GEOPHYSICS, VOL. 50, NO. 6 (JUNE 1985) P. 92&930, 7 FlGS.

Inversion of vertical seismic profiles by iterative modeling

Pierre A. Grivelet*

assumption that preprocessing parameters (i.e., relative ampli-

ABSTRACT tude of trace and wavelet) are known. The necessity to solve for
both wavelet and model adds to the nonuniqueness of the
I present an application of iterative modeling to the
solution, which is in any case inevitable due to the band-limited
inversion of vertical seismic profiles (VSPs). This
nature of the seismic data. One advantage of VSP measure-
method is derived from linear inversion which allows
ments is that they provide a good description of the downgoing
the extraction from VSP data of an impedance profile
wavelet and of its amplitude relative to the upgoing wave field.
as a function of time and thus permits the prediction of
This feature makes the inverse problem for VSP data more
impedance ahead of the drill bit. There are two steps in
tractable.
this process: first, detection of the major events on the
The first part of the paper briefly describes the model used for
seismogram which is achieved by a recursive detection
the inversion-plane waves traveling at normal incidence in a
algorithm; and second, an optimal estimate of the im-
horizontally stratified medium. The seismogram is parame-
pedances carried out by a gradient algorithm. Seismic
terized by a sparse spike train computed by a recursive detec-
data are band-limited, and consequently the solution of
tion algorithm. The inversion problem is then stated as a least-
the inversion is nonunique. This nonuniqueness is han-
squares problem, solved by a fast and simple gradient algo-
dled by assuming a piecewise-constant or blocked im-
rithm. The final section consists of both synthetic and real
pedance model and by adding a priori constraints.
examples. The synthetic examples emphasize some aspects of
Some synthetic examples are used to illustrate the
the method and the actual example proves its applicability.
method, and a field example shows a comparison be-
tween an impedance profile extracted from VSP data
with this inversion method and an impedance profile FORWARD MODEL
from well logging data. In this example the accurate
prediction of impedance values illustrates the usefulness The iterative inversion of VSPs implies calculation of many
of the method. synthetic seismograms. The forward model chosen here is the
simplest one, consisting of horizontal plane layers excited by a
compressional plane wave at normal incidence, The inversion is
therefore limited to cases where the subsurface approximately
INTRODUCHON
matches such a model.
The goal of seismic inversion is to determine the physical Consider a horizontally layered medium with plane com-
parameters of the subsurface from seismic data. This paper pressional waves traveling at normal incidence. The dis-
describes the application of iterative modeling to the inversion cretization of the subsurface in layers of equal traveltime (Gou-
of vertical seismic profiles (VSPs). VSP data are recordings of pillaud medium) allows the representation of the subsurface as
the seismic wave field at different depths in a well, after a source a time series of impedances or reflection coefficients: The reflec-
has been set off at the surface. Detailed discussions of VSP tion coefficients R, of an interface and the impedances Ai of
recording and processing are given in Gal’perin (1974), Balch et these elementary layers are linked by
al. (1982), and Hardage (1983).
Ai+l - Ai
The method presented is based on linearized inversion (Wig- Ri = (1)
Ai+l + Ai’
gins, 1972; Aki and Richards, 1980), which has been used in
various problems such as earthquake data analysis and seismic Equation (1) shows that, given the impedance of one layer (for
statics (Wiggins et al. 1976). Its basic principle is to find the absolute scaling), the parameters to be sought in the inversion
parameters giving the synthetic response which best fit the can be either the impedance profile or the reflection coefficient
actual seismogram, which here is the VSP upgoing wave. sequence.
The application of this method to surface seismic traces, as At each elementary interface, the upward u and downward d
described by Cooke and Schneider (1983), simultaneously de- propagating waves in the adjacent layers are related by (Figure
termines both an impedance profile and a wavelet, under the la)

Manuscriptreceivedby the Editor January10, 1984;revisedmanuscriptreceivedOctober 23, 1984.

*Schlumberger,12,PlacedesEtats-Unis,92541Montrouge, France.
0 1985Societyof [Link] rightsreserved.

