Introduction to Seismic Imagingamalcolm/part2.pdf · 2010-08-24 · Toward a unified theory of...
Transcript of Introduction to Seismic Imagingamalcolm/part2.pdf · 2010-08-24 · Toward a unified theory of...
Introduction to SeismicImaging
Alison MalcolmDepartment of Earth, Atmospheric and
Planetary SciencesMIT
August 20, 2010
Outline
• IntroductionI Why we image the EarthI How data are collectedI Imaging vs inversionI Underlying physical model
• Data Model• Imaging methods
I KirchhoffI One-way methodsI Reverse-time migrationI Full-waveform inversion
• Comparison of Methods
Approximate Techniques
• KirchhoffI Integral techniqueI Related to X-ray CT imagingI Generalized Radon TransformI Conventionally uses ray theory
• One-wayI Based on a paraxial approximationI Usually computed with finite
differences
One-Way MethodsPhysical Motivation
• downward continuation• imaging condition
Claerbout 71, 85
rs
One-Way MethodsPhysical Motivation
• downward continuation• imaging condition
v/2
s ss s
s ss s s s
r r r r r r r r rr
One-Way MethodsPhysical Motivation
• downward continuation• imaging condition
Claerbout 71, 85
rs
One-Way MethodsPhysical Motivation
• downward continuation
• imaging condition
r rs
V(x)
s
V(x)
A Data Model
δG(s, r, t) =
∫X
∫T
G0(r, t−t0, x)V(x)∂2t G0(x, t0, s)dxdt0
δG(s, r, ω) = −∫
X
ω2G0(r, ω, x)V(x)G0(x, ω, s)dx
G0(r, t, x)
s
G0(x, t, s)
V(x)
r
A Data Model
δG(s, r, t) =
∫X
∫T
G0(r, t−t0, x)V(x)∂2t G0(x, t0, s)dxdt0
δG(s, r, ω) = −∫
X
ω2G0(r, ω, x)V(x)G0(x, ω, s)dx
⇓∫X
ω2G−(r, ω, x)G−(s, ω, x)V(x)dx
One-Way MethodsApproximating the Wave Equation
Idea (1D, c constant):
(∂2x − ∂2
t )u = (∂x − ∂t)(∂x + ∂t)u
c not constant:
(c(x)2∂2x − ∂2
t )u = (c(x)∂x − ∂t)(c(x)∂x + ∂t)u
− c(x)(∂xc(x))∂xu
c(x) smooth ⇒ better approximation
One-Way MethodsApproximating the Wave Equation
Taylor (81), Stolk & de Hoop (05)
Multi-dimensional:Lu = f
∂z
(u
∂zu
)=
(0 1
−A 0
)(u
∂zu
)+
(0
−f
)A(z, x, ∂x, ∂t) = ∂2
x − c(z, x)−2∂2t
One-Way MethodsApproximating the Wave Equation
Taylor (81), Stolk & de Hoop (05)
Diagonalize in smooth background
∂z
(u+
u−
)=
(iB+ 00 iB−
)(u+
u−
)+
(f+f−
)b±(ξ, x, ω) = ±
√ξ2 − c0(x)−2ω2
B± FIOs
Solution:
G0 =
(G+ 00 G−
)
G0 propagator in c0
One-Way MethodsApproximating the Wave Equation
Taylor (81), Stolk & de Hoop (05)
Diagonalize in smooth background
∂z
(u+
u−
)=
(iB+ 00 iB−
)(u+
u−
)+
(f+f−
)Solution:
G0 =
(G+ 00 G−
)
G0 propagator in c0
One-Way MethodsApproximating the Wave Equation
Taylor (81), Stolk & de Hoop (05)
Relate u± to u and ∂zu(u+
u−
)= 1
2
((Q+)−1 −HQ+
(Q∗−)−1 HQ−
) (u∂zu
)
q± =
[(ω
c(z, x)
)2
− ‖ξ‖2
]−1/4
Q± differ in sub-principal partsQ± ψDOs
One-Way MethodsModeling Data Stolk & de Hoop (05)
δG(s, r, ω) ≈∫X
ω2Q∗−(r)G−(r, ω, x)V(x)G+(x, ω, s)Q+(s)dxdz
V(x) = 14Q−(z, x)δc(z, x)c0(z, x)
−3Q∗+(z, x)
One-Way MethodsModeling Data Stolk & de Hoop (05)
δG(s, r, ω) ≈∫X
ω2Q∗−(r)G−(r, ω, x)V(x)G+(x, ω, s)Q+(s)dxdz
V(x) = 14Q−(z, x)δc(z, x)c0(z, x)
−3Q∗+(z, x)
Reciprocity:ss r r
One-Way MethodsModeling Data Stolk & de Hoop (05)
δG(s, r, ω) ≈∫X
ω2Q∗−(r)G−(r, ω, x)V(x)G+(x, ω, s)Q+(s)dxdz
