity distributions between boreholes
from current and voltage data measured along the boreholes. They used a finite element
Newton-Raphson algorithm developed by Yorkey (1986). Their test results indicate
that the accuracy of image reconstruction and spatial resolution mainly depends on
data errors. Ramirez et a1 (1992) introduced resistivity tomogaphy to monitor an
underground steam injection process.

4. Three-dimensional ( 3 ~ )resistivity inversion

Full application of resistivity methods in geophysical prospecting has not been possible
because of an inability to calculate the anomalies caused by complex 3~ structures.
 Inversion techniques 673

Dieter e t a1 (1969), Bakbak (1977), Hobmann (1975), Daniels (1977), Yang and Ward
(1985). Eloranta (1986), and Beasley and Ward (1986) describe numerical modelling
methods using an integral equation approach for simple prismatic inhomogeneities.
The drawback of these methods is that they allow inhomogeneities embedded in a
homogeneous host only. A numerical modelling method to calculate the resistivity
response for arbitrary 3~ bodies has been developed by Dey and Morrison (1979)
using a finite difference scheme, and by Pridmore et a1 (1981) using the finite element
method. Inversion of ;D resistivity data is not yet well developed because of the com-
plexity and computational costs.
A ;D resistivity inversion approach using alpha centres has been reported by Petrick
et a1 (1979). In this method. the forward solution is accomplished by the alpha centres
method, and a 30 inverse algorithm is developed using the ridge regression method. This
algorithm requires less than 15 000 words of computer memory and can be used on
small computers. The use of this alpha centres method without modification (as pro-
posed by Shima 1990) is not, however, valid for a complex conductivity distribution.
Thus, the method is only useful for field data interpretation to guide drilling site choice
and to obtain a good initial guess for initializing more sophisticated and costly inver-
sion schemes.
Recently, Park and Van (1991) developed an inverse algorithm to invert pole-pole
resistivity data over ;D resistivity structures using an approach very similar to that of
Narayan (1990). However, they were able to map lateral resistivity variation more
accurately than the vertical resistivity variation.
The iterative approach described in the preceding sections consists of an initial
model guess, updated at each iteration until a specified level of convergence (or number
of iterations) is achieved in a least-squares manner applied to a so-called ‘objective
function’. The objective function is the weighted sum of the squared residuals between
observed and computed data. In initial resistivity inversion developments, weighting of
data fit by using the data variances was not employed. Incorporation of weighting by
data variance in an inversion scheme is commonly referred to as ‘chi-squared error’
in the literature. The reduced chi-squared error, which is obtained by dividing the
chi-squared error by the number of degrees of freedom, provides an estimate of the
data fit (e.g. a value close to unity indicates a good fit within the noise level specified
by the data). This variance-weighted iterative approach is straightforward but is
unequivocally valid only when all of the data are uniformly affected by the model
parameters: in practice, this is rarely encountered.
A classifcation of model parameters in terms of their effects (important, unimpor-
tant, or irrelevant) has been given (Hohmann and Raiche 1988). Theoretically. the
importance of model parameters can be described by the size of the eigenvalues of
the matrix (i.e. transpose of the square of the coefficient matrix) whose inverse is
required to develop a stable solution. The large range of eigenvalues encountered in
most resistivity problems causes oscillation in the model parameters. Inversion
schemes based on eigenvalue information (so-called singular value decomposition,
SVD, e.g., Tripp et a1 1984) require computing the singular value decomposition of a
coefficient matrix, which is computationally intensive.
A computationally attractive approach has been described previously which
involves the addition of a damping factor to the matrix (which is the transpose of
the square of the coefficient matrix) before the inversion (e.g., Marquardt 1963). This
technique has been given different names based on the criterion used for damping.
For completeness, a brief review of inversion schemes is summarized in appendix 1.
674 S Narayan et a1