924
 Inversion of VSP Profiles 925

ui = -Rid, + (1 + R,)u,+,, Despite these drawbacks, the recursion method is a fast way
to obtain a rough estimate of the impedance. Then, to get more
and (2)
out of the data, which is reasonable with a VSP where signal-
di+i = (1 - R,)d, + Riui+ Ir to-noise ratio is generally good, use an iterative method.

where Ri is the reflection coefficient at the interface. This per-

mits computation by recursion of the impulse response of the DESCRIPTION OF THE METHOD
reflection series Ri, where the impulse response is simply the
Given a model reflection coefficient sequence R, the corre-
discrete sequence of upcoming waves ui (the fact that most of
sponding impulse response I can be computed and convolved
the elementary reflections are null can be used to reduce the
with the VSP downgoing wavelet W to produce a synthetic
computing time). After the impulse response I is computed, it is
seismogram s If R is the true reflection coefficient sequence of
convolved with a wavelet W to produce the synthetic seismo-
the Earth, the synthetic seismogram $ will be equal to the
gram.
actual seismogram S, within the limits of the noise. Of course,
agreement between S and s does not identify the reflection
VSP DATA sequence R uniquely; because the data are band-limited, there
are many different models giving the same synthetic seismo-
Given a VSP after separation of downgoing and upgoing gram. However, since no other way of determining the close-
waves (Seeman and Horowitz, 1983), followed by application of ness of the modeled reflection coefficient sequence to the true
a spherical spreading correction, then at a given recording level reflection coefficient sequence is available, the error power be-
the well-known convolutional relation between the observed tween S and sis adopted as the criterion for the closenessof R
downgoing D and upgoing U wave fields is to the Earth’s reflection coefficient sequence.
U=D*I, To be a solution, R must clearly minimize the error power
between S and s, but other a priori information must also be
where I is the impulse response of the stack of levels below the invoked to reduce the degeneracy in the set of model solutions
considered depth (Figure la). The downgoing wave field D and provide a geologically acceptable result. The problem is
contains the direct pulse and the associated downgoing multi- then fundamentally stated as: “Find the time series R such that
ples generated above the recording level (I assume here that J(R) = 11S - s[[’ will be minimum under some constraints on
downgoing multiples generated by reflections below the record- R.”
ing level followed by reflections above the recording level can Unfortunately, this is a nonlinear problem for which no
be neglected). simple, direct solution is available; therefore, a linearized inver-
This relation is valid for both raw and filtered VSP data, sion must be sought. In the linearized method, this nonlinear
provided the same filters are applied to the downgoing and problem is split into a sequence of linear operations preceded
upgoing waves. Generally inversion is performed after a wave- by a search for an initial solution (Figure lb). Overall, the
shaping filter has transformed the downgoing wave in a zero- method can be summarized as follows: (1) find an initial solu-
phase wavelet. In the following, I refer to the wave-shaped D as tion R; (2) compute the corresponding synthetic seismogram
simply the “downgoing wavelet” W and to the wave-shaped U and compare it to the actual seismogram; and (3) if the desired
as the “actual seismogram” S. fit is not obtained, update R by any convenient algorithm and
The actual seismogram S is therefore not recorded at the return to step (2)
surface but is the estimated upgoing wave field at a given depth,
after some filtering. In other words, I give the equivalent of the
convolutional model S(t) = W(t) * I(t), where the observer refer-
ence is at the recording depth and both S and W are known.
Another important point is that only the interbed multiples
created below the recording depth are called “multiples.” Sur-
face multiples, such as water-bottom multiples, are a part of the
downgoing wavelet.
When cast in the described form, one-dimensional VSP in-
version is the same problem as zero-offset [i.e., common-depth-
point (CDP) stack] surface seismic inversion, and consequently
any standard inversion procedure can be used to solve for the
impedances (e.g., Cooke and Schneider, 1983 ; Bamberger et al.,
1982). For example, the most popular method is a downward
recursion which assumes I is close to the reflection series R
(Becquey et al., 1979). After an estimation of I by deconvolu-
tion, the impedance series A is computed recursively using
equation (I). This method, however, has a number of draw-
backs. First, the resulting impedance is as band-limited as the
input seismogram; second, error is accumulated through the
recursion; and, third, as soon as two large reflection coefficients
FIG. la. Schematic view of VSP recording. The downgoing
generate strong interbed multiples, identification of the impulse pulse and its multiples are reflected by the layers to be inverted.
response 1 with the reflection coefficient sequence R is no The Goupillaud equation (upper right) allows computation of
longer valid. the synthetic response.
 Grivelet