V(x) = 14Q−(z, x)δc(z, x)c0(z, x)
−3Q∗+(z, x)
Drop Qs∫X
∫Z
ω2G−(0, r, ω, z, x)G−(0, s, ω, z, x)
(E2E1a)(z, x)dzdx
E2 : b(z, r, s) 7→ δ(t)b(z, r, s)E1 : a(z, x) 7→ δ(r − s)a(z, r+s
2)
One-Way MethodsImaging Claerbout (78) Stolk & de Hoop (06)
• downward continuation
• imaging condition
rs
One-Way MethodsImaging Claerbout (78) Stolk & de Hoop (06)
• downward continuation
d(z, s, r, t) = H(z, 0)∗Q−1∗d(s, r, t)
(H(z, z0))(s, r, t, s0, r0) =
(G−(z, z0))(r, t, r0) ∗ (G−(z, z0))(s, t, s0)
H is an FIO H ∗ H is ΨDO
• imaging condition
One-Way MethodsImaging Claerbout (78) Stolk & de Hoop (06)
• downward continuation
• imaging condition
V(z, x) ≈ d(z, x, x, 0)
V(x) = 14Q−(z)δc(z, x)c0(z, x)
−3Q∗+(z)
Recall Q are ΨDOs & H∗H is ΨDO
We have again located the singularities
One-Way MethodsExample
0
5dep
th (
km)
130 135lateral position (km)
Approximate Techniques
• KirchhoffI Integral techniqueI Related to X-ray CT imagingI Generalized Radon TransformI Conventionally uses ray theory
• One-wayI Based on a paraxial approximationI Usually computed with finite
differences
‘Exact’ Techniques
• Reverse-time migration (RTM)I Run wave-equation backwardI Usually computed with finite
differencesI “No” approximations (to the acoustic,
isotropic, linearized wave-equation, for
smooth media assuming single scattering)
• Full-waveform inversion (FWI)I Iterative method to match the entire
waveformI Gives smooth part of velocity model
Reverse-Time MigrationModeling Data
δG(s, r, t) =
∫X
∫T
G0(r, t−t0, x)V(x)∂2t G0(x, t0, s)dxdt0
δG(s, r, ω) = −∫
X
ω2G0(r, ω, x)V(x)G0(x, ω, s)dx
G0(r, t, x)
s
G0(x, t, s)
V(x)
r
Reverse-Time MigrationModeling Data
δG(s, r, t) =
∫X
∫T
G0(r, t−t0, x)V(x)∂2t G0(x, t0, s)dxdt0
δG(s, r, ω) = −∫
X
ω2G0(r, ω, x)V(x)G0(x, ω, s)dx
F : δc 7→ δG is again an FIO
Reverse-Time MigrationForming an image
δG(s, r, ω) = −∫
X
ω2G0(r, ω, x)V(x)G0(x, ω, s)dx
F : δc 7→ δG is again an FIO
F∗ : δG 7→ δc is again an FIO
F∗F is ΨDO for complete coverageStolk et al. (09)
• no direct rays (transversal intersection)• no source-side caustics• c0 ∈ C∞
Reverse-Time MigrationForming an Image
Procedure:Whitmore (83), Loewenthal & Mufti (83), Baysal et al (83)
• back propagate in time
• imaging condition
G0(r, t, x)
s
G0(x, t, s)
V(x)
r
Reverse-Time Migrationan Adjoint State Method
Lailly (83,84), Tarantola (84,86,87) Symes (09)
For a fixed source, s,
(c−2∂2t − ∇2)q(x, t; s) =
∫Rs
δG(r, t; s)δ(x − r)dr
q(·, t, ·) = 0 for t > T
receivers act as sources, reversed in time
Im(x) =2
c2(x)
∫ ∫q(x, t; s)∂2
t G0(x, t, s)dtds
Reverse-Time MigrationExample Liu et al (07)
Reverse-Time MigrationExample Liu et al (07)
Reverse-Time MigrationExample Liu et al (07)
‘Exact’ Techniques
• Reverse-time migration (RTM)I Run wave-equation backwardI Usually computed with finite
differencesI “No” approximations (to the acoustic,
isotropic, linearized wave-equation, for
smooth media assuming single scattering)
• Full-waveform inversion (FWI)I Iterative method to match the entire
waveformI Gives smooth part of velocity model
