You are on page 1of 174

2-1

Seismic Inversion and AVO



Part 2
Post-stack inversion
2-2
Introduction
Inversion is the process of extracting, from seismic data,
the underlying geology which gave rise to that seismic.

Traditionally, inversion has been applied to post-stack
seismic data, with the aim of extracting acoustic
impedance.

Recently, inversion has been extended to pre-stack
seismic data, with the aim of extracting both acoustic and
shear impedance. This allows the calculation of pore
fluids.

Another recent development is to use inversion results to
directly predict lithologic parameters such as porosity and
water saturation.
2-3
Types of Inversion
Recursive: Traditional bandlimited inversion
Model Based: Iteratively updates a layered initial model.
Sparse Spike: Constrained to produce few events
Colored: Modern derivative of Recursive Inversion
Elastic Impedance: Enhancement for AVO data
LMR: Enhancement for AVO data
Simultaneous Inversion: Pre-stack inversion.
There are many types of inversion methods, such as:
We will talk about all of these methods in this course, but
in this chapter we will only discuss the first four on the
above list in detail.
2-4
Impedance
Acoustic
Shear
Elastic
Reflectivity
Wavelet
Seismic
The common forward model for all inversions can be shown as:
General Forward Model for Inversion
2-5
Impedance Reflectivity
Inverse
Wavelet
Seismic
Inversion simply reverses the procedure:
Inverse Model
Acoustic
Shear
Elastic
2-6
Acoustic Impedance
or
Shear Impedance
or
Elastic Impedance
The basic equation relating reflectivity, R, to impedance, Z,
for the i
th
interface in a stack of N layers is as follows:
Acoustic Impedance =

Shear Impedance =

Elastic Impedance =
,
1
1
i i
i i
i
Z Z
Z Z
= R
+

+
+
P
V
S
V
= Z
. ) / (
,
2
) sin 4 1 ( ) sin 8 ( ) sin 1 (
2 2 2
P S
K K
S P
V V K
V V
=
+ u u u

General Forward Model for Inversion


2-7
To then go from reflectivity to the seismic trace, we use the
convolutional model:
Seismic = Wavelet convolved
with reflectivity plus noise.
Assumptions:

(1) There are no multiples modeled.
(2) Transmission loss and geometric spreading are
ignored.
(3) Frequency-dependent absorption is ignored.
(4) The wavelet may be time varying.
From reflectivity to seismic
noise R W T + = *
2-8
The Convolutional Model
The synthetic to
the right illustrates
the equation of the
previous page for
the noise-free
case. Notice that
convolution with
the wavelet results
in a loss of
information.
2-9
The Convolutional Model
In the frequency domain,
convolution is the product of
the reflectivity spectrum and
the wavelet spectrum, or
T(f) = W(f)R(f)

This means that the seismic
trace has lost both the high
frequency and the low
frequency portions of the
spectrum.

Deconvolution and inversion
attempt to recover these lost
regions. The method of
filling in the missing data
depends on the inversion
algorithm.
2-10
Recursive Inversion
Recursive Inversion, also called bandlimited inversion, is the simplest
form of inversion. Recall that the reflection coefficient for the i
th

interface of N layers can be written:
i i
i i
i
Z Z
Z Z
= R
+

+
+
1
1
(

+
i
i
i i
R
+ R
= Z Z
1
1
1
The impedance of the i +1
st
layer can then be determined from the i
th

layer by inverting the above equation to give:
2-11
Recursive Inversion
For a single layer, as shown
above, we can write:
(

+ R - 1
R + 1
Z = Z
Z Z
Z - Z
= R
1 2
1 2
1 2
Seismic
raypath
Interface at
depth = d
Z
1
=
1
V
1
Z
2
=
2
V
2
R

t

Reflection at time
t = 2d/V
1
Geology Seismic
Surface
2-12
Recursive inversion of N layers
terms. all of product the : where
1
1
1
1
1
= H

+
[

=
,
R
R
N
N
i i
i
= Z Z
For N layers, we can start at the first layer and
compute the impedance of each successive layer by
recursively applying this formula:
This is illustrated in the next slide.
2-13
Recursive inversion
Applying recursive
inversion under ideal
conditions, we can
perfectly recover the
impedance, as
shown on the right:
2-14
Using the recursive
inversion formula
given earlier, and
assuming that = 1
(i.e. V = Z) and V
1
=
1000 m/s, compute
the inverted
velocities for (a) a
single reflector, and
(b) the reflector
convolved with a
Ricker wavelet, as
shown on the right.
Exercise 2-1 Wavelet effects
2-15
Note that we can
only recover the true
value for the change
in impedance if we
have a single spike,
which is not the
case after
convolution with the
wavelet.
V
2
= 818 m/s
V
2
= 1500 m/s
V
1
= 1000 m/s
V
3
= 1227 m/s
V
4
= 1004 m/s
V
1
= 1000 m/s
Exercise 2-1 Solution
2-16
Problems with recursive inversion
First, the wavelet lobes cause low and high impedance
zones to appear on the inverted trace which are not
geologically valid.
Second, the low frequency component of the impedance
is lost.
Third, the true impedance value is never estimated due
to the first two problems.


As shown in the previous exercise, there are three problems
with recursive inversion:
2-17
From deconvolution to inversion
Deconvolution is the process of applying the inverse
of the wavelet to the seismic trace.

If deconvolution worked perfectly (and the processing
had also been done perfectly) inversion would simply
consist of recursively inverting the reflection
coefficients. However, we rarely know enough
information to fully apply the deconvolution process.

We will therefore now look at practical approaches to
the inversion problem.
2-18
Practical inversion schemes
Now that we have looked at the issues that make
inversion difficult, such as noise problems, wavelet
tuning, and the fact that recursive inversion only works
if a wavelet is not present, let us look at practical
schemes for inversion.

We will look at the following steps:

(1) Wavelet extraction
(2) Geological model-building
(3) Integrating the geological model with the inverted
seismic data.
2-19
Wavelets and wavelet extraction
As shown in the next figure, the wavelet is defined
completely by its amplitude and phase spectra.

- Over a limited frequency range, the phase spectrum
may often be approximated by a straight line.
- The intercept of the line is the constant phase
rotation which best characterizes this wavelet.
- The slope of the line measures the time-shift of the
wavelet.

The wavelet on the next page is an example of a
minimum phase wavelet.
2-20
A typical wavelet
2-21
Wavelets in the earth vary both laterally(spatially) and temporally for
a variety of reasons:

Near surface effects (space variant)
Frequency-dependent absorption (space and time variant)
Inter-bed multiples (space and time variant)
NMO stretch
Processing artifacts

We usually assume that the wavelet is constant with time and space:
Time invariant : This means that the inversion is optimized
for a limited time window.
Space invariant : This assumes that the data has been
processed optimally to remove spatial variations in the
wavelet.
Wavelet extraction
2-22
The following methods can be used for wavelet extraction:

(1) Estimate amplitude spectrum using the the seismic data alone.
The phase is assumed known from some other source. Methods
include autocorrelation, maximum entropy spectral analysis, and
cross spectral analysis.

(2) Estimate both amplitude and phase spectra from the seismic
data alone. Methods include minimum entropy wavelet estimation
and higher order moments. Note that these methods can be quite
unstable.

(3) Estimate both amplitude and phase spectra using deterministic
measurements, such as marine signatures and VSP analysis.

(4) Estimate both amplitude and phase spectra using both seismic
and well log measurements.

(5) Estimate amplitude spectrum and a constant phase spectrum
using both seismic and well log measurements.

