You are on page 1of 24

Geophysical Prospecting 38,287-310, 1990

INVERSE THEORY APPLIED TO


MULTI-SOURCE CROSS-HOLE TOMOGRAPHY.
PART 1: ACOUSTIC WAVE-EQUATION METHOD'

R. G E R H A R D P R A T T and M. H. W O R T H I N G T O N 2

ABSTRACT
PRATT,R.G. and WORTHINGTON, M.H. 1990. Inverse theory applied to multi-source cross-
hole tomography. Part 1 : Acoustic wave-equation method. Geophysical Prospecting 38, 287-
310.
Frequency-domain methods are well suited to the imaging of wide-aperture cross-hole
data. However, although the combination of the frequency domain with the wavenumber
domain has facilitated the development of rapid algorithms, such as diffraction tomography,
this has also required linearization with respect to homogeneous reference media. This
restriction, and association restrictions on source-receiver geometries, are overcome by
applying inverse techniques that operate in the frequency-space domain.
In order to incorporate the rigorous modelling technique of finite differences into the
inverse procedure a nonlinear approach is used. To reduce computational costs the method
of finite differences is applied directly to the frequency-domain wave equation. The use of
high speed, high capacity vector computers allow the resultant finite-difference equations to
be factored in-place. In this way wavefields can be computed for additional source positions
at minimal extra cost, allowing inversions to be generated using data from a very large
number of source positions.
Synthetic studies show that where weak scatter approximations are valid, diffraction
tomography performs slightly better than a single iteration of non-linear inversion. However,
if the background velocities increase systematically with depth, diffraction tomography is
ineffective whereas non-linear inversion yields useful images from one frequency component
of the data after a single iteration. Further synthetic studies indicate the efficacy of
the method in the time-lapse monitoring of injection fluids in tertiary hydrocarbon recovery
projects.

INTRODUCTION
This paper studies wave-theoretical methods for cross-hole imaging. Ray-theoretical
traveltime tomography is an established imaging technique in cross-hole seismic
applications (La Porte et al. 1973; Dines and Lytle 1979; Wong, Hurley and West

Received March 1989, revision accepted July 1989.


' Department of Geology, Imperial College of Science, Technology and Medicine, Prince
Consort Road, London SW7 2BP, U.K.

287
288 R . GERHARD PRATT A N D M . H . WORTHINGTON

1983; Worthington 1984; Dyer and Worthington 1988). The use of wave-theoretical
methods has the potential of refining the images obtained from traveltime tom-
ography by using the wave equation rather than its high-frequency ray approx-
imation and by using the full information content of the recorded wavefield.
An important aspect of cross-hole surveys is that a large angular coverage leads
to a redundancy of information. Angular coverage, or survey aperture, is controlled
by the receiver and source array lengths, in turn related to the multiplicity of
source and receiver locations. Where the aperture is large, weak scatter approx-
imations can be used to show that a single frequency component of the data con-
tains all of the information present in any lower-frequency data component. This
can be inferred by an examination of Fig. 1, in which the wavenumber (' K space ')
coverages of the survey region for wide aperture cross-hole surveys is reproduced.
These coverages were first shown by Devaney (1984), who derived them from the
generalized projection slice theorem (Wolf 1969).
Cross-hole surveys are often designed to provide a wide-aperture data set in
order to optimize the performance of traveltime tomography. In practical situations,
limitations on the number of source and receiver locations may restrict the K space
coverages of Fig. 1. In these cases, in cases where velocity structure creates strong
focusing, and in cases where the signal-to-noise ratios are low, the use of additional
frequency components will supply additional information. However, provided the
aperture of the survey is large, a frequency-domain approach applied at several
distinct frequencies is still likely to be the optimal method for extracting this addi-
tional information. This is confirmed by the work of Lo et al. (1988) and Pratt and

(a) (b)
FIG.1. Spatial wavenumber coverages for two single-frequency wide-aperture cross-hole
seismic surveys (from Devaney 1984). The high-frequency data set (a) contains all the infor-
mation present in the low-frequncy data set (b).
INVERSE THEORY 1 289

Worthington (1988), who obtained useful images using diffraction tomography and
single-frequency data components of scale model cross-hole data.
Geophysical diffraction tomography is a wave-theoretical technique that has
received substantial interest recently (Devaney 1984; Witten and Long 1986; Wu
and Toksoz 1987; Lo et al. 1988; Pratt and Worthington 1988). Diffraction tom-
ography was originally formulated in the temporal-spatial Fourier-transform
domain (Mueller, Kaveh and Wade 1979). This formalism facilitated the develop-
ment of extremely rapid inversion algorithms (Devaney 1982; Pan and Kak 1983),
and provided a useful method for studying image resolution (Wu and Toksoz 1987).
The shortcomings of this approach are that it requires regular, Nyquist-limited
spatial sampling of the data along line arrays, and, more seriously, that it leads to
linearizations (under either the first Born approximation or Rytov’s approximation)
with respect to a homogeneous reference medium (Slaney, Kak and Larsen 1984).
It is conceivable that enough a priori information may be available in geophys-
ical cross-hole applications (from well logs, surface seismics, traveltime tomography,
etc.) to justify linearization of the waveform inversion problem. However it is
unlikely that linearization with respect to a homogeneous model will be valid except
in very specific applications.
Linearized inverse procedures based on asymptotic ray theory have been devel-
oped by many authors (Cohen and Bleistein 1979; Stolt and Weglein 1985; Taran-
tola 1984a; Miller, Oristaglio and Beylkin 1987), and some of these are sufficiently
general to be applied to cross-hole data. Here, a non-linear approach similar to that
described in Tarantola (1984b) and Mora (1987) is used, primarily because this
allows the inversion to be based on the rigorous wave-modelling technique of finite
differences (Kelly, Treitel and Alford 1976). The redundancy of the information in
multi-source cross-hole surveys is exploited by implementing non-linear inversion in
the frequency domain. This leads to additional computational benefits, since the
finite-difference solution to the frequency-domain wave equation is particularly suit-
able where the number of distinct source positions is large.
This paper presents a new 2D frequency-space domain method that can be
applied in media of arbitrary a priori complexity, and we compare images from
diffraction tomography to the images obtained from the proposed method under
different synthetic geological conditions. The method is applied here to the acoustic
wave equation in the interests of clarity and to allow direct comparisons with dif-
fraction tomography. The generalization to the elastic wave equation is described in
a companion paper (Pratt 1990). The application of the technique to scale-model
data and a discussion of the use of multi-frequency data is contained in Pratt and
Goulty (1989).

