Professional Documents
Culture Documents
1307-1318, 12 FIGS
Paul Docherty*
1307
1308 Docherty
adjacent receivers. The subsequent system to be solved has to the model parameters p through the nonlinear functional
fewer rows, is more sparse, and is better conditioned than r;,
would be the case were the actual observations used. Anal-
?i =fifi(P)~ i= 1, -** ) MA, (2)
ysis using the singular value decomposition provides insight
into the velocity-depth trade-off issue and motivates a strat- where mA is the total number of refracted arrivals recorded
egy for the inversion. It is shown that second difference for the shot at A. Linearizing about an initial model p. gives,
smoothing is effective at removing unnecessary roughness in vector notation.
from the solution, but that damping, the more common
approach, is not. The extent to which the velocity-depth
t = f. + Jhp, (3)
ambiguity can be resolved is examined using a synthetic data
example. Finally, weathering model and statics are com-
which is the familiar equation of traveltime tomography.
puted for a real data case.
Here, f, = f(po) is a vector of traveltimes through the model
MODEL DESCRIPTION po, J is the Jacobian matrix (dimension mA x n) containing
partial derivatives off, and Ap is a vector of perturbations to
Two layers make up the model: a laterally inhomogen- the model parameters.
eous weathering layer and a uniform, high speed refractor Rays corresponding to consecutive arrivals share a com-
(Figure 1). The weathering layer is divided into cells of mon path for much of their journey, as illustrated in Figure 2
constant velocity. Each cell is bounded above by the obser- for arrivals i - I and i. This feature of two-dimensional
vation surface and below by the refractor. Boundaries be- (2-D) refracted rays is reflected in the structure of the matrix
tween adjacent cells are vertical. The base of weathering is I. If J is ordered such that the ith ray occupies the ith row,
described by a series of node points, joined by straight line then many elements appearing in one row reoccur in identi-
segments. In this study a constant refractor velocity will be cal locations in the next, that is, the row vectors of J contain
assumed. many parallel components. Motivated by this observation, I
The nodes may move up or down to accommodate form the quantities 6ri = ti - ti_t and 6f,(p) = f;(p) -
changesin the weathering thickness but their x-coordinates, f;_,(p), for i = 2, * a. , mA, introduce the vectors St =
once chosen, remain fixed. The locations of the cell bound- (r, > St,, * * * > 6t,,,A)T, and Sf = (f,, Vi, ... , Sf,,,)‘, and
aries within the weathering layer are specified independently construct a new system,
of the node locations. Usually, I position each cell boundary
midway between two adjacent receivers, as in Figure 1, so St = 6fo + $‘Ap, (4)
that the velocity in the neighborhoodof any given receiver is
constant, a feature which can be exploited when tracing the where 6f, = 6f(po). The new Jacobian J_’contains partial
raypaths. (A ray leaving the base of weathering at the critical derivatives of 6f (except for its first row, which contains
angle may be assumed to remain within a single cell and, partial derivatives off,) and has the same dimensions as r.
therefore, to travel along a straight line path on its journey Typically, J’ is an order of magnitude more sparse than d. It
up to a source or receiver. The assumption is valid for a thin is also better conditioned since its rows are less parallel than
weathering layer with a large velocity contrast at its base.) those of J. Conceptually, the system (4) is obtained from
Olsen (1989) used essentially the same parameterization for equation (3) by subtracting rows corresponding to arrivals at
the weathering layer. adjacent receivers; in practice, though, equation (4) is con-
If there are n, node points and n,, weathering velocity structed directly. Figure 3 depicts the travel path required
cells, then the model can be described by a vector p of length
n = n7 + n,, + 1, given by
Ptl = 7JR,
where zi is the depth of the ith node, vj is the velocity of the FIG. 1. Parameterization of the weathering model. Cells of
jth cell, and vR is the refractor velocity. constant velocity divide up the weathering layer. Node
points joined by straight line segments describe the weath-
PROBLEM FORMULATION ering base. A constant refractor velocity is assumed.
