Sie sind auf Seite 1von 26

Non-linear three-dimensional inversion of

cross-well electrical measurements


1
Aria Abubakar
2
and Peter M. van den Berg
2
Abstract
Cross-well electrical measurement as known in the oil industry is a method for
determining the electrical conductivity distribution between boreholes from the
electrostatic eld measurements in the boreholes. We discuss the reconstruction of the
conductivity distribution of a three-dimensional domain. The measured secondary
electric potential eld is represented in terms of an integral equation for the vector
electric eld. This integral equation is taken as the starting point to develop a non-linear
inversion method, the so-called contrast source inversion (CSI) method. The CSI
method considers the inverse scattering problemas an inverse source problemin which
the unknown contrast source (the product of the total electric eld and the conductivity
contrast) in the object domain is reconstructed by minimizing the object and data error
using a conjugate-gradient step, after which the conductivity contrast is updated by
minimizing only the error in the object. This method has been tested on a number of
numerical examples using the synthetic measured data with and without noise.
Numerical tests indicate that the inversion method yields a reasonably good
reconstruction result, and is fairly insensitive to added random noise.
Introduction
Interest in cross-well tomography (imaging) of the earths electrical conductivity has
increased because of improvements in eld instrumentation, computing power and
methods of interpretation. Cross-well electromagnetic logging is a technique to
investigate the geological properties of the region between boreholes using the
electromagnetic measurements made in these boreholes, either at low frequencies (less
than 100 kHz), so-called induction logging, or at zero frequency, so-called electrical
logging. One of the important parameters to be determined from the measurements is
the conductivity because of its sensitivity to porosity, pore-uid type and saturation.
During the last decade, important progress has been made in solving the inverse
problem for cross-well congurations. Some notable contributions in this area include
2000 European Association of Geoscientists & Engineers 109
Geophysical Prospecting, 2000, 48, 109134
1
Paper presented at the 60th EAGE Conference Geophysical Division, Leipzig, June 1998. Received
October 1998, revision accepted June 1999.
2
Laboratory of Electromagnetic Research, Center for Technical Geoscience, Delft University of
Technology, PO Box 5031, 2600 GA Delft, The Netherlands.
those of Torres-Verdin and Habashy (1994) and Alumbaugh and Morrison (1995).
Torres-Verdin and Habashy (1994) developed a fast inverse solution based on a non-
linear approximation of the integral equation. This approximation is used in their
forward calculations to produce the predicted elds from the latest model update. The
approximation compares favourably with full solutions for conductivity contrast as
large as 50 to 1, for frequencies below 100 kHz. Alumbaugh and Morrison (1995)
developed a multifrequency imaging procedure for the reconstruction of two-
dimensional variations of conductivity excited by electric line sources. In their
approach, the Greens function was xed for a certain background while the unknown
electric eld was updated after each iteration. A slightly different class of iterative
methods undertakes repeated modications of the Greens function after each
iteration. This method is known as the distorted Born iterative method (Chew and
Wang 1990) and it was also used in a three-dimensional cross-well problem by
Newman (1995). Note that in this method a full forward problem must be solved in
each iteration. The necessity of solving a forward problemin each iteration was avoided
by Torres-Verdin and Habashy (1995) by using a non-linear inversion technique,
known as the iterative extended Born approximation, to investigate a two-dimensional
object with conductivity contrast. They showed that, with the same computational
efciency as the rst-order Born approximation, the extended Born approximation
enables a much wider class of two-dimensional inverse scattering problem to be solved.
All the schemes mentioned above are used to invert low-frequency electromagnetic
measurements. Unlike induction logging, electrical logging is not so widely used. In the
electrical logging technique, the conductivity distributions are determined from the
static eld measurements (the electric current or the voltage) made in the borehole.
Such a measurement system enables us to increase the investigation range while
sacricing the resolution. This technique was used by Daily and Owen (1991) to invert
resistivity data in two-dimensional cross-well congurations with the help of an
iterative modied least-squares inversion based on a nite-element forward solution of
Laplaces equation. In contrast to the work of Daily and Owen (1991), Abubakar
(1996) used the modied gradient method to reconstruct three-dimensional layered
single-well and cross-well congurations. In the modied gradient method developed
by Abubakar (1996), the vector electric current density or the scalar electric potential
eld and the conductivity contrast are updated simultaneously by a non-linear
conjugate-gradient algorithm that minimizes the error in both the object equation and
the data equation. Furthermore, it was also shown that inversion based on the vector
integral equation for the electric current density leads to substantially better results
than inversion based on the scalar integral equation for the electric potential.
Recently, we have developed a method to reconstruct a three-dimensional
conductivity distribution from the cross-well electrical measurements (Abubakar and
Van den Berg 1998). This method is called the contrast source inversion (CSI) method
and was originally introduced by Van den Berg and Kleinman (1997) to handle the
two-dimensional wave problem. Unlike most non-linear inversion methods, the CSI
method does not require any articial regularization techniques to deal with the
110 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
problems of non-uniqueness in the inversion of data. It attempts to overcome this
problem by recasting the problem as an optimization problem, in which it seeks not
only the contrast source (the product of the total eld and the conductivity contrast)
but also the conductivity contrast itself to minimize a cost functional consisting of two
terms, the L
2
errors in the data equation and in the object equation. An alternative
iterative method of solving this optimization problem is proposed, in which rst the
contrast source is updated in the conjugate-gradient direction, weighted so as to
minimize the cost functional, and then the conductivity contrast is updated to minimize
the error in the object equation using the updated contrast source. This latter
minimization can be performed analytically which allows easy implementation of the
positivity constraint for the conductivity. An ambiguity in the work of Abubakar and
Van den Berg (1998) as to whether the object error was actually reduced is removed in
the present version by performing an extra line minimization to update the
conductivity contrast in the optimization process. Furthermore, numerical tests also
indicate that, by using sources and receivers located in four boreholes around a three-
dimensional domain, we can improve the reconstruction results of Abubakar and Van
den Berg (1998). In view of the efciency of the CSI method, three-dimensional
inverse problems can now be handled with moderate computer power.
Integral representations
A theoretical model of the cross-well conguration is shown in Fig. 1. We dene an
inhomogeneous domain D with conductivity j(x) in an unbounded homogeneous
background medium with conductivity j
0
. The source is a small spherical electrode
emitting a DC current I, located in domain S. We measure the secondary electric
Non-linear 3D inversion of cross-well electrical measurements 111
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 1. Object domain D with conductivity j(x) in the unbounded homogeneous background
of conductivity j
0
.
potential elds V
s
at the various electrode locations in domain S. In the present cross-
well electrical problem we want to determine the conductivity distribution j(x) inside
the domain D from the secondary electric potential eld V
s
measurements made in
domain S.
The governing equations originating fromMaxwells equations at zero frequency are
of the form:
V E = 0; (1)
(jE) = q
ext
; (2)
where V is the electric potential eld, E is the electric eld, j is the electrical
conductivity and q
ext
is the external source. Here, =(
1
,
2
,
3
) denotes spatial
differentiation with respect to the Cartesian position vector x =(x
1
, x
2
, x
3
).
To obtain (1) and (2) in an integral form, we assume that the actual conguration in
which the eld is to be computed (the object domain D) is embedded in a medium for
which the point source solution can be determined analytically. This point source
solution {V
G
, E
G
} is also known as the Greens state. The simplest medium in this
category is the unbounded homogeneous medium with conductivity j
0
. Substituting
q
ext
=d(x) into (1) and (2), we arrive at
V
G
(x) = j
1
0
G(x) and E
G
(x) = j
1
0
G(x); (3)
where G is the scalar Greens function for the static eld, given by
G(x) =
1
4p[x[
: (4)
We dene V
p
and E
p
as the primary electric potential eld and the primary electric
eld, respectively, measured in the background conguration and excited by a small
spherical electrode emitting a DCcurrent I. The primary elds are the elds that would
be present in the entire conguration if domain D showed no contrast with the
embedding background medium (j(x) =j
0
). The small spherical electrode can be
modelled as a point source (Lovell 1993). Using q
ext
=Id(xx
s
) in (1) and (2), the
primary elds are obtained as
V
p
(x) = j
1
0
IG(x x
s
) and E
p
(x) = j
1
0
I G(x x
s
): (5)
Starting from (1) and (2), and using the spatial Fourier transform as the mathematical
tool (De Hoop 1995), we arrive at the integral equation for the scalar electric potential
eld
V(x) = V
p
(x)

