Introduction
to
Seismic Inversion Methods
Brian H. Russell
Hampson-Russell
Software
Services,
Ltd.
Calgary,
Alberta
Course Notes Series, No. 2
S. N. Domenico, Series Editor
These course notes are published without the normal SEG peer reviews. They have not been examined for accuracy and clarity. Questions or comments by the reader should be referred directly to the author.
ISBN 978-0-931830-48-8 (Series) ISBN 978-0-931830-65-5 (Volume)
Library of Congress Catalog Card Number 88-62743 Society of Exploration Geophysicists
P.O. Box 702740
Tulsa, Oklahoma 74170-2740
¸ 1988 by the Society of Exploration Geophysicists
All rights reserved. This book or portions hereof may not be reproduced in any form without permission in writing from the publisher.
Reprinted 1990, 1992, 1999, 2000, 2004, 2006, 2008, 2009
]: nl;roduc t1 on •o Selsmic I nversion •thods Bri an Russell
Table of Contents
PAGE
Part I Introduction 1-2
Part Z The Convolution Model 2-1
Part 3
Part 4
Part 5
P art 6
P art 7
2.1 Tr•e Sei smic Model
2.2 The Reflection Coefficient Series
2.3 The Seismic Wavelet
2.4 The Noise Component Recursive Inversion - Theory
3.1 Discrete Inversion
3.2 Problems encountered with real
3.3 Continuous Inversion
data
Seismic Processing Consi derati ons
4. ! I ntroduc ti on
4.2 Ampl i rude recovery
4.3 Improvement of vertical
4.4 Lateral resolution 4.5 Noise attenuation
resolution
Recursive Inversion - Practice
5.1 The recursive inversion method
5.2 Information in the low frequency component
5.3 Seismically derived porosity Sparse-spike Inversi on
6.1 I ntroduc ti on
6.2 Maximum-likelihood aleconvolution and inversion
6.3 The L I norm method 6.4 Reef Problem
I nversi on appl ied to Thi n-beds
7.1 Thin bed analysis
7.Z Inversion compari son of thin beds
Model-based Inversion
B. 1 I ntroducti on .
8.2 Generalized linear inversion
8.3 Seismic 1 ithologic roodell
ing (SLIM)
Appendix 8-1 Matrix applications in geophysics Part 8 2-2 2-6 2-12 2-18 3-1 3-2 3-4 3-8 4-1 4-2 4-4 4-6 4-12 4-14 5-1 5-2 5-10 5-16 6-1 6-2 6-4 6-22 6-30 7-1 7-2 7-4 8-1 8-2 8-4 8-10 8-14
Introduction to Seismic Inversion Methods Brian Russell
Part 9 Travel-time Inversion
g. 1. I ntroducti on
9.2 Numerical examples of traveltime inversion
9.3 Seismic Tomography
Part 10 Amplitude versus offset (AVO) Inversion
10.1 AVO theory
10.2 AVO inversion by GLI Part 11 Velocity Inversion
I ntroduc ti on
Theory and Examples Part 12 Summary 9-1 9-2 9-4 9-10 10-1 10-2 10-8 11-1 11-2 11-4 12-1
Introduction to Seismic •nversion Methods Brian Russell
PART I - INTRODUCTION
Introduction to Seismic Inversion Methods Brian Russell
I NTRODUCT ION TO SE I SMI C INVERSION METHODS
, __ _• i i _ , . , , ! • _, l_ , , i.,. _
Part i - Introduction _ . .
This course is intended as an overview of the current techniques used in
the inversion of seismic data. It would therefore seem appropriate to begin
by defining what is meant by seismic inversion. The most general definition
is as fol 1 ows'
Geophysical inversion involves mapping the physical structure and
properties of the subsurface of the earth using measurements made on
the surface of the earth.
The above definition is so broad that it encompasses virtually all the
work that is done in seismic analysis and interpretation. Thus, in this
course we shall primarily 'restrict our discussion to those inversion methods
which attempt to recover a broadband pseudo-acoustic impedance log from a
band-1 imi ted sei smic trace.
Another way to look at inversion is to consider it as the technique for
creating a model of the earth using the seismic data as input. As such, it
can be considered as the opposite of the forwar• modelling technique, which
involves creating a synthetic seismic section based on a model of the earth
(or, in the simplest case, using a sonic log as a one-dimensional model). The relationship between forward and inverse modelling is shown in Figure 1.1.
To understand seismic inversion, we must first understand the physical
processes involved in the creation of seismic data. Initially, we will
therefore look at the basic convolutional model of the seismic trace in the
time and frequency domains,
considering the thre e components
of this model:
reflectivity, seismic wavelet, and noise.Part I - Introduction
_ m i --.
Introduction to Seismic InverSion Methods Brian Russell
FORWARD MODELL I NG
i m ß
INVERSE MODELLING (INVERSION) _
, ß ß _ Input' Process: Output'
EARTH
MODEL
, MODELLING ALGORITHM SEISMIC RESPONSE i m mlm ii INVERSION ALGORITHMEARTH
MODEL
i iiFigure 1.1 Fo.•ard
' andsInverse
Model,ling
Introduction. to Seismic Inversion Methods Brian l•ussel 1
Once we have an understanding of these concepts and the problems which
can occur, we are in a position to look at the methods which are currently
ß
used to invert seismic data. These methods are summarized in Figure 1.2. The
primary emphasis of the course will be
the ultimate resul.t, as was previously
on poststack seismic inversion where
o
Oiscussed, is a pseudo-impeaance
section.