FREQUENCY-DOMAIN I N V E R S ETHEORY
NON-LINEAR
Tarantola (1984b) has derived a general method for acoustic wave-equation inver-
sion. Here we review briefly these results and transform them into the frequency
domain. The reader is referred to the original paper for a full discussion. The size of
the seismic inverse problem in arbitrary 2D media precludes direct matrix-inversion
290 R. G E R H A R D P R A T T A N D M . H . W O R T H I N G T O N

methods. Instead one can proceed by computing the gradient direction (steepest
descent with respect to model parameters) of a representative objective function,
normally an l2 norm of the data residuals. The gradient direction is then used in
iterative optimization methods (see Tarantola 1987 for a review) to minimize the
objective function.
We use as the objective function the l2 norm
S(m) = Ad*C;'Ad + Am*C;'Am, (1)
where d is the data space (in this case pressure field measurements), Ad are the data
residuals defined by Ad = do - d(m) with do the observations and d(m) the estimated
data for the current model m.The model differences, Am are defined by Am = m,
- m, where m, are the a priori model parameters. The data residuals and the model
differences are weighted by the inverses of C, and C,, the data and model space
covariance matrices.
The gradient of S(m) with respect to the model space m is

as- - D*C;'Ad
g=jj& + C;'Am,
where D is the derivative of d at the point m, or the Fr6chet derivative matrix, and *
represents the Hermitian conjugate. The gradient g can be expressed as

g= ($)+ C;' Am, (3)

where 6 2 and Si, are vectors in the 'Dual spaces' (Tarantola 1987) of the acoustic
parameters, and bulk modulus K, and bulk density p. In Appendix A we show how
these can be computed from each Fourier component of the data residuals as

1 (4)
= - 1 V[dP*(r,)G(r, r,)] . VP(r) .
P"r) s g
This is the frequency-domain expression of (A7) in Tarantola (1984b). In (4)
SP(rg) = C;'(P0(rg) - &J) are the weighted data residuals, Po(rg) is the observed
pressure field and P(r) is the forward-propagated pressure field in the current model
estimates. Pressure fields P(r) and P,(r) are the temporal Fourier components of the
time domain fields, at the circular frequency o.The frequency-domain Green's func-
tion of the acoustic wave equation G(r, rg) propagates the data residuals from the
receiver locations rg into the model space. Before propagation the complex conju-
gates of the residuals are taken; this is the frequency-domain expression of time
reversal and therefore the operation contained within the square brackets in (4) is
referred to as backpropagation.
At each iteration of a frequency-domain non-linear inversion, the gradient direc-
tion is thus calculated by taking a product of backpropagated residuals with the
forward-propagated source terms. Before multiplication, an ' adjoint ' operator is
INVERSE THEORY 1 29 1

applied to the two wavefields. The adjoint operator is parameter-dependent ; for the
density subspace the operator is simply a multiplication by io, whereas for the bulk
modulus subspace the operator is a spatial gradient. The adjoint operators serve to
resolve the computations into images of the model parameters.
In a fully iterative implementation a step length is computed that minimizes the
objective function when the parameters are varied along the gradient direction from
(4).In this way a new model estimate is generated and used in the next iteration. In
this paper we restrict the calculations to a single iteration only, and apply the
method only to examples where linearization is justified. We refer to this as imaging,
rather than inversion.

Preconditioning
Images computed using (4)invariably yield a false image of the source positions.
This is due to the multiplication by the predicted field P(r) which contains a singu-
larity at the source position. This is representative of a more general problem in full
wave inversion caused by non-uniform energy coverage of the model. This is partic-
ularly severe in cross-hole surveys, due to the source-receiver geometries and due to
the structural focusing of wave modes that often occurs. These problems could be
reduced by fully iterative implementations of non-linear inversion, but at consider-
able computational cost. Instead, preconditioning can be applied to compensate for
uneven energy distribution in the medium and to speed convergence (Mora 1987
gives a discussion of the importance of preconditioning in in seismic inversion).
To compensate for uneven energy coverage, we have devised a spatially variable
preconditioning procedure. The scheme weights each parameter estimate according
to the inverse amplitude of the source Green’s function at that point. Since the
principle of reciprocity implies that we should treat the receiver locations in the
same way, we also weight the partial images by the inverse amplitudes of the appro-
priate receiver Green’s functions. Thus we modify the gradient direction from (4)
using

where E is a damping term used to stabilize the estimates.