X
/
D
G(x x
/
)x(x
/
)
/
V(x
/
)dv; x D; (6)
where V is the total electric potential eld in D, and
x(x) =
j(x) j
0
j
0
: (7)
112 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
In (6),
/
denotes spatial differentiation with respect to x
/
. This is an integral equation
for V and V. Although these quantities are coupled by a spatial differentiation, we
prefer an integral equation in which the same unknown occurs both inside and outside
the integral. Abubakar (1996) showed that inversion based on the vector integral
equation for the electric current density (electric eld) leads to substantially better
results than inversion based on the scalar integral equation for the electric potential
eld. The integral equation for the vector electric eld is then given by
E(x) = E
p
(x)

X
/
D
G(x x
/
)x(x
/
)E(x
/
)dv; x D; (8)
where E is the total electric eld in D. Note that in (6) and (8) the target is assumed to
be localized such that the integration is over the object domain D. Outside the object
domain D the conductivity contrast x is zero, and then the total electric eld is exactly
the primary electric eld.
In the electrical logging problem we are interested in the secondary electric potential
eld V
s
in data domain S at x
R
. The secondary electric potential eld is the difference
between the total electric potential eld and the primary electric potential eld due to
the presence of the object,
V
s
= V V
p
: (9)
Using (9) in (6), and (1), the secondary electric potential eld can be given in an
integral representation form,
V
s
(x
R
) =