Wavelet extraction
2-23
A zero-phase seismic wavelet
(a) The wavelet estimated
using from the amplitude
spectrum of the seismic data.
(b) The amplitude spectrum of
the wavelet.
2-24
A general problem with wavelet extraction is that:

To extract a wavelet using logs, an optimum correlation
must be done first.

To perform correlation properly, the wavelet must already
be known.

A practical wavelet extraction procedure is as follows:

(1) Use statistical wavelet extraction to determine a preliminary
wavelet. This assumes that the approximate phase of the wavelet
is known.

(2) Stretch/squeeze the logs to tie the seismic data

(3) Extract a new wavelet using the well logs.

(4) Possibly repeat steps (2) and (3).
Wavelet extraction
2-25
Non-uniqueness why we need a model
All inversion algorithms suffer from the non-uniqueness
problem.
This means that there is more than one possible
geological model consistent with the seismic data.
The only way to decide between the possibilities is to use
other information, not present in the seismic data.
This other information is usually provided in two ways:
the initial guess model
constraints on how far the final result may deviate
from the initial guess
This means that the final result always depends on the
other information as well as the seismic data. This is
shown in the next slide.
2-26
A solution for non-uniqueness
The flowchart above shows the general procedure for inversion.
In the next few slides, we will look at building the model.
Seismic
Data
Geological
Constraints
Optimum
Section
Geological
Model
Combine
and
Invert
Final
Inversion
2-27
The initial guess model
The initial guess model consists of an impedance log, which must be measured in
2-way travel time. Since the original logs are measured in depth, a critical step is
depth-to-time conversion. The depth-to-time conversion is made using a depth-
time table which maps each depth to the two-way travel time from the datum
(surface) to that depth.
2-28
The depth-time table is usually calculated from the sonic log velocities
by:



where: t
j
= time down to layer j
d
j
= thickness of layer j
V
j
= velocity of layer j
Note: The time to an event depends on all the velocities above that layer,
including the first velocity to the surface, V
1
. That velocity is unknown
and is usually approximated by extrapolating the first measured velocity
back to the surface:

i
1 = j
j
j
i
V
d
2 = t
The initial guess model
2-29
The depth-time table calculated from the sonic log is rarely sufficient to
produce a model impedance which ties the seismic data properly
because:
The seismic datum and log datum may be different.
The average first layer velocity is not known.
Errors in the sonic log velocities produce cumulative errors in the
calculated travel-times.
The events on the seismic data may be mispositioned due to
migration errors.
The seismic data may be subject to time stretch caused by
frequency-dependent absorption and short-period multiples.
We must also correct for deviated well trajectories.

To improve the depth-time table two procedures are used:
Apply check shot corrections.
Apply manual log correlation to the seismic data
The initial guess model
2-30
Check shot corrections
The depth-time table calculated from the sonic log must be
modified to reflect the desired check shot times:
2-31
(1) Change all the
velocities in the log in such
a way that the new log will
integrate to exactly the
desired times.
(2) Change the velocities
for layers between the
first and last check shot
depth only.
(3) Do not change the
velocities in the sonic log
and use the depth-time
table for the conversion
from depth to time.
Check shot corrections - options
2-32
Log correlation
Log correlation is the process of applying a manual correction to the depth-
time curve to optimize the correlation between initial model and seismic
data. Correlation consists of selecting events on the synthetic trace and the
corresponding events on the real trace. The choice of wavelet is crucial.
2-33
(a) The picks used for
the correlation between
the synthetic and the
seismic data.
(b) The correlated
seismic after stretching
and squeezing the log.
Log correlation
2-34
The correlation of the seismic data and the synthetic from the
previous correlation. The symmetry of the correlation shows that the
zero-phase wavelet is correct. The correlation coefficient is 0.86
Log correlation
2-35
(a) The equivalent check-shot
correction for the preceding picks using
a linear interpolation.
(b) The equivalent
check-shot correction
using a spline
interpolation.
Log Correlation
2-36
Interpolating the logs
An impedance model can be built by stretching and squeezing the sonic
and density logs laterally across the seismic volume. Picking two or
more events is equivalent to applying a variable check-shot at each trace.
The material between the two picked events is stretched/squeezed.
2-37
When more than one well is entered into the model, the results are
interpolated using inverse-distance weighting. Using picked events
with multiple logs forces the inverse distance interpolation to be guided
by the picked events.
Interpolating the logs
2-38
Inversion schemes
Now that we have discussed the convolutional
model and the initial guess impedance model, we
will discuss the various inversion algorithms.
We will start with bandlimited recursive inversion.
We will then briefly discuss the new coloured
inversion method.
Next, we will discuss model-based inversion.
Finally, we will discuss sparse-spike inversion.
2-39
Now that we have discussed the general procedure for
computing a model, lets see how we can incorporate the
model using different inversion schemes. We will start
with the recursive inversion approach, shown in the next
slide. Incorporating the model involves the following
steps:

(1) Deconvolve the data using the extracted wavelet.
(2) Scale the seismic trace to true reflectivity.
(3) Recursively invert each trace, using the formula
described earlier.
(4) Add in the low frequency component of the model.
Recursive, bandlimited inversion
2-40
Step 1:

The initial background model for Recursive Inversion
is formed by filtering an impedance log from a well:
10-Hz
High Cut
Recursive, bandlimited inversion
2-41
Step 2:

The recursive equation is applied to the seismic trace.
(Note: this is almost identical to a -90 degree phase
rotation):
1 *
1
1-
i
i i
i
r
Z Z
r
+
+
=
Recursive, bandlimited inversion
2-42
Step 3:

Add the scaled inversion trace to the filtered model to
get the final result:
+
=
Recursive, bandlimited inversion
2-43
Bandlimited inversion
The resulting initial guess model for various settings of the high-cut
frequency is shown below:
2-44
Input seismic
Recursive Inversion
Recursive
Inversion produces
a result which is
bandlimited to the
same frequency
range as the input
seismic data.

Note the loss of
high frequency
detail, as
compared with the
well logs.
Recursive, bandlimited inversion
2-45
Issues in recursive inversion
We assume that the samples in the seismic trace are actually
reflection coefficients. This means that we assume that there is no
seismic wavelet.
We assume that the samples are scaled properly. Reflection
coefficients must be numbers between -1 and +1. Seismic samples
may have any amplitude.
Since the equation is applied recursively from top to bottom of the
trace, the effect of errors is cumulative.
The greatest effect of this cumulative error is in the trend or low-
frequency component of the answer. This trend is so poorly defined
that we remove the trend from the answer and replace it with the
trend from the model.
The process is called bandlimited because the final impedance
traces are defined within the same frequency band as the input
seismic data.
2-46
Scaling the result
The Convolutional Model is used as the basis for all inversion:

Trace = Wavelet * Reflectivity + Noise

In the frequency domain, this can be approximated by:

Reflectivity = Trace / Wavelet

To solve for the reflectivity, the wavelet must be known. Normally,
when a wavelet is extracted, only its shape is known; not its absolute
amplitude. Inversion requires that the absolute amplitude be known
as well. From the equation above, if the wavelet is multiplied by 2,
the resulting reflectivity will be divided by 2.