The use of damping prevents singular values from being smaller than the damping
factor and also yields stable inverse solutions. An Occam’s inversion by Constable et a1
(1987), which minimizes model roughness as well as data misfit, is gaining popularity.
In this method, the amount of damping determines the extent of smoothing. Irrespec-
tive of the choice of D (as discussed in appendix I), the selection of the appropriate
amount of damping without apriori knowledge of eigenvalues is very difficult, because
the eigenvalue magnitudes are model dependent. A method of selecting the amount of
damping is described by Oristaglio and Worthington (1980). The choice of the amount
of damping is still contentious and damping is often used on a trial and error basis.
Large amounts of damping push the solution towards the steepest ascent technique
with a higher degree of smoothing, and small amounts of damping minimize data
misfit according to the Gauss-Newton technique. Eaton (1987) proposed selection of
higher damping in the beginning and lower damping at each subsequent iteration.
This approach appears to be useful to minimize model roughness and data misfit.
The criterion of selecting lower damping at each subsequent iteration is arbitrary and
must be derived from experience.

5. Multi-dimensional resistivity inversion theory

The necessary requirements for inversion of any geophysical data set are a fast forward
algorithm for calculating theoretical data from input model parameters, and a tech-
nique for calculating derivatives of the data with respect to the model parameters.
These derivatives are known as the Jacobian or derivative matrix. The second require-
ment seems not to be readily available for ZD or 3~ inverse geophysical problems, and
no claim has yet been made for an optimum inverse method for ?D or 3~ geophysical
data. An effort has been made to circumvent this problem and to develop a multi-
dimensional resistivity inversion theory, and this development will be presented after
a brief review of methods to compute the Jacobian.
The following methods to compute the Jacobian matrix have been presented in the
published literature.
(i) The partial derivatives can be approximated by Enite differences after computing
two model responses for each model parameter value, one being the base case, one
being the case for a small change in the parameter (e.g. for { p } and for { p + Ap}).
This approach is computationally demanding.
(ii) Network sensitivity theorems may be employed (Tripp et a1 1984).
(iii) Direct differentiation of the field equations may be used (Rodi 1976). This
iterative scheme requires computation of the derivative matrix (i.e. the Jacobian
matrix) at each iteration. This computation, plus additional matrix algebra, are
required in the discrete inverse problem. When the problem is large (many discrete
data points and discretized model parameters) it becomes quite tedious to handle.
This situation can be avoided by formulating the inverse problem using Green’s func-
tions.
(iv) Direct use of Green’s functions can be implemented (Weidelt 1975). The use of
Green’s functions in the formulation of geophysical inverse problems has many differ-
ent names, such as the ‘functional space formulation’ in the inversion of heat flow data
(Shen et al 1990, Shen and Beck 1991) and the ‘adjoint state technique‘ in seismic
inverse problems (Tarantola, 1984, 1987, 1988; Tarantola et al 1988). Both these tech-
niques involve transformation of the differential equation to yield a Green’s function.
 Inversion techniques 675

Herein, the use of perturbation analysis and the reciprocity theorem of Lanczos
(1961) are proposed as a basis for an inverse formulation. This approach is somewhat
different from those which have been used to date in resistivity problems to compute the
Jacobian matrix. The use of perturbation analysis and reciprocity results in an algebraic
equation which in turn is used to construct the Jacobian matrix. The advantages of this
approach are that it gives an efficient way of calculating the partial derivatives of data
with respect to the model parameters, and the sensitivity of measurements is propor-
tional to the power dissipated in the anomalous zone. The sequence for typical inver-
sion schemes is briefly summarized in the form of a Row chart (figure I), where an
asterisk indicates the section for which we propose a somewhat new approach.
The equations governing the DC response due to a point current source I s ( x , y , z )are
given by (e.g. Telford et al 1976)
V+,Y,4 = -P(x>Y>=)J(x,Y,4 (1)
and
V . J ( X , Y , ~=
) I,(x,y,z), (2)
where $(x,y.z) is the electrical potential, J(x,y , z) is the current density and p ( x , y , = )is
the 3~ resistivity distribution. The basic equations (1) and (2) governing the potential $
and current density J around a point source I can be rewritten as an operator D acting
on V with source p
DV=P (3)
where