Full-waveform InversionProblem Setup Tarantola (87), Virieux & Operto (09)
Recall our initial formulation:
Lu := (∇2 −1
c2∂2
t )u = f
u = 0 t < 0
∂zu|z=0 = 0
FWI attempts to solve for c directly given u,f
no splitting of cbut band limited data ⇒ smooth solution
Full-waveform InversionProblem Setup Tarantola (87), Virieux & Operto (09)
Recall our initial formulation:
Lu := (∇2 −1
c2∂2
t )u = f
Compute:
minc‖Lu − d‖L2((S,R)×[0,T])
Full-waveform InversionSome Issues Symes (09)
Compute:
minc‖Lu − d‖L2((S,R)×[0,T])
• Objective function appears non-convexCan we restrict the domain of models sothat it is? e.g. cmin < c < cmax
• What space must c be in for objectivefunction to be differentiable?
• Data are finite bandwidth – we cannotresolve structures on all scales
Full-waveform InversionWork-around 1: Partial Linearization Symes (09)
Idea: Extend δc(x) from X to XExample: δc(x) 7→ δc(x, h) = δ(h)δc(x)
x
s
form image here
r
x + h/2x − h/2 image here
form extended
Claerbout (85) suggested this extension
Full-waveform InversionWork-around 1: Partial Linearization Symes (09)
Idea: Extend δc(x) from X to XExample: δc(x) 7→ δc(x, h) = δ(h)δc(x)
Now extendF[δc] 7→ δG
toF[δc(x, h)] 7→ δG
s.t. F is ‘invertible’ (I − F†F is smoothing)
Find c0 s.t. supp (δc(x, h)) ⊂ {(x, h)|h = 0}Stolk & de Hoop (2001), Stolk et al (2005) show invertibility
Full-waveform InversionWork-around 2: Separation of Scales
Pratt (99), Virieux (09)
minc‖Lu − d‖L2((S,R)×[0,T]) not convex
⇒ ‖cinitial − ctrue‖ must be small
Solve for G(ω) from ωmin to ωmax
No proof optimization converges
Full-waveform InversionWork-around 2: Separation of Scales, example
Virieux (09)
Outline
• IntroductionI Why we image the EarthI How data are collectedI Imaging vs inversionI Underlying physical model
• Data Model• Imaging methods
I KirchhoffI One-way methodsI Reverse-time migrationI Full-waveform inversion
• Comparison of Methods
Comparison of MethodsKirchhoff vs One-way Fehler & Huang (02)
Kirchhoff
Comparison of MethodsKirchhoff vs One-way Fehler & Huang (02)
Phase-shift
Comparison of MethodsKirchhoff vs One-way Fehler & Huang (02)
Split-step
Comparison of MethodsKirchhoff vs RTM Zhu & Lines (98)
Comparison of MethodsKirchhoff vs RTM Zhu & Lines (98)
Kirchhoff
Comparison of MethodsKirchhoff vs RTM Zhu & Lines (98)
Reverse-time
Comparison of MethodsOne-Way vs RTM Farmer (06)
Model
Comparison of MethodsOne-Way vs RTM Farmer (06)
One-way
Comparison of MethodsOne-Way vs RTM Farmer (06)
Reverse-time
Comparison of MethodsOne-Way vs RTM Farmer (06)
One-way
Comparison of MethodsOne-Way vs RTM Farmer (06)
Reverse-time
Comparison of MethodsReal Data Farmer 2006
Kirchhoff
Comparison of MethodsReal Data Farmer 2006
One-way
Comparison of MethodsReal Data Farmer 2006
Reverse-time
Summary of Methods
• Kirchhoff requires a lot of rays(or a simple model)
• One-way doesn’t handle turning waves
• RTM separation of up/down-going wavesis challenging
c0 is key• FWI optimization doesn’t convergence
from an arbitrary starting model
. . . and we never spoke about the source . . .
