Geophysical Prospecting, 1996, 44, \31-1 52

Rapidleast-squares inversionof apparent

resistivity pseudosectionsby a
quasi-Newtonmethodl
M . H . L o k e 2a n d R . D . B a r k e r 3

Abstract
A fast inversion technique for the interpretation of data from resistivity
tomography surveys has been developed for operation on a microcomputer. This
technique is based on the smoothness-constrainedleast-squaresmethod and it
produces a two-dimensional subsurface model from the apparent resistivity
pseudosection. In the first iteration, a homogeneous earth model is used as the
starting model for which the apparent resistivity partial derivative values can be
calculated analytically. For subsequent iterations, a quasi-Newton method is used
to estimate the partial derivatives which reduces the computer time and memory
space required by about eight and twelve times, respectively, compared to the
conventional least-squaresmethod. Tests with a variety of computer models and
data from field surveys show that this technique is insensitive to random noise and
converges rapidly. This technique takes about one minute to invert a single data set
on an 80486DX microcomDuter.

Introduction

Electrical tomography surveys can be used to map areas with complex subsurface
geology where conventional resistivity sounding or profiling surveys are inadequate
(Griffiths and Barker 1993). Such surveys normally employ a number (25 or more)
of electrodes laid out on a straight line with a constant spacing (Fig. 1). A
computer-controlled system is then used to select automatically the active
electrodes used for each measurement (Griffiths, Turnbull and Olayinka 1990).
The data from these surveys are commonly arranged and contoured in the form of
a pseudosection (Hallof 1957), which gives an approximate picture of the
subsurface resistivity. However, an inversion of the data is required to obtain a
more accurate model of the subsurface resistivity.

I
Paper presented at the 56th EAEG meeting, June 1994, Vienna, Austria. Received June 1994, revision
oencntprl TlhF IOA<

2
School of Physics, Universiti Sains Malaysia, 11800 Penang, Malaysia. Forrnerly ar 3.
3
School of Earth Sciences,The University of Birmingham, Edgbaston, Birmingham B15 2TT, United
Kinedom.

aO 1995 European Association of Geoscientists & Engineers 131

 132 M.H. Loke and R.D. Barker

Slation 3

lr :" Tr r*" lz r. fz ReslstlultU

leptop
computor
Station 2 motgr

P2

Station I

C1 P1 P2 C2 tlectrodes
llata
I€U8I
n = l
n = 2
n = 3
n = 4
n = 5 5'
n = 6
E.
Figure 1. Sequence
of measurements
to build up a pseudosection.

One approach used to interpret the data is trial-and-error forward modelling

with two-dimensional (2D) earth models using the finite-difference or finite-
element method. Flowever, this approach is time consuming and requires an
experienced geophysicist. Furthermore the model obtained could suffer interoreter
bias.
Thus much research has been carried out on the use of automatic inversion
techniques (Smith and Vozoff 1984; Tripp, Hohmann and Swift l9g4). These
commonly use a non-linear optimization technique to improve a simple starting
model. The Gauss-Newton least-squares optimization method has been widely
used and it has proved to be a robust technique which converges rapidly (deGroot-
Hedlin and Constable 1990; Sasaki 1992). However the main disadvantage of
this
technique is the large amount of computing time required. For this reason it
is
commonly carried out on mainframe or workstation computers. The computing
resources needed are often not conveniently available, particularly to
small
geophysical companies involved in mineral, hydrological and engineering
surveys,
and so it is not practical to carry out this type of 2D modelling in the field
during
the course of a survey. Thus, for 2D resistivity surveys to be more widely
used, an
inversion technique which is sufficiently fast when used on commonly
available
microcomputers is needed.
The least-squaresand other gradient-based optimization techniques require
the
calculation of the apparent resistivity partial derivarives (Lines and Treitel l9g4).
The calculation of the partial derivatives can be time consuming. The use of
a
quasi-Newton technique (Broyden 1972) which avoids the direct calculation
of the
partial derivatives is examined in this paper.
Brief descriptions of the Gauss-Newton least-squares and quasi-Newton
methods are given in the following section. This is followed by a discussion on
ways to reduce the computational time required by the finite-difference or finite-

e 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting,

44, l3I-152
 Rapid least-squares intersion 133

element method to calculate the apparent resistivity values. A simple inversion

scheme using the quasi-Newton method is then described. The proposed technique
is completely automatic and it does not even require the user to supply a starting
model. Finally, the results from several tests with this technique, including
comparisons with the conventional least-squaresmethod, are given.

Method
Smoothness-constrained
least-squaresmethod
The model used, which consists of a number of 2D rectangular blocks, is shown in
Fig. 2. We have adopted the model used by Barker (1992), where the blocks are
equal in number to the data points in the apparent resistivity pseudosection and are
arranged in a similar manner. The depths to the centres of the interior blocks are
placed at the median depth of investigation (Edwards 1977) for the different
electrode spacings used. The median depth of investigation is about 0.5 times the
electrode spacing for the \Tenner array. In some cases,slightly better results can be
obtained by using a model with thinner blocks near the surface and thicker blocks
near the bottom (Loke 1994).
The conventional Gauss-Newton least-squaresmethod (Lines and Treitel 1984)
can be used to determine the resistivity of the rectangular blocks (the model
parameters) that will minimize the differences between the calculated and measured
apparent resistivity vaiues. $7e start with an initial model for the subsurface and try
to improve this model in an iterative manner. The smoothness-constrained
modification to the Gauss-Newton method (deGroot-Hedlin and Constable 1990)
leads to the following system of normal equations:

: ITg;,
(JIJr+ .\oc'rc)po (1)

Electrode number
8 9 1 0 1 r 1 2 1 3 1 4 l s

Figure 2. Arrangement of rectangular blocks used in the 2D model.