X
/
D
G(x
R
x
/
)x(x
/
)E(x
/
)dv: (10)
Equation (10) is an integral equation of the second kind, and it relates the conductivity
contrast x to the secondary electric potential eld V
s
(measurement data). In forward
modelling, the conductivity contrast x is known and the secondary electric potential
eld V
s
can be calculated after the total electric eld E in the object domain D has been
obtained by solving (8). We observe that the integral equation (8) is a singular integral
equation, in which the grad-div operator acts on a normalized vector potential, dened
as the spatial convolution of the Greens function and the product of the conductivity
contrast and the total electric eld. Numerical implementation of such an integral
equation must be carried out carefully. This discretization procedure is discussed in
Appendix A.
In the inverse problemto solve for x, the eld values V
s
are known only over a limited
set of space points which lie outside the object domain D, and one must solve for both
the conductivity contrast x and the total electric eld E inside D. This inverse problem
is not linear.
Forward problem
In Appendix A we present the discretized version of the integral equation for the
electric eld given in (8). We discretize the object domain D using a mesh uniformly
Non-linear 3D inversion of cross-well electrical measurements 113
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
subdivided in the x
1
-, x
2
- and x
3
-directions with spacings Dx
1
, Dx
2
and Dx
3
. From the
results of Appendix A, by substituting (A16) into (A11)(A13), and using the results
in (A10), we obtain a linear system of equations for E
1
, E
2
and E
3
when the
conductivity contrast x is known. This linear system of equations can be written
compactly in an operator notation. The operator expression LE is directly obtained
from the left-hand side of (A10) of Appendix A.
Since the matrix operator consists of spatial convolutions we can use advantageously
fast Fourier transform (FFT) routines (Zwamborn and Van den Berg 1994). However,
we then need an iterative solution, and the conjugate-gradient method seems to be one
of the most efcient methods. With this so-called conjugate-gradient FFT technique
we are able to solve complex three-dimensional problems efciently. Furthermore, it
also gives the basis of our solution of the inverse problem. We observe that the matrix
describing this linear system of equations is non-symmetric. Therefore we also need
the adjoint operator in order to set up the conjugate-gradient scheme (Van den Berg
1981). The matrix form of this adjoint operator is given in Appendix B. In such
a scheme we also need the denition of the norm and inner product. In view of
the denition of the discretized eld quantities in Appendix A, the norm on D is
dened as
k Ek
2
D
= Dx
1
Dx
2
Dx
3
X
3
k=1
X
M
m=1
X
N
n=1
X
P
p=1
E
k;m;n;p
E
k;m;n;p
: (11)
With these denitions we are nowable to apply a conjugate-gradient iterative scheme
to solve the operator equations
(LE)
k;m;n;p
= E
p
k;m;n;p
; k |1; 2; 3; (12)
for m=1, . . ., M, n =1, . . ., N, and p =1, . . ., P. Once the normalized error,
F
1
2
=
krk
D
kE
p
k
D
; (13)
where
r = E
p
LE; (14)
is small enough, the approximate solution of E is substituted in (10), to arrive at the
secondary potential eld at the receiver points:
V
s
(x
R
) = Dx
1
Dx
2
Dx
3
X
3
k=1
X
M
m=1
X
N
n=1
X
P
p=1
G
R
k
(x
R
1
x
1;m
; x
R
2
x
2;n
; x
R
3
x
3;p
)x
m;n;p
E
k;m;n;p
;
(15)
where
G
R
k
(x
R
1
x
1;m
; x
R
2
x
2;n
; x
R
3
x
3;p
) =
R
k
G(x
R
1
x
1;m
; x
R
2
x
2;n
; x
R
3
x
3;p
); (16)
in which
R
k
denotes the spatial differentiations with respect to x
R
k
.
114 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Inverse problem
We now assume that the inhomogeneous domain D is irradiated successively by
a number (i =1, . . ., I ) of known primary electric elds. For each excitation, the
forward scattering problem may be reformulated as the domain integral equation,
see (12),
LE
(i )
= E
p;(i )
; on D; (17)
where the operator L depends on the conductivity contrast x. To show this explicitly,
we rewrite this equation as
E
(i )
K
D
xE
(i )
= E
p;(i )
; on D: (18)
In (18), K
D
xE
(i )
follows from (A11)(A13) by replacing A
k;m, n, p
with A
(i )
k;m, n, p
.
Equation (18) is referred to as the object equation that holds in the object domain D.
Furthermore, in the inverse problem the secondary potential eld is known or
measured at the measurement points x
R
. We assume that all the measurement points
are located in the data domain S, outside D. We can also write (15) in the shorthand
notation,
K
S
xE
(i )
= V
s;(i )
; on S; (19)
where V
s, (i )
follows from (15) after replacing E
k;m, n, p
with E
(i )
k;m, n, p
. Equation (19) is
referred to as the data equation that holds in the data domain S.
The data equation contains both the unknown total electric eld and the unknown
conductivity contrast, but they occur as a product which can be considered as a
contrast source that produces the secondary electric potential eld at the measurement
points; hence there is no unique solution to the problem of inverting the data equation
by itself. The CSI method attempts to overcome this difculty by recasting the problem
as an optimization, in which we seek not only the contrast sources but also the
conductivity contrast itself to minimize a cost functional consisting of two terms, the L
2
errors in the data equation and in the object equation, rewritten in terms of the
conductivity contrast and the contrast sources rather than the electric elds. An
alternative method of solving this optimization problemiteratively is proposed in which
rst the contrast sources are updated in the conjugate-gradient step, weighted so as to
minimize the cost functional, and then the conductivity contrast is updated to minimize
the error in the object equation using the updated sources. This latter minimization can
be carried out analytically which allows an easy implementation of the positivity
constraint for the conductivity.
To this end we introduce the contrast source W
(i )
as follows:
W
(i )
= xE
(i )
: (20)
The data equation becomes
K
S
W
(i )
= V
s;(i )
; on S: (21)
Non-linear 3D inversion of cross-well electrical measurements 115
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Substituting (18) into (20), we obtain an object equation for the contrast source rather
than for the eld, namely,
W
(i )
xK
D
W
(i )
= xE
p;(i )
; on D: (22)
These latter two equations are the basic equations for the CSI method.
The CSI method consists of an algorithm to construct sequences of contrast sources
{W
(i )
k
} and of conductivity contrasts {x
k
} which iteratively reduce the value of the cost
functional
F =
S
i
kV
s;(i )
K
S
W
(i )
k
2
S
S
i
kV
s;(i )
k
2
S

S
i
kxE
p;(i )
W
(i )
xK
D
W
(i )
k
2
D
S
i
kxE
p;(i )
k
2
D
; (23)
where the norm on S is given by
kV
(i )
k
2
S
=
X
X
R
V
(i )
(x
R
)V
(i )
(x
R
); (24)
and the norm on D is given by (11). The normalization is chosen so that both terms are
equal to one if the contrast source W
(i )
is zero. The rst term measures the error in the
data equation, and the second term measures the error in the object equation. This is a
quadratic functional in W
(i )
, but it is highly non-linear in x. Note that the object
equation acts as a regularization for the data equation, and we have not employed any
other regularization techniques.
The algorithm involves the construction of sequences {W
(i )
k
and {x
k
}, k =1, 2, . . ., in
the following manner. Dene the data error and the object error at the kth step to be
r
(i )
k
= V
s;(i )
K
S
W
(i )
k
and r
(i )
k
= x
k
E
(i )
k
W
(i )
k
; (25)
where
E
(i )
k
= E
p;(i )
K
D
W
(i )
k
: (26)
Now suppose W
(i )
k1
and x
k1
are known; we update W
(i )
as follows:
W
(i )
k
= W
(i )
k1
a
(i )
k
w
(i )
k
; (27)
where a
(i)
k
is a constant, and w
(i )
k
is the PolakRibiere gradient direction,
w
(i )
0
= 0; w
(i )
k
= w
(i )
k