D= [v. "1 0 I:[=. and P= [;I. (4)

Figure 1. Flow chart for a typical resistivity inversion algorithm. The new method proposed
in the text enters in step (*).
676 S Narayan et a1

A perturbation analysis can be done to determine the sensitivity of the potential field to
changes in resistivity at depth (figure 2). Perturbing the resistivity gives
(D+ 6D). ( V f 6V) = 0 (5)
where

6D = [ ]
AP
0
0
0
and

Expanding this equation and neglecting second order terms yields

D.6V = -6D. V. (7)
Comparison of equations (3) and (7) reveals that changes in conductivity act as equiva-
lent sources of current for changes in potential fields at the surface.
Using the above relation, the concept of reciprocity, and the generalized Green’s
identity (Lanczos 1961), the changes in the potential field at the surface due to changes
in the conductivity of each discretized block (figure 2) is given by (Narayan 1990)

64 =
ITSpJ- J‘dr.

In the above relation, J’ is the current density of a discretized block after interchanging
(8)

the current and potential dipole. Using equation (S), the sensitivity of a surface
measurement to the changes in resistivities of a 3~ structure discretized into small
blocks can be expressed by the algebraic sum of sensitivities to individual blocks.
For a particular set of measurements, the sensitivity of the potential field over a 3~ dis-
cretized body can be written as

The current densities can be obtained for each block by a forward scheme, so that the
above equation reduces to

Figure 2. Perturbation analysis and inverse formulation for multi-dimensional resistivity

problems. Vertical cross section of the 3D body at depth is shown here. Reciprocity can be
established by interchanging the current and potential dipoles.
 Inversion techniques 677

RESISTIVITY INVERSION

1. Select field data in the form of apparent resistivity or voltage difference.

2. Select initial model compatible with the gross aspect of the field data.
3. Discretize the initial model into numerous blocks.
4. Compute the voltage at each nodal point of the discretized blocks due to ea
current electrode and also to compute the voltage at each nodal point after
interchanging the location of current and potential electrodes for each set of
measurements using forward algorithm.
5. Compute the logarithmic difference of field and calculated data.
6. Compute current density J (i.e. current density due to the current dipole for a
given position of current and potential dipoles) and current density J' (i.e.
current density due to the current dipole after interchanging the position of
current and potential dipoles) in x , y, and z directions for each block in the
discretized model.
7. Equation (8) is equivalent to a matrix equation as follows:
[Difference of field and calculated data] =[coefficient matrix] * [model para-
meters] i.e.

[@INX I =[ C M l ~ x . ~ [ ~1.f l ~ x (1)

8. Linearize the potential field data (which is a function, 4 ( p ) ) about an initial
value po, use log weighting as explained below. Arrange the computed data
into a matrix form as shown in step 7. The number of rows and columns in
the coefficient matrix indicates the number of data points and unknowns of
the model parameters respectively. Then, compute model parameters.

This can be reduced to

where

9. Update old model parameters, compute new resistivity data, and minimize the
field data in a least squares manner until a best fit is obtained.

Figure 3. Sequence of resistivity inversion scheme for multi-dimensional structures.

678 S Narayati et a1

This relationship can be reduced for ZD and I D structures by assuming that the current
density in the strike direction of the ZD body is negligible (i.e. the average current density
in the strike direction is zero), in which case 64 for ZD bodies is:

64 for ID bodies is:

The detailed sequence of this inversion process is illustrated in figure 3. The details of
this theoretical investigation for the specific ZD case, including linearization, matrix
formulation and resolution of model parameters etc, are described by Narayan
(1990).