. . . or multiple-scattering . . .
Edip Baysal, Dan D. Kosloff, and John W. C. Sherwood.
Reverse time migration.Geophysics, 48(11):1514–1524, 1983.
J. F. Claerbout.
Imaging the Earth’s Interior.Blackwell Sci. Publ., 1985.
Jon F. Claerbout.
Toward a unified theory of reflector mapping.Geophysics, 36(3):467–481, 1971.
P. A. Farmer.
Reverse time migration: Pushing beyond wave equation.EAGE Annual Meeting, 2006.
Michael C. Fehler and Lianjie Huang.
Modern imaging using seismic reflection data.Annual Review of Earth and Planetary Sciences, 30(1):259–284, 2002.
P. Lailly.
The Seismic Inverse Problem as a Sequence of Before-Stack Migrations, pages 206–220.Conference on Inverse Scattering: Theory and Applications. SIAM, Philadelphia, 1983.
P. Lailly.
Migration Methods: Partial but Efficient Solutions to the Seismic Inverse Problem.Inverse Problems on Acoustic and Elastic Waves. SIAM, Philadelphia, 1984.
Faqi Liu, Guanquan Zhang, Scott A. Morton, and Jacques P. Leveille.
Reverse-time migration using one-way wavefield imaging condition.SEG Technical Program Expanded Abstracts, 26(1):2170–2174, 2007.
Dan Loewenthal and Irshad R. Mufti.
Reversed time migration in spatial frequency domain.Geophysics, 48(5):627–635, 1983.
C. C. Stolk and M. V. de Hoop.
Modeling of seismic data in the downward continuation approach.SIAM J. Appl. Math., 65(4):1388 – 1406, 2005.
C. C. Stolk and M. V. de Hoop.
Seismic inverse scattering in the downward continuation approach.Wave Motion, 43(7):579–598, 2006.
Christiaan C. Stolk, Maarten V. de Hoop, and William W. Symes.
Kinematics of shot-geophone migration.Geophysics, 74(6):WCA19–WCA34, 2009.
W. W. Symes.
The seismic reflection inverse problem.Inverse Problems, 25:123008, 2009.
A. Tarantola.
Inverse Problem Theory.Elsevier, 3 edition, 1987.
Albert Tarantola.
Inversion of seismic reflection data in the acoustic approximation.Geophysics, 49(8):1259–1266, 1984.
Albert Tarantola.
A strategy for nonlinear elastic inversion of seismic reflection data.Geophysics, 51(10):1893–1903, 1986.
M. E. Taylor.
Pseudodifferential Operators.Princton University Press, Princton, NJ, 1981.
J. Virieux and S. Operto.
An overview of full-waveform inversion in exploration geophysics.Geophysics, 74(6):WCC1–WCC26, 2009.
N. D. Whitmore.
Iterative depth migration by backward time propagation.In Expanded Abstracts, page S10.1. Society of Exploration Geophysicists, 1983.
Jinming Zhu and Larry R. Lines.
Comparison of kirchhoff and reverse-time migration methods with applications to prestack depthimaging of complex structures.Geophysics, 63(4):1166–1176, 1998.
[10, 11, 2] [17, 12, 8] [3, 19, 9] [1, 6, 7] [15, 16, 14] [13, 5, 20] [4, 18]