There is a computational cost in using ( 5 ) instead of (4), since the receiver
Green’s functions are not normally explicitly calculated (the summation over recei-
ver positions can be applied before backpropagation). However, in multi-source
problems this is not a severe penalty. Rather than doing a forward propagation and
a backpropagation for each source position, the same wavefields can be computed
by calculating all source and receiver Green’s functions by forward propagation,
and then applying appropriate source amplitudes (or conjugate data residualsj. For
a survey with the same number of sources and receivers, the same number of propa-
292 R. GERHARD PRATT AND M. H. WORTHINGTON

gations is involved, although there are additional overheads in computing the esti-
mates before summation. In all examples that follow spatially variable
preconditioning is applied.

Forward modelling
The generation of synthetic data from a given set of model parameters is an
essential requirement of all inverse methods. This implies an ability to compute the
Green’s functions (impulse response) of the wave equation. The authors of diffrac-
tion tomography were able to use the analytical free-space Green’s functions of the
acoustic wave equation to formulate inversion formulas. In the development of Born
inversion (Cohen and Bleistein 1979) and of the inverse generalized Radon trans-
form method of Miller et al. (1987), these authors used first-order asymptotic
assumptions of geometrical optics, in which the phase and amplitude terms of the
Green’s functions were determined numerically by ray-tracing methods.
In this paper we model the frequency-domain wave equation by the method of
finite differences. In this way arbitrary models of 2D media can be used, and all
wave types are generated (and used in the inversion), including interface waves,
guided waves, diffracted waves and multiply reflected waves. We have already
shown in the introduction that a few frequency components suffice to image wide-
aperture cross-hole data. For this reason we have developed a 2D wave propagator
that solves the frequency-domain wave equation, rather than the time-domain
equation.
The frequency-domain wave equation is a boundary value problem, in which the
wave equation and all its boundary conditions must be satisfied simultaneously. In
Appendix B we develop the finite difference equations for the frequency-domain
inhomogeneous acoustic wave equation on a 2D grid. This development leads to a
large system of coupled linear equations which can be expressed by the matrix
equation

Fp= -S.

In (6) p is a vector whose elements are single Fourier components of the pressure
field everywhere on the grid, s are the Fourier components of the source functions
and F is a large, sparse matrix whose elements depend only on the frequency and on
the properties of the media (the elements of F, p and s are given explicitly in Appen-
dix B). The solution of this equation solves the forward-modelling problem.
Due to the absorbing boundary conditions used, the matrix F is complex valued
and non-symmetric. The solution of linear systems involving matrices such as F is a
highly specialized topic in numerical methods (Concus, Golub and O’Leary 1976;
Vorst 1981). In order to preserve the sparsity of F only a partial factorization was
feasable. Matrix ‘fill-in ’ of zero-valued elements is usually carefully avoided.
With the advent of large vector computers the problem of matrix fill-in is miti-
gated by the efficiency with which the rows and columns of such matrices can be
INVERSE THEORY 1 293

manipulated. It is now feasible on certain machines to factor the band-limited


matrix F for realistic model sizes into band-limited upper and lower triangular
matrices (LU decomposition). Matrix decomposition is achieved by evaluating a
large number of scalar products of vectors, a process that can be efficiently carried
out in pipelined computer architecture. Moreover, many of the computations are
decoupled, allowing the algorithm to take full advantage of multi-processor parallel-
ism. Solution vectors are then found using the stored factors by the relatively cheap
processes of forward reduction and back substitution.

Multi-source forward modelling and inversion


The procedure described above is not cost-effective for generating one frequency
component of the solution for a single source position. Since F (and its factors)
depend only on the frequency and on the media properties, the real advantage of the
approach is that solutions can be generated for additional source positions at
minimal extra cost by forward reduction and back substitution. Numerical tests we
have carried out for a wide range of model sizes indicate that forward reduction and
back substitution can be achieved in less than 5% of the time it takes to initially
factor the matrix. Factoring times are extremely hardware-dependent (25 min for a
model with 60000 grid points on an F P S - ~ ~ ~ / M Ascientific
X computer equipped with
two matrix accelerator boards and 3 megawords of central memory).
Since backpropagation involves using the conjugates of the data residuals as
source functions (see (4)), this can also rapidly be computed using the factored F
matrix. This also facilitates the preconditioning scheme described above, where the
amplitudes of the source and receiver Green’s functions are required. For these
reasons this modelling procedure is ideally suited to be applied to the multi-source
(wide-aperture) cross-hole inverse problem.

NON-LINEAR
I N V E R S I O Nv. D I F F R A C T ITOMOGRAPHY
ON
Figure 2 schematically illustrates the two inverse procedures of diffraction tom-
ography and non-linear inversion, and shows their basic similarities. The essential
difference is that in the non-linear approach the wave propagation steps are com-
puted numerically, allowing arbitrarily complex models to be used and enabling an
iterative technique to be applied. This overcomes the restrictive assumption required
in diffraction tomography of a homogeneous background medium. However, there
is a price to be paid: in the non-linear approach iteration is necessary even when the
problem is linear (i.e. when weak scatter approximations are justified).
Multiple iterations are required when using the non-linear formulation even in
linear problems because the gradient direction of the objective function (2) does not
provide sufficient information. For linear problems the 1, norm of (1) is a quadratic
function of the model parameters. To minimize a quadratic function second deriv-
atives (the elements of the ‘inverse’ Hessian matrix) are required. In non-linear
inversion the inverse Hessian is effectively applied by using multiple iterations to
294 R. GERHARD PRATT AND M . H. WORTHINGTON