6. Linearization and maeix formulation

Non-linear resistivity problems can be linearized using a Taylor series expansion and
written in a matrix form as AG = AAP. Equations (IO), (ll), and (12) can be written
in this form, where AG is the difference between the field and calculated responses (i.e.
64), A is the derivative of data with respect to the model parameters (known as the
system or coefficient matrix, equal to the sum of the current densities in each block),
and AP is the difference between the initial guess and the calculated model parameters
(i.e., 6p). It is better to deal with logarithmic variations, as the data and model
parameters are weighted uniformly. In this case, the damped least-squares solution is
given by
AIn(p) = (ATWA + kl)-'ATWAln(G), (13)
where W is a squared diagonal matrix containing the reciprocal of the data variances.
The choice of k is critical because it controls both the speed of convergence and the final
solution. A damping which is the product of the misfit error and an estimate of a cut-off
eigenvalue (A) is used. The trace of ATWA is the sum of eigenvalues, and from the
trace, the average eigenvalue is calculated as, A, = Trace/N. This gives a damping fac-
tor k =fAavee, where E is the RMS error,fis a fraction of the average eigenvalue, and
may range from 0 to 1. The resolution of model parameters can be studied by comput-
ing the resolution matrix (R) as follows:
+
R = (ATWA kl)-'ATWA. (14)
If the resolution matrix is an identity matrix (I) or close to an identity matrix, para-
meters are well resolved. Otherwise, the model parameters are poorly resolved.

7. Model result

On the basis of the inverse theory outlined above. a ZD inverse resistivity algorithm has
been developed and tested for several theoretical models as well as for real field data
(Narayan 1990). The objective in testing the ?D inverse resistivity algorithm was to
 Inversion techniques 679

study the rate of convergence from a poor initial guess, the resolution of model
parameters, and the simplicity and flexibility of the algorithm (i.e. delineation of
various boundaries and resolution of multiple conductors etc). One of the important
results describing delineation and resolution of multiple conductors embedded iu the
homogeneous halfspace is presented here.
The synthetic model used for this simulation was the response over three l o a m
bodies embedded in a homogeneous l00Qm halfspace. Figure 4 shows the true
model and its resistivity pseudosection, which has been used for testing of the ?D
version of the resistivity inverse algorithm. The data are computed by the forward
scheme of Madden (1967) and are noise free. The initial model guess for inverting
this set of data consisted of 30am blocks in a constrained 100C2m halfspace
(figure 5). The RMS error went down from about 160% to about 10% after 6 iterations
(figure 6). The rate of convergence is rather fast in the beginning. This algorithm takes
3 minutes of CPU time for 25 iterations on a VAX. The final inverted result is shown in
figure 5. The result indicates that the algorithm is capable of delineating the multiple
conductors separated by a unit dipole length. It is also clear that the algorithm is
capable of delineating various boundaries of the anomalous body. The resolution
matrix of the model was analysed, which indicates that the model parameters of the
top, left and right boundary blocks of the anomalous region are not well resolved.
This may be due to the lack of proper consideration for lumping the discretized
blocks.
The resistivity inverse algorithm is also applied to interpret field data (Hallof 1974)
containing geologic noise. The field data were measured over a partially delineated
disseminated sulphide mineralization near Safford, Arizona with dipole separation of
1000 feet for different N-spacings, as shown in Hallof (1974). It is evident from the
field data that there is an abrupt change in resistivity data from east to west. Hallof

11
2
I U "h"
/ \
IUuhmm
2
'I J

6J

S.

10 '

Figure 4. Cross section of 2 0 resistivity model (top) and synthetic resistivity pseudosection
over the model (bottom).
680 S Narayan et a1

I ' _ ' I ' ' ' ' ' ' ' ' '
01 ; L 11

-?
4 .