We will start by looking at the most contanon methods of poststack
inversion, which are based on single trace recursion. To better unUerstand
these recurslye inversion procedures, it is important to look at the
relationship between aleconvolution anU inversion, and how Uependent each
method is on the deconvolution scheme Chosen. Specifically, we will consider
classical "whitening" aleconvolution methods, wavelet extraction methods, and
the newer sparse-spike deconvolution methods such as Maximum-likelihood
deconvolution and the L-1 norm metboa.
Another important type of inversion method which will be aiscussed is
model-based inversion, where a geological moael is iteratively upUated to finU
the best fit with the seismic data. After this, traveltime inversion, or
tomography, will be discussed along with several illustrative examples.
After the discussion on poststack inversion, we shall move into the realm
of pretstack. These methoUs, still fairly new, allow us to extract parameters
other than impedance, such as density and shear-wave velocity.
Finally, we will aiscuss the geological aUvantages anU limitations of
each seismic inversion roethoU, looking at examples of each.
Introduction to Selsmic Inversion Methods Brian Russell
SE
I SMI
C I NV
ERSI
ON
.MET•OS ,,,
POSTSTACK INVERSION PRESTACK INVERSIONMODEL-BASED
I RECURSIVE
INVERSION
• ,INVE
SION
- "NARROW
BAND TRAVELTIME INVERSION!TOMOGRAPHY)
SPARSE-
SPIKE
WAV
EF
I EL
D
NVERSIOU
i LINEAR METHODS ,, i i --I METHODS
]
Figure 1.2 A summary of current inversion techniques.
Introduction to Seismic Inversion Methods Brtan Russell
PART
2 - THE CONVOLUTIONAL
MODEL
Introduction to Seismic Inversion Methods Brian Russell Part 2 - The Convolutional Mooel
2.1 Th'e Sei smi c Model
The most basic and commonly
used one-Oimensional
moael for the seismic
trace is referreU to as the convolutional moOel, which states that the seismictrace is simply the convolution of the earth's reflectivity with a seismic
source function with the adUltion of a noise component. In equation form,
where * implies convolution,
s(t) : w(t) * r(t) + n(t)s
where
and
s (t) = the sei smic trace, w(t) : a seismic wavelet, r (t) : earth refl ecti vi ty, n(t) : additive noise.
An even simpler assumption
is to consiUer the noise component
to be zero,
in which case the seismic tr•½e is simply the convolution of a seismic wavelet
with t•e earth ' s refl ecti vi ty,
s(t) = w{t) * r(t).
In seismic processing
we deal exclusively with digital data, that is,
data sampled
at a constant
time interval.
If we consiUer
the relectivity to
consist of a reflection coefficient at each time sample (som• of which can be
zero), and the wavelet to be a smooth function in time, convolution can be
thought of as "replacing" each reflection. coefficient with a scaled version of
the wavelet and summing
the result. The result of this process
is illustrated
in Figures 2.1 and 2.Z for both a "sparse"
and a "dense"
set of reflection
coefficients. Notice that convolution with the wavelet tends to "smear" the
reflection coefficients. That is, there is a total loss of resolution, which
is the ability to resolve closely spaced reflectors.
Introduction to Seismic Inversion Nethods Brian Russell WAVELET:
(a) ' * • •
: -' ':'
REFLECTIVITY Figure 2.1 TRACE:Convolution
of a wavelet with a
(a) •avelet. (b) Reflectivit.y.
sparse" reflectivity.
(c) Resu 1 ting Sei smic Trace.(a) (b') ! . i : !
!
:
:
i i , ß : i ! i i '?t * c o o o o oFi õure 2.2 Convolution of a wavelet with a sonic-derived "dense"
reflectivity. (a) Wavelet. (b) Reflectivity. (c) Seismic Trace
, i , ß .... ! , m i i L _ - '
Introduction to Seismic Inver'sion Methods Brian Russell
An alternate, but equivalent, way of looking at the seismic trace is in
the frequency domain. If we take the Fourier transform of the previous
ß
equati on, we may write
S(f) = W(f) x R(f),
where
S(f) = Fourier transform
of s(t),
W(f) = Fourier transform of w(t),
R(f) = Fourier transform of r(t), ana f = frequency.
In the above equation we see that convolution becomes multiplication in
the frequency domain. However, the Fourier transform is a complex function,
and it is normal to consiUer the amplitude and phase spectra of the individual
components. The spectra of S(f) may then be simply expressed
esCf) = e
w
where
(f) + er(f),
I •ndicates
amplitude
spectrum,
and
0 indicates phase spectrum. .
In other words, convolution involves multiplying the amplitude spectra
and adding the phase spectra. Figure 2.3 illustrates the convolutional model
in the frequency domain. Notice that the time Oomain problem of loss of
resolution becomes one of loss of frequency content in the frequency domain.
Both the high and low frequencies of the reflectivity have been severely
reOuceo by the effects of the seismic wavelet.
Introduction to Seismic Inversion Methods Brian Russell
AMPLITUDE SPECTRA PHASE SPECTRA
w (f) I I -t- R (f) i i , I ! i. iit |11 loo s (f) I i! I i i
Figure 2.3 Convolution in the frequency domain for the time series shown in Figure 2.1.
