least squares kirchho depth migration with preconditioningkstanton/files/lsmig.pdf · running head:...
TRANSCRIPT
Least squares Kirchhoff depth migration with
preconditioning
Aaron Stanton∗
∗University of Alberta,
Department of Physics,
4-183 CCIS,
Edmonton AB T6G 2E1
(April 2, 2013)
Running head: Least Squares Migration
ABSTRACT
In this article the Kirchhoff operator is used for least squares migration (LSM) of 2D seismic
data. The algorithm produces Common Image Gathers (CIGs) that when passed to the
forward operator fits the input data in a least squares sense. The inverse problem is shown
to be underdetermined. Preconditioning is added to the algorithm to constrain the solution
to be smooth. This is achieved via a simple five-point triangle filter to smooth the data
along the offset dimension of the model, along with a band-pass filter which aims to sup-
press low and high-frequency artifacts. A modified Conjugate Gradients algorithm is shown
that allows for preconditioning operators to be easily added and allows for their contribu-
tion to the model update to be easily measured. This makes setting the hyper-parameter
straightforward. Two examples are shown using synthetic data, the first is a simple linear
velocity model with a single cup-shaped reflector. In this case the LSM provides superior
results to the adjoint operator, while the LSM with preconditioning offers a slight improve-
1
ment to the smoothness of the solution. The second example uses acoustic finite difference
synthetic data from the Marmousi model. This model is more complex than the first. The
least squares approach to migration offers an improved result compared to using the adjoint
operator. In this example preconditioning offers a very minor improvement, suggesting a
more advanced migration operator and preconditioning operator are required.
2
INTRODUCTION
Seismic migration is a key step in the processing of seismic data. It is responsible for trans-
forming reflected energy into a meaningful coordinate system for structural interpretation
and parameter inversion (Tarantola, 1984). Migration can also be thought of in terms of
downward continuation– a survey consisting of source and receivers on the earths surface is
successively re-datumed to deeper recording surfaces (Schneider, 1978).
Kirchhoff migration, using the adjoint Kirchhoff operator, aims to solve the migration
problem by integration. Because we record discrete data with a finite aperture and aim to
solve for a discrete earth model the integrals are replaced with finite discrete summations.
This is the primary source of error for these methods. Another problem with Kirchhoff
methods relates to the artifacts created by spreading amplitudes over travel time curves.
While typically the bulk of the energy constructively sums in place of true reflectors, a great
deal of energy persists in the form of ”smiles” or migration artifacts.
Least squares migration (LSM) can attempt to deal with the above problems. Instead
of simply using the adjoint operator, the inverse is sought by attempting to match the data
predicted by the model with the observed data. Significant work has been done in the field
of LSM. Nemeth et al. (1999) use a Kirchhoff operator to perform LSM of incomplete surface
seismic and Vertical Seismic Profile (VSP) data. They find that the migration results are
more focussed and suffer less from acquisition footprint compared with the adjoint operation.
Kuehl and Sacchi (2003) use LSM with a Double Square Root (DSR) operator to express
reflectivity as a function of subsurface position and ray parameter to be used for further
Amplitude Versus reflection Angle (AVA) inversion. Guitton et al. (2007) used LSM to
attenuate low frequency Reverse Time Migration (RTM) artifacts due to crosscorrelation of
3
diving-waves, head-waves, or backscattered waves. The attenuation of artifacts is achieved
via preconditioning using prediction-error filters.
This paper considers least squares prestack depth migration using a Kirchhoff operator.
Two synthetic data examples are provided, the first involves a single reflector that has been
sampled by a small number of shots modelled using a forward Kirchhoff operator, while the
second involves a more complicated reflectivity structure with input shot gathers modelled
using an acoustic finite difference approach. In both cases it is clear that LSM offers an
improved result over application of the adjoint operator. A simple preconditioning strategy
is used to impose smoothness across offsets. In the first example the desired CIGs contain
a single reflector with a flat, constant amplitude response with offset. The preconditioning
is successful in this scenario. In the second more complicated scenario, the gathers are
not flat, and display amplitude variations with offset. In this case the preconditioning is
less successful in attenuating migration artifacts, although a computational advantage is
achieved, as the misfit is lowered more quickly than the case when no preconditioning is
used.