-6
30 ( I S M ) 30 (353)0hm-m 30 (185.1)
8. -8

10 . M (Mori ohm-m
-10

I2 - - 12
9'
' ' ' ' , ' ' ' ' ' ' ' ' ' I '
r
z-
9
5z :-
jo
-.
__9
D

"

Figure 5. Inversion of 2~ synthetic resistivity data (figure 4). Resistivity obtained after
inversion is indicated in the parenthesis. Resistivity outside the parenthesis is the starring
model parameters. Symbolf is used to fix the resistivity during the inversion.

10 t5 m 25
Numbxafirendcms

Figure 6. Per cent of RMS error versus number of iterations.

 68 1

Figure 7. Inversion of field resistivity data.

10
1 11 21

m r a f m n r

Figure 8. Per cent of RMS error versus number of iterations.

682 S Narayan et a1

(1974) postulated the geologic section using the available well data, gravity data and
resistivity data. The disseminated mineralization has also been encountered in two
drilled holes as shown in the geologic section (Hallof 1974). During the inversion of
field data, the anomalous region producing the resistivity response was discretized
into numerous blocks and an intrinsic resistivity of 10ohmm was given to them in
the starting model. The final resistivity parameters obtained after inversion are
shown in figure 7. Figure 8 illustrates the rate of convergence of this inverse algorithm
applied on the field data. The result obtained on the-interpretation of field data has RMS
error of 16% and is consistent with the model proposed by Hallof (1974). Correlating
the final model obtained by the inversion of field data with the geologic section
postulated by Hallof (1974), it is found that the zone of sulphide mineralization was
resolved reasonably well. The boundaries on the top and sides of the sulphide minera-
lization are particularly well delineated.

8. Conclusions

The 2~ inverse resistivity algorithm developed by Narayan (1990) on the basis of an

inverse formulation discussed herein is stable. and is capable of satisfactorily resolving
anomalous bodies embedded in a homogeneous host rock. The algorithm has also been
applied to field data and the inverted result is consistent with the available geologic and
well data. The experience derived from working with this inverse algorithm is that the
rate of convergence is a function of the amount of information obtained from observed
geologic data. To increase the rate of convergence of the model, an initial model guess
should be compatible with the gross aspects of the field data.
The multi-dimensional inverse formulation developed here using reciprocity and
perturbation analysis has the following advantages.
(i) This approach to inverse formulation is a linearized form of a class of non-linear
problems, and is an iterative process.
(ii) It is independent of any modelling algorithm, and requires accurate solutions to
a given 2~ or 3~ source.
(iii) Analysis shows that the sensitivity of measurements at the surface is directly
proportional to the amount of power dissipated across the heterogeneous region.
(iv) The derivatives of data with respect to the model parameters, which is one of
the needed inputs of a generalized inverse technique, can be computed easily. This is
because the inverse formulation done using perturbation analysis itself produces a
linear form of a non-linear problem.

Appendix. A review of typical inversion schemes discussed in the text

Consider a set of data containing N measurements over a model containing M para-

meters, which can be described in the form of the following equation:

The non-linear relationship between the data and model parameters can be linearized
using a Taylor series expansion which, after neglecting second and higher-order
 Inversion techniques 683

terms, yields

In the above relation, di is the ith measurement and d," is the ith calculated datum.
Rewriting equation (A?) in matrix form:
6d = A6p (A3)
where
6d=dj-d," 6p = p i - p i 0 and

aF.
A - .- 2= Jacobian or coefficient matrix. (A4)
'i - ap,
The solution of equation (A3) is given by various workers as follows:

Least-squares solution
The least-squares solution using linear inverse theory is (Menke 1984)
6p = (ATA)-'AT6d (A51
where the estimate of data misfit is
N
e = 6dT6d= c ( d i - d!)2.
i= I

Weighted least-squares solution

This approach is also known as Gauss-Newton's method of least-squares minimization
(Menke 1984).
6p = (ATWA)-IATW6d (A7)
where W is a squared diagonal matrix containing the reciprocal of the data variances.
The estimate of model misfit (i.e. reduced chi-squared error) is