PARAMETERIZATION

The parameters to be inverted are the impedances of the

elementary layers below the recording depth (Figure la), where
the elementary traveltime is usually between 1 and 4 ms (two-
way time). It is not necessary, however, to find a solution for
each elementary layer since this sampling is usually beyond the
resolution of the data. For example, the inversion of 1 s of data
I FiEFLECTNrPl
ESTMATION I would require optimizing the error power J over 250 variables,
which could be both costly and hazardous. It is more realistic
to assume a blocky impedance profile with only a small set of
nonzero reflection coefficients. The assumption here is that the
Earth consists of only a small number of distinct homogeneous
layers in which the impedance is constant. This is a rough but
useful parameterization and is equivalent to considering only
the major events in the seismogram.
An important consequence of this parameterization is that
estimating the major events correctly gives some information
on the low-frequency component of the impedance profile.
Indeed, even if the wavelet is band-limited, the amplitude of an
isolated reflection is representative of the impedance contrast
across some interface. Therefore, the first step of the inversion
procedure is to locate the major reflection events in the upgoing
VSP wave field and to give an initial value to the corresponding
reflection coefficients. This is performed automatically using
the following recursive detection algorithm.
Given the wavelet W (the downgoing wave), the initial seis-
I -DANCE
profile I mogram S, (the upgoing wave), and the initial estimate of the
impulse response I, with I,(t) = 0 for all t, then (a) find the time
FIG. lb. Flow chart of iterative inversion. After the detection delay r of maximum correlation between S and W, i.e., r such
step (reflectivity estimation), the reflection series is updated to that
achieve a best fit between actual and synthetic seismograms.
S,(t + r)W(t) dt
IS
is maximum, and (b) set I,, 1 = I, + 6, where 6 is a spike at
time ‘c, of amplitude

j S,(t + r)W(t) dt
s W2dt ’

and (c) set S,, 1 = S, - 6 * W and return to step (a). When the
residual error 11S, - I, * W I( is low enough or when the desired
number of events has been found, the process is stopped.
The detection algorithm estimates the impulse response, not
the reflection coefficient series directly. Using its output for a
recursive computation of the impedance would give unrealistic
results due to interbed multiples (e.g., Figure 3) and to the fact
that the algorithm only finds a sequence of locally optimal
solutions. As shown in Figure lc, however, this sequence I is a
good first guess for the reflection coefficient series, since the
locations in time of the actual reflection series should line up
with detected events in the impulse response. A further search
for a global optimum will greatly improve the solution.

GLOBAL MINIMIZATION

After the detection step, the final inversion involves mini-

mizing the error power between the actual and the synthetic
seismogram. The N detected reflection coefficients are denoted
by C = (ci), i = 1, N. I switch notation R to notation C to
FIG. lc. VSP event detection. The upgoing wave contains both
primary and multiple events which are detected to give the represent the N reflection coefficients sought. The goal, then, is
initial solution for reflection series. to minimize the function
 Inversion of VSP Profiles 927

J(C) = 11s - I * w 112 Nevertheless, the initial reflection series does have nonzero
coefficients at the right times, among some other artificial ones.
over all possible values of ci . Since I is a polynomial of degree
Figure 2c shows, however, that after 12 iterations the exact
N in (ci), J is a squared polynomial of degree 2N, which implies
impedance profile has been restored. This example shows that
that it is a convex function of C. [A function of one variablef(x)
the detection has found a guesswell-suited to the minimization.
is said to be convex if the line segment drawn between two
If events are detected at the correct times, irrespective of their
points on the graph of the function never lies below the graph.]
amplitudes and the existence of some false events, the mini-
The property of convexity ensures that there are no local
mization still tends to converge to the correct solution.
minima in the function J (Polak, 1971).
The synthetic model for the second example (Figure 3a)
Of the available optimization methods, a gradient method is
consists of a single high-impedance bed within a low-
the natural choice for this problem since gradient methods
impedance medium. The impedance contrast (a ratio of 1 to 4)
avoid matrix inversions and the inherent problems of ill-
is exaggerated in order to give clearly visible multiples. In this
conditioning when dealing with large matrices. The algorithm
case, the initial estimate of the impedance (Figure 3b) contains
used here is the conjugate gradient method. Let VJ be the
artificial steps corresponding to the detection of these multiples.
gradient of J, i.e., the vector of partial derivatives ~J/iTc,. Then,
However, since the global minimization accounts for multiples,
given the initial solution Co for the reflection coefficients (the
these steps progressively disappear with successive iterations,
superscript indicates the iteration number) and two auxiliary
until the exact solution is found by the seventh iteration.
vectors Y” and coo, the conjugate gradient method can be
implemented by the following algorithm:
CONSTRAINTS