A good way to determine the scaling of the wavelet automatically is
to force the root-mean-square amplitude of the initial guess synthetic
to be equal to the root-mean-square amplitude of the real trace.
2-47
Scaling the Result
Good Scaling:





Scaling too low:




Scaling too
high:
2-48
The density component
The coefficients, a and b, can be derived by a least-squares fit using all
the wells in your area. Note that the equation is linearized by taking the
logarithm of both sides.
25 . 0
23 . 0 V =
b
aV =
As an arbitrary relationship, we can use the generalized Gardners
equation:
As we have seen, post-stack inversion can only give us the impedance,
not the velocity. If a density log is not available, a common
approximation is to use Gardners equation:
V b a ln ln ln + =
Note that post-stack inversion is incapable of deciding whether a
particular impedance change is a change in velocity, or a change in
density, or both. This procedure assumes that the change is distributed
between the two.
2-49
Examples of bandlimited inversion
In the next few slides, we shall see several
example of bandlimited, recursive inversion.
In our first example, we will use a wedge
model. Although this model is simple, we
know what the right answer should be and
can therefore judge the effectiveness of the
method.
In the second example, we will look at a
carbonate reef from Alberta.
In the reef example, we want to see how well
the method can image the porosity.
2-50
Wedge model
A good illustration of inversion is provided by using the simple wedge
model. The figure on the left shows the basic seismic model with the
velocity log inserted at CDP 45. The wavelet and its amplitude spectrum
are on the right.
2-51
Wedge model
Here is the initial velocity model for the wedge inversion, created by
picking the four major events and squeezing the velocity log. The
colour bar is on the right.
2-52
Wedge model
The above figure shows the bandlimited recursive inversion of the
wedge model. The extra events seen throughout the inversion are due
to the wavelet sidelobes.
2-53
Reef example
Next, we will consider the inversion of a pinnacle reef from Alberta,
Canada. We will first look at one line from a 3D survey that goes over
the discovery well. The seismic section is shown above, with the
synthetic seismogram inserted at the well location. The zone of interest
is at 1150 ms.
2-54
Reef example
Above is shown the initial model for the reef inversion. Notice that the
velocities from the well have be stretched and squeezed laterally using
three picked events. The box indicates the zone displayed in the
following inversions.
2-55
Reef Example
Here is the bandlimited recursive inversion around the zone of interest
using a 10 Hz high-cut filter. Notice the reef porosity below 1160 ms.
2-56
Reef Example
Here is the bandlimited recursive inversion around the zone of interest
using a 20 Hz high-cut filter. Notice that there is more effect from the
model than there was using a 10 Hz high-cut.
2-57
Model based inversion
Model based inversion also follows from the convolutional model:

seismic = wavelet * reflectivity + noise

We assume that:
the seismic trace is known
the wavelet is known
the noise is uncorrelated and random

Model based inversion differs from bandlimited inversion in that the
reflectivity is defined as that sequence which fits the data best. That
is, if we can find a reflectivity which convolves with the wavelet to give
a good approximation to the seismic trace, we assume that this is the
right answer. The theory of this method will be discussed in the next
section.

In practice, Model Based Inversion starts with an initial guess and
improves on it by a series of steps.
2-58
Synthetic
Seismic
Seismic
Data
Update
Impedance
Inversion =
Model
Model based inversion
The above flowchart shows the general implementation of model
based inversion.
Wavelet Acoustic
Impedance
Difference
Small
Error?
Yes
No
2-59
Step 1:

The initial background model for Model Based Inversion is
formed by blocking an impedance log from a well:
The user specifies the layer size
in ms.

All the layers are originally set
to the same size (in time).
Model Based Inversion
2-60
Step 2:

Using the blocked model, and the known wavelet, a
synthetic trace is calculated.
This is compared with the
actual seismic trace.

By analyzing the errors or
mis-fit between synthetic
and real trace, each of the
layers is modified in
thickness and amplitude to
reduce the error.

This is repeated through a
series of iterations.
Synthetic Seismic
Model Based Inversion
2-61
Input seismic
Model Based Inversion
Model Based
Inversion
produces a
broad-band, high
frequency result.

The problem is
that the high
frequency detail
may be coming
from the initial
guess model,
and not from the
seismic data.
Model Based Inversion
2-62
Recursive Inversion
Model Based Inversion
This is a
comparison
between Recursive
and Model Based
Inversion.

Generally, the
Model Based gives
more detail, but the
results are actually
quite similar.
Model Based Inversion
2-63
Issues in Model Based Inversion:

(1) Because the wavelet is known, its effects are removed from the
seismic during the calculation. For example, the seismic does not
have to be zero-phase, as long as the wavelet has the same phase as
the seismic.
(2) Errors in the estimated wavelet will cause errors in the inversion
result.
(3) The effective resolution of the seismic is enhanced.
(4) The result can be very much dependent on the initial guess model.
This can be alleviated by filtering the model.
(5) There is a non-uniqueness problem, as with all inversion.

Model Based Inversion
2-64
Model Based Inversion Issues
(1) Dependence on the wavelet.
Inversion using the correct wavelet:







Inversion using the wrong wavelet:
2-65
Model Based Inversion Issues
(2) Non-uniqueness. For a given wavelet, all these results fit the trace
about equally well:
2-66
Inversion diagnostics
A good diagnostic for the quality of the final model based
inversion result is the error plot, which is the difference between
the seismic data and the final synthetic. A typical error plot is
shown below:
2-67
Examples of model based inversion
In the next few slides, we shall see several
example of model based inversion.
In our first example, as in bandlimited inversion,
we will use a wedge model. Since we know what
the right answer should be, we can therefore
judge the effectiveness of the method.
In the second example, we will look at a
carbonate reef from Alberta.
Again, in the reef example, we want to see how
well the method can image the porosity.
We will also look at the effect of changing
several of the key parameters in model based
inversion.
2-68
Wedge model
The above figure shows the model based inversion of the wedge
model. The result is better than bandlimited recursive inversion, but
starts to break down on the left, when thin bed tuning effects are
encountered.
2-69
This is the original
model:
This model was
created by deleting
the horizon at the
base of the wedge:
Wedge Model
Next, we will consider the effect of changing the initial model, as
shown above.
2-70
Original Model and Inversion
Inversion with horizon removed
Wedge model
2-71
Reef example
Here is the model based inversion around the zone of interest using a
6 ms block size. Notice that there is more detail than in recursive
bandlimited inversion and that the boundaries are more clearly
indicated.
2-72
Reef example
Here is the model based inversion around the zone of interest using a
2 ms block size. Notice that there is more detail than in the previous
inversion, but that this detail may be coming from the model rather
than the seismic.
2-73
Here is the full 3D inversion
of the reef structure, where
the top figure shows a
structure slice averaged
over the reef porosity, and
the figure on the right
shows the inverted result
superimposed on the reef
structure.
Reef example
2-74
Sparse-spike inversion
Sparse-spike inversion attempts to recover a sparse estimate of the
earths reflectivity from the seismic trace. There are two separate
approaches in the literature for performing sparse-spike inversion, the
maximum-likelihood (ML) approach (Chi, C. Y., Mendel, J. M. and
Hampson, D., 1984, A computationally fast approach to maximum-
likelihood deconvolution: Geophysics, 49, 550-565 ) and the Linear
Programming (LP) approach (Oldenburg, D. W., Scheuer, T. and Levy,
S., 1983, Recovery of the acoustic impedance from reflection
seismograms: Geophysics, 48, 1318-1337). Briefly:
Maximum-likelihood sparse-spike inversion performs the
sparse-spike estimate in the time domain by assuming that the
reflectivity has a Poisson distribution.
Linear programming sparse-spike inversion performs the
sparse-spike estimate in the frequency domain by assuming
that the central frequency range is correct and the upper and
lower frequencies constrain the spike series to be sparse.
2-75
Sparse Spike Inversion
Sparse Spike
Inversion assumes
that the actual
reflectivity can be
thought of as a series
of large spikes
embedded in a
background of small
spikes.
Sparse Spike Inversion
assumes that only the
large spikes are
meaningful. It finds the
location of the large
spikes by examining
the seismic trace.
2-76
Sparse Spike
Inversion builds up
the reflectivity
sequence one spike
at a time. Spikes are
added until the trace
is modeled accurately
enough.