Introduction to Seismic Inversion Methods Brian Russell
2.g The Reflection Coefficient Series
l_ _ ,m i _ _ , _ _ m_ _,• , _ _ ß _ el
of as the res within the ear
compres si onal i ropedance to re impedances by
coefficient at fo11 aws:
'The reflection coefficient series (or reflectivity, as it is also called)
described
in the previous
section
is one
of the fundamental
physical
concepts
in the seismic method. Basically, each reflection coefficient may be thought
ponse of the seismic wavelet to an acoustic impeUance change
th, where acoustic impedance is defined as the proUuct of velocity and Uensity. Mathematically, converting from acoustic
flectivity involves dividing the difference in the acoustic
the sum of the acoustic impeaances. This gives t•e reflection
the boundary between the two layers. The equation is as
•i+lVi+l - iVi
Zi+l- Z
i
i • i+1 where and r = reflection coefficient,
/o__ density,
V -- compressional velocity, Z -- acoustic impeUance, Layer i overlies Layer i+1.We must also convert from depth to time by integrating the sonic log transit times. Figure •.4 shows a schematic sonic log, density log, anU resulting acoustic impedance for a simplifieU earth moael. Figure 2.$ shows
the result of converting
to the reflection
coefficient
series
and
integrating
to time.
It should be pointed out that this formula is true only for the normal incidence case, that is, for a seismic wave striking the reflecting interface at right angles to the beds. Later in this course, we shall consider the case of nonnormal inciaence.
Introduction to Seismic Inversion Methods Brian Russell
STRATIGRAPHIC SONIC LOG
SECTION •T (•usec./mette) 4OO SHALE ... DEPTH ß ß ß ß ß ß SANOSTONE . . - .. ,
'
I ! !_1 ! ! ! UMESTONE I I I ! I ! I 1 LIMESTONE 2000111 30O 200I
3600 m/s _v-- I
V--3600
J
V= 6QO0I
loo 2.0 , 3.0 OENSITY LOG. ß •Fig. 2.4. Borehole
Log Measurements.
mm mm rome m .am ,mm mm m --- mm SHALE ... OEPTH
•---'-
[
SANDSTONE . . ... , ! I !11 I1 UMESTONE I I 1 I I I II i ! I 1 i I i 1000m SHALE •.--._--.---- • •.'• LIMESTONE 2000 m ACOUSTIC IMPED,M•CE (2• (Y•ocrrv x OEaSn• REFLECTWrrY V$ OEPTH VS TWO.WAY TIME 20K -.25 O Q.2S -.25 O + .2S I I v ' I - 1000 m -- NO ,• , .. - 20o0 m I SECONDFig. 2.5. Creation of Reflectivity Sequence.
IntroductJ on 1:o Sei stoic Inversion Herhods Bri an Russell
Our best method of observing seJsm•c impedance and reflectivity is •o
derlye them from well log curves. Thus, we may create an impedance curve by
multiplying together •he sonic and density logs from a well. We may •hen
compute the reflectivlty by using •he formula shown earlier. Often, we do not
have the density log available• to us and must make do with only the sonJc. The
approxJmatJon of velocJty to •mpedance 1s a reasonable approxjmation, and
seems
to hold well for clas;cics and carbonates (not evaporltes, however).
Figure 2.6 shows the sonic and reflectJv•ty traces from a typJcal Alberta well
after they have been Jntegrated to two-way tlme.
As we shall see later, the type of aleconvolution and inversion used is
dependent on the statistical assumptions which are made about the seismic
reflectivity and wavelet. Therefore, how can we describe the reflectivity seen
in a well? The traditional answer has always been that we consider the
reflectivity to be a perfectly random sequence and, from Figure •.6, this
appears to be a good assumption. A ranUom sequence has the property that its
autocorrelation is a spike at zero-lag. That is, all the components of the
autocorrelation are zero except the zero-lag value, as shown in the following
equati on-
t(Drt = ( 1 , 0 , 0 , ...
)
t
zero-lag.
Let us test this idea on a theoretical random sequence, shown in Figure
2.7. Notice that the autocorrelation of this sequence has a large spike at
ß
the zeroth lag, but that there is a significant noise component at nonzero lags. To have a truly random sequence, it must be infinite in extent. Also on this figure is shown the autocorrelation of a well log •erived reflectivity. We see that it is even less "random" than the random spike sequence. We will discuss this in more detail on the next page.
IntroductJon to Se•.s=•c Inversion Methods Br•an Russell
RFC
F•g. 2.6. Reflectivity
sequence
derived
from
sonJc
.log.
RANDOM SPIKE SEQUENCE
WELL LOG DERIVED REFLECT1vrrY
AUTOCORRE•JATION OF RANDOM SEQUENCE
AUTOCORRELATION OF REFLECTIVITY
Fig. 2.7. Autocorrelat4ons of random and well log
der4ved
spike sequences.