set Y”+l = -VJ(C”);

set &+l= y+l + 9”; where 8” is the ratio of two The two synthetic examples are textbook ones showing the
basic steps of the method: Since~they usesnoise-freesseismo-
dot products :
grams generated by only a few reflection coefficients, the inver-
t3”=(Yn+1 sion is straightforward and does not need extra information-
- y”)*(y”+l); in effect, the nonuniqueness of the inversion is absent (or very
Y” * Y”

find the scalar h minimizing J(C” + hw”+ )‘; and

compute C”+’ = C” + Ace”+’ (updated reflection
series)and return to step (a).

An interesting property of this algorithm is that, theoreti-

cally, it converges in fewer than N iterations if N is the number
of variables (here, the number of reflection coefficients). The
procedure requires at each step the computation of the partial
derivatives of J versus C and the computation of the synthetic
seismogram for each evaluation of the function J.
For the partial derivatives, it is easy to show that

dJ
-=-2(S-I*W).
hi
Lg*w>,
where al/&, is a time series. In fact, this time seriesis calculated
in the same way as the impulse response I, using the derivative
of the interface equation (2). Computation of the N derivatives
of J thus involves computation of N synthetic seismograms.

SOME SYNTHETIC EXAMPLES

The two noise-free synthetic examples presented illustrate

the two stages of the method: the detection and the mini-
mization. The first example (Figure 2) shows what happens
when the impedance model contains a thin bed, i.e., a bed of
thickness smaller than the apparent wavelet resolution; the
second example (Figure 3) shows how the method handles
strong interbed multiples.
The synthetic model for the first example is shown in Figure
2a, along with the corresponding seismogram. The reflections
from the first two interfaces are well separated and are therefore
easy to locate. The events generated by the low-impedance thin FIG. 2. Synthetic example with a thin bed. (a) An impedance
bed are, however, less clearly resolved. Figure 2b shows the profile and the corresponding seismogram. (b) The initial solu-
tion is far from the exact impedance due to the approximate
result of the detection step on these data. The result is correct detection of the thin bed doublet. (c) The result of twelve
for the two first events, but rather unclear for the thin bed. iterations. The low-impedance thin bed has been restored.
928 Grivelet

low) in these cases. In real examples, of course, the situation is Equality constraints
different. Even a simple representation of the major events on 1
s of real data requires 50 reflection coefficients or more, and the Given a reflection sequence C corresponding to the im-
nonuniqueness of the problem increases considerably. In such pedance profile A, then
cases, iterative inversion without any a priori knowledge of the
low-frequency trend of the impedance profile often gives bad ci = Ai+t - Ai
results. For example, a generally decreasingprofile can be ob-
A t+lAi’ +

tained in an area where the compaction trend implies increasing Alternatively, given C and impedance of the first layer A,, the
impedances. rest of the impedances are obtained by
Nevertheless, it is possible to invert real VSPs if the inversion
is constrained in a geologically realistic manner. These con- i 1fC.
Ai= A,n 2
straints can come from prior knowledge of the subsurface, such j=l 1 - cj’

as a known general compaction trend, or in a given devel-

If A, is a constrained impedance, the relation on the reflection
opment field, an easily locatable seismic marker bed. These
coefficients is as follows
constraints can be incorporated into the inversion by forcing
the impedance profile to go through the desired value (equality p-’
n
1 Ap= constant,
constraint) or to lie inside a given domain around this value i=, 1 -c,=A,
(proximity constraint). The following sections briefly show how
these types of constraints can be incorporated into the iterative Because of this relation, the N variables C(i) are constrained to
inversion method described above. a subspace of dimension (N - 1). The variable C, can be ex-
cluded from the minimization, since it can be expressed in terms
of C, , i = 1 to p - 1, A, and A,, as follows:

It5
1$- 4
(4)
1 - c, p-i 1 + Ci’
time ISI l-I 1 -c;
j=i
0 0.5 1.0

In addition, the iterative inversion must obtain the gradient of

J with respect to the N - 1 free variables. Consider the differ-
ential of J
(GmlCm’)
(h/s) x 1000 IIMPEDANCE
PROFILE I (5)

From equation (4) 1 obtain

SEISMOGRAM /

which, when substituted into equation(S), yields

IPULSE
/ RESPONSE
(6)

Thus, for the constrained gradient,

0 1.0
.1,..

25* dJ SJ 1 - C;
_=-_ (74
SC, (7ci FC, 1 - c;

for i < p and

?J* ?J
_=- (7’4
SC, ac,

for i > p.
The minimization is then performed using the same algo-
rithm with this new value of the gradient vector. This procedure
_-
; ____ _. --of----- A is easily generalized to any number of constraints. Each new
equality constraint reduces the number of free variables,
15 --c- thereby speeding up the convergence and, in principle, reducing
I * f f- the nonuniqueness of the solution.

FIG. 3. Synthetic example with strong interbed multiples. (a) An

impedance profile and the seismogram to be inverted; the Proximity constraints
arrows indicate the interbed multiples. (b) The influence of
those multiples is progressively canceled through iteration. With a proximity constraint, the impedance of the con-
 Inversion of VSP Profiles 929

strained layer is forced to lie between A - 6A and A + 6A. This die (S)
can be done using a penalty function. A new “power” function 1.0 1.5 2.0 2.5
1
is defined by adding to J a term which is minimal when the
impedance matches the desired constraint. The type of penalty
function employed here is generally called “an interior penalty
function.” The interested reader can find a good discussion of
this and other penalty function methods in Polak (1971).
A proximity constraint does not reduce the number of free
variables and consequently does not speed up the convergence.
However, it monitors the natural trend of the data by checking
if the resulting impedance value lies in the upper or lower part
of (A - &A, A + &A).
Finally, note that an equality or proximity constraint affects
the solution globally, i.e., both above and below the con-
strained layer. This is easily seen from equation (3), which
indicates that the partial derivative aJ/dc, depends upon the
complete impulse response I of the current model and therefore
on the values of all the other reflection coefficients (and im-
plicitly, the impedances). This is important, since otherwise the
results of the inversion method would be reliable (in terms of
absolute values of the impedance) only above a constrained
point, and the inversion for impedance below TD would
depend on providing a good guess for the impedance of the
deepest layer.

FIELD EXAMPLE FIG. 4. Field example VSP data; total well depth is around
4 300 m. Inversion of a shallow level of this VSP allows com-
Figures 4 to 7 illustrate the iterative inversion of an actual parison with logging data.
VSP. Here attention is restricted to the impedances above TD,

inversio
level +

2000

I I I

2.0 2.5
time (S)
FIG. 5. VSP upgoing waves with principal-component log analysis. The good match shows that comparison of VSP and logging
data makes sense.
 930 Grivelet

1lN~VER~;tM&l,LT
1

pzKq
SYNTHETIC SEISMOGRAM
FROM LOGS

.-
~-

z :: L I ~1-L I ~__~I

FIG. 6. Field example inversion result; major features of the FIG. 7. Field example inversion without constraints; although
impedance log are reproduced by the inversion of VSP. The the shape of the exact profile is still recognizable, the resulting
two arrows on the inversion result indicate the constraints. The impedance profile shows an anomalous increase between 1.6
result below the last constraint is fairly good. and 1.8 s.