The amplitudes of
the impedance blocks
are determined using
the Model Based
Inversion algorithm.

Sparse Spike Inversion
2-77
For maximum-likelihood sparse-spike inversion needs to know the
following information in order to perform inversion:

Maximum Number of Spikes

This tells the algorithm the maximum number of allowable
spikes per trace, and is normally the same as the total
number of samples in the window to be inverted.

Spike Detection Threshold

As each spike is added, its amplitude is compared with the
average amplitude of all spikes detected so far. When the
new amplitude is less than a specified fraction of the
average, the algorithm stops adding spikes.
Sparse-spike inversion
2-78
Input seismic
Sparse Spike Inversion
Sparse Spike
Inversion
produces a
broad-band,
high frequency
result.

Sparse Spike Inversion
2-79
Model Based Inversion
Sparse Spike Inversion
Sparse Spike
Inversion
produces a result
which is similar
to Model Based
Inversion.

The main
difference is that
the very thin
layers are
missing.
Sparse Spike Inversion
2-80
Examples of ML sparse-spike inversion
In the next few slides, we shall see several
example of ML sparse-spike inversion.
In our first example, as in bandlimited inversion,
we will use a wedge model. Since we know what
the right answer should be, we can therefore
judge the effectiveness of the method.
In the second example, we will look at a
carbonate reef from Alberta.
Again, in the reef example, we want to see how
well the method can image the porosity.
We will also look at the effect of changing
several of the key parameters in sparse-spike
inversion.
2-81
Wedge model
The above figure shows the sparse-spike inversion of the wedge
model. The result is similar to model based inversion but shows some
vertical striping. However, it has done better in the thin bed tuning
region.
2-82
Reef example
Here is the sparse-spike inversion around the zone of interest using a
100 spikes and an amplitude threshold of 5%. The result is comparable
to model based inversion, but shows more vertical blocking.
2-83
Reef example
Here is the sparse-spike inversion around the zone of interest using a 100
spikes and an amplitude threshold of 15%. The result is definitely less
realistic than the previous result using a threshold of 5%.
2-84
Sparse-spike inversion summary
Sparse Spike Inversion

Puts events only where the seismic demands.

Attempts to produce the simplest possible model consistent
with the data.

Often produces fewer events than are known to be geologically
true.
Less dependent on initial-guess model

Model-based Inversion

Puts events where the initial guess model (user) demands.

Produces the closest model to the initial guess, which is also
consistent with the seismic data.

Can produce higher-resolution results than supported by
seismic alone.
Subject to non-uniqueness. Dependent on initial-guess model.

2-85
Colored Inversion
Colored Inversion is a modification of Recursive Inversion, which was
originally described by Lancaster and Whitcombe of BP at the 2000 SEG
Convention.

In this process, there is a single operator, O, which is applied to the seismic
trace to transform it directly into the inversion result:
* I O S =
The authors defined the operator, O, in the frequency domain.

By examining transforms between seismic data and actual inversion
results, they concluded that the operator phase is -90 degrees.
2-86
The amplitude spectrum of the operator is derived this
way:
Using a set of wells
from the area, the
amplitude spectra of the
acoustic impedance for
all the wells are plotted
on a log-log scale.

As predicted by theory,
we can fit a straight line
which represents the
desired output
impedance spectrum.
Log(Frequency)
L
o
g
(
I
m
p
e
d
a
n
c
e
)

Amplitude Spectrum of Acoustic Impedance
Colored Inversion
2-87
Then, using a set of seismic
traces from around the wells,
the average seismic spectrum is
calculated.
Spectrum of Seismic Data
Frequency (Hz)
Operator Spectrum
From the two preceding spectra, the
operator spectrum is calculated.
This has the effect of shaping the
seismic spectrum to the impedance
spectrum within the seismic band.
Colored Inversion
2-88
Colored Inversion Operator
Time (ms)
Putting together the derived
amplitude spectrum with the -
90 degree phase shift
produces the Colored
Inversion Operator.

This is applied to all the
seismic traces by
convolution.
Colored Inversion
2-89
Input seismic
Colored Inversion
Colored Inversion
produces a result very
similar to Recursive
Inversion.

One difference is that
the scale is relative
Acoustic Impedance,
with positive and
negative values.

-3000
0
+3000
Colored Inversion
2-90
Recursive Inversion
Colored Inversion



-3000
0
+3000
4600
8300
12000
Relative AI
Absolute AI
Colored Inversion
2-91
Issues in Colored Inversion

(1) Very little dependence on the initial model, except to determine the
general impedance trend.
(2) Very fast to apply.
(3) Very simple with few user parameters.
(4) Assumes the data is zero-phase.
(5) Produces a relative impedance result.


Colored Inversion
2-92
AVO Inversion
The basic convolutional model assumes zero-offset data.

Conventional inversion should not be applied to data with AVO effects,
since changes in V
P
/V
S
are not explicitly accounted for.

To extend inversion to handle AVO data, these algorithms are currently
used:
(1) Elastic Impedance
(2) Lambda-Mu-Rho (LMR)
(3) Simultaneous Inversion
AVO inversion will be discussed later in this course.
2-93
Blackfoot case study
We will now illustrate the preceding concepts using a channel sand case
study from Alberta, the Blackfoot 3-D survey, with the following
parameters:

Location: Southern Alberta, Canada
Recorded: October, 1995
Target: the Glauconitic member of the Mannville group.
Reservoir: sand channel at depth of 1550 m.
13 wells tie the 3-D volume.

The following slides, on the geology of the area, are taken from the paper:
Integrated geological and geophysical case study, and Lam rock
parameter extractions using AVO analysis on the Blackfoot 3C-3D
seismic data, southern Alberta, Canada by Dufour et al., Geophysics
Vol 67, no 1, Jan-Feb, 2002, p 27-37.
2-94
The Blackfoot survey
This map shows the location of the Blackfoot survey area, with the portion
used in this study outlined in red. The objective, a Glauconitic channel
within the Lower Cretaceous Mannville formation, is shown running north-
south on the map.
Alberta
Calgary
N
Channel
2-95
Blackfoot case study
This figure shows the
paleogeography of
the Lower Cretaceous
during the formation
of the Glauconitic
incised-valley system.
The system was
fluvial at the present
Canada/U.S. border
and estuarine at the
Hoadley barrier bar.
The boreal sea was
retreating northward.
(Dufour et al.)

2-96
Blackfoot case study
This figure shows the schematic stratigraphy of the Blackfoot area,
showing three different incised valleys. The relative age is also
indicated, where 30 is oldest and 40 is youngest. (Dufour et al.)
2-97
Blackfoot case study
The wave-dominated facies model that was proposed to explain the
formation of the incised valley. The model is divided into three
environments: marine-dominated, mixed-energy, and river-dominated.
(Dufour et al.)
2-98
Blackfoot case study
An index map of the area showing seismic cross-line 95, and
two east-west cross-sections. The wells are also indicated.
2-99
Blackfoot case study
Seismic cross-line 95 from the PP data, showing a clear
indication of the three valleys. (Dufour et al.)
2-100
Blackfoot case study
Extracted amplitude slices from the PP data, where (a) was extracted
from the upper valley (40), and (b) from the lower valley (30). The white
outlines shown the outline of the valley and the anomalous amplitudes
are defined by the red outlines. (Dufour et al.)
(a) (b)
2-101
Blackfoot case study
This southern east-west geological cross-section and facies
interpretation. (Dufour et al.)
2-102
Blackfoot case study
The synthetic southern east-west geological cross-section showing the
expected seismic response of the channels. (Dufour et al.)
2-103
Blackfoot case study
This northern east-west geological cross-section and facies
interpretation. (Dufour et al.)
2-104
Blackfoot case study
The synthetic northern east-west geological cross-section showing the
expected seismic response of the channels. (Dufour et al.)
2-105
Blackfoot case study
- Now that we have had a detailed look at the
regional stratigraphy of the area, we will apply the
inversion techniques discussed in the previous
section to the seismic volume, and see if we can
get a more complete picture of the channel
system.