O 1996EuropeanAssociationof Geoscientists& Engineers,GeophysicalProspecting,44, 131-152

134 M.H. Loke and R.D. Barker

where i is the iteration number, J; is the Jacobian matrix of partial derivatives, g;

is the discrepancy vector which contains the differences between logarithms of
rhe measured and calculated appafent resistivity values, .\; is the damping factor
and p; is the perturbation vecror to the model parameters for the ith iteration.
The 2D flatness filter C is used to constrain the smoothness of the perturbations
to the model parameters to some constant value (Sasaki 1992). The logarithms
of the apparent and model resistivity values are normally used in resistivity
problems.
The value of the damping factor ) depends on the level of random noise present
in the data (Sasaki 1992). A larger value of .\ is used for higher levels of noise. For
blocks of equal size, the partial derivatives associatedwith a block become smaller
with increasing depth of the block. The amplitude of the elements of the flatness
filter matrix C are increased by_about 10% for each deeper row to stabilize the
inversion process. The matrix C'C is symmetric and very sparse. The computing
time and memory space required can be significantly reduced by storing the
elements of this matrix in a sparse matrix format (George and Liu 1981).
The conventional least-squaresinversion method may be divided into three main
steps. The first step is to calculate the apparent resistivity values for the present
model. The second step is to calculate the Jacobian matrix J of partial derivatives.
The apparent resistivity and partial derivative values can be calculated using the
finite-difference or finite-element methods (Smith and Vozoff 1984; Tripp et al.
1984). The third step is to solve the above system of linear equations. A number of
numerical techniques, such as the modified Gram-Schmidt, Cholesky decomposi-
tion and singular value decomposition methods (Golub and van Loan 1989) can be
used. The time taken to calculate the partial derivatives can be much longer than
that taken by the other steps when there is a moderate (less than 500) number of
data points.

Quasi-Newton methods
The Jacobian matrix J must be recalculated in each iteration in the conventional
Gauss-Newton method. Quasi-Newton methods avoid the recalculation of the
Jacobian matrix by using an updating method. The Jacobian matrix for subsequent
iterations can be estimated by using an updating formula if the Jacobian matrix Je
for the starting model in the first iteration is available. The partial derivatives can
be calculated analytically by using a homogeneous earth model as the starting
model (McGillivray and Oldenburg 1990; Park and Van 1991). A very fast
technique to obtain an approximate initial model for the subsurface with one
iteration of the least-squaresmethod by using a homogeneous earth starting model
is described by Loke and Barker (1995). A detailed description of the calculation of
the partial derivatives for the homogeneous earth model is also given in the same
paper.
The Jacobian matrix J; for the ith iteration is replaced by an approximation B;.

O 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44, l3l-152
 Rapid least-squaresinaersion 135

The quasi-Newtonmethod (Broyden 1965)usesthe following updating equation:

B,+t : B; + u;pl, Q)
where

u;: (Ay; -Bip)lpTpt,

AY;:Yt+r-Yi,

and y; is the model response for the ith iteration. The approximation of the
Jacobian matrix B;*1 for the (i * l)th iteration is calculated using the Jacobian
matrix approximation B;, the parameter perturbation vector p; and the change in
the model response Ay; for the ith iteration. The updating matrix .t;pf, i. a rank-
one matrix and Broyden's method is a rank-one quasi-Newton method (Scales
1985). The flop count for this updating method is proportion to ,2 (we use the
definition by Sheriff (1991) where one flop is equal to one floating point
multiplication and addition), where z is the number of data points. Thus it is
much faster than the direct recalculation of the Jacobian matrix where the flop
count is proportional to n3. However, the convergence rate of Broyden's method is
slower than the Gauss-Newton method. It has a super-linear convergence rate
compared to the quadratic convergence of the Gauss-Newton method (Burden,
Faires and Reynolds 1981). In many applications the decreasein the convergence
rate is more than compensated for by the reduction in the computational time for
each iteration.
In order to calculate the parameter perturbation vector for the ith iteration using
the quasi-Newton method, the following system of normal equations is used:
(BIB'+)crc)p,:Bfgr. (3)
It should be noted that several other quasi-Newton updating techniques, some of
which might be more efficient than Broyden's method for the 2D inversion of
resistivity data, have been proposed (Dennis and Schnabel 1983). Broyden's
method was chosen because it has been widely used, it is stable and has a
reasonably fast convergence rate for many non-linear problems (Broyden 1972;
More and Trangenstein 1976).
The time taken to solve the system of normal equations can also be reduced by
using matrix updating techniques (Golub and van Loan 1989). The solution of the
system of equations must be recalculated after each iteration in the conventional
Gauss-Newton method. As an example, the Cholesky method (Schwarz et al.
1973) uses the decomposition

JtJ: RtR, (4)

where R is an upper triangular matrix. The estimated Jacobian matrix B;*1 differs
from the matrix B,. for the previous iteration by a rank-one matrix. In this case,it is
possible to calculate the matrix R;11 for the (i * l)th iteration from the R; matrix in

O 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44, l3l-152
136 M.H. Loke and R.D. Barker

the ith iteration (Golub and van Loan 1989). The flop count of the matrix updating
technique is proportionalto n2, whereasthe Cholesky method requires about2nsi3
flops.
However, in most non-linear optimization schemes, the damping factor ,\ is
normally changed after each iteration (Scales 1985), especially during the first few
iterations. Standard matrix updating techniques cannot be used if the damping
factor is changed since the flatness filter matrix C in the smoothness-constrained
least-squares equation (3) has off-diagonal elements. The following updating
method was used when the damping factor changes by making use of the fact that
the solution of the least-squaresequation by the Cholesky method consists of two
main parts. In the first part, the matrix product

H: BTB (s)
is calculated. This operation requires about nt 12 flops when the number of data
points and model parameters are the same. We can make use of Broyden's updating
formula to construct an updating formula for H as follows:

Bl,B,*, : (Bi + uipf )r(Br + tr,pf)

(6)
: BfB' + (Bful)pl + p1(Bfq)r + (ulu1)p1pl

which can be rewritten as

Hi+r : HI + (BIur)pI + p'(elui)r t (of.tt)p'pf. (7)

Thus the H111 matrix can be calculated from the fI1 matrix in the previous
iteration. The flop count for this updating method is proportion to n2. This merhod
is much faster than the direct calculation of Bf*rB,*r when n is large. The second
part of the Cholesky method, i.e. the factorization of B1f1B1*1 to determin€ Ri*r,
involves n3f 6 flops. This updating method reduces the time taken by the Cholesky
method by nearly three times. A similar partial updating scheme can also be
devised for the modified Gram-Schmidt method.
Some care must be taken in using the Cholesky method when the number of data
points and model parameters is large. The round-off errors for the Cholesky
method are much larger than those for the modified Gram-Schmidt or singular
value decomposition methods if the variables are stored in arrays with the same
precision (such as single precision arrays). The elements of the BTB matrix are
stored in an 8-byte double precision array (Lawson and Hanson 1974) to ensure
that the Cholesky method has the same order of accuracy as the other methods.
O$y slightly more than half the elements are stored in the array since the matrix
B'B is symmetric. The Cholesky method is used for most data sets where the
sparsity of the flatness filter matrix product CTC can be more easily exploited. The
calculations are repeated for some data sets using the modified Gram-Schmidt
method to ensure that round-off errors are not a serious oroblem. The differences

(O 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44,

 Rapid least-squaresinaersion 137

Elsctrodes
El E" E3 E4 E5 E6 E7 Es Es
Node
numbgrS

Figure 3. Part of finite-differencemesh with 139 bv 18 nodes together with the node
numbersfor the electrodes.

in the results obtained bv the two methods are less than 0.106, which is not
sisnificant.

Optimizing the finite-difference method

when the partial derivatives are nor required, the number of floating-point
operations required by the finite-difference or finite-element technique can be
reduced to a bare minimum so that the results are just sufficient to calculate the
apparent resistivity values. The potential distribution $(x,y,e) due to a current
source l(x,y,z) in a2D earth model is given by Poisson'sequation
-V.fo(r, z)Y$(x,y,z)]: I(x,y,z). (S)
\Yhile the conductivity distribution o(r, z) is2D, the current source is 3D (point).
The 3D potential distribution d(*,y,e) is reduced to a 2D rransformed potential
Q(x, ke, z) by taking the Fourier transform of the above equation in the y-direction.
This gives the equation
- V . f o ( x , z ) Y $ ( x , k o , z \ l J -k o o ( x , r ) d \ * , k o , z ) : I ( x , k o , z ) ,
(9)
where A, is the Fourier transform variable. The finite-difference discretization
(Dey and Morrison 1979) of (9) for each node in the 2D mesh (Fig. 3) gives rise to
the matrix equation

46: ", (10)

where the capacitancematrix A is an N x N symmetric sparse banded matrix (,A1is
the number of nodes), @ is a coiumn vector with the transformed potentials and v is
the current vector. Equation (10) can be solved using the band-cholesky method
(Martin and lTilkinson 1965). This method first carries out the decomposition.

A : R'r'R (11)

(() 1996 European Association of Geoscientists & Bngineers, ()eophysical Prospecting, 44, I3l-152
138 M.H. Loke and R.D. Barker

In order to determinethe potentialvector d, t10) is solvedby using the forward

substitutionstep
(1)\
Rtq: to
followed by the backward substitution step

Rd: q. (1 3 )
The transformed potential $(x,ko,z) is obtained for a number (usually five to ten)
of An values and an inverse Fourier transform is used to obtain the potential
dG,y,z). When the electrodesare placed along a single profile, the inverse Fourier
transform reduces to the intesral

d(,,o,z): (rI r) ff ,O,ko,z) dk,, (14)

which is evaluated by numerical quadrature. The partial derivative of the potential

f is obtained by differentiating (10) with respect to the conductivity of the block,
say oj, which gives

AQ, : *, (1s)
where
w: -A'Q.

The vector w may be viewed as a fictitious current vector given by the product of
the capacitancematrix differentjal A'and the potential vector @ iSasaki 19891.
Figure 3 shows part of a mesh used in this research to discretize the subsurface.
The electrodes are placed 4 horizontal nodes apart. The nodes are numbered along
columns (Dey and Morrison 1979) starting from the top left node. To the left of
the first electrode there are 1 t horizontal nodes where the spacing between the
nodes is progressively increased so that the left edge of the mesh is sufficiently far
from the first electrode. A similar arrangement is used to the right of the last
electrode. The mesh used has 18 nodes in the vertical direction. Thus for a survey
with 30 electrodes,the mesh has 139 by 18 nodes giving a total of 2502 nodes. To
determine the apparent resistivity values measured with a multielectrode survey, it
is most convenient to calculate first the potentials due to a single current electrode
at the first electrode E,1.After that, the potentials due to a current electrode at E2
are determined. This process is repeated for the other electrodes. The Cholesky
decomposition in (11) is only carried out once since the capacitancematrix A does
not depend on the location of the current electrode. The forward and backward
substitution steps in (12) and (13) are repeated to calculate the potentials due to
each electrode.
Only the potentials at the nodes where the electrodes are located are needed to
calculate the apparent resistivity values. Thus, for a single current electrode in a
3Q-electrodearray, it is only necessaryto calculate the potentials at 29 nodes (out of

O 1996 European Association of Geoscientists & Engineers, GeophysicalProspecting, 44, l3I-152