in order to compare the estimated impedance profile with the CONCLUSION

impedance profile obtained from sonic and density logs. Figure
4 shows the raw VSP data, while Figure 5 shows the upgoing Inversion by iterative modeling is well-adapted to vertical
waves after predictive deconvolution and wavelet shape nor- seismic profiling, provided a good estimate is obtained for the
malization. Also shown in Figure 5 is a facies log obtained by downgoing wavelet (by velocity filtering) and the inversion
extraction of the principal components of a complete set of logs algorithm is suitably constrained. The simple algorithm de-
(Wolff and Pelissier, 1982). The good correlation between the scribed here appears to work well on both synthetic and real
facies log and the upcoming VSP waves indicates that the data; moreover, the computation is very fast. Typically, an
assumption of plane layers is reasonable for this area. inversion of 1 s of data sampled at 1 ms with fifty variables
The trace marked by the arrow in Figure 5 was inverted to takes around 5 minutes of CPU time on a VAX computer.
obtain the impedances below this level. The result is shown in Experience with this algorithm, as checked against subse-
Figure 6, where the two constrained impedances are also quent well-logging control, indicates that when the VSP is good
marked by arrows. The final synthetic seismogram of the inver- (high S/N ratio, continuity of reflectors) and the subsurfack
sion matches the actual upgoing wave in the VSP very well, and matches the hypothesis of small dips and a (near) zero-offset
the final impedance profile looks like a blocky version of the log source, then the prediction of absolute values of impedance
impedance. The fit is particularly good between 1.7 and 1.9 s below TD is possible.
and between 2.0 and 2.2 s. Note also that when the synthetic
seismogram computed from the log information fits well with
REFERENCES
the actual upgoing wave, then the impedance profiles are also
similar. In contrast, when these two seismograms differ, then Aki, K., and Richards,P., 1980,Quantitativeseismology:W. H. Free-
the impedances differ. This suggests that the discrepancies are man and Co.
BambergerA., Chavent,G., Hemon, C. H., and Lailly, P., 1982,Inver-
caused by differences between sonic and VSP measurements sion of normal incidence seismograms: Geophysics, 47,757-770.
(e.g., in the depth of investigation) rather than by the inversion Balch, A. H., Lee, M. W., Miller, J. J., and Taylor, R. T., 1982, The use
method. of vertical seismic profiles in seismic investigations of the earth:
Geophysics, 47,906918.
However, the main point here concerns the constraints. Becquey, M., Lavergne, M., and Willm, C., 1979, Acoustic impedance
Figure 7 shows the result of inversion without constraints. The logs computed from seismictraces: Geophysics, 44, 1485-1501.
impedance shows roughly the same relative changes, but be- Cooke, D. A., and Schneider, W. A., 1983, Generalized linear inversion
of reflection seismicdata: Geophysics, 48,665-676.
tween 1.6 and 1.8 s there is an increasing trend which leads to a Gal’perin, E. I., 1974, Vertical seismic profiling: Sot. of Expl. Geophys.
very high-impedance layer around 1.75 s. The shape of the Hardage, B. A., 1983, Vertical seismic profiling: Geophysical Press.
Polak, E., 1971, Computational methods in optimization: Academic
curve after 2.0 s is still the same, but the impedances are lower Press, Inc., 44-66.
thari the exact values. The two constraints are then located near Seeman, B., and Horowitz, L., 1983, Vertical seismic profiling: separa-
strong reflections, at 1.73 and 1.89 s (Figures 5-6) which are tion of upgoing and downgoing acoustic waves in a stratified
medium: Geophysics, 48,555-568.
supposed to be the well-known seismic markers. The immediate Wiggins, R., 1972, The general linear inverse problem: implications of
effect of the constraints (Figure 6) is to suppress the anomalous surface waves and free oscillations for earth structure: Rev. of Geo-
increase of impedance between 1.6 and 1.8 s, but the resulting phys. and Space Phys., 10,251-285.
Wiggins, R. A., Larner, K. L., and Wisecup, R. D., 1976, Residual
impedance below these constraints is also better: the absolute statics analysis as a general linear inverse problem: Geophysics, 41,
value of impedance after 2.0 s is nearly exact. This shows that 922-938.
with a few added data (the constraints), the gain of information Wolff, M., and Pelissier-combescure, J., 1982, Faciolog: automatic
electrofacies determination: Presented at the Sot. Prof. Well Log
is large. Analysts, 23rd Annual Logging Symp., Corpus Christi, paper-FF.