- We will start by tieing the well data to the seismic,
and then apply the various inversion techniques
that were discussed.
2-106
Blackfoot case study
The map on the left
shows the survey area
from the previous map
in more detail, with both
the wells and the
position of a seismic
line that intersects three
of the wells. This
seismic line is shown in
the next slide.
2-107
Blackfoot case study
The figure above shows one of the seismic lines from the 3D
volume, with both the sonic logs from the three intersecting wells,
and two picked seismic events.
2-108
Blackfoot case study
The figure above shows a blowup of the seismic line from the
previous slide. The channel sand is at the base of each sonic log.
2-109
Blackfoot case study
The figure above shows the visual tie between the synthetic seismogram from
well 08-08 and the corresponding seismic trace. The next two slides show the
seismic wavelet, which was zero-phase, and the cross-correlation function.
2-110
Blackfoot case study
The figures on the right
show the wavelet used
for the previous seismic
tie, where (a) shows the
time domain response
of the wavelet and (b)
shows its frequency
domain response. Note
that this is a zero-phase
wavelet, and that the
amplitude spectrum
was estimated from the
seismic data.
(a)
(b)
2-111
Blackfoot case study
The figure above shows the cross-correlation between the synthetic
seismogram from well 08-08 and the corresponding seismic trace. Note
that the correlation coefficient is 0.69 and there appears to be a phase shift.
In the next slides, we will re-extract the wavelet using a match filter between
the well and seismic data.
2-112
Blackfoot case study
The figure above shows the visual tie between the synthetic seismogram from
well 08-08 and the corresponding seismic trace, using a non-zero phase wavelet.
The next two slides show the wavelet and the cross-correlation function.
2-113
Blackfoot case study
The figures on the right
show the wavelet used
for the previous seismic
tie, where (a) shows the
time domain response of
the wavelet and (b)
shows its frequency
domain response. This
is a nonzero-phase
wavelet, estimated using
a match filter between
the seismic trace and
well log derived
reflection coefficients.
(a)
(b)
2-114
Blackfoot case study
The figure above shows the new cross-correlation between the synthetic
seismogram from well 08-08 and the corresponding seismic trace. Note
that the correlation coefficient is now 0.704 and there is a good phase tie.
2-115
Blackfoot case study
The figure above shows the initial model used for the inversions that will
follow. The impedance from each of the wells has been interpolated,
and the two seismic picks have been used to stretch the logs laterally.
2-116
Blackfoot case study
The map on the left
shows the model
impedance values
computed over the
complete map, using a
10 ms arithmetic
average centered 20
ms below the Lower
Mannville pick. (This
should correspond to
the channel zone).
2-117
Blackfoot case study
The figure above shows the model-based inversion for one of the
seismic lines in the 3D survey. A block size of 4 ms has been used.
2-118
Blackfoot case study
The figure above shows the residual error between the seismic and
synthetic seismic data for model-based inversion for one of the
seismic lines in the 3D survey.
2-119
Blackfoot case study
The map on the left
shows the model-based
inversion result
computed over the
complete map, using a
10 ms arithmetic
average centered 20
ms below the Lower
Mannville pick. (This
should correspond to
the channel zone).
2-120
Blackfoot case study
The figure above shows the recursive bandlimited inversion for one of
the seismic lines in the 3D survey. A high frequency cutoff of 15 Hz
has been used in the inversion.
2-121
Blackfoot case study
The map on the left
shows the bandlimited
inversion result
computed over the
complete map, using a
10 ms arithmetic
average centered 20
ms below the Lower
Mannville pick. (This
should correspond to
the channel zone).
2-122
Blackfoot case study
The figure above shows the linear programming sparse-spike inversion
for one of the seismic lines in the 3D survey. A sparseness of 100%,
constraint frequency of 15 Hz, and window length of 128 samples have
been used for the inversion.
2-123
Blackfoot case study
The map on the left
shows the linear
programming sparse-
spike inversion result
computed over the
complete map, using a 10
ms arithmetic average
centered 20 ms below the
Lower Mannville pick.
(This should correspond
to the channel zone).
2-124
Blackfoot case study
The figure above shows a comparison of the three inversion results,
where (a) shows the model-based inversion, (b) shows the bandlimited
inversion, and (c) shows the sparse-spike inversion. The channel is
best represented in (a).
(a) (b) (c)
2-125
Blackfoot case study
The figure above shows a comparison of the two inversion results and
the amplitude envelope, where (a) shows the model-based inversion,
(b) shows the amplitude envelope, and (c) shows the sparse-spike
inversion. The channel is best represented in (a).
(a) (b) (c)
2-126
Conclusions
This has been a overview of the post-stack
inversion approach and three specific methods
which are used for the inversion itself.
We first discussed the common factors in all three
methods, which involved getting a good wavelet
estimate and building an accurate model.
We then considered the simplest method,
bandlimited recursive inversion, with examples.
After this, we looked at model based inversion, and
looked at examples.
Finally, we discussed sparse-spike inversion, and
looked at examples.
In general, the model based approach produces the
most geologically consistent results.
2-127
Appendix Post-stack inversion theory
In this appendix, we will discuss the theory
of post-stack inversion in more detail.
We will focus on the theory of model-based
inversion.
The theory will be illustrated with several
numerical examples and exercises.
2-128
Recall that the basic equation relating reflectivity, R, to
impedance, Z, for the i
th
interface in a stack of N layers is as
follows (elastic impedance will be discussed later):
Acoustic Impedance =

Shear Impedance =

Elastic Impedance =
,
1
1
i i
i i
i
Z Z
Z Z
= R
+

+
+
P
V
S
V
Acoustic Impedance
or
Shear Impedance
or
Elastic Impedance
= Z
. ) / (
,
2
) sin 4 1 ( ) sin 8 ( ) sin 1 (
2 2 2
P S
K K
S P
V V K
V V
=
+ u u u

General Forward Model for Inversion


2-129
A useful approximation to R can be derived by noting that:
Reflectivity approximations
.
2
where ,
2
1
1
1 i i
i
i
i
i i
i i
i
Z Z
Z
Z
Z
Z Z
Z Z
= R
+
=
A
~
+

+
+
+
) (
) (
)) ( ln(
) (
) (
1 )) ( ln(
t Z
t dZ
t Z d
dt
t dZ
t Z dt
t Z d
= =
From basic calculus, we know that:
). ln (ln
2
1
2
ln
1 i i
i
i
Z Z
Z
R =
A
~
+
Replacing the derivative d with the difference operator A gives:
2-130
The forward model using matrices
Reflectivity can therefore be derived from the logarithm of
impedance using the matrix operation given by:
DL R
Z L
L
L
L
R
R
R
i i
N N
=
=
(
(
(
(

(
(
(
(

=
(
(
(
(

: or
). ln( where ,
5 . 0 5 . 0 0 0
0
0 5 . 0 5 . 0 0
0 0 5 . 0 5 . 0
2
1
1
2
1

Note that the difference matrix has N-1 rows and N


columns, since there are N layers but only N-1 reflection
coefficients.
2-131
Example
As an example for the three layer case, we can write the
matrix equation as:
( )
( )
(
(
(