THEORY
The Kirchhoff operator
The Kirchhoff operator is currently the most widely used migration operator. One of
the main reasons for the algorithm’s popularity lies in its simplicity. Kirchhoff migration
involves integrating trace amplitudes over a reflectivity model. After travel times and
Kirchhoff weights have been calculated, the migration process can be written as a trace by
trace process (Audebert, 2001). This means that a tiny amount of memory is required for
4
this process, and perhaps more importantly the task is highly parallelizable. The speed
of this algorithm also makes it possible to perform iterative velocity analysis. For these
reasons the adjoint Kirchhoff operator is currently the most popular migration algorithm
in commercial processing flows.
In the case of LSM the forward and adjoint migration operators are required. Given
the simplicity of the Kirchhoff adjoint operator, the forward operator is straightforward to
define.
In addition to its simplicity, the Kirchhoff operator also suffers from artifacts that could
benefit from LS migration with preconditioning that aims to mitigate these artifacts as will
be shown later in this paper.
The basis of Kirchhoff migration is the Huygen-Fresnel principle which states when a
disturbance reaches a point, this point behaves as a secondary source, and that summing
the waves from this source can be used to determine the form of the wave at a later time.
The forward Kirchhoff operator can be written
d(s, g, t) =∑Nx
∑Nz
m(x, z)K(s, g, x, z, t)
where d(s, g, t) is the data, m(x, z) is the model, and K(s, g, x, z, t) are the Kirchhoff
weights. Here t is the travel time that is obtained via ray tracing through the velocity
model. The adjoint, or migration operator can be written
m(x, z) =∑
Ns∗Ngd(s, g, t)K(s, g, x, z, t).
Algorithm 1 shows an implementation of the Kirchhoff operator that was used for the
synthetic data experiments in this paper.
The main elements of the 2D Kirchhoff depth migration operator described in Algo-
5
Algorithm 1 Kirchhoff operatorprocedure Kirchhoff(d,m, v, sx, gx, dt, Ntrace, Nx, Nz , fwd, adj) . fwd diffracts, adj migrates
for itrace = 1 : Ntrace do
t← ray tracing(sx(itrace), gx(itrace), V )
it = bt/dtc
b = (t− it ∗ dt)/it ∗ dt
a = 1− b
for ix = 1 : Nx do
for iz = 1 : Nz do
K ← kirchoff weights(sx(itrace), gx(itrace), V, ix, iz)
if fwd then
d(itrace,itime) = d(itrace,itime) + a*m(ix,iz)*K
d(itrace+1,itime) = d(itrace,itime) + b*m(ix,iz)*K
end if
if adj then
m(ix,iz) = m(ix,iz) + (a*d(itrace,itime) + b*d(itrace,itime + 1))*K
end if
end for
end for
end for
end procedure
rithm 1, functions ”ray tracing” and ”kirchhoff weights” are taken from the open source
SeismicUnix code sukdmig2d.c (Liu, 1995).
Least squares migration
Given the forward Kirchhoff operator, L, data can be generated from a reflectivity model.
Given recorded data, we may want to collapse diffractions to the position where reflection
occurred. To do this we should use the inverse of the forward operator, L−1. Typically
for imaging the inverse is approximated by the adjoint operator, LT . While the adjoint
can produce reasonable structural images, the produced model is not true amplitude in the
sense that applying the forward operator will not recover the recorded data to a reasonable
degree, especially if the acquisition geometry contains significant holes, or irregular geometry
(Nemeth et al., 1999). LSM aims to recover the model, m, such that the application of
6
the forward operator reproduces the recorded data in a least squares sense. The system of
equations for LSM is
Lm = d + n
where d is the observed data and n is noise. The minimum norm solution comes from
minimizing the cost function
J = ||Lm− d||22 + λ2||m||22,
where J is the cost to be minimized, m is the reflectivity model to be recovered, and λ is
a trade-off parameter which controls the under-fitting or over-fitting to the observed data,
d.