Introductlon to Sei smic Inversion Methods Brian Russel 1
Therefore, the true earth reflectivity cannot be considered as being
truly random. For a typical Alberta well we see a number
of large spikes
(co•responding
to major lithol ogic change)
sticking up above
the crowd. A good
way to describe this statistically is as a Bernoulli-Gaussian
sequence. The
Bernoulli part of this term implies a sparseness in the positions of the
spikes and the Gaussian
implies a randomness
in their amplitudes. When we
generate such a sequence, there is a term, lambda, which controls the
sparseness of the spikes. For a lambda
of 0 there are no spikes, and for a
lambda
of 1, the sequence
is perfectly Gaussian in distribution.
Figure 2.8
shows a number of such series for different values of lambda. Notice that atypical Alberta well log reflectivity
would have a lambda
value in the 0.1 to
0.5 range.I ntroducti on to Sei smic I nversi on Methods Brian Russell It tl I I I LAMBD^•0.01 i I I •11 I 511 t •tl I (VERY SPARSE) 11 311 I LAMBDA--O. 1 4# I 511 I #1 I TZIIE (KS ! 1,1
::.
•"• •'•;'"
' "";'•'l•'
"••'r'•
LAMBDAI0.5- • "(11
TX#EI
(HS) LAMBDA-- 1.0 (GAUSSIAN:]EXAMPLES
OF REFLECTIVITIES
Fig. 2.8. Examples of reflectivities using lambda
factor to be discussed in Part 6.
, , m i ß i
Introduction to Seismic Inversion ,Methods Brian Russell
2.3 The Seismic Wavelet
-- _ ß • ,
Zero Phase and Constant Phase Wavelets
m _ m _ m ß m u , L m _ J
The assumption tha.t there is a single, well-defined wavelet which is convolved with the reflectivity to produce the seismic trace is overly simplistic. More realistically, the wavelet is both time-varying and complex in shape. However, the assumption of a simple wavelet is reasonable, and in
this section we shall consider several types of wavelets and their
characteristics.
First, let us consider the Ricker wavelet, which consists of a peak and
two troughs, or side lobes. The Ricker wavelet is dependent only on its dominant frequency, that is, the peak frequency of its a•litude spectrum or the inverse of the dominant period in the time domain (the dominant period is
found by measuring
the time from trough to trough). Two Ricker wave'lets are
shown in Figures 2.9 and 2.10 of frequencies 20 and 40 Hz. Notice that as the anq•litude spectrum of a wavelet .is broadened, the wavelet gets narrower in the
time domain,
indicating
an increase
of resolution. Our ultimate
wavelet
would
be a spike, with a flat amplitude spectrum. Such a wavelet is an unrealistic goal in seismic processing, but one that is aimed for.
The Rtcker wavelets of Figures 2.9 and 2.10 are also zero-phase, or
perfectly symmetrical. This is a desirable character. tstic of wavelets since
the energy is then concentrated at a positive peak, and the convol'ution of the
wavelet with a reflection coefficient will better resolve that reflection. To
get an idea of non-zero-phase wavelets, consider Figure 2.11, where a Ricker
wavelet has been rotated by 90 degree increments, and Figure 2.12, where the
same wavelet has been shifted by 30 degree increments. Notice that the 90
degree rotation displays perfect antis•nmnetry, whereas a 180 degree shift
simply inverts the wavelet. The 30 degree rotations are asymetric.
Introduction to Seismic Inversion Methods Brian Russell
Fig.
Fig.
2.9. 20 Hz Ricker Wavelet'.
•.10. 40 Hz Ricker wavelet.
Fig. 2.11. Ricker wavelet rotated by 90 degree increments
Fig.
Part 2 - The Convolutional Model
2.12. Ricker wavelet rotated by 30 degree increments
Introduction to Seismic Inversion Methods Brian Russell
Of course, a typical seismic wavelet contains a larger range of
frequencies than that shown on the Ricker wavelet. Consider the banapass
fil•er shown
in Figure 2.13, where we have passed a bana of frequencies
between 15 and 60 Hz. The filter has also had cosine tapers applied between 5
and 15 Hz, and between 60 and 80 Hz. The taper reduces the "ringing" effect
that would be noticeable if the wavelet amplitude spectrum was a simple
box-car. The wavelet of Figure 2.13 is zero-phase, and would be excellent as
a stratigraphic wavelet. It is often referred to as an Ormsby wavelet.
Minimum Phase Wavelets
The concept of minimum-phase is one that is vital to aleconvolution, but
is also a concept that is poorly understood. The reason for this lack of
understanding is that most discussions of the concept stress the mathematics
at the expense of the physical interpretation. The definition we
use of minimum-phase is adapted from Treitel and Robinson (1966):
For a given set of wavelets, all with the same amplitude spectrum,
the minimum-phase
wavelet
is the one
which
has the sharpest
leading
edge. That is, only wavelets which have positive time values.The reason that minimum-phase concept is important to us is that a
typical wavelet in dynamite work is close to minimum-phase. Also, the wavelet
from the seismic instruments is also minimum-phase. The minimum-phase
equivalent of the 5/15-60/80 zero-phase wavelet is shown in Figure 2.14. As
in the aefinition used, notice that the minimum-phase wavelet has no component
prior to time zero and has its energy concentrated as close to the origin as
possible. The phase spectrum of the minimum-wavelet is also shown.
I•troduct•on to Sei stoic !nversion
Nethods.
Br•an Russell
ql Re• R Zero Phase I•auel•t 5/15-68Y88 {•
0.6
f1.38 - Trace 1
iii
- e.3e
... ,
• ...