w
(i )
k
; w
(i )
k
w
(i )
k1
)
D
kw
(i )
k1
k
2
D
w
(i )
k1
; k 1; (28)
where w
(i )
k
is the gradient (Frechet derivative) of the cost functional F with respect to
W
(i )
evaluated at W
(i )
k1
and x
k1
. Explicitly, the gradient is found to be
w
(i )
k
= n
S
K
+
S
r
(i )
k1
n
D;k1
(r
(i )
k1
K
+
D
x
k1
r
(i )
k1
); (29)
where
n
S
=
X
i
kV
s;(i )
k
2
S
!
1
and n
D;k1
=
X
i
kx
k1
E
p;(i )
k
2
D
!
1
: (30)
116 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
The various operators are dened as (omitting the iteration index k1)
(K
+
S
r
(i)
)
k;m;n;p
=
X
X
R
G
R
k
(x
R
1
x
1;m
; x
R
2
x
2;n
; x
R
3
x
3;p
)r
(i)
(x
R
); (31)
(K
+
D
xr
(i)
)
k;m;n;p
= C
(i)
k;m;n;p
; (32)
where C
(i)
k;m,n,p
is obtained from (B3) by replacing r
k;m,n,p
in (B4)(B6) with
x
m, n, p
r
(i)
k;m,n,p
. The motivation for choosing the PolakRibiere gradient direction
rather than the FletcherReeves gradient direction is the presence of the conductivity
contrast x itself in the cost functional in (23) which is also updated during the
optimization process and hence disturbs the orthogonality properties of the gradients.
Moreover, the PolakRibiere gradient direction is more robust when the difference
between the successive gradients becomes small.
With the update directions completely specied, the constant a
(i)
k
is determined to
minimize the cost functional
F
k
= n
S
X
i
kr
(i)
k1
a
(i)
k
K
S
w
(i)
k
k
2
S
n
D;k1
X
i
kr
(i)
k1
a
(i)
k
(w
(i)
k
x
k1
K
D
w
(i)
k
)k
2
D
;
(33)
and is found to be
a
(i)
k
=
n
S
r
(i)
k1
; K
S
w
(i)
k
)
S
n
D;k1
r
(i)
k1
; w
(i)
k
x
k1
K
D
w
(i)
k
)
D
n
S
kK
S
w
(i)
k
k
2
S
n
D;k1
kw
(i)
k
x
k1
K
D
w
(i)
k
k
2
D
: (34)
Once the contrast source W
(i)
k
is determined, the eld E
(i)
k
is obtained using (26). To
update x we proceed in two steps. First we determine x
k
to minimize the numerator of
the second term in (23), which we write as

F
D
=
X
i
kr
(i)
k
k
2
D
=
X
i
kxE
(i)
k
W
(i)
k
k
2
D
: (35)
Minimizing (35) with respect to x, we arrive at
x
k
=
S
i
(E
(i)
k
W
(i)
k
)
S
i
[E
(i)
k
[
2
: (36)
As noted by Van den Berg and Kleinman (1997), it is not certain whether this choice
will necessarily reduce the error in the second term of (23). We ensure that this choice
is reduced by a line minimization. To this end we write
x
k
= x
k1
v( x
k
x
k1
) = x
k1
vd
k
; (37)
and choose the real parameter v to minimize
F
D
= n
D;k
X
i
kr
(i)
k
k
2
D
=
S
i
kx
k
E
(i)
k
W
(i)
k
k
2
D
S
i
kx
k
E
p;(i)
k
2
D
=
av
2
bv c
Av
2
Bv C
; (38)
Non-linear 3D inversion of cross-well electrical measurements 117
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
where
a =
P
i
|d
k
E
(i)
k
|
2
D
; A =
P
i
|d
k
E
p;(i)
|
2
D
;
b =
P
i
x
k1
E
(i)
k
W
(i)
k
; d
k
E
(i)
k
)
2
D
; B =
P
i
x
k1
E
p;(i)
; d
k
E
p;(i)
)
2
D
;
c =
P
i
|x
k1
E
(i)
k
W
(i)
k
|
2
D
; C =
P
i
|x
k1
E
p;(i)
|
2
D
:
This is the quotient of two quadratics which, using elementary analysis, may be shown
to attain its minimum when
v =
(aC Ac)