Difference data
I
4.
Difference data -
Convolutional filter
1
Backpropagation filter Numerical backpropagation

I
1 Parameter dependant adjoint operator

Convolution with
1
Zero lag cross-correlation
forward propagated with forward propagated
source terms source terms

1
Step length calculation
and update of model
FIG.2. Diffraction tomography and non-linear inversion algorithms compared. The essential
difference is that in non-linear inversion wavefield propagation can be achieved numerically,
making an iterative solution possible. Due to the absence of the equivalent of a convolutional
filter, in non-linear inversion iteration is necessary even in linear problems.

reduce the objective function to a minimum. In diffraction tomography the convolu-


tional filter modifies the data in such a way as to reconstruct a band-limited version
of the scattering object exactly (Wu and Toksoz 1987). Thus the inverse Hessian is
implicit in the algorithm.
The differences between the two approaches to cross-hole imaging are illustrated
by the following synthetic study. Figure 3 shows the model used, which consisted of
a number of point scatterers located one half wavelength apart. The scatterers were
weak ( 5 % ) perturbations to the density (distributed in the shape of the Greek letter
p ) and to the acoustic velocity (distributed in the shape of the letter c). Two refer-
ence media were used, the first was a homogeneous velocity field of 2000 m/s and
the second consisted of a 30% increase in velocity with depth from 1700 m/s at the
top of the model to 2300 m/s at the bottom. A simulation of a 200 Hz cross-hole
seismic survey was carried out using 62 sources located in a vertical well at the
left-hand edge of the model and 62 receivers 188 m away in a similar well at the
right-hand edge of the model. The spatial sample interval of the source and receiver
array elements was 5 m, or one half wavelength at 2000 m/s.
Figures 4 and 5 are intended to illustrate the algorithmic implementation of
frequency-domain non-linear inversion and its application to the model shown in
Fig. 3. In Fig. 4a we show the frequency-domain wavefield (the solution to (6)) in the
INVERSE THEORY 1 295

FIG.3. Model used to demonstrate multi-source frequency-domain non-linear inversion and


to compare the algorithm with diffraction tomography. The scatterers making up the Greek
letter p consist of 5% perturbations to the bulk density, the scatterers making up the letter c
are 5% perturbations to the acoustic velocity. The background velocities used were (i) a
homogeneous field (2000 m/s) and (ii) a linear velocity gradient from 1700 m/s at the top to
2300 m/s at the bottom of the model. The model is 188 m wide and 326 m deep.

homogeneous reference medium for a single-source position. Figure 4b is the pertur-


bation to the solution, or the difference wavefield, introduced by the presence of the
point scatterers. The point scatterers act as virtual sources for the difference wave-
field. The difference field is shown for the purposes of illustration. It can only be
computed because of our advance knowledge of the ' unknown perturbations '.
In a real experiment all we can measure is the total field at the receiver locations.
We assume we can extract the difference field (i.e. the data residuals) from the total
field, although we emphasize that this computation is a non-trivial problem in cross-
hole seismic data processing. The data residuals are then backpropagated from the
receiver locations into the current model, generating the terms in the square
296 R. G E R H A R D PRATT A N D M . H . W O R T H I N G T O N

(Cl) (b) (C)


FIG.4. Monofrequency wavefields at 200 Hz relating to the model in Fig. 3, using a homoge-
neous background velocity field. Panel a is the forward propagated wavefield in the reference
medium using a single source, panel b is the difference wavefield due to the presence of the
perturbation shown in Fig. 3 and panel c is the backpropagation of the difference field from
the receiver locations. These wavefields are the real parts of the complex-valued solutions to
the frequency-domainacoustic wave equation.

brackets in (4). Figure 4c shows the result of this calculation. The backpropagated
residuals come into focus in the region of the anomalies and diverge again on the
other side.
The next step is to apply the adjoint operators of (4) to both the forward propa-
gated source terms (Fig. 4a) and the backpropagated residuals (Fig. 4c) and multiply
the results forming density and bulk modulus images. Each source-receiver partial
image is then preconditioned using (5). Since several researchers have pointed out
that a parameterization in terms of the bulk modulus and density is poorly resolved
(Mora 1987; Tarantola 1984c), we use the following derivation to re-evaluate the
gradient direction in terms of the velocity and the density (see Mora 1987, equation
(37)):

(7)
I N V E R S E THEORY 1 297

FIG.5. Partial images from a single-source position at 200 Hz and all 62 receiver positions of
the acoustic velocity (a) and the density (b), after a single iteration of non-linear inversion.