for the calculation of Sfi. It is made up of three legs and, as taken as the traveltime pick closest to the beginning of the
indicated, the traveltime along one of the legs is considered line; its variation is not known, and this equation should be
negative. I will refer to this travelpath as a “delta raypath,” weighted conservatively. Letting d’ = 6t - Sf, and introduc-
while recognizing that it is not a physical ray. Bahorich et al. ing W = diag(llo;) to express the uncertainties in the
(1982) used delta raypaths in formulating the ABCD method observations, equation (4) becomes
of refraction statics. Note that the first equation (i = 1) in (4)
represents a true arrival-the refracted arrival closest to the Wd’ = WJ’Ap. (5)
shot.
The experiment described above is not standard practice Surveys are often carried out using the split-spread geom-
in production surveys. Typically, there are many shots, and etry. As such a survey progresses,shots are located on both
the spread length for each shot is less than the total length of sides of any given receiver pair. Shots to the left of the
the line. As illustrated in Figure 4, the pair of receivers that receiver pair can be used to recreate the experiment de-
recorded arrivals i - 1 and i from the shot at A will record scribed above; shots to the right enable the reverse experi-
arrivals from a number of different shots in a production ment to be formulated. In this, a shot is set off near the far
survey. Notice from the figure that the delta ray for each of end of the line (at point B, say) while refractions are
these shots is the same. Likewise, the traveltime St;, com- recorded all the way back down the line to the beginning. A
puted from the data, should also be the same for each shot. delta ray corresponding to this experiment is shown in
Any variation in the value of 6ti from one shot to the next is Figure 5. If there are mB refracted arrivals for Shot B, then
a result either of noise in the data or of picking errors. mB - 1 delta ray equations and one full ray equation (for the
(Variation in 6ti does not depend, for example, on the shot refraction closest to B) can be appended to the system (5).
depths, which may not be known accurately, nor on the Setting m = mA + mg, then (5) is a system of m equations
absolute time of each shot-should one be unintentionally in n unknowns, where m 5 2 x (number of receivers on the
early or late. Similarly, it is independent of whether peaks or line). Had the problem been formulated in terms of the
troughs are chosen for picking the data, as long as one is original data set, the number of equations might have been as
consistent.) great as (number of shots) x (number of traces/shot).
Given a number of independent realizations of the quan- As it stands, the system (5) contains equations corre-
tity 6ti, we can estimate its standard deviation (T;. Choosing sponding to, at most, two full raypaths. In some cases I add
one of the observed values for 6tj (the median, say) and more full rays to the system, say one or two full rays from
using it as the ith datum in the system (4), it is possible to each shot. An example would be when the shots do not
construct the hypothetical experiment described above. In occupy the same locations as the receivers. In this case, rays
addition, Ui can be used to weight the ith equation. Thus, the leaving the shot locations traverse parts of the model not
multiplicity of arrivals at each pair of adjacent receivers in penetrated by any ray emerging at a receiver. As with I 1
the production data set is replaced by a single traveltime and above, the difficulty with using full raypaths such as these is
a weighting factor. The first observation t , in equation (4) is that the uncertainties in the observed traveltimes are not
known. Care should be taken to ensure that the picks used
are accurate.
Because of the different types of parameters involved
i-l i (depths and velocities) it is advisable to apply some form of
A A column scaling before attempting a’ solution of equation (5).
Following Wiggins (1972), the condition of the system can be
improved by introducing a new set of variables x given by
x = Shp, (6)
i-1 i
A A
A +
FIG. 4. Arrivals from a number of different shots. The delta FIG. 5. The delta raypath corresponding to a shot at point B,
raypath is the same for all shots. which is located at the opposite end of the line to A.
1310 Docherty
--*......,.z.,l Knee
=-....,. Singular
s...
k..._ t Values
-...............“,“.“..