=
(
(
(

=
(

2 3
1 2
3
2
1
2
1
ln ln
2
1
ln ln
2
1

5 . 0 5 . 0 0
0 5 . 0 5 . 0
Z Z
Z Z
L
L
L
R
R
2-132
Exercise 2-2
Assume that a three layer earth has these impedances:
Z
1
= Z
3
= 5500 m/s*g/cc, and Z
2
= 4500 m/s*g/cc.
(Note: L
1
= L
3
= lnZ
1
= 8.612, and L
2
= lnZ
2
= 8.412)
Compute the reflectivities using matrix multiplication:
(

=
(
(
(

=
(
(
(

=
(

___
___
____
____
____
5 . 0 5 . 0 0
0 5 . 0 5 . 0

5 . 0 5 . 0 0
0 5 . 0 5 . 0
3
2
1
2
1
L
L
L
R
R
Plot the result on the next page.
2-133
Exercise 2-2 Plot
4000 5000 6000
Impedance (m/s*g/cc)
t
1
t
2
Reflectivity
0 0.05 0.1 -0.05 -0.1
Z
2
= 4500
Z
1
= 5500
Z
3
= 5500
2-134
Compute the full solution with the matrix multiplication:
(

=
(
(
(

=
(
(
(

=
(

1 . 0
1 . 0
612 . 8
412 . 8
612 . 8
5 . 0 5 . 0 0
0 5 . 0 5 . 0

5 . 0 5 . 0 0
0 5 . 0 5 . 0
3
2
1
2
1
L
L
L
R
R
Exercise 2-2 Answer
2-135
Exercise 2-2 Plot
4000 5000 6000
Impedance (m/s*g/cc)
t
1
t
2
Reflectivity
0 0.05 0.1 -0.05 -0.1
Z
2
= 4500
Z
1
= 5500
Z
3
= 5500
2-136
To then go from reflectivity to the seismic trace, we use the
convolutional model:
Seismic = Wavelet convolved
with reflectivity plus noise.
Assumptions:

(1) There are no multiples modeled.
(2) Transmission loss and geometric spreading are
ignored.
(3) Frequency-dependent absorption is ignored.
(4) The wavelet may be time varying.
From reflectivity to seismic
noise R W T + = *
2-137
Convolution
Convolution can be conveniently represented by the following matrix
equation, where the j
th
column in the matrix represents the wavelet to
be convolved with the j
th
reflection coefficient.
(
(
(
(

(
(
(
(
(
(
(

=
(
(
(
(

M
K
K
K
N
R
R
R
W
W W
W
W
W
T
T
T

2
1
1
1
1
2
1
0 0
0
0
0 0
In the full convolution shown above, N = K + M -1, where N is the output
length, K is the wavelet length and M is the input length. In other words,
the output is longer than the input by the wavelet length minus one sample.
2-138
Example
Here is an example in which the length of the wavelet is K = 3, the
length of the reflection coefficient series is M = 4, and therefore the
length of the seismic trace is N = K + M 1 = 6:
(
(
(
(
(
(
(
(

=
(
(
(
(

(
(
(
(
(
(
(
(

6
5
4
3
2
1
4
3
2
1
3
2 3
1 2 3
1 2 3
1 2
1
0 0 0
0 0
0
0
0 0
0 0 0
T
T
T
T
T
T
R
R
R
R
W
W W
W W W
W W W
W W
W
2-139
Here is a pictorial example of convolution using a
simple wavelet and three layer earth. Note that
the reflections are quite far apart compared with
the wavelet length.
Convolving with a wavelet
Wavelet
-1 0 1 -2
0
2
4
6
8
2
t
i
m
e

(
m
s
)

Impedance
(m/sg/cc)
Reflectivity Seismic Trace
3500 5000 0 0 0.1 - 0.1 0.2 - 0.2
Geology
t
i
m
e

(
m
s
)

0
2
4
6
8
10
12
14
- 0.1 0.1
2-140
Wavelet tuning
Impedance
(m/sg/cc)
Reflectivity Seismic Trace
3500 5000 0 0 0.1 - 0.1 0.3
- 0.3
Geology
t
i
m
e

(
m
s
)

0
2
4
6
8
10
Wavelet
-1 0 1 -2
0
2
4
6
8
2
t
i
m
e

(
m
s
)

Here is an example in which the reflections have
been moved much closer together compared with
the previous example (they are now one sample
apart). Note that the result is a tuned response,
in which the result looks like a phase-rotated
wavelet. Let us now do the calculations.
2-141
Exercise 2-3: Convolving with a wavelet
In exercise 2-2, we found that reflection coefficients were -0.1 and +0.1.
If we let W = (-1, 2, -1) and the reflectors be three samples apart, we
can convolve W with R using the matrix multiplication: T = WR.
Compute the seismic values and graph the seismic trace on the right,
putting zeros at times 0 and 14 ms.
(
(
(
(
(
(
(
(

=
(
(
(
(

(
(
(
(
(
(
(
(

__
__
__
__
__
__
1 0
0
0
1 0
1 0 0 0
2 1 0 0
1 2 1 0
0 1 2 1
0 0 1 2
0 0 0 1
.
.
Amplitude
0 0.1 0.2 -0.1 -0.2
0
2
4
6
8
10
12
14
t
i
m
e

(
m
s
)

2-142
Exercise 2-3 Solution
Seismic Trace
0 0.1 0.2 -0.1 -0.2
0
2
4
6
8
10
12
14
t
i
m
e

(
m
s
)

(
(
(
(
(
(
(
(

+
=
(
(
(
(

(
(
(
(
(
(
(
(

1 . 0
2 . 0
1 . 0
1 . 0
2 . 0
1 . 0
1 0
0
0
1 0
1 0 0 0
2 1 0 0
1 2 1 0
0 1 2 1
0 0 1 2
0 0 0 1
.
.
2-143
Notice that the trace looks like the one below, and both wavelets are
separate. In other words, convolution has scaled the wavelet by the
value of each reflection coefficient (in this case equal amplitude,
opposite sign), and shifted the scaled wavelet to the location of the
reflection coefficient.
Convolving with a wavelet
Seismic Trace
0 0.1 0.2 -0.1 -0.2
First scaled
wavelet
Second scaled
wavelet
0
2
4
6
8
10
12
14
t
i
m
e

(
m
s
)

(
(
(
(
(
(
(
(

+
=
(
(
(
(

(
(
(
(
(
(
(
(

1 . 0
2 . 0
1 . 0
1 . 0
2 . 0
1 . 0
1 0
0
0
1 0
1 0 0 0
2 1 0 0
1 2 1 0
0 1 2 1
0 0 1 2
0 0 0 1
.
.
2-144
The effect of zeros
In the previous convolution, notice that a zero reflection coefficient has
no effect on the resulting seismic trace. We can therefore remove the
zero reflection coefficients and the corresponding columns in the
wavelet matrix to get the equivalent operation written below:
(
(
(
(
(
(
(
(

=
(

(
(
(
(
(
(
(
(

1 . 0
2 . 0
1 . 0
1 . 0
2 . 0
1 . 0
1 0
1 0
1 0
2 0
1 0
0 1
0 2
0 1
.
.
2-145
Exercise 2-4 Wavelet tuning
Now, we will let the coefficients be one sample apart. Re-compute the
seismic trace using the matrix method just described. Draw the new
trace on the right, again putting zeros in the first and last samples.
How does this new trace compare with the original trace?
(
(
(
(

=
(

(
(
(
(

_
_
_
_
1 0
1 0
1 0
2 1
1 2
0 1
.
.
Amplitude
-0.1 0 0.1 -0.2 -0.3
0
2
4
6
8
10
0.2 0.3
t
i
m
e

(
m
s
)

2-146
(
(
(
(

+
=
(

(
(
(
(

1 . 0
3 . 0
3 . 0
1 . 0
1 . 0
1 . 0
1 0
2 1
1 2
0 1
Exercise 2-4 Solution
Amplitude
-0.1 0 0.1 -0.2 -0.3
0
2
4
6
8
10
0.2 0.3
t
i
m
e