'
2be
1 Trace I
Fig. 2.13. Zero-phase bandpass
wavelet.
Reg 1) min,l• wavelet •/15-68/88 hz
18.00 p Trace I
Reg E wayel Speetnm
'188.88
•
Trace 1
0.8
188
Fig. 2.14. Minim•-phase equivalent
of zero-phase wavelet
shown in Fig. 2.13.
_
! m,m, i m
Part 2 -Th 'e Convolutional Model
i
Introduction to Seismic Inversion Methods Brian Russell
Let us now look at the effect of different wavelets on the reflectivity
function itself. Figure 2.15 a anU b shows a number of different wavelets
conv6lved with the reflectivity (Trace 1) from the simple blocky model shown
in Figure Z.5. The following wavelets have been used- high
zero-phase (Trace •),
low frequency
zero-phase
(Trace ½), high
minimum phase (Trace 3), low frequency minimum phase (Trace 5).
figure, we can make the fol 1 owing observations:
frequency frequency
From the
(1) Low freq. zero-phase wavelet: (Trace 4)
- Resolution of reflections is poor.
- Identification of onset of reflection is good.
(Z) High freq. zero-phase wavelet: (Trace Z)
- Resolution of reflections is good.
- Identification of onset of reflection is good.
(3) Low freq. min. p•ase wavelet- (Trace 5)
- Resolution of reflections i s poor.
- Identification of onset of reflection is poor. (4) High freq. min. phase wavelet: (Trace 3)
- Resolution of refl ec tions is good.
- Identification of onset of reflection is poor.
Based on the above observations, we would have to consider the high
frequency,
zero-phase
wavelet
the best, and the low-frequency,
minimum
phase
wavelet the worst.
(a)
Introduction to Seismic Inversion Methods Brian Russell
!ql Reg R Zer• Phase Ua•elet •,'1G-•1• 14z
F
- •.• ['
'
•,3 Recj B miniilium phue ' '
17 .•
q2 Reg C Zero Phase 14aue16(' ' •'le-3•4B Hz
e
q• Reg 1) 'minimum phase " •,leJ3e/4e h• '
8
e.e •/••/'•-•"v--,._,,
-r
e.•
' ' " s•e
,m ,,''
Tr'oce [b) Fig. 7002.15. Convolution of four different wavelets shown in (a) with trace I of (b). The results are shown on traces 2 to 5 of (b).
Introduction to Seismic Inversion Methods Brian Russell
g.4 Th•N.
oi se. C
o.
mp.o•ne
nt
-
The situation that has been discussed so far is the ideal case. That is,
.
we have interpreted every reflection wavelet on a seismic trace as being an
actual reflection from a lithological boundary. Actually, many of the
"wiggles" on a trace are not true reflections, but are actually the result of
seismic noise. Seismic noise can be grouped under two categories-(i) Random Noise - noise which is uncorrelated from trace to trace and is
•ue mainly to environmental factors.
(ii) Coherent Noise - noise which is predictable on the seismic trace but is unwanted. An example is multiple reflection interference.
Random noise can be thought of as the additive component n(t) which was
seen in the equation on page 2-g.
Correcting for this term is the primary
reason for stacking our •ata.
Stacking actually uoes an excellent job of
removing ranUom noise.Multiples, one of the major sources of coherent noise, are caused
by
multiple "bounces"
of the seismic signal within the earth, as shown
in Figure
2.16. They may be straightforward, as in multiple seafloor bounces
or
"ringing", or extremely
complex, as typified by interbed multiples. Multiples
cannot be thought of as additive noise and must be modeled
as a convolution
with the reflecti vi ty.Figure
generated by the simple blocky model
this data, it is important that
Multiples may be partially removed
powerful elimination technique.
aleconvolution, f-k filter.ing, wil 1 be consi alered in Part 4.
2.17
shown on Figure •. 5. the multiples be
by stacking, but
Such techniques and inverse velocity stacking.
shows the theoretical multiple sequence which would be
If we are to invert
effectively removed. often require a more include predictive
These techniques
Introduction to Seismic Inversion Methods Brian Russell
Fig. 2.16. Several multiple generating mechanisms.
TIME TIME
[sec) [sec)
0.7 0.7
REFLECTION R.C.S.
COEFFICIENT WITH ALL
SERIES MULTIPLES
Fig. 2.17.
Refl ectivi ty sequence
of Fig.
and without mul tipl es.
Part 2 - The Convolutional Model
2.5. with
.
PART 3 - RECURS IVE INVERSION - THEORY
m•mmm•---' .• ,- - - ' •- - _ - - _- _
•ntroduct•on to SeJsmic Znversion Methods Brian Russell
PART 3 - RECURSIVE INVERSION - THEORY
3.1 Discrete Inversion
, ! ß , , •
In section 2.2, we saw that reflectivity was defined in terms of
acoustic impedance changes. The formula was written:
Y•i+lV•+l
' •iV! 2i+
1' Z
i
ri-- yoi'+lVi+l+
Y•iVi
-- -Zi..+l
+ Z
i
where r -- refl ecti on coefficient,
/0-- density,
V -- compressional velocity, Z -- acoustic impedance,
and Layer i overlies Layer i+1.