The minimum norm solution to this cost function is
m̂ = (LTL + λ2I)−1LTd.
Typically the size of the model is greater than the size of the data which makes this an
underdetermined problem. The next section discusses methods to precondition the problem
in order to narrow down the range of solutions.
Preconditioning
To obtain a smooth solution a derivative operator can be applied to the model. Sharp
changes in the model are then penalized in the cost function resulting in a smoother set of
solutions. This is written:
J = ||Lm− d||22 + λ2||Dm||22,
where D can be referred to as a ”bad pass” operator, such as a derivative across the
offset dimension of Common Image Gathers (CIGs).
7
An alternative way to write this is via a change of variables:
z = Dm
giving
J = ||LD−1z− d||22 + λ2||z||22.
If the derivative operator, D, acts as a high-pass operator then its inverse, D−1 acts as
a low-pass operator (Wang, 2005). It becomes clear that operators placed into the right
hand side (model-space part) of the cost function are ”bad-pass” operators that emphasize
parts of the model we wish to penalize, while terms on the left hand side (data-space) are
”good-pass” operators that emphasize parts of the model we wish to reinforce.
This cost function can be incorporated into a Conjugate Gradients framework such as
given by Scales (1987), by simply concatenating the preconditioning operator, D, with the
migration operator, L, to form the new operator L̃ = LD−1, with adjoint L̃T = (D−1)TLT .
An alternative formulation is to incorporate one or more regularization terms directly
into the Conjugate Gradients algorithm∗. Algorithm 2 shows a typical implementation of
Conjugate Gradients, but with two operators, A, and B. A clear advantage to writing the
algorithm in this way is that the contribution of each operator to the model update step
∆ =< As,As > +p < Bs,Bs > can be controlled by the hyper-parameter p. Considering
B to be a ”good-pass” operator such as a smoothing operator, setting p = 0 provides the
minimum-norm solution, while setting p to a value that makes the contribution of the two
inner products of comparable magnitude provides a regularized solution. Finally, setting p
to be a very large value makes the contribution of < Bs,Bs > larger than < As,As >,
restricting the update to be mainly controlled by the smoothness of the model.
∗via personal communication with M. D. Sacchi
8
The good pass operator used in this study is a five-point triangle filter, P, which is
applied across offsets within each Common Image Gather. Unlike a rectangular filter, a
triangle filter gives higher weight to the central point on the filter compared to the edges.
When smoothing a signal of length 8 the filter can be written in matrix form as
P =
1 0 0 0 0 0 0 0
14
12
14 0 0 0 0 0
19
29
13
29
19 0 0 0
0 19
29
13
29
19 0 0
0 0 19
29
13
29
19 0
0 0 0 19
29
13
29
19
0 0 0 0 0 14
12
14
0 0 0 0 0 0 0 1
, with adjoint
PT =
1 14
19 0 0 0 0 0
0 12
29
19 0 0 0 0
0 14
13
29
19 0 0 0
0 0 29
13
29
19 0 0
0 0 19
29
13
29 0 0
0 0 0 19
29
13
14 0
0 0 0 0 19
29
12 0
0 0 0 0 0 19
14 1
.
This operator has very few non-zero entries, so writing the smoothing operator in oper-
ator form makes this forward adjoint pair computationally inexpensive.
To avoid low and high frequency artifacts in the solution the smoothing operator is
concatenated with a band-pass filter. This operator can be written Q = F−1TF. Where
9
F and F−1 are the forward and inverse Fourier transforms respectively and T is a diagonal
matrix (TT = T) that weights spectral coefficients outside of the pass-band to zero. Since
the Fourier transform is orthogonal (FT = F−1), the adjoint of this operator is written
QT = (F−1TF)T = FTTT(F−1)T = F−1TF. This implies that the band-pass operator is
self adjoint. The concatenation of the smoothing operator, P, and the band-pass operator,
Q, is written B = PQ, with adjoint BT = QPT. To ensure that the forward adjoint
pair are not written incorrectly, the dot product test is performed. Two vectors of random
numbers are created, one with the dimension of the data, d1, and one with the dimension
of the model, m1. d1 is passed through the concatenation of Adjoint operators to generate
m2, while m1 is passed through the concatenation of Forward operators to generate d2.