(
m
s
)

2-147
Wavelet tuning
(
(
(
(

=
(

(
(
(
(

1 . 0
3 . 0
3 . 0
1 . 0
1 . 0
1 . 0
1 0
2 1
1 2
0 1
Note that the resulting trace is tuned, meaning that the two zero-
phase wavelets have been merged into a single ninety-degree
wavelet (Phase will be discussed shortly).
Amplitude
-0.1 0 0.1 -0.2 -0.3
0
2
4
6
8
10
0.2 0.3
t
i
m
e

(
m
s
)

2-148
Deconvolution Theory
) ( ) ( e e
T
WR T WR T J
T
= =
To deconvolve the seismic trace, we minimize the squared error
between the observed seismic and the modeled seismic, or:
The solution is given by:
matrix. identity the is and
factor whitening - pre a is where
, ) (
1
I
T W I W W R
T T



+ =
In the above equation, W
T
W is the autocorrelation of the wavelet and
W
T
T is the cross-correlation of the seismic trace with the wavelet. This
is often referred to as the Weiner-Levinson equation, and can be
solved using the Levinson recursion algorithm.
2-149
Deconvolution example
We can deconvolve the seismic trace exactly if we know the wavelet
exactly and also the positions of the reflection coefficients. This can
be done using the generalized inverse solution R = (W
T
W)
-1
W
T
S (no
pre-whitening is needed here), as shown below:
(

=
(

=
(

=
(
(
(
(
(
(
(
(

(
(
(
(
(
(
(
(

1 . 0
1 . 0
6 . 0
6 . 0
6 / 1 0
0 6 / 1
6 . 0
6 . 0
6 0
0 6
1 . 0
2 . 0
1 . 0
1 . 0
2 . 0
1 . 0
1 2 1 0 0 0
0 0 0 1 2 1
1 0
2 0
1 0
0 1
0 2
0 1
1 2 1 0 0 0
0 0 0 1 2 1
1
1
R
R
2-150
Exercise 2-5 Deconvolution
Next, deconvolve the seismic trace from exercise 2-4,
using the starting point shown below:
=
(
(
(
(

(
(
(
(

1 . 0
3 . 0
3 . 0
1 . 0
1 2 1 0
0 1 2 1
1 0
2 1
1 2
0 1
1 2 1 0
0 1 2 1
1
R
(

=
(


a c
b d
bc ad d c
b a
Note
1
:
1
2-151
Exercise 2-5 solution
Again, we can deconvolve the seismic trace exactly if we know the
wavelet matrix exactly (that is, if we know the wavelet values and the
positions of the reflection coefficients).
(

=
(

=
(


=
(
(
(
(

(
(
(
(

1 . 0
1 . 0
0 . 1
0 . 1
6 4
4 6
20
1
0 . 1
0 . 1
6 4
4 6
1 . 0
3 . 0
3 . 0
1 . 0
1 2 1 0
0 1 2 1
1 0
2 1
1 2
0 1
1 2 1 0
0 1 2 1
1
1
R
R
2-152
Recursive Inversion
Recall that the reflection coefficient for the i
th
interface of N layers can
be written as shown on the left below. The impedance of the i+1
st
layer
can then be determined from the i
th
layer by inverting the reflection
equation to give the equation on the right, called recursive inversion.
(

+
+
+
i
i
i i
i i
i i
i
R
+ R
= Z Z
Z Z
Z Z
= R
1
1
1
1
1
For N layers, we can start at the first layer and compute the impedance
of each successive layer by recursively applying the inversion formula:
terms. all of product the where
1
1
1
1
1
= H

+
[

=
,
R
R
N
N
i i
i
= Z Z
2-153
Linearized inverse approximation
( ) ( )
.
5 3
2
1
1
ln : since
,
5 3
2
1
1
ln ln ln
5 3
1
1
5 3
1
1
1
(

+ + + =
(

+
(

+ + + =
(

+
+ =


=

x x
x
x
x
R R
R
R
R
Z Z
N
i
i i
i
N
i
i
i
N
Taking the logarithm of each side of the previous equation, we get
| |
1
1
1
1
*
ln exp 2 ln ln Z L Z R Z Z L
N N
N
i
i N N
+ = ~ =

=
Since reflection coefficients are usually of the order 0.1 or less, we can
drop the higher terms in the series to give:
2-154
Matrix implementation
The logarithm of impedance can be derived from the reflectivity using the
matrix operation given by:
HR L
Z Z L
R
R
R
L
L
L
i i
N
N
=
=
(
(
(
(

(
(
(
(

=
(
(
(
(
(

*
1
*
1
2
1
*
*
3
*
2
: or
). ln( ) ln( where ,
2 2 2
0
2 2
0 0 2

Note that the integrating matrix H has N rows and N columns, unlike the
difference matrix, since we must assume the first layer value is known.
2-155
L versus L*
Note that when we estimated reflectivity from impedance, we
let the symbol L
i
represent the logarithm of impedance, or:
). ln( ) ln(
1
*
Z Z L
i i
=
) ln(
i i
Z L =
However, when we estimated impedance from reflectivity, we
used the symbol L
i
*, which was equal to L
i
L
1
, or:
This is because the process of transforming from impedance
removes the low frequency DC component, as can be seen
by the simple example on the next page.
2-156
The loss of low frequencies
If we consider the simple case where the impedance (and
logarithm of impedance) is equal to a constant, for example
1, we get:
(
(
(
(

=
(
(
(
(

(
(
(
(

=
(
(
(
(

0
0
0

1
1
1
5 . 0 5 . 0 0 0
0
0 5 . 0 5 . 0 0
0 0 5 . 0 5 . 0
1
2
1

N
R
R
R
That is, the result is equal to zero. In other words, we have
destroyed the dc component, or zero frequency, and can
never recover it.
2-157
The forward and inverse approaches
If we invert the H matrix, we get the D matrix with the first column
missing.
(
(
(

=
(
(
(

(
(
(

(
(
(

=
(
(
(

(
(
(

(
(
(

=
(
(
(

3
2
1
3
2
1
*
4
*
3
*
2
1
1
2 2 2
0 2 2
0 0 2
5 . 0 5 . 0 0
0 5 . 0 5 . 0
0 0 5 . 0

2 2 2
0 2 2
0 0 2
: that so ,
5 . 0 5 . 0 0
0 5 . 0 5 . 0
0 0 5 . 0
2 2 2
0 2 2
0 0 2
R
R
R
R
R
R
L
L
L
As expected, H and D are inverses of each other. Also, it doesnt
matter that we use L* rather than L.
R HR H DL L H R = = = =
1 * * 1
For N = 3, we get:
2-158
Exercise 2-6
In a three-layer earth, assume that we have the
following values for the two reflection coefficients

R
1
= -0.1 and R
2
= +0.1.

On the next two pages, compute the impedances in the
three layers using both methods just discussed, and
then plot the results.

Note: You will need to know the first layer impedance.
Assume that Z
1
= 5500 m/s*g/cc and lnZ
1
= 8.612.