If we have the true reflectivity available to us, it is possible to recover the a.coustic impedance by inverting the above formula. Normally, the inverse' formulation is simply written down, but here we will supply the
missing steps for completness. First, notice that:
Also
Ther'efore
Zi+l+
Z
i
Zi+
1- Z
t
2 Zi+
1
I + ri- Zi+l
+ Zi + Zi+l
+ 2i Zi+l
+ Zi
I- ri--
Zi+l+
Zi+l+ Z i
Z
i
Zi+
Zi+l+ Z i
1- Z
i
Zi+l+ Z i
2 Zf[
Zi+l
Z i
l+r.
1
1
Part 3 - Recursive Inversion- Theory
ill, ß , I
Introduction to Seismic Invers-•on Methods Brian Russell pv-e- TIME (sec] 0.7 REFLECTION COEFFICIENT SERIES RECOVERED ACOUSTIC IMPEDANCE
Fig. 3.1,
Applying
the recursive
inversion
formula
to a
simple, and exact, reflectivity.
, ! ß
!ntroductt on to Se1 smJc ! nversi on Methods Brian Russell •9r• ;• • •;• • • •-•• 9rgr•t-k'k9r9r• •-;• ;• ...
Or, the final •esult-
Zi+[= Z
ß
l+r i
.
This is called the discrete recursive inversion formula and is the basis
of many current inversion techniques. The formula tells us that if we know the acoustic impedance of a particular layer and the reflection coefficient at
the base of that layer, we may recover the acoustic impedance of the next
layer. Of course we need an estimate of the first layer impedance to start us
off. Assume we can estimate this value for layer one. Then
l+rl
,
Z2:
Zl i r 1
Z3=
Z
2 11
+ r 2
- rand so on ...
To find the nth impedance from the first, we simply write the formula as
Figure 3.1 shows the application of the recursive formula to the "
reflection coefficients derived in section 2.2. As expected, the full acoustic impedance was recovered.
Problems encountered with real data
• ß , m i i • i ! m
When the recursive inversion formula is applied to real data, we find
that two serious problems are encountered. These problems are as follows-
(i) Frequency Bandl imi ti ng _ ß
Referring back to Figure 2.2 we see that the reflectivity is severely
bandlimited when it is convolved with the seismic wavelet. Both the
low frequency components and the high frequency components are lost.
Introduction to Seismic Inversion Methods Brian Russell
0.2
0
V•) 'V,•
•R R = +0.2V
o: 1000
m
Where:
--• V,• = 1000 i-o.t
- 1500 m - •ec'. (a) - 0.1 '•0.2R•
R=
{ASSUME
j•: l)
R•=
-0.1
R =+0.2 R: -0.1V
o= 1000 m
-'+ ¾1
= 818 m
ii•.Figure 3.2 Effect of banUlimiting on reflectivity, where (a) shows
single reflection coefficient, anU (b) shows bandlimited
refl ecti on coefficient.
i i m i m I
I __ ___ i _
Introduction to Seismic Inversion Methods Brian Russell
(ii) Noise
The inclusion of coherent or random noise into the seismic 'trace will
make the estimate• reflectivity deviate from the true reflectivity.
To get a feeling for the severity of the above limitations on recursire
inversion, let us first use simple models. To illustrate the effect of
bandlimiting, consider Figure 3.Z. It shows the inversion of a single spike (Figure 3.2 (a)) anU the inversion of this spike convolved with a Ricker
wavelet (Figure 3.2 (b)). Even with this very high frequency banUwidth
wavelet, we have totally lost our abil.ity to recover the low frequency
component of the acoustic impedance.
In Figure 3.3 the model derived in section Z.2 has been convolved with a
minimum-phase wavelet. Notice that the inversion of the data again shows a
loss of the low frequency component. The loss of the low frequency component is the most severe problem facing us in the inversion of seismic data, for it
is extremely Oifficult to directly recover it. At the high end of the ß
spectrum, we may recover much of the original frequency content using
deconvolution techniques. In part 5 we will address the problem of recovering
the low frequency component.
Next, consider the problem of noise. This noise may be from many sources, but will always tend to interfere with our recovery of the true
reflectivity.
Figure 3.4 shows the effect of adding the full multiple
reflection train (including transmission losses) to the model reflectivity.
As we can see on the diagram, the recovered acoustic impedance has the same
basic shape as the true acoustic impedance, but becomes increasingly incorrect
with depth. This problem of accumulating error is compoundeU by the amplitude
problemns introduced by the transmission losses.
Introduction to Seismic Invers,ion Methods Brian Russell TIME Fig. TIME (see) Fig. 0.? RECOVERED ACOUSTIC IMPEDANCE REFLECTION SYNTHETIC COEFFICIENT (MWNUM-PHASE SERIES WAVELET) pv-•, INVERSION OF SYNTHETIC
3.3. The effect of bandlimiting on recurslye inversion.
0.7
TIME
(re.c)
REFLECTION RECOVERED R.C.S. RECOVERED
COEFFICIENT ACOUSTIC WITH ALL ACOUSTIC
SERIES IMPEDANCE MULTIPLES IMPEDANCE
3.4. The effect of noise on recursive inversion.
Introduction to Seismic Inversion Methods Brian Russell
3.3 Continuous Inversion
A logarithmic relationship is often used to approximate the above formulas. This is derived by noting that we can write r(t) as a continuous function in the following way:
Or
r(t) - Z(t+dt)
- Z{t) _ 1 d Z(t)
ß - Z(t+dt) + Z(•) - •' z'(t)
! d In Z(t)
r(t) = •
dt
The inverse formula is thus-
t
Z(t)
= Z(O)
exp
2y r(t) dt.