(aC Ac)
2
4(aB Ab)(bC Bc)
p
2(aB Ab)
: (39)
This completes the denitions of the conductivity contrast x if there is no a priori
information. If it is known that j is a positive quantity, this information may easily be
incorporated by choosing x to be 1 if (37) yields a value less than 1.
The last point in the description of the algorithm is the choice of the starting values
W
(i)
0
. Observe that we cannot start with W
(i)
0
=0 since then x
0
=0 and the cost
functional F in (23) is undened for k =1. Therefore we choose as starting values the
values that minimize only the data error (the rst term in (23)),
W
(i)
0
=
|K
+
S
V
s;(i)
|
2
D
|K
S
K
+
S
V
s;(i)
|
2
S
K
+
S
V
s;(i)
: (40)
Note that K
*
S
V
s,(i)
is the back-propagation of the data from the data domain S into the
object domain D, and is often called a back-propagation of the eld data. With this
initial estimate for W
(i)
0
, the conductivity contrast estimate x
0
is obtained using (26) and
(36). This completes the description of the inversion algorithm.
Numerical examples
We demonstrate three examples of inversion of synthetic data. Although we realize
that the ultimate goal is to invert eld data, we show the results of our inversion
scheme using synthetic data. The synthetic data were simulated by solving the
forward scattering problem with the conjugate-gradient method in which all discrete
convolutions are calculated using FFTroutines. In order to avoid the possibility of an
inverse crime, we also carry out some experiments with noise added to these
synthetic data. In addition, in our last example, we generate the data by a different
type of integral equation with a discretization different from that used in the inverse
problem, namely, the scalar electric potential integral equation in (6) using a ner
spatial discretization grid, where the mesh size is half the size employed in the
inversion scheme. The reconstruction was nearly identical to the reconstruction by
inversion of the synthetic data generated by solution of the vector electric eld
integral equation.
118 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Reconstruction of a homogeneous anomaly
Figure 2 shows the test (object) domain and the set-up of the sources and receivers.
We used 42 stations; 21 stations are located in the borehole on the left-hand side of
the test domain and the other 21 stations are located on the other side. The distance
between the two boreholes is 100 m. The vertical distance between two stations is 5 m
with each station serving successively as a point source and all other stations acting as
point receivers. Thus for the reconstruction of the prole we have used 1764 data
points.
Figure 3a illustrates one of the three-dimensional test model examples we have used
to investigate the performance of the inversion scheme. The model consists of a
homogeneous anomaly of conductivity 0.9 S/m embedded in a 0.3 S/m background.
The dimensions of the anomaly are 30 30 30 m
3
. In this example we take a test
domain of 28 28 28 subdomains with a side length of 2.5 m; hence the total number
of conductivity contrast unknowns is equal to 21 952, and the dimensions of the test
domain are 70 70 70 m
3
. Note that the test domain is much larger than the actual
anomaly. Figure 3 shows, from top to bottom, the conductivity distribution (j) in the
(x
1
, x
2
)-plane at different x
3
-levels. The x
3
-levels, from top to bottom, are 27.5,
12.5, 2.5, 2.5, 12.5 and 27.5 m.
Figure 3b is a plot of the conductivity distribution inverted after 512 iterations with
the present inversion method using the synthetic data. Although the total number of
iterations is large, note that we do not solve a full forward problem in each iteration of
the inversion procedure. The resolution of the reconstructed object is very low, and
further iterations do not improve the results. After 256 iterations the error in the
conductivity has decreased to ERR
k
=0.30708. The error in conductivity at the kth
Non-linear 3D inversion of cross-well electrical measurements 119
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 2. Sourcereceiver set-up with 42 sources and receivers located in two boreholes.
iteration is given by
ERR
k
=
kj
k
j
exact
k
2
D
kj
exact
k
2
D
: (41)
The square root of the cost functional F
k
in (33) has now decreased to 0.94%. We
observe clearly that although the error with respect to the data is very small, the
conductivity contrast is not well reconstructed; this is due to the insensitivity of the data
to the changes in the conductivity contrast. Figure 4 shows the conductivity of the
original prole and reconstruction after 512 iterations in the (x
2
, x
3
)-plane at x
1
=0.
Abubakar and Van den Berg (1998) showed that if we include some a priori
120 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 3. Conductivity distribution (j) in the (x
1
,x
2
)-plane for different x
3
-levels of a three-
dimensional homogeneous anomaly. (a) The original prole; (b) the reconstructions after 512
iterations.
information about the location of the anomaly, we can improve the reconstruction
results signicantly. On the other hand, this a priori information is very often not
present. If we have a three-dimensional well coverage about the target, it might be
expected that the results would be greatly improved and the need for prior information
greatly reduced. In order to improve the reconstructed prole, we now use 16 sources
and 28 receivers located in four boreholes. The boreholes are located at (0, 50, x
3
),
(0, 50, x
3
), (50, 0, x
3
) and (50, 0, x
3
) as shown in Fig. 5. The vertical positions of the
sources in a borehole are x
3
= 15, 5, 5 and 15 m. The vertical positions of the
receivers in a borehole are x
3
= 30, 20, 10, 0, 10, 20 and 30 m. Thus for the
reconstruction of the prole we have used only 448 data points.
The original prole is shown again in Figs 6a and 7a. The reconstruction after 512
iterations is shown in Figs 6b and 7b. We observe that by using fewer sources and
receivers located in four boreholes, better reconstruction results are obtained than
Non-linear 3D inversion of cross-well electrical measurements 121
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 4. Conductivity distribution (j) in the (x
2
,x
3
)-plane at x
1
=0 of a three-dimensional
homogeneous anomaly. (a) The original prole; (b) the reconstructions after 512 iterations.
those shown in Figs 3b and 4b. The position of the anomaly is now accurately
determined, but the conductivity value of the anomaly is still lower than the original
one. The error in the conductivity has decreased to ERR
k
=0.28976 and the square
root of the cost functional F
k
has now decreased to 1.07%. The computation time for
one iteration in the inversion scheme of this example was 210 seconds on a personal
computer with a 400 MHz Pentium processor.
Reconstruction of a vertically layered conguration
As the second example we consider a more complicated example. The measurement
set-up used is shown in Fig. 5. Thus for the reconstruction of the prole we use 448
data points. The conguration which is to be reconstructed is embedded in a
homogeneous background medium of conductivity 0.3 S/m. A 35 35 35 m
3
test
domain is used and discretized into 14 14 14 subdomains. Thus each subdomain
again has the dimensions 2.5 2.5 2.5 m
3
, and the total number of conductivity
contrast unknowns is 2744.
The original prole which is used to produce the synthetic data is shown in Fig. 8a.
This gure shows, from top to bottom, the original prole in the (x
1
, x
2
)-plane at
different x
3
-levels. The x
3
-levels from top to bottom are 13.75, 11.25, 1.25, 1.25,
122 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 5. Sourcereceiver set-up with 16 sources and 28 receivers located at four boreholes
around the three-dimensional test domain.
11.25 and 13.75 m. The model consists of three vertical layers. Each layer has the
same dimensions of 30 10 30 m
3
. The conductivity of the dark blue vertical layer is
0.1 S/m, that of the blue vertical layer is 0.3 S/m, and that of the red vertical layer is
0.9 S/m. Thus, the conductivity contrasts x from left to right in Fig. 8a are 0.667, 0
and 2.
Figure 8b shows the conductivity distribution that was obtained after 256 iterations.
We observe that the reconstruction of the middle (blue) vertical layer is not as good as
the other two vertical layers. In Fig. 9 we show the conductivity prole of the original
conguration and reconstructions after 256 iterations in the (x
2
, x
3
)-plane at x
1
=0. We
note that after 256 iterations, the error in the conductivity has decreased to
ERR
k
=0.22450, the square root of the cost functional F
k
has decreased to 0.78%,
and further iterations do not improve the reconstruction results. The computation time
Non-linear 3D inversion of cross-well electrical measurements 123
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 6. As Fig. 3, but using the measurement set-up shown in Fig. 5.
for one iteration in the inversion scheme of this example was 10 seconds on a personal
computer with a 400 MHz Pentium processor.
Reconstruction of a horizontally layered conguration
We have considered two examples that are quite far removed from actual cases of
practical cross-well electrical logging problems. Thus, as the next example, we consider
a more realistic conguration. The conguration consists of three different layers in the
vertical direction (x
3
-direction) with dimensions of 30 30 10 m
3
. The conductivity
of each layer varies from 0.9 S/m (red) to 0.3 S/m (blue) and 0.1 S/m (dark blue). The
inversion is performed by discretizing a test domain of 35 35 35 m
3
into a grid of
14 14 14 subdomains, and hence the total number of conductivity contrast
unknowns amounts to 2744. The source and receiver set-up is the same as before.
124 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 7. As Fig. 4, but using the measurement set-up shown in Fig. 5.
We use 16 sources and 28 receivers located in four boreholes. The boreholes are
located at (0, 50, x
3
), (0, 50, x
3
), (50, 0, x
3
) and (50, 0, x
3
) as shown in Fig. 5. The
vertical positions of the sources in a borehole are x
3
= 15, 5, 5 and 15 m. The
vertical positions of the receivers in a borehole are x
3
= 30, 20, 10, 0, 10, 20 and
30 m. Thus for the reconstruction of the prole we use 448 data points. The
conductivity of the background medium is 0.3 S/m.
The original prole which is used to produce the synthetic data is shown in
Fig. 10a. This gure shows, from top to bottom, the conductivity distribution in
the (x
1
, x
2
)-plane at different x
3
-levels. The x
3
-levels from top to bottom are 13.75,
11.25, 1.25, 1.25, 11.25 and 13.75 m. Figure 10b shows the reconstruction
after 256 iterations. We observe that after 256 iterations we have already obtained a
Non-linear 3D inversion of cross-well electrical measurements 125
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 8. Conductivity distribution (j) in the (x
1
,x
2
)-plane for different x
3
-levels of a vertically
layered conguration. (a) The original prole; (b) the reconstructions after 256 iterations.
good reconstruction result. We note that after 256 iterations, the error in the
conductivity has decreased to ERR
k
=0.20412, the square root of the cost functional
F
k
has decreased to 0.53%, and further iterations do not yield substantial
improvement of the reconstruction results. In order to investigate the vertical
resolution, we show in Fig. 11 the conductivity of the original prole and the
reconstructed prole after 256 iterations in the (x
2
, x
3
)-plane at x
1
=0 m. We
observe that the resolution of the reconstruction of the middle layer is lower than the
others.
So far, the results presented are for noiseless synthetic data. In order to simulate a
more realistic eld experiment we have added 5%, 10% and 20% random noise to the
synthetic data. In our analysis we have taken the secondary electric potential eld
V
s,(i)
(x
R
) as the data quantities (see (15)); then the noisy secondary electric potential
126 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
10 0 10
15
10
5
0
5
10
15
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
10 0 10
15
10
5
0
5
10
15
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
(a)
(b)
0.9
0.9
Conductivity
(S/m)
Conductivity
(S/m)
x
3
(m)
x
3
(m)
x
2
(m)
Figure 9. Conductivity distribution (j) in the (x
2
,x
3
)-plane at x
1
=0 of a vertically layered
conguration. (a) The original prole; (b) the reconstructions after 256 iterations.
eld is given by
V
s;(i)
noise
(x
R
) = [1 bf
noise
]V
s;(i)
(x
R
); (42)
where f
noise
is a random number between 1 and 1 and b=0.05, 0.1 or 0.2.
The 5% and 10% noise do not alter the reconstruction results signicantly. The
results of the reconstruction after 256 iterations with noisy data (20% noise) are shown
in Figs 12b and 13b. By comparing Figs 10 and 11 with Figs 12 and 13, we observe that
the 20% noise only altered the region with positive contrast (the red layers). We
note that after 256 iterations, the error in the conductivity of the reconstruction
Non-linear 3D inversion of cross-well electrical measurements 127
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 10. Conductivity distribution (j) in the (x
1
,x
2
)-plane for different x
3
-levels of a
horizontally layered conguration. (a) The original prole; (b) the reconstructions after 256
iterations.
with noisy data has decreased to ERR
k
=0.20950, the square root of the cost functional
F
k
has decreased to 10.08%, and further iterations do not improve the reconstruction
results.
Conclusions
We have demonstrated that the present inversion algorithm can reconstruct a three-
dimensional electrode logging problem over a wide range of conductivity contrasts
using moderate computer power. The main advantage of the present non-linear
inversion method is that we do not have to solve a full forward problem in each
iteration, and we do not need an articial regularization technique to deal with the non-
uniqueness problem of the inversion of data. Without using any a priori information
128 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
10 0 10
15
10
5
0
5
10
15
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
10 0 10
15
10
5
0
5
10
15
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.9
(a)
(b)
x
3
(m)
x
3
(m)
x
2
(m)
Conductivity
(S/m)
Conductivity
(S/m)
Figure 11. Conductivity distribution (j) in the (x
2
,x
3
)-plane at x
1
=0 of a horizontally layered
conguration. (a) The original prole; (b) the reconstructions after 256 iterations.
(except for the positivity of the conductivity), the present algorithmcan reconstruct the
unknown conductivity contrast up to an acceptable level of accuracy. Because of the
insensitivity of the data with respect to changes in the conductivity contrast, the use of a
three-dimensional well coverage about the target is crucial. Furthermore, the numerical
tests indicate that this inversion algorithmusing synthetic data with 20%noise still gives
reasonably good reconstruction results. We expect that the present inversion scheme is
capable of handling a realistic three-dimensional inverse problem.
Future research will be directed at using the present inversion method to reconstruct
the conductivity distribution from the measurements in a single-borehole congura-
tion. In addition, we also plan to test our inversion method for the non-static case,
namely, the induction logging problem.
Non-linear 3D inversion of cross-well electrical measurements 129
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Figure 12. As Fig. 10, but with 20% noise.
Appendix A
Discretization procedure
Before discussing the discretization procedure of the integral equation for the electric
eld, we rst write (8) as
E
k
B
k
= E
p
k
; x
k
D; k |1; 2; 3; (A1)
where the vector B
k
is given by
B
k
=
k
[
1
A
1