The inner products < d1,d2 > and < m1,m2 > are compared and found to agree within
machine precision.
SYNTHETIC DATA EXAMPLES
Simple synthetic
The algorithm is first tested on a simple reflectivity model consisting of a cup shaped
reflector in a simple background velocity shown on the right in Figure 1. Using the known
reflectivity the forward model is used to generate 6 shot gathers, shown on the left in
Figure 1. Adjoint, LSM, and LSM with preconditioning are applied to the input data. In
the case of preconditioning the hyper-parameter, p, is selected by first allowing it to be
equal to one, then monitoring the difference in magnitude between the two inner products
that contribute to the model update. p = 1000 was then chosen to equalize the contribution
of these two terms, and the code was restarted. The adjoint operator is applied to these
10
gathers resulting in the selected Common Image Gathers (CIGs) shown on the left in Figure
2. Here 8 gathers are shown with increasing offset to the right within each gather. The
minimum norm solution after 20 iterations of Conjugate Gradients is shown in the middle
of the figure, while the right figure shows the gathers after LSM with preconditioning
(again with 20 iterations). It is apparent that the amplitudes of the artifacts in the adjoint
are of the same level as the flat reflector amplitudes, while the LSM results have higher
amplitudes corresponding to the flat reflector. Between the two LSM results it is apparent
that the use of preconditioning provides slightly less artifacts compared with the minimum
norm solution. Figure 3 shows the results of adjoint (left), LSM (middle), and LSM with
preconditioning (right) after stacking CIGs. It is clear that LSM mitigates much of the
migration artifacts seen in the adjoint stack (left), while preconditioning (right) provides a
slightly clearer result compared with the minimum norm solution (middle). Figure 4 shows
the data predicted from the migrated model. The data predicted from the adjoint (left)
does not come close to modelling the original data, while the minimum norm LSM result
(middle) matches the original data quite closely. The LSM with preconditioning (right) also
matches the original data closely, but appears to be slightly noisier than that obtained using
the minimum norm solution. Figure 5 shows the misfit versus iteration number for both
LSM (a) and LSM with preconditioning (b). Here the curves are similar in both these cases,
but it is clear that the minimum norm solution reaches a lower misfit than the least squares
solution with preconditioning. Figure 6 shows the amplitude spectrum of the adjoint stack
(a) and the coefficients of the band-pass operator that was applied in the preconditioning
(b).