;lmia = 1.2X1 Oe5
I ...‘.“........l.,,ll._
Node
Depths
Model
Eigenvectors
Weathering
Velocities
_ Refractor
Velocity
FIG. 6. Singular values and associated model eigenvectors from a small synthetic example. Each eigenvector contains node
depth and weathering velocity components (separatedby gap in the plot), and a single refractor velocity component. To the left
of the knee, depth and weathering velocity components appear in-phase. To the right of the knee, depth-and weathering velocity
components are mostly out-of-phase, indicating velocity-depth trade-off. Trade-off is most apparent for the long wavelengths,
at the small singular value end of the spectrum.
2-D Refraction Tomography 1311
(4 (b)
FIG. 7. Selected eigenvectors from Figure 6 with the weathering velocity components plotted on top of the depth components.
The component corresponding to the refractor velocity has not been plotted. (a) Eigenvectors l-10. Velocity and depth
components are in-phase. (b) Eigenvectors 92-101. Velocity and depth components are largely out-of-phase.
1312 Docherty
The matrix C in the objective function is used to express (ATA+ PCJ*C+ pQ*Q)x= A*d. (12)
any constraints on the solution. Here, only simple con-
straints will be implemented, in which certain parametersare Following Scales et al. (1990), the effects of second
held close to their initial values (which are assumed to be difference smoothing can be made more apparent by replac-
correct). To constrain the ith parameter in this way, a row is ing A in equation (12) by its SVD and by setting l3 = 0. Then,
included in C which contains all zeroes except for a one in with A = WV*, where U and V are orthogonal matrices
__-
the ith column (Lawson and Hanson, 1974). If n, constraints
containing data and model eigenvectors, respectively, and n
are specified, then the dimension of C is n, x IZ. In terms of
is a diagonal matrix of singular values, solving for x gives
the SVD of A, the purpose of C is to raise up the small
singular values, thus reducing the long wavelength ambigu-
ity. x = V[A2
_ _ + /.L(W)*(QV)]-‘~U*~. (13)
The multipliers p and p in equation (10) balance the
goodness of fit of the data versus the smoothness of the Multiplication by D tends to accentuate roughness. In equa-
solution versus the degree to which the constraints are tion (13), a rough eigenvector in the jth column of V gives
satisfied. In practice, it is not difficult to find a satisfactory rise to a large entry in the jth member of the diagonal of
value for p (see the examples), and I will assume that its (Dy)*(DY). If the column vectors of V are approximately
value is now fixed. To find the smoothest solution, the full sinusoidal, in addition to being orthogonal, then the columns
nonlinear problem is solved with a series of different values of Dy will be approximately orthogonal also. In this case,
for p. For a given value of CL,the iterations are stoppedif (1) (Dv *(DV) is nearly diagonal, with large entries correspond-
X2 falls below m, (2) the procedure converges, that is, ing to the high-frequency sinusoids. This is illustrated in
changes in the model are becoming very small, (3) some Figure 8, which is a plot of the matrix (DWJ*(LW) for the
maximum number of iterations allowed has been exceeded. eigenvectors of Figure 6. The large values clustered along
The smoothest solution is obtained at the largest value of p the main diagonal occur at positions corresponding to the
for which the procedure converges and, in addition, pro- locations of the high-frequency eigenvectors in Figure 6.
duces a X2 of m. (Choosing the mean or median value of each Noting the inverse in equation (13), it is clear that the effect
delta traveltime, we might expect a better fit to the data than of second difference smoothing is to downweight contribu-
X 2 = m. Even so, given a limited number of first-arrival tions from the high-frequency eigenvectors; thus, smoothing
times from which to form the delta traveltimes, m is the the solution. Unlike damping, second difference smoothing
value I accept for X * .) does not require that the high frequencies be associatedwith
Minimization of equation (10) produces the linear system the smallest singular values.
11
6
1
-4
-_-
(DV)T(W
FIG. 8. Plot of the matrix @.))*(I)~ for the eigenvectors )! in Figure 6 (D is the second difference matrix). Large values,
indicated by bright colors, occur along the main diagonal at locations corresponding to the high-frequency eigenvectors.