2
A
2

3
A
3
]; (A2)
in which the normalized vector potential A
k
is given by
A
k
(x
1
; x
2
; x
3
) =

X
/
D
G(x
1
x
/
1
; x
2
x
/
2
; x
3
x
/
3
)x(x
/
1
; x
/
2
; x
/
3
)E
k
(x
/
1
; x
/
2
; x
/
3
)dv; (A3)
130 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
10 0 10
15
10
5
0
5
10
15
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
10 0 10
15
10
5
0
5
10
15
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.9
Conductivity
(S/m)
Conductivity
(S/m)
x
3
(m)
x
3
(m)
(a)
(b)
x
2
(m)
Figure 13. As Fig. 11, but with 20% noise.
and the Greens function is given by
G(x
1
; x
2
; x
3
) =
1
4p(x
2
1
x
2
2
x
2
3
)
1
2
: (A4)
We assume that the domain D as shown in Fig. 1 is a rectangular domain with
boundaries along the x
1
-, x
2
- and x
3
-directions. We discretize the domain into a
rectangular mesh. The mesh is uniformly spaced in the x
1
-, x
2
- and x
3
-directions. The
rectangular subdomains with a width of Dx
1
in the x
1
-direction, Dx
2
in the x
2
-direction,
and Dx
3
in the x
3
-direction
1
2
are given by
D
m;n;p
= (x
1
; x
2
; x
3
) R
3
|x
1;m