 Rapid least-squares
inztersion 739

a total of 2502 nodes). As an example, consider the casewhen the current electrode
is at E1 in Fig. 3 and the data level is one (Fig. l). The current vector v then only
has a single non-zero element, u1ee.Since the first 198 elements of the auxiliary
vector q in (12) are zero, only the 199th to tlre 25o2nd elements of the q vector
have to be directly calculated. For the backward substitution step in (13), it is only
necessary to carry out the calculations for the 2l7th (where the electrode E2 is
located) to the 2502nd elements of the f vector. When the current electrode is at
82, the backward substitution step is carried out for the 235th node (with the E3
electrode) onwards. Only the potentials at the nodes with numbers higher than or
equal to the first electrode to the right ofthe current electrode are needed. By using
the principle of reciprocity, the calculation of the potentials at the electrodes to the
left of the current electrode can be avoided. Thus the times taken for the forward
and backward substitution steps are progressively reduced as the current electrode
is further away from the first electrode. In comparison, to determine the partial
derivatives, the time taken by the backward substitution step cannot be reduced
since the potentials at all the nodes must be calculated. They are needed in the
calculation of the elements of the fictitious current vector w in (15) which further
increasesthe computing time.
The times taken for the forward and backward substitution steps become shorter
as the node number of the current electrode becomes larger. This property can be
more fully exploited by a judicious use of sparse marrix techniques (George and
Liu 1981) to renumber the nodes so that the nodes with the electrodes end up with
even higher numbers. For example, by using the alternate-diagonal method (woo
et al. 1976) and placing the electrodes on the even columns of nodes, all the
electrodes will have node numbers that are higher than l25l (compared to 199 for
E1 in Fig. 3). Even larger reductions in the computing time can be achieved by
using more sophisticated sparsematrix techniques such the one-way dissection and
nested dissectionmethods (George and Liu 1981).
The number of inverse Fourier transforms needed to calculate all the partial
derivatives is much larger than that required to calculate the apparent resistivity
values only. For a survey with 30 electrodes using a \Tenner array with 8 levels of
data in the pseudosection) 435 inverse Fourier transforms are needed to calculate
the apparent resistivity values. In comparison, 45408 inverse Fourier transforms
are required to calculate the partial derivatives. An optimized finite-difference
program takes about 14 seconds to calculate the apparent resistivity values on a
33 MHz 80486DX microcomputer. In comparison, if the partial derivatives are
calculatedas well, it takes about 130 seconds.

Practical procedure

The main steps in a simple but effective algorithm using the quasi-Newton
inversion method are described below. It is assumed that the only data available are
the measured apparent resistivity values.

A 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44, 131-152
l4O M.H. Loke and R.D. Barker

Step (a). Firstly some parameters that will guide the inversion process have to be
chosen. An initial damping factor ,\e and a minimum damping factor )- for the
least-squaresmethod have to be selected. \We have found that values of 0.20 and
0.04 for .\e and ,\- give satisfactory results for most field data sets. The maximum
number of iterations for the quasi-Newton inversion method should also be
selected.
Some type of convergence criterion, such as a minimum rms error value, should
be set. For field data sets where the exact noise level is usually not known, it is
more convenient to use the change in the rms Pttot €i. This is given by
ei : ( e t - e ; * r )l e i , (16)

where €; and €;11 or€ the rms errors for the ith and (i + 1)th iterations. Normally
the inversion processis stopped when e; is less than 57o.
Step (b). A homogeneous earth model is used as the starting model for the first
iteration. The logarithm of the resistivity ro of this model is calculated by taking
the average of the logarithms of the measured apparent resistivity values J using
the equation
1 n
r":ltt, (r7)
n-

where n is the number of data points.

The Jacobian matrix 86 is then calculated for the electrode array used from the
partial derivatives which have been precomputed and stored in a disc file (Loke and
Barker 1995). The least-squaresequation (3) is then solved to determine the model
parameter perturbation vector po. An estimate 11 of the resistivity of the blocks is
given by
11:rs*p6. (18)
The apparent resistivity values for the new model are then calculated using the
finite-difference method.
Step (c). Broyden's method is used to update the approximation of the Jacobian
matrix B1*1 for the next iteration. Normally if the rms error has decreasedin the
previous iteration, the damping factor is reduced by half if it is still larger than the
minimum value )- set in step (a). If the rms error is larger than that for the
previous iteration, a line search using quadratic interpolation is used in an attempt
to find the optimum step size for the model parameter perturbation vector p1. This
will usually result in a model with a lower rms error. If this still does not work, the
damping factor is increased and the vector p; is recalculated.
Step (d). The parameter perturbation vector pt+r is calculated by using the
estimated Jacobian matrix B;*1 in (3), after which a new model is obtained.
Steps (c) and (d) are repeated until the program converges or the maximum
number of iterations is reached. A fairly simple strategy is used to change the
damping factor after each iteration in this algorithm. This algorithm has worked

O 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44' 731-152
 Rapid least-squaresinxersion 141

B s c t a n g u l abr l o c k m o d o l
a) Measured Epparenl resistiuitU
0 80 160 24n 520 400 480 560 m
N
I
2
5
4
5
6
-t

I
f---l f-: l-l fi:::l Flljll F:-i-l ffi ffi ffi f U'nil I l e c t r o d e Spacing = 2B
100 110 120 150
(Clmf
ResistluitU
l n u e r s om o d E l s
h) 6 a u s s - N e u t o nm e t h o d
Iteratlon4 - rms error l.0i 1. s?. IlmE taksn - 691 s
0 80 160 24D 520 400 480 560
7.5
20
5 0
40
50
60
?0
80
f__-l t:l-l f:.rl f=.l:l E-!:-'::l
ffi ffi ' t E
r00 150 160 90 I
BeslstiultU (clml

c) 0 u a s i - N e t u t o nm e t h o d
Iteratlon 5 - rms error 1.01 T l m et a k e n - 9 t s
0 80 1 6 0 240 520 400 480 560 m
?.5
2l
50
{0
50
iiiii
ijii:
!
60
?t
80
f---l f:l f::a1 EilEl F:-l E€ ffi E I
1 0 0 1 5 0 1 6 0 1 9 0
Reslstlultg (o mI

Figure 4. (a) $Tenner array pseudosection due to a wide rectangular block. Inverse models
obtained with (b) the Gauss-Newton method and (c) the quasi-Newton method. Note that a
different contour interval is used for the apparent resistivity pseudosection. The depths to
the centre of each row of the model blocks are shown in (b) and (c). The outline of the block
is also shown for comparison.

lal 1996 European Association of Geoscientists & Engineers, GeophysicalProspecting,44, l3l-l12

142 M.H. Loke and R.D. Barker

fairly well in all the tests that we have carried out with computer-generated and
field data sers. More sophisticated algorithms (Scales 1985) could be used for more
problematical data sets.