2-159
Exercise 2-6 Calculations (1)
4000 5000 6000
Impedance (m/s*g/cc) Reflectivity
0 0.05 0.1 -0.05 -0.1
=
(

1
1
1 2
1
1
R
+ R
= Z Z
=
(

2
2
2 3
1
1
R
+ R
= Z Z
cc g s m Z / / 5500
1
=
2-160
Exercise 2-6 Solution (1)
Reflectivity
0 0.05 0.1 -0.05 -0.1
4500
1 . 1
9 . 0
5500
1
1
1
1
1 2
=
(

=
(

R
+ R
= Z Z
5500
9 . 0
1 . 1
4500
1
1
2
2
2 3
=
(

=
(

R
+ R
= Z Z
cc g s m Z / / 5500
1
=
4000 5000 6000
Impedance (m/s*g/cc)
2-161
Exercise 2-6 Calculations (2)
4000 5000 6000
Impedance (m/s*g/cc) Reflectivity
0 0.05 0.1 -0.05 -0.1
Z 612 . 8 ln
1
=
| | = + =
= =

=
1 2 2
1
1
2
ln exp
2
Z L Z
R L
i
i
| | = + =
= =

=
1 3 3
2
1
3
ln exp
2
Z L Z
R L
i
i
2-162
Exercise 2-6 Solution (2)
Reflectivity
0 0.05 0.1 -0.05 -0.1
Z 612 . 8 ln
1
=
| | | | 4500 412 . 8 exp ln exp
2 . 0 2
1 2 2
1
1
2
= = + =
= =

=
Z L Z
R L
i
i
| | | | 5500 612 . 8 exp ln exp
0 2
1 3 3
2
1
3
= = + =
= =

=
Z L Z
R L
i
i
4000 5000 6000
Impedance (m/s*g/cc)
2-163
Model Based Inversion Theory
Recall from our earlier discussion of forward modeling that:
impedance. the of logarithm the and
matrix, difference a ty, reflectivi the where
, ) 2 (
=
= =
=
L
D R
DL R
ty. reflectivi the and
matrix, wavelet the trace, seismic the where
, ) 1 (
=
= =
=
R
W T
WR T
Combining (1) and (2) gives:
, ) 3 ( ML WDL T = =
where M = the product of the wavelet and difference matrices.
2-164
Model Based Inversion Theory
) ( ) ( e e
T
ML T ML T J
T
= =
To solve for the impedance, we minimize the squared error
between the observed seismic and the modeled seismic,
or:
This is given by:
matrix. identity the is I and
factor whitening - pre a is where
, ) (
1

T M I M M L
T T
+ =
2-165
Using the numerical examples given in Exercises 3-1 and 3-3, compute the
trace using the forward model just described:
(
(
(
(

=
(
(
(

(
(
(
(

=
(
(
(

(
(
(
(

=
(
(
(
(

___
___
___
___
612 . 8
412 . 8
612 . 8
___ ___ ___
___ ___ ___
___ ___ ___
___ ___ ___
612 . 8
412 . 8
612 . 8
5 . 0 5 . 0 0
0 5 . 0 5 . 0
1 0
2 1
1 2
0 1
4
3
2
1
T
T
T
T
Exercise 2-7
Remember that Z
1
= Z
3
= 5500, and Z
2
= 4500.
2-166
1 . 0
3 . 0
3 . 0
1 . 0
612 . 8
412 . 8
612 . 8
5 . 0 5 . 0 0 . 0
0 . 1 5 . 1 5 . 0
5 . 0 5 . 1 0 . 1
0 . 0 5 . 0 5 . 0
612 . 8
412 . 8
612 . 8
5 . 0 5 . 0 0
0 5 . 0 5 . 0
1 0
2 1
1 2
0 1
4
3
2
1
(
(
(
(

=
(
(
(

(
(
(
(



=
(
(
(

(
(
(
(

=
(
(
(
(

T
T
T
T
Exercise 2-7 Answers
2-167
The results of the previous exercise are as follows:

1 . 0
3 . 0
3 . 0
1 . 0
612 . 8
412 . 8
612 . 8
5 . 0 5 . 0 0 . 0
0 . 1 5 . 1 5 . 0
5 . 0 5 . 1 0 . 1
0 . 0 5 . 0 5 . 0
4
3
2
1
(
(
(
(

=
(
(
(

(
(
(
(



=
(
(
(
(

=
T
T
T
T
ML T
Numerical inversion
Direct inversion gives us:
(
(
(

(
(
(



=
(
(
(

5 . 0
0 . 1
5 . 0
5 . 1 5 . 2 0 . 1
5 . 2 0 . 5 5 . 2
0 . 1 5 . 2 5 . 1
) (
1
3
2
1
1
L
L
L
T M M M L
T T
2-168
Unfortunately, the matrix just given has a zero determinant, and
therefore no inverse. However, it was mentioned that we can stabilize
the solution using a small amount of prewhitening, or:
Numerical inversion
matrix. identity the and ng prewhiteni where
, ) (
1
= =
+ =

I
T M I M M L
T T


(
(
(

=
(
(
(

(
(
(



=
(
(
(

=
(
(
(


067 . 0
133 . 0
067 . 0
5 . 0
0 . 1
5 . 0
51 . 1 5 . 2 0 . 1
5 . 2 01 . 5 5 . 2
0 . 1 5 . 2 51 . 1
1
*
3
*
2
*
1
3
2
1
L
L
L
L
L
L
For example, letting = 0.01 gives:
2-169
Obviously, the solution just given has failed to recover the
absolute values of the impedances, (which is why we wrote
L = L*). This is due to the problem discussed earlier, in
which the low frequency component is not recoverable.
There are two ways to solve this problem:
Recovering the low frequencies

(1) Adding back the first layer impedance, as discussed
earlier.

(2) Start with a model of the impedance and use an
iterative inversion technique such as the gradient
descent or conjugate gradient method to solve for the
final impedance.
2-170
Let us add back the first layer impedance and see if we get
the correct result:
Adding back the first layer impedance
Note that we also had to set the first value to zero, since the
computed values have a small bias.
5500 ) 067 . 0 612 . 8 067 . 0 exp(
4500 ) 067 . 0 612 . 8 133 . 0 exp(
5500 ) 067 . 0 612 . 8 067 . 0 exp(
612 . 8 ) 5500 ln( ,
067 . 0
133 . 0
067 . 0
3
2
1
1
*
3
*
2
*
1
= + =
= + =
= + =
= =
(
(
(

=
(
(
(

Z
Z
Z
L
L
L
L
2-171
Recall that our forward model was:
The gradient descent algorithm
This equation can be solved by minimizing the function:
T ML WDL = =
Since this was an overdetermined system, we first multiply
by the transpose of M to get the symmetric equation:
. and , , : where
, or
L x T M b M M A
b Ax T M ML M
T T
T T
= = =
= =
,
2
1
) ( b x Ax x x
T T
= |
. ) ( b Ax g x = = V |
with the negative gradient given by:
2-172
The gradient descent algorithm solves for x iteratively by first
making an initial guess x
0
. The starting gradient is thus:
The gradient descent algorithm
0 0
Ax b g =
We next update x by adding a scaled version of the gradient:
. where ,
1 1
1 1
1 1



= + =
k
T
k
k
T
k
k k k k k
Ag g
g g
g x x o o
We therefore approach the true value of x in N steps.
k k
Ax b g =
The new gradient is now:
2-173
Although the gradient descent algorithm converges after
the first iteration for simple problems, in general it is not
the best algorithm to use.

This is because the convergence is usually very slow for
large, sparse matrices of the kind we are dealing with in
seismic inversion.

In a later section we will discuss the conjugate gradient
algorithm, which converges much faster and is the
preferred algorithm for this type of problem.
Limitations of gradient descent
2-174
Conclusions
In this chapter, we looked at the theory of post-
stack inversion.
We discussed forward modelling, convolution,
deconvolution, and the theory of model-based
inversion.
We found that using the full matrix inverse
approach to model-based inversion does not allow
us to recover the low frequency component of the
initial impedance.
Recovery of the low frequency component can be
done by building an initial model and using the
conjugate gradient approach to iterate towards a
solution.

You might also like