11
Marmousi finite difference synthetic
This example follows the advice of Gray et al. (2001), namely, that a more accurate forward
modelling operator should be used compared to the migration operator being tested. The
shot gathers are modelled using an acoustic finite difference algorithm and the Marmousi
subsurface model. The velocity model is shown in Figure 7. Both LSM and LSM with pre-
conditioning were run with 20 iterations. For LSM with preconditioning a hyper-parameter
of 20000 was chosen to equalize the action of the migration and regularization operators on
the model update. Figure 8 shows selected CIGs after adjoint (left), LSM (middle), and
LSM with preconditioning (right). All three figures appear similar, but with slightly less
noise from artifacts on the results after LSM and LSM with preconditioning. What these
CIGs also show is that the velocity structure is likely too complicated for ray-based imaging
methods, and a wave-equation based method would be better suited for this dataset. This
is further discussed in Gray et al. (2001). Figure 9 shows stacked data after adjoint (top),
LSM (middle), and LSM with preconditioning (bottom). The amplitudes of reflectors are
more balanced on the LSM result compared with the adjoint result, while the result with
preconditioning appears to have very slightly less noise, especially in the deeper 2-3km of
the section. What is also apparent is that the uppermost reflector at depth ≈ 0km has
been attenuated in the preconditioned result. This is likely due to the fact that the pre-
conditioning is a smoothing operator applied across offsets. Because the reflector has an
average fold of approximately 2 at this depth (see Figure 8), the preconditioning is asking
for solutions that are more continuous across offset, thereby lowering the amplitude of this
uppermost reflector. This suggests that either a more accurate velocity field is required to
flatten this event, or a more accurate migration operator should be used for such a compli-
cated velocity structure. Yet another alternative is to use a more complex preconditioning
12
strategy that can smooth along the local dips in the data (see for example Rebollo and
Sacchi (2010)). Figure 10 shows selected input shot gathers (upper left), residual after
using the adjoint migration operator (upper right), residual after LSM (bottom left), and
residual after LSM with preconditioning (bottom right). It is clear that the adjoint does
a poor job in recovering the input data, while LSM and LSM with preconditioning have
a much lower amplitude residual. Figure 11 shows the misfit versus iteration number for
both LSM (a) and LSM with preconditioning (b). A positive effect of preconditioning in
this example is the steepness of the curve for early iterations. The preconditioning has the
effect of clustering eigenvalues in the LTL term which stabilizes the inversion (Wang, 2005).
Figure 12 shows the amplitude spectrum of the adjoint stack (a) and the coefficients of the
band-pass operator that was applied in the preconditioning (b).
CONCLUSIONS
The two synthetic examples considered in this paper demonstrate that a least squares
formulation of Kirchhoff depth migration is effective in providing an improved image of
the subsurface. Preconditioning is added to the LSM using a 5 point triangle smoothing
operator across offsets along with a band-pass operator that attempts to limit low and
high frequency artifacts in the solution. In the first example a significant improvement is
achieved through the use of preconditioning, while in the second example the improvement
due to preconditioning is less. This was likely due to the complexity of the velocity model
used, suggesting a wave-equation migration operator would be better suited to migrate such
data. A drawback of this method is that the entire dataset must be loaded to memory. For
extremely large datasets this would be impractical. A topic of further research could be how
to implement an inversion procedure that can utilize ”warm starts” from a shot-by-shot LS
13
migration procedure.
ACKNOWLEDGMENTS
I would like to thank Zhenyue Liu for contributing the 2d Kirchhoff depth migration code
in the open-source SeismicUnix package. This code provided the basic framework for de-
veloping the LSM code.
14
REFERENCES
Audebert, F., 2001, 3-D prestack depth migration: why Kirchhoff?: Technical Report 80,
Stanford Exploration Project.
Gray, S. H., J. Etgen, J. Dellinger, and D. Whitmore, 2001, Seismic migration problems
and solutions: Geophysics, 66, 1640.
Guitton, A., B. Kaelin, and B. Biondi, 2007, Least-squares attenuation of reverse-time-
migration artifacts: GEOPHYSICS, 72, S19–S23.
Kuehl, H., and M. D. Sacchi, 2003, Least-squares wave-equation migration for avp/ava
inversion: Geophysics, 68, 262–273.
Liu, Z., http://www.seismicunix.com/w/Sukdmig2d.
Nemeth, T., C. Wu, and G. T. Schuster, 1999, Least-squares migration of incomplete re-
flection data: Geophysics, 64, 208–221.
Rebollo, R. C., and M. D. Sacchi, 2010, Time domain least-squares prestack migration:
GeoCanada Expanded Abstracts, 4.
Scales, J. A., 1987, Tomographic inversion via the conjugate gradient method: Geophysics,
52, 179–185.
Schneider, W., 1978, Integral formulation for migration in two and three dimensions: Geo-
physics, 43, 49–76.
Tarantola, A., 1984, Linearized inversion of seismic reflection data: Geophysical Prospect-
ing, 32, 998–1015.
Wang, J., 2005, 3-d least-squares wave-equation avp/ava migration of common- azimuth
data: PhD thesis, University of Alberta, least squares migration.