The partial derivatives are easily obtained from

V P
yielding
SE = 2cpsff - *($)sj,
s j c = c2SZ .+ sj. (9)
These two components of the preconditioned gradient direction for the svnthetic
example, evaluated using (4), (5) and (9) for a single-source position, are depicted in
Fig. 5.
The calculation of the gradient direction of the 12-norm is completed by
summing similar partial images from all 62 source positions. These results (Figs 6a
298 R. GERHARD PRATT AND M. H. WORTHINGTON

(a) (b) (C)


FIG.6. Images at 200 Hz formed using all 62 sources and all 62 receivers. Panels a and b are
the velocity and density images after a single iteration of non-linear inversion. Panel c is an
image of the object function formed using diffraction tomography. Synthetic data were gener-
ated using the constant velocity background medium.

and 6b) can be compared with the diffraction tomography images formed using the
same monofrequency difference data (Fig. 6c). The diffraction tomography images
were calculated using the filtered backpropagation algorithms for line sources and
for line receivers derived by Wu and Toksoz (1987, equations (41) and (42)). The
comparison shows that diffraction tomography gives sharper, better defined images
than one iteration of non-linear inversion. As discussed in the previous section this
is because, even in linear problems, non-linear inversion requires multiple iterations
to converge. These images also indicate that the density anomalies are poorly
resolved from the velocity anomalies.
This comparison is very similar to that given by Wu and Toksoz (1987) between
holography and diffraction tomography. Apart from important differences due to
the application of the adjoint operators, the first iteration of frequency-domain non-
linear inversion is identical to holographic imaging.
Figure 7 shows similar comparison, but in this case the reference medium was
the one containing the linear velocity gradient (Fig. 3). The result is a severely
degraded diffraction tomography image, whereas a single iteration of the non-linear
inversion produces well-focused images.
Diffraction tomography requires regular spatial sampling along line arrays, due
to its implementation in the spatial Fourier domain. Furthermore, the sampling
must satisfy the Nyquist criterion (that the source and receiver intervals be less than
INVERSE THEORY 299

c
r
-
c
c
c
c
F
,
c-

(a) (b) (C)


FIG.7. The same images as in Fig. 6, but in this case synthetic data were generated using the
linear velocity gradient as the reference velocity field. Exactly the same difference field was
used to form all three images.

one half the wavelength). If the Nyquist criterion is not satisfied, spatial under-
sampling in diffraction tomography causes high wavenumber data components
(generated by energy travelling nearly vertically) to alias into low-wavenumber data.
These components are incorrectly filtered by the convolutional filter, and also back-
propagate in the wrong directions. Non-linear inversion places no restriction on
source-receiver geometries other than the 2D one.
In Fig. 8 data from the homogeneous reference medium are used again, but here
the images have been formed using the 300 Hz frequency component of the survey.
The non-linear inversion images show a significant improvement in resolution, and
also yield a visible image of the p-shaped anomaly in the density estimate. At 300
Hz the spatial sample interval (5 m) is 0.66 wavelengths, thus violating the Nyquist
criterion. The diffraction tomography image is severely corrupted as a result.

C O 2 FLOOD-FRONT
DETECTION
EXAMPLE
In order to apply the proposed technique to a more realistic example, we studied the
reservoir engineering problem of detecting and mapping the progress of a CO,
injection flood. In the model shown in Fig. 9, based on a real experiment
(Worthington et al. 1989), COz is injected into a permeable dolomitic reservoir,
sandwiched between two impermeable anhydrite sequences, for the purposes of ter-
tiary oil recovery. The progress of the flood front is monitored by conducting a
300 R . GERHARD PRATT AND M. H. WORTHINGTON

(a) (b) (C)


FIG.8. Images as in Fig. 6 (with synthetic data generated using a homogeneous reference
velocity field). In this case data at 300 Hz were used to form the images. The non-linear
inversion images are improved, however, the violation of the spatial Nyquist criteria causes
aliasing and corruption of the diffraction tomography image.

time-lapse (pre-flood and post-flood) cross-hole seismic survey. Since the properties
of injection fluids are often chosen so as to match those of the fluids already in
place, we have assumed that the only change to the media properties is a 2% change
in the compressional velocities. We also assume that an accurate representation of
the wavefield perturbation due to the presence of the injection fluid can be obtained
by subtracting the two data sets, and that an accurate a priori model of the gross
velocity structure has been obtained.
The time-domain common-source gathers in Figs 9b and 9c computed using a
time-domain finite-difference approximation of the acoustic wave equation, show
that for source location within the low-velocity reservoir little change is expected in
the direct wave. Most of the perturbation is contained in the later, high-amplitude,
guided waves. If the information contained in these waveforms is to be exploited,
more sophisticated techniques than traveltime tomography will be required.
The results after one iteration of non-linear inversion for this problem are shown
in Fig. 10. The algorithm has succeeded in providing a useful image of the COz
injection fluid. There are some aberrations that would be resolved in a fully iterative
implementation. Specifically, focusing of acoustic energy due to the wave guide
effect of the low velocity layer has caused a false image of the original layering of the
reference model. The false image of the layering was a more severe problem when no
INVERSE THEORY 1 301

5.5 kmls

5.5 bmls

MODEL COMMON SOURCE GATHER DIFFERENCE FIELD

(a) (b) (C)


FIG.9. Model of a CO, flood-front detection time-lapse experiment. In the central layer (the
reservoir) CO, has been injected from a well at the left-hand edge of the model. A 90 Hz
cross-hole seismic survey is carried out to map the progress of the injection fluids, using 43
sources in the injection well and 43 receivers in the monitoring well 48 m away at the right-
hand edge. The model and the injection fluid perturbation are shown in panel a. Panel b
shows a time-domain common-source gather from a source located within the reservoir, and
panel c shows the perturbation to the wave forms induced by the presence of the injection
fluids.

preconditioning was used (Fig. lob), although the resolution of the image of the
flood front appears to suffer slightly from preconditioning.
The issue of preconditioning is clearly an important one, and future research
may well yield more effectivemethods of preconditioning. Ideally, the images would
be spatially deconvolved with the impulse response of the imaging technique. The
impulse response is spatially variable and this approach is likely to be computa-
tionally expensive.