Applications

Some of the results from the tests that we have carried out are given in this section.
A comparison between the models obtained (as well as the computer time and
memory required) by the conventional Gauss-Newton method and the quasi-
Newton method is made in the first two examples. The tests were carried out on a
33 MHz 80486DX IBM PC compatible microcomputer with 4 megabytes of
memorY.

Example 1 : Rectangular bloch model

Figure 4a shows the apparent resistivity pseudosection due to a wide rectangular

block model with a resistivity of 500Om embedded in a medium of 100Om. The
initial and minimum damping factors were set at 0.05 and 0.01, respectively, in the

ffectangular block model

.o. m ethod ( 5%noi se)

0uasi-NetUlon
m ethod ( 5%noi se)
- - . s . - - .G a u s s - N e u t o n
--+-. QuEsi-NeuJton m ethod ( 0%noi se)
--+{- Gauss-NeuJton m ethod ( 0%noi se)

1 2 3 4 5 6 ? A 9 1 0 1 1 1 2
Itsratlon numbcr

Figure 5. Error curves for the inversion of the rectangular block pseudosection data using
the Gauss-Newton and quasi-Newton methods.

O 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44' 731-152
 Rapid least-squaresint-tersion 743

Table 1. A comparison of the number of iterations and the rime taken by the Gauss-
Newton (GN) and quasi-Newton (QN) methods to converge for different data sets. The
exact rms error of the final model and the memory required by the finite-difference
subroutine are also shown.

rms Time Memory

Inversion No. of errol taken required
Data set method iterations (%) (s) (Kilobytes)

Rectangular block GN 4 0.96 691 2370

model (07o noise) QN 5 0.97 91 210
Rectangular block GN 3 4.80 553 2370
model (5% noise) QN 3 4.85 60 2t0
Stud Farm GN 5 5.07 i008 2700
survey QN 5 5.00 108 240

inversion of this data set. Rather small damping factors can be used since this data
set is noise free. The rms error convergence limit was set at 1o/r. The models
obtained by the Gauss-Newton and quasi-Newton methods are shown in Figs 4b
and c, respectively. There are no significant differences between the two models.
The shapesof both models agree reasonably well with the actual shape of the block.
The highest model resistivity value near the centre of the rectangular block is about
250 om, which is less than the true value of 500 flm. This is partly a result of
equivalence (Keller and Frischknecht 1966), where a thicker body with a lower
resistivity contrast can give rise to the same anomaly as a thinner block with a
higher resistivity contrast.
The quasi-Newton method takes 5 iterations to converge compared to 4
iterations for the Gauss-Newton method. Figure 5 shows the error curves for both
methods in the inversion of this data set. In the first 3 iterations the Gauss-
Newton converges at a slightly faster rate than the quasi-Newton method.
However, after the 4th iteration the difference in the rms errors of the two methods
is less than 0.2o/o. The slightly slower convergence rate of the quasi-Newton
method is more than offset by the shorter time per iteration that it requires. The
computer time and memory required by the two methods are shown in Table 1.
The quasi-Newton method takes about 91 seconds to reduce the rms error to less
than 17o compared to 691 seconds for the Gauss-Newton method. The finite-
difference subroutine for the quasi-Newton program requires about 210 kilobytes
to store all the arrays involved. In comparison, a similar subroutine for the Gauss-
Newton program (which also calculatesthe partial derivatives) requires about 2370
kilobytes.
Gaussian noise (Press et al. 1988) with an amplitude of 5oh was added to the
apparent resistivity data to study the effect of noise on the inversion results. The
resulting pseudosection (Fig. 6a) shows severe distortions. Since this data set has
some noise, larger values of 0.20 and 0.04 were used for the initial and minimum

O 1996 European Association of Geoscientists & Engineers, GeophysicalProspecting, 44, l3l-152

144 M.H. Loke and R.D. Barker

blockmodsl
Bectangular
a) Measured aPParent resisliuitU
80 1 6 0 240 520 400 480 560
N

1
2
5
4
5
6
'l

s
- T:-]-l [:.:::l filtil EnE:| Effi ffi E| Il U nn i il t tI l e c t r o d e sPacing = 20
1 0 0 1 5 0 1 6 0 1 9 0
[o m)
Reclsllultg
lnuercomodsls
b) 6suss-Neuton method
Iteratlon 5 - rms error 4.8% TImetaken - 555 s
240 520 400 {80 560 m

7.5

tr 2r
50
40
50
60
?0
80
't00 150 160 190
B€slstlultg (o ml

C) 0uasi-Neuton method
I t o r a t l o n5 - r m s e r r o r 4 . 9 * Tlme takon - 60 s
n Rn t6ft 240 520 400 480 560 m

7.5

E z t
30
- 40
i so
E 60
?0
80
T---l f : :l [i:X flj]l F!j-l ffi ffi ffi I
100 150 160 190
R e s l s t l u l t g( o m l
Figure 6. (a) Wennerarray pseudosection due to a wide rectangularblock wlth 5o/orandom
noise. Models obtained with (b) the Gauss-Newton method and (c) the quasi-Newton
method. The outline of the block is also shown for comparison.

