Professional Documents
Culture Documents
10.1190/1.2967501
Manuscript received by the Editor 22 May 2008; published online 1 October 2008.
1
New York University, Department of Mathematics, Courant Institute of Mathematical Science, New York, New York, U.S.A. E-mail: kourkina@
earthlink.net.
2
The University of Texas at Austin, Bureau of Economic Geology, John A. and Katherine G. Jackson School of Geosciences, Austin, Texas, U.S.A. E-mail:
sergey.fomel@beg.utexas.edu.
3
University of California at Berkeley, Department of Mathematics, Berkeley, California, U.S.A. E-mail: sethian@math.berkeley.edu.
© 2008 Society of Exploration Geophysicists. All rights reserved.
VE205
VE206 Cameron et al.
The objective of this work is to find an efficient method for build- 3兲 Velocity is arbitrary in two dimensions. Equation 2 is a conse-
ing a velocity model from time-migration velocity. We establish new quence of the truncated Taylor expansion for the traveltime
ray-theoretical connections between time-migration velocity and around the surface point x0. Velocity Vm共x0,t0兲 is a certain kind
seismic velocity in two and three dimensions. These results are of mean velocity, and we establish its exact meaning in the next
based on the image-ray theory and paraxial ray-tracing theory section.
共Popov and Pšenčik, 1978; Červený, 2001; Popov, 2002兲. 4兲 Velocity is arbitrary in three dimensions. Equation 2 is heuristic
The results can be viewed as an extension of the Dix formula 共Dix, and not a consequence of the truncated Taylor expansion. To
1955兲 to laterally inhomogeneous media. We show that the Dix ve- write an analog of traveltime approximation 2 for the three-di-
locity is seismic velocity divided by the geometric spreading of im- mensional case, we use the relation 共Hubral and Krey, 1980兲
age rays. Hence, we use the Dix velocity instead of time-migration
⌫ ⳱ 关V共x0兲R共x0,t0兲兴ⳮ1 , 共5兲
velocity as a more convenient input. We develop a numerical ap-
proach to find 共1兲 seismic velocity from the Dix velocity, and 共2兲 where ⌫ is the matrix of the second derivatives of the travel-
transition matrices from time-domain coordinates to depth-domain times from a subsurface point x to the surface, R is the matrix of
coordinates. We test the approach on synthetic and field data exam- radii of curvature of the emerging wavefront from the point
ples. This approach is complementary to more traditional velocity- source x, and V共x0兲 is the velocity at the surface point x0.
estimation methods. It can be used as the first step in a velocity mod-
el-building process. For convenience, we prefer to deal with matrix K ⬅ ⌫ⳮ1, which is,
according to equation 5,
冉 冊
root equation
T̂共x0,t0,s,r兲 ⳱ 冑 t20 Ⳮ
兩x0 ⳮ s兩2
2
Vm 共x0,t0兲
Ⳮ 冑t20 Ⳮ
兩x0 ⳮ r兩2
2
Vm 共x0,t0兲
,
det
t0
K共x0,t0兲
冑冕
In this section, we establish theoretical relationships between
t0 time-migration velocity and seismic velocity in two and three di-
1 mensions.
Vm共t0兲 ⳱ V2共z共t兲兲dt. 共3兲 The seismic velocity and Dix velocity are connected through the
t0
0 quantity 兩Q兩, the geometric spreading of image rays. The quantity
兩Q兩 is a scalar in two dimensions and a 2 ⫻ 2 matrix in three dimen-
In this case, the Dix inversion formula 共Dix, 1955兲 is exact. We sions. The simplest way to introduce Q is the following. Trace an im-
formally define the Dix velocity VDix共t兲 by inverting equation 3 age ray x共x0,t兲; x0 is the starting surface point, and t is the traveltime.
as follows: Call this ray central. Consider a small tube of rays around it.All these
Velocity estimation using time migration VE207
rays start from a small neighborhood dx0 of the point x0 perpendicu- tion 10. Moreover, let us eliminate the differentiation in q using the
lar to the earth’s surface. Thus, they represent a fragment of a plane definition of Q and rewrite it in the time-domain coordinates x0,t0.
wave propagating downward. Indeed, Q ⳱ dq/dx0, hence d/dq ⳱ 共dx0 /dq兲共d/dx0兲 ⳱ Qⳮ1d/dx0.