15
APPENDIX A
CONJUGATE GRADIENTS WITH MORE THAN ONE OPERATOR
Algorithm 2 Conjugate Gradients with more than one operatorprocedure CG(d,Niter, p) . solves ||Ax− d||22 + p||B−1x||22
x = 0
r1 = d
g = AT r1 + BT r2
s = g
γ =< g, g >
γold = γ
while k ≤ Niter do
ss1 = As
ss2 = Bs
∆ =< ss1, ss1 > +p < ss2, ss2 >
α = γ/∆
x = x + αs
r1 = r1 − αss1
r2 = r2 − αss2
g = AT r1 + BT r2
γ =< g, g >
β = γ/γold
γold = γ
s = g + βs
k + +
end while
return x . The solution is x
end procedure
16
LIST OF FIGURES
1 Data and velocity model used for the first example.
2 A selection of Common Image Gathers (CIGs) after adjoint (left), LSM (middle),
and LSM with preconditioning (right).
3 Stack after adjoint (left), LSM (middle), and LSM with preconditioning (right).
4 Predicted data after application of adjoint (left), LSM (middle), and LSM with
preconditioning (right).
5 Misfit vs. Iteration number for LSM (a), and LSM with preconditioning (b).
6 Amplitude spectrum for stacked data (a) and band-pass filter applied as part of
the preconditioning operator (b). Note that the band-pass filtering operator is self adjoint.
7 Smooth velocity model used for the migration.
8 A selection of Common Image Gathers (CIGs) after adjoint (left), LSM (middle),
and LSM with preconditioning (right).
9 Stack after adjoint (top), LSM (middle), and LSM with preconditioning (bottom).
10 A selection of shot gathers representing input data (upper left), residual after
predicting data using the adjoint (upper right), residual after using the LSM result to
predict data (bottom left), and residual data after using LSM with preconditioning to
predict data (bottom right).
11 Misfit vs. Iteration number for LSM (a), and LSM with preconditioning (b).
12 Amplitude spectrum for stacked data (a) and band-pass filter applied as part of
the preconditioning operator (b). Note that the band-pass filtering operator is self adjoint.
17
0
0.5
1.0
1.5
2.0
2.5
Tim
e (
s)
2 4 6 8
0
500
1000
1500
2000
2500
De
pth
(m
)
0 1000 2000 3000Distance (m)
Figure 1: Data and velocity model used for the first example.Stanton & Sacchi –
18
0
500
1000
1500
2000
2500
De
pth
(m
)
0.2 0.4 0.6 0.8 1.0x104
0
500
1000
1500
2000
2500
De
pth
(m
)
0.2 0.4 0.6 0.8 1.0x104
0
500
1000
1500
2000
2500
De
pth
(m
)
0.2 0.4 0.6 0.8 1.0x104
Figure 2: A selection of Common Image Gathers (CIGs) after adjoint (left), LSM (middle),
and LSM with preconditioning (right).
Stanton & Sacchi –
19
0
500
1000
1500
2000
2500
De
pth
(m
)
1000 2000 3000Midpoint (m)
0
500
1000
1500
2000
2500
De
pth
(m
)
1000 2000 3000Midpoint (m)
0
500
1000
1500
2000
2500
De
pth
(m
)
1000 2000 3000Midpoint (m)
Figure 3: Stack after adjoint (left), LSM (middle), and LSM with preconditioning (right).
Stanton & Sacchi –
20
0
0.5
1.0
1.5
2.0
2.5
Tim
e (
s)
2 4 6 8
0
0.5
1.0
1.5
2.0
2.5
Tim
e (
s)
2 4 6 8
0
0.5
1.0
1.5
2.0
2.5T
ime
(s)
2 4 6 8
Figure 4: Predicted data after application of adjoint (left), LSM (middle), and LSM with
preconditioning (right).