Generalized damped least-squares solution

The generalized damped least-squares can be given by
6p = (ATWA + XCTC)-'ATW6d ('49)
where X is the damping factor. The generalized damped least-squares inversion
approach has different names depending on the nature of the C and the criteria for
selection of the damping coefficients. The important ones are
(i) Marquardt-Levenburg inversion. If the matrix C is considered as the identity
matrix I, then the Marquardt-Levenburg inverse solution (Marquardt 1963) is
61, = (ATWA + XITl)ATW6d. (A10)
684 S Navayan et a1

then the solution given by equation (A9) is called Occam’s inversion or smooth inver-
sion. Matrix C is called smoothing operator by Constable et ai (1987).

References

Assai H M and Mahmoud S F 1987 A new inversion technique for apparent resistivity measurements IEEE
Twm. Georei. Remore Sensing GE-25 7-10
Bdckus G E and Gilbert J F 1967 Numerical applications of a formalism for geophysical inverse problems
Geophys. J. Asrron. Soc. A 266 123-92
- 1968 The resolving power of gross earth data: Ceophys. J . R . Astron. Soc. 16 169-205
-1969 Uniqueness in the inversion of inaccurate gross earth data Phil. Trans. R. Soc. A 266 123-92
Bakhak M R 1977 Threedimensional numerical modelline- in resistivitv . - PhD thesis University of
..Drosvecting
California, Berkeley
Beasley C Wand Ward S H 1986 Three-dimensional mises-a-la-masse modelling applied to mapping.fracture
~ ~~ ~~

mnes Geophysics 51 98-1 13

Bibby J M and Risk G F 1973 Interpretation of dipole-dipole resistivity surveys using a hemispheroidal
model Geophysics 38 719-37
Coggon J H 1971 Electromagnetic and electrical modelling by the finite element method Geophysics 36 132
Constable S C, Parker R Land Constable C G 1987 Occam’s inversion: A practical algorithm for generating
smooth models from electromagnetic sounding data Geophysics 52 289-300
Daily W and Owen E 1991 Cross-borehole resistivity tomography Geophysics 56 1228-35
Daniels J I 1977 Three-dimensional resistivity and induced polarization modelling using buried electrodes
Geophysics 42 1006-19
-1978 Interpretation of buried electrode resistivity data using a layered earth model Geophysim 43 988-
1001
- 1981 Interpretation of hole to surface resistivity measurements at Yucca Mountain, Nevada Test Site
USGS openfile report no 81. p 1336
- 1983 Hole to surface resistivity measurements Geophysics 48 87-97
Dcy A and Morrison H F 1976 Resistivity modelling for arbitrarily shaped two-dimensional structures, Part I:
Theoretical formulation Lnweme Berkeley laboratory reporl no LBL-5223
-1979 Resistivity modelling for arbitrarily shdped three-dimensional structures Geophysics 44 753-80
Dieter K. Paterson N R and Grant F 5 1969 IP and resistivity Iype curves for three-dimensional bodies Geo-
physic3 34 615-32
EdtOn P A 1987 Three-dimensional electromagnetic inversion PhD thesis,University of Utah
Eloranta E H 1986 Potential field of a stationary electric current using Fredholm’s integral equations of the
second kind Geophys Prospecting 34 856-72
Evjen H M 1938 Depth factors and resoiving power of electrical measurements Ceophysim 3 78-95
Fox R C. Hohmann G W. Killpack T J and Rijo L 1980 Topographic effects in resistivity and induced-polar-
ization surveys Geophysics 45 75-93
Fry A W Band Neuman M R 1985 Impedance-computed tomography algorithm and system Appf. Op. 24
3985-92
Glenn W E and Ward S H 1976 Statistical evaluation of electrical sounding methods: Part 1, Experiment
design Geoplr.vsics 41 1207-2 .~
 Inversion techniques 685