0
The preceding
approximation
is valid if r(t) <10.3• which is usually the
case. A paper by Berteussen and Ursin (1983), goes into much more detail on
the continuous versus discrete approximation. Figures 3.5 and 3.6 from their
paper show that the accuracy of the continuous inversion algorithm is within
4% of the correct value between reflection coefficients of -0.5 and +0.3.
If our reflection coefficients are in the order of + or - 0.1, an even
simpler
approximation
may
be made
by dropp'ing
the logarithmic
relationship:
t
1 d Z(t)
•_==•
Z(t)
--2'Z(O)
fr(t) dt
r(t)
--• -dr
VO
Introduction to Seismic Inversion Methods Brian Russell Fig. 3.5 m i ,, ,m I I IIIII
I + gt
½xp
(26•)
Difference
-1.0 0.0 0.14 -0.14 -0.9 0.05 0. I? -0.12 -0.8 0.11 0.20 -0.09 -0.7 0.18 0.25 -0.07 -0.6 0.25 0.30 -0.05 -0.5 0.33 0.37 -0.04 ' -0.4 0.43 0.45 --0.02 -0.3 0.• 0.•5 --0.01 -0.2 0.667 0.670 -0.003 -0.1 0.8182 0.8187 --0.0005 0.0 1.0 1.0 0.0 0.1 1.222 1.221 0.001 0.2 1.500 1.492 0.008 0.3 1.86 1.82 0.04 0.4 2.33 2.23 o.1 0.5 3.0 2.7 0.3 0.6 4.0 3.3 0.7 0.7 5.7 4.1 1.6 0.8 9.0 5.0 4.0 0.9 19.0 6.0 13.0 1.0 co 7.4 •oNumerical c•pari son of discrete and continuous i nversi on.
(Berteussen and Ursin, 1983)
Fig. 3.6
$000
O} m
MPEDANCE
(O
I SCR.
)
r-niL
${300
o-•
O
I FFERENCE
SO0 O I FFERENCE ( SCALED UP )
T •'•E t SECONOS
C•pari son between
impedance
c•putatins based
on a
discrete and a continuous
seismic •del.
(Berteussen and Ursin, 1983)
Introduction'to Seismic Inversion Methods Brian Russell
PART 4 - SEISMIC PROCESSING CONSIDERATIONS
•ntroduction to Seismic •nvers•on Methods B.r. ian Russell
4.1 Introduction
Having looked at a simple model'of the seismic trace, anu at the
recursire inversion alogorithm in theory, we will now
look at the problem of
processing
real seismic eata in order to get the best results from seismic
inversion.
We may group the key processing
problems into the following
categories:( i ) Amp 1 i tu de rec o very.
(i i) Vertical resolution improvement.
(i i i ) Horizontal resol uti on improvement.
(iv) Noise elimination.
Amplitude problems are a major
consideration
at the early processing
stages
and we will look at both deterministic amplitude
recovery
and surface
consistent residual static time corrections. Vertical resolution improvement
will involve a discussion of aleconvolution and wavelet processing techniques. In our discussion of horizontal resolution we will look at the resolution
improvement
obtained in migration, using a 3-D example. Finally, we will
consider several approaches
to noise elimination, especially the elimination
of multi pl es.
Simply stateu, to invert our
one-dimensional model given in the
approximation of this model (that
band-limited reflectivity function)
these considerations in minU. Figure 4.1
be useU to do preinversion processing.
seismic data we usually assume the
previous section. And to arrive at an
is, that each trace is a vertical,
we must carefully process our data with shows a processing flow which could
Introduction to Seismic Inversion Methods Brian Russell
INPUT RAW DATA
DETERMINISTIC AMPLITUDE CORRECTIONS ,. _•m mlm SURFACE-CONS ISTENT
DECONVOLUTIO,
N FOLLOWED
BY HI GH RESOIJUTI.ON DECON i i SURFACE-CONS I STENT AMPt:ITUDE ANAL'YSIS SURFACE-CONSI STENT STATI CS ANAIJY SISVELOCITY ANAUYS IS
APPbY STATICS AND VEUOCITY
MULTIPLE ATTENUATION
STACK
ß •
MI GRATI ON
,
Fig. 4.1.
Simpl
i fied i nversi
on
processing
flow.
ll , ß ' ß I , _ i 11 , m - -- m _ • • ,11
Inl;roducl:ion 1:o SeJ smlc Invers1 on Nethods BrJ an Russell
4.2 Am.p'l i tu. de.. P,.ecovery
The most dJffJcult job in the p•ocessing of any seismic line is
ß•econst•ucting
the amplJtudes
of the selsmJc
t•aces as they would
have been
Jf
the•e
were no dJs[urbJng inf'luences present. We normally make the
simplJfication
that the distortion of the seJsmic
amplJtudes
may
be put into
three main categories' sphe•Jcal
divergence, absorptJon, and t•ansmJssion
loss. Based on a consideration of these three factors, we may wrJte aown an
approximate
functJon for the total earth attenuation-
Thus, data, the
formula.