CONCLUSIONS
Wave-theoretical imaging of acoustic cross-hole data using frequency-domain tech-
niques has been described. In order to provide a suitably general imaging technique
we have developed a new finite-difference modelling method for the frequency-
302 R . GERHARD PRATT AND M . H. WORTHINGTON

(4 ( b)
FIG.10. Result of a single iteration of non-linear inversion using the model in Fig. 9. A single
frequency component (90 Hz) from all 43 sources and all 43 receivers was used. Panel a is the
velocity image with preconditioning, panel b is the same image without preconditioning.

domain 2D wave equation in arbitrary media. The technique has been developed
specifically to facilitate wave-equation imaging of wide-aperture cross-hole data. By
taking full advantage of recent developments in computer hardware, the method can
be used to model and image data from a large number of source positions.
We have concentrated on acoustic wave applications. This has simiplified the
problem allowed us to make direct comparisons with diffraction tomography.
Clearly cross-hole experiments need to be modelled and inverted using the elastic
wave equation. The methods we use can be applied to any 2D equation of motion.
Pratt (1990) applies the technique using the elastic wave equation. In more general
problems transversely isotropic media could also be handled. Fully three-
dimensional implementations await further developments in modelling techniques
and in computer hardware, such as those recently reported by Reshef et al. (1988).

APPENDIXA
This appendix derives the adjoint of the frequency-domain acoustic wave equation.
It follows the derivation in Tarantola (1984b) for the time-domain equation,
although for clarity here we do not consider the source function as an unknown.
INVERSE THEORY 1 303

The linearized forward problem is


6d = Ddm,
where d, m and D have the same meaning as given in the text. The adjoint problem
corresponding to this is
6m = D*6d, (‘42)
where 6m is a vector in the dual of the model space (Tarantola 1987). By substitut-
ing 6d = C;lAd we can use (A2) to calculate the first term in the gradient of the
objective function in (2). In continuous-form equations (Al) and (A2) are

6d(D) = 1M
dM [-]
amw)
6m(M)

and
Gm(M)= dD[-] ad(D) *6d(D).
dm(M)
If we can write down the linearized integral expression for the wave equation
that gives the data perturbations in terms of the model perturbations, as in (A3), and
identify the FrCchet kernel (the term within the square brackets), we can then also
evaluate the adjoint expression (A4). In the following steps this is achieved using the
integral solution to the wave equation in terms of Green’s functions, following lin-
earization by using a Born approximation.
The frequency-domain acoustic wave equation is

where P(r) and S(r) are Fourier components of the pressure field and source func-
tions at the circular frequency o.
At this stage the assumption is often made that the source function can be rep-
resented by a spatial delta function centred at the source location rs, that is that
S(rJ = 6(r - rJS. When this is done, often the implicit dependence of the pressure
field on the source location is made explicit by using P(r) = P(r, rJ. We do not use
this notation, and the reader should bear in mind that there is an implied depen-
dence of P(r) (and later 6P(r)) on the source location.
By perturbing the bulk modulus and the density using
K(r) + K(r) + 6K(r),
P(r) + P(r) + 6P(r),
we can derive a new wave equation for the perturbed wavefield oP(r):

where the ‘secondary sources’ AS(r) are given within the Born approximation by
304 R . GERHARD PRATT A N D M. H . WORTHINGTON

(the higher order terms 0(6F(r), 6p(r))2are dropped under the Born approximation).
The integral solution to (A5) is

P(r) =
s, dr‘G(r, r’)S(r’),

where G(r, r‘) is the frequency-domain Green’s function, or the response to a point
(‘49)

source at r’. Formally G(r, r’) satisfies

G(r, r‘) = -6(r - r’).

Similarly the solution to (A7) is

6P(r) = -
1” dr’G(r, r’)AS(r’)

or, substituting for AS(r’) from (A8),

6P(r) = --w2IVdr‘G(r‘, r) -

where we have invoked the reciprocity of the Green’s function. Equation (A12) can
be simplified using the same steps as in Tarantola (1984b, (A5a), (A5b)) yielding

6P(r) = -w21v drf6K(r‘)[G(r’, r) “1 + s,


K2(r‘)
dr’dp(r’)[ VG(r’, r) -
vp(r’)]
p2(r‘)
(A13)

(the gradient is taken with respect to the first spatial variable of the Green’s
function). In (A13) the terms within the square brackets define the Frkchet kernels in
(A3). The data residuals in (A3) are given by 6P(rg),the value of 6P(r) at the receiver
locations re. Hence the adjoint expressions for each component of 6m = [6R(r),
@(r)] are

(the integral over the data space has been replaced by a summation over source and
receiver locations). The asterisk here indicates complex conjugation. By substituting
the weighted residuals C; ‘(P0(rg)- P(rg)) for 6P(rg) in (A14) and rearranging the
terms as in Tarantola (1984b, (A7))we obtain (4).