Consider a fragment of the wavefront defined by this ray tube at Therefore, system 9 becomes
冉 冊
time t0. Let dq be the fragment of the tangent to the front at the point
x共x0,t0兲 reached by the central ray at time t0 and bounded by the ray 1 共fQ兲x0
tube 共Figure 1兲. Then, in two dimensions, Q is the derivative
Qt0 ⳱ 共fQ兲2 P, P t0 ⳱ ⳮ . 共13兲
fQ Q x0
Q共x0,t0兲 ⳱ dq/dx0. In three dimensions, Q is the matrix of the deriv-
atives Qij共x0,t0兲 ⳱ dqi /dx0j, i, j ⳱ 1,2, where derivatives are taken Eliminating P in system 13, we get the following partial differen-
along certain mutually orthogonal directions e1, e2 共Popov and Pšen- tial equation 共PDE兲 for Q,
冉 冊 冉 冊
čik, 1978; Červený, 2001; Popov, 2002兲.
The time evolution of the matrices Q and P is given by Q t0 1 共fQ兲x0
⳱ⳮ . 共14兲
f 2Q 2
冢 冣冉 冊
fQ Q
冉冊
0 V20I t0 x0
d Q Q
⳱ 1 , 共9兲 The initial conditions are Q共x0,0兲 ⳱ 1, Qt0共x0,0兲 ⳱ 0. Equation 14
dt P ⳮ V 0 P simplifies in terms of the negative reciprocal of Q as follows. Intro-
V0
duce y ⳱ ⳮ1/Q. Then equation 14 becomes
冉 冊 冉冉 冊 冊
where V0 it the velocity at the central ray at time t, V
⳱ 共 2V/ qi q j兲i,j⳱1,2, and I is the 2 ⫻ 2 identity matrix. The absolute y t0 y f
value of det Q has a simple meaning: it is the geometric spreading of 2 ⳱ y . 共15兲
f t0 f y x0
image rays 共Popov and Pšenčik, 1978; Červený, 2001; Popov, 2002兲. x0
The matrix ⌫, introduced in the previous section, relates to Q and P In the expanded form, equation 15 is
as ⌫ ⳱ PQⳮ1. Hence, K ⳱ QPⳮ1.
In Cameron et al. 共2007兲, we prove that y t0t0 y t0 f t0 f x0x0 f x0 y x2
ⳮ2 ⳱y ⳮ y x0 ⳮ y x0x0 Ⳮ 0
冑
2 3 .
V共x共x0,t0兲,z共x0,t0兲兲 f f f f y
VDix共x0,t0兲 ⬅ 共t0Vm
2
共x0,t0兲兲 ⳱
t0 兩Q共x0,t0兲兩 共16兲
共10兲
in two dimensions, where Vm共x0,t0兲 is the time-migration velocity; 3D case
and Equation 11 can be rewritten in the following form,
V ⳱ 冑det F共det Q兲2 ,
4
共K共x0,t0兲兲 ⳱ V共x共x0,t0兲兲共Q共x0,t0兲QT共x0,t0兲兲ⳮ1 共17兲
t0
where F is the left-hand side of equation 11. As in the 2D case, we re-
共11兲 write system 9 in time-domain coordinates 共x0,t0兲. Then we get
in three dimensions, where K is defined by equation 6 and can be de- Qt0 ⳱ V2P, 共18兲
termined from equation 7.
F⬅ 共K共x0,t0兲兲. 共12兲
t0
Furthermore, we imply that t0 denotes the one-way traveltime along
the image rays. Finally, we assume that the domain does not contain
caustics; i.e., the image rays do not cross on the interval of time we
consider.