Hallaf P 1974 The IP phase measurement and inductive coupling Geopkysim 39 650-65
Hoerl A E and Kenndrd R W 1970a Ridge regression: Biased estimation for non-orthogonal problems Teeh-
no,nr,iric.~12 55-67
- 1970b Ridge regression: Applications to nonorthogonal problems Terl?nomerries12 69-82
Hohmdnn G W 1975 Three-dimensional induced polarization and electromagnetic modelling Geophysics 40
309-15
Hohmann G W and Raiche A P 1988 Inversion of controlled-source electromagnetic data, Eleerromagnefic
,nerhodr in gmpkysics vol 1 ed M N Nabighien (Investigation in Geophysics 3) (Society of Exploration
Geophysicists) pp 469-503
Holcombe H T and Jiracek G R 1984 Three-dimensional terrain corrections in resistivity surveys Geophysics
49 439-52
Hoversten G M er a1 1982 Comparison of five least squares inversion techniques in resistivity sounding
Geopkys. Prospeeling 30 688-715
Inman J R 1975 Resistivity inversion with ridge regression Geophysics 40 788-817
Inman J R, Ryu J and Ward S H 1973 Resistivity inversion Geophysics 38 1088-108
Jackson D D 1972 Interpretation of inaccurate, insufficient, and inconsistent data Geophys. J. R. Astron. Soc.
28 97-109
Jepsen A F 1969 Numerical modelling in resistivity prospecting PhD rhesis University of California, Berkeley
Jupp D L B and Vozoff K 1975 Stable iterative methods for the inversion of geophysical data Geopkys. J . R.
Astron. Soc. 42 957-76
LaBrecque D J and Ward S H 1988 Two-dimensional inversion of cross-borehole resistivity data using
multiple boundaries Proc. 58rk Ann. lm. Meetrng Society of Explororion Geophysicists Expanded
Abstracts pp 194-7
Lanczos C 1961 Linear Differmiid Opeiafom (New York Van Nostrand)
Lee T 1975 An integrai equation and its solution for some two- and three-dimensional problems in resistivity
and induced polarization: Geophys. J . R. Asfron. Soc. 42 81
Lines L R and Treitel S 1984 A review of least squares inversion and iis application to geophysical problems
Geopkys. Prospecting 32 159-86
Madden T R 1967 Calculations of induced polarization anomalies for arbitrary two-dimensional resistivity
structure Proe. Symp. on IP, Universify of Califomm. Berkeley
Marcuello J P and Queralt P 1985 Inversion with an alternative error function in resistivity measuements
Geoexploration 23 527-36
Marquardt D W 1963 An algorithm for least squares estimation of non-linear parameters J . Soc. Industrial
Appl. M a d . 11 431-41
-1970 Generalized inverses. ridge regression, biased linear estimation, and non linear estimation Techno-
mvfrics 12 591-612
McPhar Geophysics 1966 Catalogue of resistivity and IP nrodel darn (Ontario: McPhar Geophysics, Ltd)
Menke W 1984 Geophysical data analysis Discrere Inverse nieory (New York Academic)
Mufti I R 1976 Finite-difference resistivity modelling for arbitrarily shaped two-dimensional structures
Geopkysies 41 62
Narayan S 1990 Two-dimensional resistivity inversion MSc Thesis. University of California, Riverside,
CA
Oristaglio M L a n d Worthington M H 1980 Inversion of surface and borehole electromagnetic data for two-
dimensional electrical conductivity models Geophys. Prospecring 28 633-57
Park S K and Van G P 1991 Inversion of pale-pole data for 3-D resistivity structure beneath arrays of
electrodes Geophysics 56 951-60
Parker R L 1984 The inverse problem of resistivity sounding Geophysics 49 '2143-58
Pekeris C L 1940 Direct method of interpretation in resistivity prospecting Gcophysim 5 31-42
Pelton W H, Rijo L a n d Swift C M Jr 1978 Inversion of two dimensional resistivity and induced polarization
data Geophysics 43 788-803
Petrick W R, Pelton W H a n d Ward S H 1977 Ridge regression inversion applied to crustal resistivity sound-
ing data from South Africa Geophysics 42 995-1006
Petrick W R Jr, Sill W R and Ward S H 1979 Three-dimensional resistivity inversion using alpha centers
University of Utah, Depr of Geology and Geophysics, Report no DE-AC07-79ET/27002
Pridmore D F, Hohmann G W, Ward S H and Sill W R 1981 An investigation of finite element modelling far
electrical and eleclromagnetic data in three dimensions Geophysics 46 1009-24
Ramirez A, Daily W. LaBrecque D.Owen E and Chesnut D 1992 Monitoring an underground steam injection
process using electrical resistance tomography Wafer Resource Res. 3 1608-14
686 S Narayan et a1