damping factors and the convergence limit was set at 570. Both the quasi-Newton
and Gauss-Newton methods took 3 iterations to converge. The models obtained
by both methods show relatively minor distortions (Figs 6b and c).
The error curves (Fig. 5) show that the convergence rates for both methods are

Prospecting, 44' 131-152

O 1996 European Association of Geoscientists & Engineers, Geophysical
 Rapid least-squaresinxersion 145

StudFarmsuruog
a) Measured apparent resistiuitU
0 100 200 300 400 500 600 ?00 800 m
H

''
2

6
Unit tlectrode Spacing = 25m
?5 50 100 200
Rarlstlultg(clm)
lnu8rsemodsls
b) 6 a u s s - N e u t o nm e t h o r l
I t e r a t l o n5 - r m s e r r o r 5 . 1 % T l m et a k e n - l 0 l 0 s
0 100 z t l0 500 400 500 60U ?00 8u0

9.4

25.0
57.5
50.0
62.5
75.0

?s 50 100 200
B e c l s t l u l t g[ o m l
C) Quasi-Neuton method
Iteratlon 5 - rms error 5.0% TIme taken - | 09 s
0 100 200 500 400 500 600 700 800
9.4
F
25.0
57.5
50.0
62,5
75.0

25 50 100 200
Reslstlultg (cl m!

Figure 7. (a) Stud Farm survey apparent resistivity pseudosection. Models obtained with
(b) the Gauss-Newton and (c) the quasi-Newton methods. Locations of boreholes and
observed depths to weathered microdiorite are indicated.

almost the same for the noisy data. There are no significant differences between the
rms errors for the two methods from the 3rd iteration onwards. However, the
quasi-Newton method requires only about one-ninth of the time taken by the
Gauss-Newton method (Table 1).

A 1996EuropeanAssociationof Geoscientists
& Engineers,Geophysical
Prospecting,
44, 131-152
146 M.H. Loke and R.D. Barker

$tud Farmsurueu

Q u a $ i - N e u r t o nm e t h o d
G a u s s - N e u r t o nm e t h o d

6 ? 8 9
Iteratlonnumbsr
Figure 8. Error curves for the inversion of the Stud Farm survey data using the Gauss-
Newton and quasi-Newton methods.

Example 2: Stud Farm survey

This survey was carried out to map the thickness of the clay overburden along a
proposed tunnel route to join a new microdiorite quarry at Stud Farm to another
quarry at Cliffe Hill (Leicester) which is about 2km away. The microdiorite is
expected to have a resistivity of the order of several hundred ohm-metres, whereas
the overburden has a resistivity of 25 to 80Om (Turnbull 1986).
Figure 7a shows the apparent resistivity pseudosection from this survey where an
electrode spacing of 25 m was used. The initial and minimum damping factors were
set at 0.20 and 0.04 for this data set (as well as for the following examples). The
error curves of the Gauss-Newton and quasi-Newton methods in the inversion of
this pseudosection are shown in Fig. 8. The rms error for the Gauss-Newton
method at the lst iteration is higher than that for the quasi-Newton method. This
is probably caused by small differences between the homogeneous earth model
partial derivatives calculated numerically in the Gauss-Newton method and those
calculated analytically in the quasi-Newton method. The rms error decreases

e 1996 European Association of Geoscientists & Engineers, GeophysicalProspecting, 44, l3l-152

 inaersion 147
Rapid least-squares

rapidly in the first 4 iterations for both methods. After the 5th iteration, the relative
change in the rms error is less than 5o/o.The models obtained at the 5th iteration
by the two methods (Fig. 8) do not show any significant differences.
Both models show an undulating bedrock with a resistivity of over 200 f)m
covered by overburden material with a resistivity of less than 80 Om (Figs 7b and

BlueFarmsurusu
a) Measured appsrent resistiuitU
0 I t6 24 52 4 8 m
N
1
2
5
4
5
6
?
f---l F-.:1 [:].T *i-I'l F:;-=]@ ffi E I Ileclrode SPacing
90 100 110 120
Reslstlultg (c) m)

lnuersomodel
b) Q u s s i - N e r u t o nm e t h o d
f taratlon 5 - rms eJroJ 2.2* Tlm€ taken - 66 s
0 I 16 24 4 0 48 m

0 . 75
2.O
5 5.0
! +.u
i 5.0
E 6.0
7.0
f----'l F-::] Fr:.] Ir:ifl E:-::=l
ffi ffi E I
90 100 110 12tJ
Beslstlultg (() ml

c) Geologlcam
l odel
24 sand 46

0.?5 soil

2.O
5.0
4.0
Ihfllk
5.0 'rock
ridges'
6.0
?.0

Figure 9. (a) Blue Farm survey apparent resistivity pseudosection. (b) Model section
obtained at the 5th iteration with the quasi-Newton method. (c) Schematic diagram of the
subsurface geology from surface and borehole surveys.