2D case
Consider a set of image rays coming to the surface. Suppose we
are tracing them all backward in time along with the quantities Q and
P. Let us eliminate the unknown velocity V in system 9 using equa- Figure 1. Illustration for the definition of geometric spreading.
VE208 Cameron et al.
冉 冊
1
Qt
V2 0 t0
1
⳱ ⳮ Qⳮ1关ⵜ共Qⳮ1 ⵜ V兲T兴Q.
V
共20兲
spectral Chebyshev method.
Finite-difference method
The initial conditions are Q共x0,0兲 ⳱ I2, Qt0共x0,0兲 ⳱ 0. The required This method was inspired by the Lax-Friedrichs method for hy-
input 冑det F is well approximated by the squares of the Dix velocity perbolic conservation laws 共Lax, 1954兲 because of its total variation
obtained from the 3D prestack time migration. We emphasize that diminishing property. We use the “Lax-Friedrichs averaging” and
despite the fact that Q is a matrix in three dimensions, scalar data are wide five-point stencil in space. The scheme is given by
enough for its computation.
PnjⳭ1 Ⳮ Pnjⳮ1 ⌬t 1
PnⳭ1
j ⳱ ⳮ
2 4⌬x Vnj
冉 冊
Cauchy problem for elliptical equations
VnjⳭ2 ⳮ Vnj Vnj ⳮ Vnjⳮ2
Equations 14 and 20 reveal the nature of the instabilities in the ⫻ ⳮ , 共21兲
problem at hand. These PDEs are elliptic. The physical setting al-
QnjⳭ1 Qnjⳮ1
lows us to pose only a Cauchy problem for them, which is known to
be ill-posed. Furthermore, the PDEs involve not only the Dix veloci- 1 1 ⌬t n 2 n
ty, but also its first and second derivatives. This dependency leads to ⳮ ⳱ⳮ Ⳮ 共共f j 兲 P j Ⳮ 共f jnⳭ1兲2 PnⳭ1
j 兲,
QnⳭ1
j Qnj 2
high sensitivity of the solutions to input data.
Nonetheless, we found two ways for solving these PDEs numeri- 共22兲
cally on the required, and relatively short, interval of time: namely, a
finite-difference scheme inspired by the Lax-Friedrichs method and where V ⬅ fQ.
a spectral Chebyshev method. A detailed analysis of the problem We impose the following boundary conditions Qn0 ⳱ Qnxⳮ1
n
⳱ 1,
shows that these methods work because of P0 ⳱ Pnxⳮ1 ⳱ 0 corresponding to the straight boundary rays. We set
n n
We solve the eikonal equation with an unknown right-hand side example, there is a significant difference between the two velocities
coupled with the orthogonality relation caused by the geometric spreading of image rays. The middle part of
the velocity model might not be recovered properly. The true struc-
1 ture should include a salt body visible in the image. The inability of
兩 ⵜ t 0兩 ⳱ , ⵜ t0 · ⵜ x0 ⳱ 0. 共24兲
V共x0共x,z兲,t0共x,z兲兲 our method to recover it exactly shows the limitation of the proposed
approach in the areas of significant lateral velocity variations, which
The orthogonality relation means that the image rays are orthogonal
invalidate the assumptions behind time migration 共Robein, 2003兲.
to the wavefronts. Such time-to-depth conversion is very fast and
Figure 5 compares three images: poststack depth-migration im-
produces the outputs directly on the depth-domain grid.
age using Dix velocity, poststack depth-migration image using the
velocity estimated by our method, and prestack time-migration im-
EXAMPLES age converted to depth with our algorithm. The evident structural
improvements in Figure 5b in comparison with Figure 5a, in particu-
Synthetic data example lar near salt flanks, and a good structural agreement between Figure
Figure 2a shows a synthetic velocity model. The model contains a 5b and c, serve as an indirect evidence of the algorithm’s success. An
high-velocity anomaly that is asymmetric and decays exponentially. ultimate validation should come from prestack depth-migration ve-
The corresponding Dix velocity mapped from time to depth is shown locity analysis, which is significantly more expensive.
in Figure 2b. There is a significant difference between both the value
and shape of the velocity anomaly recovered by the Dix method and
the true anomaly. The difference is explained by taking into account
geometric spreading of image rays. Figure 2c shows the velocity re-
covered by our method and the corresponding family of image rays.