Rodi W L 1976 A technique for improving the accuracy of finite element solutions for MT data Geophys. J. R.
Astron. Soc. 44 483-506
_.
S c u m E F 1972 Comouter calculations of resistivitv nseudosections of a buried snherical conductor bodv
Gcoph.vs. Prosprerinz 20 605-25
Shen P Y, Wane K and Beck A E 1990 Two-dimensional inverse modeling of crustal thermal regime with
applicationto East European geotraverses J. Geophys. Res. 95 19 903125.
Shen P Y and Beck A E 1991 Least squares inversion of borehole temperature measurements in functional
space J . Geophys Res. 96 19 965-79
Shima H 1990 Two-dimensional automatic resistivity inversion technique using alpha centers Geophysics 55
682-4
Slichter L B 1933, The interpretation of the resistivity prospecting method for horizontal structures Physics 4
307-22
Smith N C and Vozoff K 1984 Two dimensional DC resistivity inversion for dipole-dipole data IEEE Trans.
Gemeience Remote Sensing 22 21-8
Stevenson A F 1934 On the theoretical determination of earth resistance from surface potential measurements
Physics 5 114-24
Tarantola A 1984 Inversion of seismic rdection data in the acoustic approximation Geophysics 49 1259-66
- 1987 h w m e Problem Theory (New York Elsevier)
__ 1988 Theoretical background for the inversion of seismic waveforms including elasticity
~. and attenuation
Pure Appl. Geophys. 128 365-99
Tarantola A, Jobert G. Trezeyet D and Denelle E 1988 The non-linear inversion ofseismic waveforms can be
performed either by time extrapolation or by depth extrapolation Geophys. Prospecting 36 383-416
Telford W M. Geldart L P, Sheriff R E and Keys D A 1976 Applied Geophysics (New York Cambridge
University Press)
Tom - L T and Yam- C H 1990 Incornoration of tonography . - . .into two-dimensional resistivity inversion
Geophysics 55 354-61
Tripp A C. Hohmann G W and Swift C M 1984 Two dimensional resistivity inversion Geophysics 49 708-
1717
Van Nostrand R G and Cook K L 1966 Interpretation of resistivity data: USGS Prof. Poper 499.
Vozoff K 1958 Numerical resistivity analysis: horizontal layers Geophysics 23 536-56
Vozoff K and Jupp D L B 1975 Joint inversion of geophysical data Geophys. J . R. Astron. Soc. 42 977-91
Weidelt P 1975 Inversion of two-dimensional conductivity structures Phys. Earth Planet. Int. 10 282-91
Wiggins R A 1972, The general linear inverse problem: Implication of surface waves and free oscillations on
earth structure Rev. Geophys. Space Sci. 10 251-85
Yang F W and Ward S H 1985 Single and cross-borehole resistivity anomalies of thin ellipsoids and shperoids
Gcopltys. 50 637-55
Yorkey T J 1986 Comparing reconstruction methods for electrical resistivity tomography PhD rhesis
University of Wisconsin
Zohdy A A R 1972 Automatic interpretation of resistivity sounding curves using modified Dar Zarrouk
functions Proc. 42nd Inr. SEC Annunl MePring ( A n o i , & z ,C A )

View publication stats