APPENDIXB
Here we give the finite-difference scheme used to numerically solve the frequency-
domain acoustic wave equation (A5). In 2D, this equation can be written in terms of
INVERSE THEORY 1 305

the Cartesian coordinates x and z as

The wavefield P(x, z), the density p(x, z), the bulk modulus K(x, z) and the source
terms S(x, z) are discretized on a regular 2D grid. For example, P(x, z) is represented
by the set of discrete values (Pi,j-: i = 1, 2, 3, ..., N,; j = 1, 2, 3, . . ., Nz},where
x = (i - l)Ax, z = (j- 1)Az and Ax, Az are the discretization intervals.
By using second-order finite-difference approximations to the partial derivatives
in (Bl), each Pi,j is coupled to only four of its eight neighbours, P i - l ,j , Pi,j - l ,
Pi+1, and Pi,j + . The differencing schemes used are Fourier-transformed versions
of those given in Kelly et al. (1976). These are best depicted by the finite-difference
stars (e.g. Claerbout 1976)

and

where the averages pi* 1/2 and p j * 1/2 are given by

To approximate each term on the left-hand side of (Bl) the appropriate star is
overlayed on the grid with its centre at (i, j). The values of the star are multiplied by
the values of P at the points over which they lie, and the resulting products are
summed. For example at the point (i,j)

(B5)
Pi- I / Z
306 R. G E R H A R D P R A T T A N D M . H . W O R T H I N G T O N

This procedure is repeated at all grid points, leading a large system of coupled linear
equations.
The full scheme for the interior grid points is given by the star

where the five non-zero coefficients are given by

Boundary conditions
The boundary conditions used at the edge of the grid profoundly affect the solu-
tion. This is particularly true for frequency-domain solutions, since the equations
represent the steady-state, or infinite-time, response of a monofrequency oscillator.
In time-domain applications boundary reflections can be eliminated by time win-
dowing. In the frequency domain this is not possible and the use of effective absorb-
ing boundary conditions is essential.
To absorb outward-propagating waves we use the frequency-domain ' 45" ' par-
axial wave propagator of Kjartanson (1979). For example, at the bottom edge we
use the downgoing wave equation
i a a2 R(x, z) +- a2 a
R(x, z ) + 2im(x, z ) - R(x, z) = 0,
2 4 x , z ) a Z axz axz az2

where R(x, z) = eim(x*


)',' m(x, z ) - w/u(x, z) and u(x, z ) is the local velocity, given
by

Equation (B8) can be approximated using the following finite-difference stars :


I N V E R S E THEORY 1 307

1 - 1 1 2 1 - 1

a
--+-
1
ax2 2Ax2
0 0 0

The six non-zero elements of the finite-difference star in (B4) for the downgoing
wave equation to be used at the bottom row of the grid are therefore
-1 i 1 2
BE. .=---- ~ x 2 E misj’
A x 2 ~ zm i , j
+
I, J

1 i 1 2
DD. .E---- + - mi, j ,
‘,I Ax2Az miVj Ax2 Az ’

1 i 1
A A . . = C C i , j= --
1. I
, , +
Ax2Az 2mi,j 2Ax
-1 i2 1
AD. . = C D . .=--+-
1. J ‘,I Ax2Az 2mi3 2Ax2’
Similar boundary conditions can be formulated for the other three edges by
rotating the finite-difference star appropriately. Corner points are treated using the
rotated A1 conditions of Clayton and Engquist (1977) for the scalar wave equation
which also fit the finite-difference star, (B4).

Matrix expression of the finite-difference equations


The action of the finite-difference star, equation (B4) upon the wavefield at the
point (i, j ) must equal the source term at (i, j). To solve the complete system that
308 R. G E R H A R D P R A T T A N D M . H. W O R T H I N G T O N

FIG.11. The structure of the frequency-domain finite-difference matrix. Each non-zero


element is shown here as a black circle. This structure is known as ‘tri-diagonal with fringes ’.
The matrix contains (N, x N,)* elements of which only 9 x N, x N, elements are non-zero.
The number of elements contained within the non-zero band limits is N, x 2(N: + N,).

approximates the wave equation everywhere on the grid, the linear system of
coupled equations are recast into matrix form. First the 2D indices (i,j ) are mapped
onto a single index k using k = (i - 1)N, + j . The new index k thus runs from 1 to
N, . N,. In this way the 2D arrays Pi, and Si, are replaced with the column
vectors p and s. Explicitly, the components of p are
PT = (P1 9 PZ? P39 ...? PNx.Nx)
-
-(pl,l, pZ,l, p3,19...9 P N x , l ? p 1 , 2 , p 2 , 2 , p3,29*..3 PNx’ N,). 0314)
The finite-difference coefficients represented by (B6)then map onto the finite
difference matrix F schematically shown in Fig. 11. The elements of F depend only
on the media properties and on the frequency. Thus the original boundary value
problem is expressed by
Fp= -S (B15)
to be solved for the pressure field p.
The matrix F is clearly extremely sparse, since each point on the grid is coupled

to only its nearest neighbours. The full matrix contains (N, x N,)’ elements of
which only 9 x N, x N, elements are non-zero. The band-limited matrix contains
+
N , x 2(NZ N,) elements, this is the number of (complex-valued) elements that
INVERSE T H E O R Y 1 309

need to be stored in order t o factor the matrix into upper and lower triangular
matrices as described in the text.