An analogous 3D example is provided in Cameron et al. 共2007兲.
a)
Figure 3. 共a兲 Seismic image from the North Sea obtained by prestack
time migration using velocity continuation 共Fomel, 2003兲. 共b兲 Cor-
responding time-migration velocity.
b) a)
c)
b)
b) REFERENCES
Bevc, D., J. L. Black, and G. Palacharla, 1995, Plumes: Response of time mi-
gration to lateral velocity variation: Geophysics, 60, 1118–1127.
Boyd, J. P., 2001, Chebyshev and Fourier spectral methods, 2nd ed.: Dover
Publications.
Cameron, M., S. Fomel, and J. Sethian, 2007, Seismic velocity estimation
from time migration: Inverse Problems, 23, 1329–1369.
Červený, V., 2001, Seismic ray method: Cambridge University Press.
Dix, C. H., 1955, Seismic velocities from surface measurements: Geophys-
c) ics, 20, 68–86.
Fomel, S., 2003, Time-migration velocity analysis by velocity continuation:
Geophysics, 68, 1662–1672.
Hatton, L., K. L. Larner, and B. S. Gibson, 1981, Migration of seismic data
from inhomogeneous media: Geophysics, 46, 751–767.
Hubral, P., 1977, Time migration—Some ray theoretical aspects: Geophysi-
cal Prospecting, 25, 738–745.
Hubral, P., and T. Krey, 1980, Interval velocities from seismic reflection time
measurements: SEG.
Kim, Y. C., W. B. Hurt, L. J. Maher, and P. J. Starich, 1997, Hybrid migration:
A cost-effective 3-D depth-imaging technique: Geophysics, 62, 568–576.
Larner, K. L., L. Hatton, B. S. Gibson, and I. S. Hsu, 1981, Depth migration
Figure 5. Migrated images of the field data example. 共a兲 Poststack of imaged time sections: Geophysics, 46, 734–750.
Lax, P. D., 1954, Weak solutions of hyperbolic equations and their numerical
migration using Dix velocity. 共b兲 Poststack migration using estimat- computation: Communications in Pure and Applied Mathematics, 7,
ed velocity. 共c兲 Prestack time migration converted to depth with our 159–193.
algorithm. Popov, M. M., 2002, Ray theory and Gaussian beam method for geophysi-
cists: Editora da Universidade Federal da Bahia.
Popov, M. M., and I. Pšenčik, 1978, Computation of ray amplitudes in inho-
CONCLUSIONS mogeneous media with curved interfaces: Studia Geophysica et Geoda-
etica, 22, 248–258.
We have applied the recently established theorem that the Dix ve- Robein, E., 2003, Velocities, time-imaging and depth-imaging in reflection
locity obtainable from time-migration velocity is the true interval seismics: EAGE.
Sethian, J., 1996, Fast marching level set method for monotonically advanc-
velocity divided by the geometric spreading of image rays to pose ing fronts: Proceedings of the National Academy of Sciences, 93,
the corresponding inverse problem. 1591–1595.
We have suggested a set of numerical algorithms for solving the ——–, 1999, Fast marching methods: SIAM Review, 41, 199–235.
Siliqi, R., and N. Bousquié, 2000, Anelliptic time processing based on a shift-
problem numerically. We have tested these algorithms on a synthetic ed hyperbola approach: 70thAnnual International Meeting, SEG, Expand-
data example with laterally heterogeneous velocity and demonstrat- ed Abstracts, 2245–2248.
Yilmaz, O., 2001, Seismic data analysis: SEG.
ed that they produce significantly better results than simple Dix in- Yilmaz, O., I. Tanir, and C. Gregory, 2001, A unified 3-D seismic workflow:
version followed by time-to-depth conversion. Moreover, the Dix Geophysics, 66, 1699–1713.