2-D Refraction Tomography 1313
m/S
I 450
1400
1350
W Rehoctwvelocay- 5ooO.
1300
1250
1200
1150 z
1100
1050
1000
950
900
850
c) Mroclor~- 3ss7
d) RClrOCtQ
velocity - 4008.
m/s
1450
1400
1350
I 300
I250
1200
1150
I100
1050
1000
950
900
850
800
750
~~
600
550
500
FIG. 9. Synthetic example. Vertical scale measures elevations, horizontal scale distance, both in meters. Color indicates
weathering velocity in m/s. The refractor velocity is given at the top of each plot. (a) True model used to generate synthetic first
arrivals. The weathering velocity is constant at 1000 m/s. Random noise in the range 28 ms was added to the arrival times. (b)
Initial model input to inversion. The weathering velocity is 500 m/s. (c) Result of the inversion when only the node at x = 0
is constrained and no smoothing is applied. (d) Result obtained with the same constraint as in (c) but with second difference
smoothing added. (e) Inversion result using more node depth constraints and with second difference smoothing. (f) Same
constraints as in (e), but with smoothing replaced by damping.
2-D Refraction Tomography 1315
conjugate gradient method of Scales (1987). The cost of a figures indicates the location of the riverbed. The continuity
single run was a few minutes on an IBM 3090. of the reflectors through the zone of weak data beneath the
Statics calculated using the model in Figure 11b produced riverbed appears better in the stack obtained from the
the stacked section shown in Figure 12a. For comparison, tomography statics. Also, there is a difference in structure
Figure 12b is the stack obtained using statics from a more between the two sections. By accounting for the lower
conv_entional,delay time method, which assumed a con- velocity riverbed material, the tomographic approach has
stant velocity in the weathering layer. The letter R in the produced more of a structural high at that location.
CONCLUSIONS
Distance (m)
0 2000 4000 6000 6000
The limited aperture provided by refracted raypaths is the
causeof a fundamental long wavelength ambiguity in the 2-D
refraction tomography problem. Gradual increases (decreas-
es) in weathering velocity accompanied by gradual increases
(decreases)in weathering thickness have little effect on the
observations; this velocity-depth trade-off shows up clearly
in the singular value decomposition. In general, it is neces-
sary to add a small number of constraints to obtain the
0.6 correct weathering model. An exception, for which no
constraints may be required, is the case of rapidly varying
topography.
The use of delta rays in formulating the problem results in
FIG. 10. First-arrival time picks from a split-spreaddynamite
survey. The line crossesa dry riverbed in the neighborhood a smaller, sparser matrix for inversion. Finally, second
of coordinate 3808; distortions in curves are due to river- difference smoothing, but not damping, leads to smooth
banks. solutions.
a) v2 - 4000.
Ilvr
965 t
930
895
860
825
790
755
720 X
685 W v2 - 4070.
650
615
580
545
510
475
440
405
370
335
300
FIG. 11. (a) Fist guessat the weathering model for the picks in Figure 10. The model has a constant weathering velocity equal
to 508 m/s and an average weathering thickness of 50 m. (b) Tomographic solution for the weathering model indicating higher
velocity material away from the riverbed.
1316 Docherty
Distance (km)
0 R 4 8
FIG. 12. (a) Stacked section obtained using tomography statics computed from the model in Figure 1lb. (b) Stacked section
obtained using statics calculated by a delay time method that assumed a constant velocity for the weathering layer. R marks
the location of the riverbed in each section.
2-D Refraction Tomography 1317
b) Distance (km)
0 R 4 8
ACKNOWLEDGMENTS REFERENCES
This work was carried out at the Amoco Research Center Bahorich, M. S., Coruh, C., Robinson, E. S., and Costain, J. K.,
in Tulsa, during the period 1987-89. I would like to thank 1982, Static corrections on the southeastern Piedmont of the
United States: Geophysics, 47, 1540-1549.