Stanton & Sacchi –
21
0
2000
4000
6000
8000
10000
12000
14000
16000
0 2 4 6 8 10 12 14 16 18 20
Mis
fit
Iteration number
(a)
10000
20000
30000
40000
50000
60000
70000
80000
0 2 4 6 8 10 12 14 16 18 20
Mis
fit
Iteration number
(b)
Figure 5: Misfit vs. Iteration number for LSM (a), and LSM with preconditioning (b).
Stanton & Sacchi –
22
0 0.02 0.04 0.06 0.08 0.10 0.12(1/m)
0.2
0.4
0.6
0.8
1.0
1.2
x10 -3
Am
plit
ude
(a)
0
0.2
0.4
0.6
0.8
1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
Filt
er
Am
plit
ud
e
Frequency (1/m)
(b)
Figure 6: Amplitude spectrum for stacked data (a) and band-pass filter applied as part of
the preconditioning operator (b). Note that the band-pass filtering operator is self adjoint.
Stanton & Sacchi –
23
0
1000
2000
3000
Depth
(m
)
0 2000 4000 6000 8000Distance (m)
Figure 7: Smooth velocity model used for the migration.Stanton & Sacchi –
24
0
500
1000
1500
2000
2500
3000
De
pth
(m
)
2000 2200 2400 2600 2800 3000
0
500
1000
1500
2000
2500
3000
De
pth
(m
)
2000 2200 2400 2600 2800 3000
0
500
1000
1500
2000
2500
3000
De
pth
(m
)
2000 2200 2400 2600 2800 3000
Figure 8: A selection of Common Image Gathers (CIGs) after adjoint (left), LSM (middle),
and LSM with preconditioning (right).
Stanton & Sacchi –
25
0
1000
2000
3000
Depth
(m
)
2000 3000 4000 5000 6000 7000 8000 9000Midpoint (m)
0
1000
2000
3000
Depth
(m
)
2000 3000 4000 5000 6000 7000 8000 9000Midpoint (m)
0
1000
2000
3000
Depth
(m
)
2000 3000 4000 5000 6000 7000 8000 9000Midpoint (m)
Figure 9: Stack after adjoint (top), LSM (middle), and LSM with preconditioning (bottom).
Stanton & Sacchi –26
0
0.5
1.0
1.5
2.0
2.5
Tim
e (
s)
0.2 0.4 0.6 0.8 1.0x104
0
0.5
1.0
1.5
2.0
2.5
Tim
e (
s)
0.2 0.4 0.6 0.8 1.0x104
0
0.5
1.0
1.5
2.0
2.5
Tim
e (
s)
0.2 0.4 0.6 0.8 1.0x104
0
0.5
1.0
1.5
2.0
2.5
Tim
e (
s)
0.2 0.4 0.6 0.8 1.0x104
Figure 10: A selection of shot gathers representing input data (upper left), residual after
predicting data using the adjoint (upper right), residual after using the LSM result to predict
data (bottom left), and residual data after using LSM with preconditioning to predict data
(bottom right).
Stanton & Sacchi –
27
1.4e+11
1.6e+11
1.8e+11
2e+11
2.2e+11
2.4e+11
2.6e+11
2.8e+11
3e+11
3.2e+11
3.4e+11
0 2 4 6 8 10 12 14 16 18 20
Mis
fit
Iteration number
(a)
2e+11
2.5e+11
3e+11
3.5e+11
4e+11
4.5e+11
5e+11
5.5e+11
0 2 4 6 8 10 12 14 16 18 20
Mis
fit
Iteration number
(b)
Figure 11: Misfit vs. Iteration number for LSM (a), and LSM with preconditioning (b).
Stanton & Sacchi –
28
0 0.02 0.04 0.06 0.08 0.10 0.12(1/m)
0.5
1.0
1.5
Am
plit
ude
(a)
0
0.2
0.4
0.6
0.8
1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
Filt
er
Am
plit
ud
e
Frequency (1/m)
(b)
Figure 12: Amplitude spectrum for stacked data (a) and band-pass filter applied as part of
the preconditioning operator (b). Note that the band-pass filtering operator is self adjoint.
Stanton & Sacchi –
29