1
2
Dx
1
< x
1
< x
1;m

1
2
Dx
1
;

x
2;n

1
2
Dx
2
< x
2
< x
2;n

1
2
Dx
2
; x
3;p

1
2
Dx
3
< x
3
< x
3;p

1
2
Dx
3

; (A5)
for m=1, . . ., M, n =1, . . ., N, p =1, . . ., P, where
x
1;m
= x
1;
1
2
(m
1
2
)Dx
1
; m = 1; . . . ; M;
x
2;n
= x
2;
1
2
(n
1
2
)Dx
2
; n = 1; . . . ; N;
x
3;p
= x
3;
1
2
(p
1
2
)Dx
3
; p = 1; . . . ; P;
(A6)
in which x
1;1/2
is the lower x
1
-bound of the contrasting domain D, x
2;1/2
is its lower x
2
-
bound, and x
3;1/2
is its lower x
3
-bound. In each subdomain D
m,n,p
with centre points
(x
1;m
, x
2;n
, x
3;p
), we assume the conductivity contrast x to be constant, with the same
value as that at the centre point, x
m, n, p
=x(x
1;m
, x
2;n
, x
3;p
). We assume that the
boundary of the object domain D lies completely outside the contrasting object
(target). Using this spatial discretization grid, the pertaining eld quantities are dened
as follows:
E
k;m;n;p
= E
k
(x
1;m
; x
2;n
; x
3;p
); (A7)
E
p
k;m;n;p
= E
p
k
(x
1;m
; x
2;n
; x
3;p
); (A8)
B
k;m;n;p
= B
k
(x
1;m
; x
2;n
; x
3;p
); (A9)
for k {1, 2, 3}. Then (A1) is discretized as
E
k;m;n;p
B
k;m;n;p
= E
p
k;m;n;p
; (A10)
for m=1, . . ., M, n=1, . . ., N, and p =1, . . ., P, where B
k
, givenin (A2), is computed with
the nite-difference rule (Abramowitz and Stegun 1970), and the results are given as
B
1;m;n;p
= (Dx
1
)
2
(A
1;m1;n;p
2A
1;m;n;p
A
1;m1;n;p
)

1
4
(Dx
1
Dx
2
)
1
(A
2;m1;n1;p
A
2;m1;n1;p
A
2;m1;n1;p
A
2;m1;n1;p
)

1
4
(Dx
1
Dx
3
)
1
(A
3;m1;n;p1
A
3;m1;n;p1
A
3;m1;n;p1
A
3;m1;n;p1
);
(A11)
Non-linear 3D inversion of cross-well electrical measurements 131
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
B
2;m;n;p
= (Dx
2
)
2
(A
2;m;n1;p
2A
2;m;n;p
A
2;m;n1;p
)

1
4
(Dx
1
Dx
2
)
1
(A
1;m1;n1;p
A
1;m1;n1;p
A
1;m1;n1;p
A
1;m1;n1;p
)

1
4
(Dx
2
Dx
3
)
1
(A
3;m;n1;p1
A
3;m;n1;p1
A
3;m;n1;p1
A
3;m;n1;p1
);
(A12)
B
3;m;n;p
=(Dx
3
)
2
(A
3;m;n;p1
2A
3;m;n;p
A
3;m;n;p1
)

1
4
(Dx
1
Dx
3
)
1
(A
1;m1;n;p1
A
1;m1;n;p1
A
1;m1;n;p1
A
1;m1;n;p1
)

1
4
(Dx
2
Dx
3
)
1
(A
2;m;n1;p1
A
2;m;n1;p1
A
2;m;n1;p1
A
2;m;n1;p1
):
(A13)
Next we have to replace the continuous representation of the normalized vector
potential A
k
by a discrete one. In order to handle the singularity of the Greens function,
we rst take the spherical mean of the normalized vector potential. We integrate A
k
over
a spherical domain with centre at the point (x
1;m
, x
2;n
, x
3;p
). The radius of these patches
is taken to be
1
2
Dx =
1
2
min(Dx
1
, Dx
2
, Dx
3
). The results are divided by the volume of the
spherical domain of radius
1
2
Dx. We may then write
A
k;m;n;p
= A
k
(x
1;m
; x
2;n
; x
3;p
)
=

X
/
D
G(x
1;m
x
/
1
; x
2;n
x
/
2
; x
3;p
x
/
3
)x(x
/
1
; x
/
2
; x
/
3
)E
k
(x
/
1
; x
/
2
; x
/
3
)dv; (A14)
where we have interchanged the order of integrations, such that
G(x
1
; x
2
; x
3
) =
6
p(Dx)
3