Adam Gersztenkorn, Freeman Gilbert, Sam Gray, Ken Bube, K. P., Jovanovich, D. B., Langan, R. T., Resnick, J. R.,
Kelly, Larry Lines, John Scales, Sven Treitel, and Don Shuey, R. T., and Spindler, D. A., 1985, Well-determined and
poorly determinedfeatures in seismicreflection tomography:Part
Wagner for many stimulating discussions on inversion; and II: 55th Ann. Intemat. Mtg., Sot. Expl. Geophys., Expanded
Mike Bahorich, Doug Haun, Marty Williams for help with Abstracts, 608-610.
refraction statics. The comments of Mike Powers, Terry Constable, S. C.? Parker, R. L., and Constable, C. G., 1987,
Occam’s inversion: A practical algorithm for generating smooth
Watt, and two anonymous reviewers were very much appre- models from electromagnetic sounding data: Geophysics, 52,
ciated. 289-300.
1318 Docherty
de Amorim, W. N., Hubral, P., and Tygel, M., 1987, Computing Palmer, D., 1980, The generalized reciprocal method of seismic
field statics with the help of seismic tomography: Geophys. refraction interpretation: Sot. Expl. Geophys.
Prosp.,35, 907-919. Russell. B. H., 1989, Statics corrections-A tutorial: Can. Sot.
Dobrin, M., 1976, Introduction to geophysical prospecting: Mc- Expl. Geophys. Recorder, 16-30.
Graw-Hill Book Co. Scales, J. A., 1987, Tomographic inversion via the conjugate gradi-
Farrell, R. C., and Euwema, R. N., 1984, Refraction statics: Proc. ent method: Geophysics, 52, 179-185.
IEEE, 72, 1316-1329. Scales, .I. A., Docherty, P., and Gersztenkorn, A., 1990, Regular-
Hampson, D., and Russell, B., 1984, First-break interpretation ization of nonlinear inverse problems: Imaging the near-surface
using generalized linear inversion: J. Can. Sot. Expl. Geophys., weathering layer: Inverse Problems, 6, 115-131.
20, 40-54. Schneider, W. A., and Shih-Yen Kuo, 1985, Refraction modeling for
Hollingshead, G. W., and Slater, R. R., 1979, A novel method of static corrections: 55th Ann. Internal. Mtg., Sot. Expl. Geophys.,
deriving weathering statics from first-arrival refractions: Pre- Expanded Abstracts, 295-299.
sented at the 49th Ann. Internat. Mtg., Sot. Expl. Geophys. Stork, C., and Clayton, R. W., 1986, Analysis of the resolution
Lanczos, C., 1961, Linear differential operators: D. Van Nostrand between ambiguous velocity and reflector nosition for traveltime
co. tomography: 36th Ann. Internat. Mtg., Sot. Expl. Geophys.,
Lawson, C. L., and Hanson, R. J.. 1974, Solving least squares expanded Abstracts. 545-550.
problems: Prentice-Hall, Inc. White, D. J., 1989, Two-dimensional seismic refraction tomogra-
Lines, L. R., and Treite!, S., 1984, Tutorial: A review of least- phy: Geophys. J., 97, 223-245.
squares inversion and tts application to geophysical problems: Wiggins, R. A., 1972, The general linear inverse problem: Implica-
Geophys. Prosp., 32, 159-186. tion of surface waves and free oscillations for earth structure:
Musgrave, A. W., 1967, Seismic refraction prospecting: Sot. Expl. Reviews of Geophysics and Space Physics, 10, 251-285.
Geophys. Wiggins, R. A., Larner, K. L., and Wisecup, R. D., 1976, Residual
Olsen, K. B., 1989, A stable and flexible procedure for the inverse statics analysis as a general linear inverse problem: Geophysics
modeling of seismic first arrivals: Geophys. Prosp., 37, 455465. 41, 922-938.