REFERENCES
CLAERBOUT, J.F. 1976. Fundamentals of Geophysical Data Processing. McGraw-Hill Book Co.
CLAYTON, R. and ENGQUIST, B. 1977. Absorbing boundary conditions for acoustic and elastic
wave equations. Bulletin of the Seismological Society of America 67, 1529-1540.
COHEN,J.K. and BLEISTEIN, N. 1979. Velocity inversion procedure for acoustic waves. Geo-
physics 44, 1077-1085.
CONCUS, P., GOLUB,G. and OLEARY,D. 1976. A generalized conjugate gradient method for
the numerical solution of elliptic pde’s. In: Sparse Matrix Computations. Bunch, J. and
Rose, D. (eds.). Academic Press, Inc.
DEVANEY, A.J. 1982. A filtered backpropagation algorithm for diffraction tomography. Ultra-
sonic Imaging 4, 336-350.
DEVANEY, A.J. 1984. Geophysical diffraction tomography. IEEE Transactions on Geoscience
and Remote Sensing GE-22,3-13.
DINES,K.A. and LYTLE,R.J. 1979. Computerised geophysical tomography. Proceedings of the
IEEE 67,471-480.
DYER,B.C. and WORTHINGTON, M.H. 1988. Some sources of distortion in tomographic veloc-
ity images. Geophysical Prospecting 36,209-222.
KELLY,K.R., TREITEL, S. and ALFORD,R.M. 1976. Synthetic seismograms: a finite difference
approach. Geophysics 41,1,2-27.
KJARTANSSON, E. 1979. Constant Q-wave propagation and attenuation. Journal of Geophys-
ical Research 84,47374748.
LAPORTE, M., LAKSHMANAN, J., LAVERGNE, M. and WILLM,C. 1973. Seismic measurements by
transmission - Application to civil engineering. Geophyscial Prospecting 21, 146158.
Lo, T., TOKSOZ,M.N., Xu, S-H. and Wu, R-S. 1988. Ultrasonic laboratory test of geophysical
tomographic reconstruction. Geophysics 53,947-956.
MILLER,D.E., ORISTAGLIO, M. and BEYLKIN, G. 1987. A new slant on seismic imaging: migra-
tion and integral geometry. Geophysics 52,943-964.
MORA,P.R. 1987. Nonlinear two-dimensional elastic inversion of multioffest seismic data.
Geophysics 52, 121 1-1228.
MUELLER, R.K., KAVEH,M. and WADE,G. 1979. Reconstructive tomography and applications
to ultrasonics. Proceeedings of the IEEE 67,567-587.
PAN,S.X. and KAK,A.C. 1983. A computational study of reconstruction algorithms for dif-
fraction tomography : interpolation versus filtered backpropagation. IEEE Transactions
on Acoustics, Speech and Signal Processing ASSP-31, 1262-1275.
PRATT,R.G. and WORTHINGTON, M.H. 1988. The application of diffraction tomography to
cross-hole seismic data. Geophysics 53, 12841294.
PRATT, R.G. 1990. Inverse theory applied to multi-source cross-hole tomogaphy. Part 2:
Elastic wave-equation method. Geophysical Prospecting, submitted.
PRATT,R.G. and GOULTY, N.R. 1989. High resolution tomography using the wave equation:
results with physical data. 59th SEG meeting. Dallas, U.S.A., Expanded abstracts, 70-74.
RESHEF,M., KOSLOFF,D., EDWARDS,M. and HSIUNG,C. 1988. Three-dimensional acoustic
modelling by the Fourier method. Geophysics 53, 1175-1 183.
SLANEY, M., KAK,A.C. and LARSEN, L.E. 1984. Limitations of imaging with first-order diffrac-
tion tomography. IEEE Transactions on Microwave Theory and Technology MIT-32,
860-873.
310 R. G E R H A R D P R A T T A N D M. H. W O R T H I N G T O N

STOLT,R.H. and WEGLEIN, A.B. 1985. Migration and inversion of seismic data. Geophysics 50,
2458-2472.
TARANTOLA, A. 1984a. Linearized inversion of seismic reflection data. Geophysical Prospecting
32,998-1015.
TARANTOLA, A. 1984b. Inversion of seimic reflection data in the acoustic approximation. Geo-
physics 49, 1259-1266.
TARANTOLA, A. 1984c. The seismic reflection inverse problem. In: Inverse Problems of Acous-
tic and Elastic Waves. Santosa, F., Pao, Y.H., Symes, W. and Holland, Ch. (eds.). Society
ofIndustrial and Applied Mathematics, Philadelphia.
TARANTOLA, A. 1987. Inverse Problem Theory: Methods for Data Fitting and Parameter Esti-
mation. Elsevier Science Publishers Co.
VANDER VORST,H. 1981. Iterative solution methods for certain sparse linear systems with a
non-symmetric matrix arising from PDE problems. Journal of Computational Physics 44,
1-19.
WITTEN,A.J. and LONG,E. 1986. Shallow applications of Geophysical Diffraction Tom-
ography. IEEE Transactions on Geoscience and Remote Sensing GE-24,654662.
WOLF,E. 1969. Three-dimensional structure determination of semi-transparent objects from
holographic data. Optical Communications 1, 153-156.
WONG,J., HURLEY,P. and WEST,G. 1983. Crosshole seismology and seismic imaging in
crystalline rocks. Geophysical Research Letters 10,686689.
WORTHINGTON, M.H. 1984. An introduction to geophysical tomography. First Break 2, (I),
20-25.
WORTHINGTON, M.H., EAST,R.J.H., KERNER, C.K. and ODONOVAN, A.R. 1989. Limitations of
ray theoretical tomographic imaging of cross-hole seismic data for monitoring EOR
flooding. Scientific Drilling 1,47-53.
Wu, R. and TOKS~Z, M.N. 1987. Diffraction tomography and multisource holography
applied to seismic imaging. Geophysics 52, 11-25.

You might also like