[X
/
[<
1
2
Dx
G(x
1
x
/
1
; x
2
x
/
2
; x
3
x
/
3
)dv;
=
3
4pDx
[x[ = 0;
1
4p(x
2
1
x
2
2
x
2
3
)
1
2
[x[ >
1
2
Dx:
8
>
>
>
<
>
>
>
:
(A15)
In fact, G(x
1
, x
2
, x
3
) is the mean value of the Greens function over a spherical domain
with centre at (x
1
, x
2
, x
3
). Note that this weakening of the singularity is different from
the technique used by Van Bladel (1991), where the spatial differentiations act on the
Greens function directly, while we rst compute the normalized vector potential A
k
,
in which the Greens function has been weakened by taking its spherical mean, and
subsequently the differentiations are carried out numerically on the normalized
vector potential A
k
. This has proved (Zwamborn and Van den Berg 1994) to yield an
efcient, stable and accurate algorithm. After this weakening procedure, we are now
able to compute the integral of (A14) numerically. In view of the functional
properties of E
k
, we approximate the integral of (A14) using a midpoint rule. We then
132 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
arrive at
A
k;m;n;p
= Dx
1
Dx
2
Dx
3
X
M
m
/
=1
X
N
n
/
=1
X
P
p
/
=1
G
mm
/
;nn
/
;pp
/ x
m
/
;n
/
;p
/ E
k;m
/
;n
/
;p
/ ; (A16)
for m=0, . . ., M1, n =0, . . ., N1, and p =0, . . ., P1, where
G
mm
/
;nn
/
;pp
/ = G(x
1;m
x
1;m
/ ; x
2;n
x
2;n
/ ; x
3;p
x
3;p
/ ): (A17)
Note that A
k;m, n, p
are discrete convolutions in m
/
, n
/
, and p
/
and can be computed
efciently by FFT routines (Press et al. 1992).
Appendix B
Adjoint operator
The adjoint operator is dened through the relationship
r; LE)
D
= L
+
r; E)
D
; (B1)
where r and E are both in the same vector space (see (14)). Substituting the expression
of the operator LE in the left-hand side of (B1) and interchanging the various
summations, the adjoint operator is recognized as
(L
+
r)
k;m;n;p
= r
k;m;n;p
x
m;n;p
C
k;m;n;p
; (B2)
for m=1, . . ., M, n =1, . . ., N, and p =1, . . ., P, in which
C
k;m;n;p
= Dx
1
Dx
2
Dx
3
X
M1
m
/
=0
X
N1
n
/
=0
X
P1
p
/
=0
G
m
/
m;n
/
n;p
/
p
F
k;m
/
;n
/
;p
/ ; (B3)
where
F
1;m;n;p
= (Dx
1
)
2
(r
1;m1;n;p
2r
1;m;n;p
r
1;m1;n;p
)

1
4
(Dx
1
Dx
2
)
1
(r
2;m1;n1;p
r
2;m1;n1;p
r
2;m1;n1;p
r
2;m1;n1;p
)

1
4
(Dx
1
Dx
3
)
1
(r
3;m1;n;p1
r
3;m1;n;p1
r
3;m1;n;p1
r
3;m1;n;p1
); (B4)
F
2;m;n;p
= (Dx
2
)
2
(r
2;m;n1;p
2r
2;m;n;p
r
2;m;n1;p
)

1
4
(Dx
1
Dx
2
)
1
(r
1;m1;n1;p
r
1;m1;n1;p
r
1;m1;n1;p
r
1;m1;n1;p
)

1
4
(Dx
2
Dx
3
)
1
(r
3;m;n1;p1
r
3;m;n1;p1
r
3;m;n1;p1
r
3;m;n1;p1
); (B5)
F
3;m;n;p
= (Dx
3
)
2
(r
3;m;n;p1
2r
3;m;n;p
r
3;m;n;p1
)

1
4
(Dx
1
Dx
3
)
1
(r
1;m1;n;p1
r
1;m1;n;p1
r
1;m1;n;p1
r
1;m1;n1;p1
)

1
4
(Dx
2
Dx
3
)
1
(r
2;m;n1;p1
r
2;m;n1;p1
r
2;m;n1;p1
r
2;m;n1;p1
): (B6)
Non-linear 3D inversion of cross-well electrical measurements 133
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134
Since according to (B3), m
/
runs from 0 to M1, n
/
runs from 0 to N1 and p
/
runs
from 0 to P1, we set in (B4)(B6)
r
k;m;n;p
= 0; m = 1; 0; M 1; M 2; np;
r
k;m;n;p
= 0; n = 1; 0; N 1; N 2; mp;
r
k;m;n;p
= 0; p = 1; 0; P 1; P 2; mn:
(B7)
Note that C
k;m, n, p
is discrete convolution in m
/
, n
/
, and p
/
and these convolutions can
also be computed efciently by FFT routines (Press et al. 1992).
References
Abramowitz M. and Stegun I.A. 1970. Handbook of Mathematical Functions, p. 884. Dover
Publications Inc.
Abubakar A. 1996. Non-linear inversion in electrode logging using the modied gradient method.
MSc thesis, Delft University of Technology, The Netherlands.
Abubakar A. and Van den Berg P.M. 1998. Three-dimensional non-linear inversion in cross-well
electrode logging. Radio Science 33, 9891004.
Alumbaugh D.L. and Morrison H.F. 1995. Theoretical and practical considerations for crosswell
electromagnetic tomography assuming a cylindrical geometry. Geophysics 60, 846870.
Chew W.C. and Wang Y.M. 1990. Reconstruction of two-dimensional permittivity distribution
using the distorted Born iterative method. IEEE Transactions on Medical Imaging 9, 218225.
Daily W. and Owen E. 1991. Cross-borehole resistivity tomography. Geophysics 56, 12281235.
De Hoop A.T. 1995. Handbook of Radiation and Scattering of Waves, pp. 719729. Academic
Press Inc.
Lovell J.R. 1993. Finite element methods in resistivity logging, pp. 4748. PhD thesis, Delft
University of Technology, The Netherlands.
Newman G. 1995. Crosswell electromagnetic inversion using integral and differential equations.
Geophysics 60, 899911.
Press W.H., Teukolsky S.A, Vetterling W.T. and Flannery B.P. 1992. Numerical Recipes in Fortran:
The Art of Scientic Computing, pp. 490525. Cambridge University Press.
Torres-Verdin C. and Habashy T.M. 1994. Rapid 2.5-dimensional forward modeling and
inversion via a new non-linear scattering approximation. Radio Science 29, 10511079.
Torres-Verdin C. and Habashy T.M. 1995. A two-step linear inversion of two-dimensional
electrical conductivity. IEEE Transactions on Antennas and Propagation 43, 405415.
Van Bladel J. 1991. Singular Electromagnetic Fields and Sources. Clarendon Press, Oxford.
Van den Berg P.M. 1981. Iterative computational techniques in scattering based upon the
integrated square error criterion. IEEE Transactions on Antennas and Propagation 32, 1063
1071.
Van den Berg P.M. and Kleinman R.E. 1997. Contrast source inversion method. Inverse Problems
13, 16071620.
Zwamborn A.P.M. and Van den Berg P.M. 1994. Computation of electromagnetic elds inside
strongly inhomogeneous objects by the weak-conjugate gradient FFT method. Journal of the
Optical Society 11, 14141421.
134 A. Abubakar and P.M. van den Berg
2000 European Association of Geoscientists & Engineers, Geophysical Prospecting, 48, 109134

Das könnte Ihnen auch gefallen