O 1996 EuropeanAssociationof Geoscientists& Engineers,GeophysicalProslecting' 44' 131-152

148 M.H. Lokeand R.D. Barker

c). The microdiorite is known to be deeply weathered in places, so a gradational

increase in the resistivity from the overburden to the fresh bedrock is expected.
The two regions with higher resistivity material near the surface at the 450 and
650m marks are probably due to boulders and partially weathered material. The
depths to the weathered bedrock at four boreholes near the survey line are also
shown. There is good agreement between the microdiorite topography in the
models and the borehole depths at locations BHl and BH2. The model bedrock
interface is significantly deeper than the borehole depths at BH3 and BH4. One
possible reason is that the boreholes record the top of the very weathered
microdiorite and not the higher resistivity bedrock shown in the sections. Another
possible reason is that the boreholes were sunk into core boulders (Turnbull 1986).
Furthermore, it is known from other boreholes in this area that the bedrock
topography is not 2D which could also account for differences between the models
and the boreholes.
Blue Farm surueU

Reslstlultg (o m)
0 100 120

Boreholelog
Q u a s i - N e u J l o nm o d e l

:
i

l o

5.0
0epth io
(ml
o

4.0

Figure 10. Borehole resistivity log from a survey at Blue Farm. The model resistivity
obtained by the quasi-Newton method along the borehole is also shown for comparison.

e 1996EuropeanAssociationof Geoscientists& Engineers,GeophysicalProspecting,44, l3l-152

 Rapid least-squares
inaersion 149

The quasi-Newton method requires only about one-ninth of the computing time
needed by the Gauss-Newton method to converge (Table l). perhaps more
importantly, the finite-difference subroutine in the quasi-Newton progr€rm requires
only about 240 kilobytes of memory compared to 2700 kilobytes for the Gauss-
Newton program. It is possible to reduce the memory space required by saving
some of the arrays temporarily in a disc file but this would substantially increase
the computing time. This pseudosection (with 35 electrodes) is about the largest
that can be efficiently processedby the Gauss-Newton program in a computer with
only 4 megabytes of memory. In comparison, we have successfully processed
pseudosectionswith 64 electrodes using the quasi-Newton program.

Example 3: Blue Farm surzsey

Figure 9a shows the apparent resistivity pseudosection from one of the surveys by
Andrews (1993) at Blue Farm, cambridgeshire. This was part of a study of the

Mflgusi Biuer ore bodU

a) Measured apparent resistiuitU

r.r I 122 244 566 m

f---l F-l r::::l tT:-rl F:r-':l Effi ffi E I Unit Ilectrode Spocing = 50m
10 20 40 s0
Bestsiluttg (f) m)
b) 0uasi-Neuton model
Iteratlon 6 - rms Error 8.0% llme taken 52 s

0uerburden
4.6

: 14

i zs tountrg
rocks

34
0rebodU
r---l F-:r r:.:tl ai-ri:Tl
E=:I+.:l
ffi ffi ffi r
10 40 160 640
Beststlultg(o m)

Figure 11. (a) Dipole-dipole apparent resistivity pseudosecrion for the Magusi River
orebody survey. (b) Model section obtained at the 6th iteration by the quasi-Newton
method. Note that the apparent resistivity pseudosection uses a diflerent contour interval.
The boundaries of the orebody and overburden (Edwards 1977) are also shown.

aA 1996EuropeanAssociationof Geoscientists& Engineers,GeophysicalProspecting,44, 131-152

150 M.H. Loheand R.D. Barker

movement of nitrates through chalk covered with thin soil. The apparent resistivity
pseudosection shows several prominent areas with high resistivity values of up to
120 f,)m which extend to the 5th data level. Between the high resistivity zones there
are areas with lower apparent resistivity values.
The model produced by the quasi-Newton method at the 5th iteration (Fig. 9b)
shows that the source of the zones of high and low apparent resistivity values lies
near the surface. This agrees with the known surface geology and borehole
measurements. The geological and borehole data indicate that the resistivity
anomalies are caused by variations in the thickness of a sand layer (with a
maximum thickness of 2 m) overlying an undulating chalk bedrock (Andrews
1993). A schematic diagram of the subsurface geology is shown in Fig. 9c. Figure
10 shows the resistivity log obtained in a borehole at the 18m mark. The borehole
measurements were made using a Wenner array with a spacing of 0.3 m. The
subsurface resistivity reaches a maximum value of about 120f)m at a depth of
about 3 m after which it slowly decreaseswith depth. The model resistivity values
at the borehole location show the same general trend but with a smaller resistivity
range. This is probably becausethe borehole measurements are mainly affected by
the subsurface material within a distance of 0.5m from the borehole. llowever, the
model gives the average resistivity for a much larger volume of the subsurface.

Example 4: Magusi Rizter orebodY

The apparent resistivity data for the Magusi River orebody survey (Edwards 1977)
is used as an example of the inversion of a dipole-dipole array field data set. A
resistivity and IP survey was carried out over a massive sulphide orebody using
dipole lengths of 100,200 and 300 feet (30,60 and 90m). Figure 11 shows the
apparent resistivity pseudosection for the data collected with a dipole length of
30 m and the model obtained by the quasi-Newton method at the 6th iteration
(after which there is only a small change in the rms error). Although the minimum
electrode spacing of 30 m (and thus the expected resolution of the model) is about
the same as the width of the ore body, there is good agreement between the inverse
model and the true shape of the ore body (Edwatds 1977). The orebody has a much
lower resistivity (about 10ftm) compared to the country rocks (over 100Qm).

Conclusions

A fast and powerful technique using a quasi-Newton optimization method is

presented for the automatic inversion of resistivity pseudosections.The computer
time and memory spacerequired to produce a satisfactory 2D subsurface model are
reduced by about eight and twelve times respectively, by avoiding the direct
calculation of the partial derivatives. It has proved to be a robust method which
converges rapidly. This method is able to resolve complex geological structures
which cause overlapping anomalies in the pseudosection. The inversion of a single

Proslecting, 44, 137-152

O 1996 European Association of Geoscientists & Engineers, Geophysical
 Rapid least-squares
inaersion l5l

data set takes about one minute on an 80486DX microcomputer. This makes it
possible to carry out the interpretation of the data during the course of a survev
using a field laptop microcomputer.
Research is being carried out into different ways to improve the inversion
method further, using more sophisticated quasi-Newton techniques (Dennis and
Schnabel 1983) and different schemes to determine the optimum damping factor
and step size for the model perturbation vector (Scales 1985; deGroot-Hedlin and
constable 1990). The reduction in the computer time and memory required
achieved by using the quasi-Newton method might ultimately prove to be an even
more important factor in the inversion of 3D resistivity data which involves
thousands of data points.

Acknowledgements

\7e thank Dr R. Andrews for providing the Blue Farm data and Dr J. Turnbull for
providing the Stud Farm data. One of us (MHL) thanks The Association of
Commonwealth Universities and Universiti Sains Malavsia for the scholarship
provided.

References
Andrews R. 1993. The impact of sewagesludgeapplicationof nitrate leachingfrom arableland
on the unconfinedchalk aquifer of East Anglia, England. PhD thesis, The University of
Birmingham.
Barker R.D. 1992.A simple algorithm for electricalimagingof the subsurface.First Break
10, 53-62.
Broyden C.G. 1965. A class of methods for solving nonlinear simultaneousequations.
Mathematicsof Computarion19, 577-593.
Broyden C.G. 1972. Quasi-Newtonmethods. In: Numerical Methodsfor Unconstrained
Optimization(ed. W. Murray), pp. 87-106. AcademicPress,Inc.
Burden R.L., FairesJ.D. and ReynoldsA.C. 1981.NumericalAnalysis.Prindle, \trfebber&
Schmidt.
deGroot-HedlinC. and ConstableS.C. 1990.Occam'sinversionto generatesmooth,two-
dimensionalmodelsfrom magnetotelluricdata.Geophysics 55, l6L3-1624.
Dennis J.E. and SchnabelR. 1983.NumericalMethodsfor (Jnconstrained Obtimizationand
NonlinearEquations.Prentice-Hall, Inc.
Dey A. and Morrison H.F. 1979. Resistivity modelling for arbitrarily shaped two-
dimensionalstructures.Geophysical Prospecting 27, 106-136.
Edwards L.S. 1977. A modified pseudosectionfor resistivity and induced-polarization.
Geophgsics 42, 1020-1036.
GeorgeA. and Liu J.$7. 1981.ComputerSolutionof Large SparsePositiaeDefiniteSgstems.
Prentice-Hall Inc.
Golub G.H. and van Loan C.F. 7989.Matrix Computations. The JohnsHopkins University
Press.

O 1996 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44, 731-152
152 M.H. Loke and R.D. Barker

in
Griffiths D.H. and Barker R.D. 1993. Two-dimensional resistivity imaging and modelling
areas of complex geology. Journal of Applied Geophysics29' 2lL-226'
Griffiths D.H., Turnbull J. and Olayinka A.L 1990. Two-dimensional resistivity mapping
with a computer-controlled arcay. First Break 8, l2l-129'
Hallof P.G. 1957. On the interpretation of resistiaitg and induced polarization measurements.
PhD thesis, M.I.T.
Keller G.V. and Frischknecht F.C. 1966. Electrical Methods in Geophysical Prospecting'
Pergamon Press, Inc.
Lawson C.L. and Hanson R.J. 1974. Solaing Last-squaresProblems.Prentice-Hall Inc.
Lines L.R. and Treitel S. 1984. Tutorial: A review of least-squares inversion and its
application to geophysical problems. GeophysicalProspecting32' 159-186'
Loke M.H. 1994. The inttersion of two-dimensional apparent resistiaity data. P}:D thesis, The
University of Birmingham.
Loke M.H. and Barker R.D. 1995. Least-squares deconvolution of apparent resistivity
pseudosections. Geophysics60, 1682-1690.
Martin R.S. and \Tilkinson J.H. 1965. Symmetric decomposition of positive definite band
matrices. Numerische Mathematik 7, 355-361
McGillivray P.R. and Oldenburg D.S7. 1990. Methods for calculating Frechet derivatives
and sensitivities for the non-linear inverse problem: A comparative study. Geophysical
Pr ospecting 38, 499- 524.
More J.J. and Trangenstein J.A. 1976. On the global convergence of Broyden's method.
Mathematics of Computation 30, 523-540.
Park S.K. and Van G.P. 1991. Inversion of pole-pole data for 3-D resistivity structures
beneath arrays of electrodes. Geophysics56' 951-960.
Press S7.H., Flannery 8.P., Teukolsky S.A. and Vetterling V.T. 1988. Numerical Recipesin
C. Cambridge lJniversity Press.
Sasaki Y. 1989. Two-dimensional joint inversion of magnetoteliuric and dipole-dipole
resistivity data. Geophgsics54, 254-262.
Sasaki Y. 1992. Resolution of resistivity tomography inferred from numerical simulation.
Geophysical Prospecting 40, 453-464.
Scales L.E. 1985. Intoduction to Non-linear Optimization. MacMillan Publishing Co.
Schwarz H.R., Rutishauser H., Stiefel E. and Hertelendy P.1973. Numerical Analysis of
Symmetric Matices. Prentice-Hall Inc.
Sheriff R.E. 1991. Encyclopedic Dictionary of Exploration Geophysics. Society of Exploration
Geophysicists.
Smith N.C. and Vozoff K. 1984. Two-dimensional DC resistivity inversion for dipole-dipole
data. IEEE Transactionson Geoscienceand Remote Sensing22' 2l-28.
Tripp A.C., Hohmann G.$/. and Swift Jr. C.M. 1984. Two-dimensional resistivity
inversion. Geophysics49, 1708-17 17.
Turnbull J. 1936. A fi.eld and interpretation technique for resistiztity suraeying ooer two-
dimensionalstructures.PhD thesis., The University of Birmingham.
\7oo P.T., Eisenstat M.H., Schultz M.H. and Sherman A.H. 1976. Application of sparse
matrix techniques to reservoir simulation. In'. Sparse Ma*ix Computations (eds J.R.
Bunch and D.J. Rose), pp. 427-438. Academic Press, Inc.

O 1995 European Association of Geoscientists & Engineers, Geophysical Prospecting, 44' l3l-152