At: AO*
(
b / t) * exp(-at),where t = time,
A
t = recorded
amplitude,
A
0 = true ampl
i tude,
anU
a,b = constants.
if we estimate the constants in the above equation from the seismic
true amplitudes
of the data coulU be recovered by using the inverse
The deterministic amplitude correction and trace to trace mean
scaling will account
for the overall gross changes in amplitude. However,
there may still
be subtle (or even not-so-subtle) amplitude problems
associated
with poor surface conditions or other factors. To compensate
for
these effects, it is often advisable to compute and apply surface-consistent
gain corrections. This correction involves computing
a total gain value for
each trace and then decomposing this single value in the four components
Aij=
Six
Rj
x G
k x MkX
•j,
where A = Total amplitude factor,
S = Shot component, R: Receiver component, G = CDP component, and
M = Offset component, X = Offset distance,
i,j = shot,receiver pos., k = CDP position.
Introduction to Seismic .Inversion Methods Brian Russell
SURFACE
SUEF'A•CONS
Ib'TEh[O{
AND T |tV•E : ,Ri L-rE R ßFig. 4.2. Surface and sub-surface geometry and
surface-consistent decomposition. (Mike Graul).
, ,
Introduction to Seismic Inversion Methods Brian Russell
Figure 4.g (from Mike Graul's unpublished course notes) shows the
geometry
used for this analysis. Notice that the surface-consistent
statics
anti aleconvolution
problem
are similar. For the statics problem, the averaging
can be •1one by straight summation. For the amplitude problem we must
transform the above equation into additive form using the logarithm:
In Aij=
In S
i + In Rj
+ In G
k + lnkMijX•.
The problem can then be treated exactly the same way as in the statics
case. Figure 4.3, from Taner anti Koehler (1981), shows
the effect of doing
surface consistent amplitude and statics corrections.4.3 I•mp.
rov.
ement_
o.[_Ver.
t.i.ca.1..Resoluti
on
Deconvol ution is a process by which an attempt is made to remove the
seismic wavelet from the seismic trace, leaving an estimate of reflectivity.
Let us first discuss the "convolution" part of "deconvolution" starting with
the equation for the convolutional modelIn the
st-- wt* r t
where
frequency domain
st = the sei smic
trace,
wt= the seismic
wavelet,
rt= reflection coefficient series,
* = convol ution operation.
S(f) • W(f) x R(f) .
The deconvol ution
procedure and consists
reflection coefficients.
fol 1 owl ng equati on-
rt: st* o
process is simply the reverse of the convolution
of "removing" the wavelet shape to reveal the
We must design an operator to do this, as in the
where Or--
operator
-- inverse
of w
t .
Part 4 - Seismic Processing Considerations
Introduction to Seismic Inversion Methods Brian Russell ii 11 ß 1' i ii '..,•' •, ," " " ß d.
Preliminary stack bet'ore surface consistent static and ompli- lude corrections.
ß Stock with surface consistent static and amplitude cor- rections.
Fig. 4.3. Stacks with and without surface-consi stent
corrections. (Taner anu Koehler, 1981).
Part 4 - Seismic Processing Considerations
ß ,
Introduction to Seismic Inversion Methods Brian Russell
In the frequency domain, this becomes R(f) = W(f) x 1/W(f) .
After this extremely simple introduction, it may appear that the
deconvolution
problem should be easy to solve. This is not the case, and the
continuing research into the problem testifies
to this.
There are two main
problems. Is our convolutional
model
correct, and, if the model is correct,
can we derive the true wavelet from the data? The answer to the firstquestion is that the convolutional
model
appears
to be the best model
we have
come
up with so far. The main problem is in assuming
that the wavelet does
not vary with time. In our discussion we will assume that the time varying
problem
is negligible within the zone of interest.
The second
problem is much more severe, since it requires solving the
ambiguous
problem
of separating a wavelet and reflectivity sequence
when only
the seismic trace is known. To get around this problem, all deconvolution or
wavelet estimation programs
make certain restrictive assumptions,
either about
the wavelet or the reflectivity. There are two classes of deconvolution
methods: those which make restrictive phase assumptions and can be considered ,
true wavelet processing techniques only when these phase assumptions are met,
and those which do not make restrictive phase assumptions and can be
considered as true wavelet processing methods. In the first category are (1) Spiking deconvolution,
(2) Predictive deconvolution,
(3) Zero phase deconvoluti on, and
(4) Surface-consi stent deconvoluti on.
Introduction to Seismic Inversion Methods Brian Russell
(a)
Fig. 4.4 A comparison of non surface-consistent and surface-consistent
decon on pre-stack data. {a) Zero-phase deconvolution.
{b) Surface-consistent soikinB d•convolution.
(b),
Fig. 4.5 Surface-consistent decon comparison after stack. (a) Zero-phase aleconvolution. (b) Surface-consistent
deconvol ution.
'--'- , ß , ,• ,t ß ß _ , , _ _ ,, , ,_ , ,
Introduction to Seismic Invers. ion Methods Brian Russell
In the second category are found
(1) Wavelet estimation using a well
(Hampson and Galbraith 1981)
1 og (Strat Decon).
(2) Maximum-1 ikel ihood aleconvolution. (Chi et al, lg84)
Let us surface-consi stent surface-consi stent
components. We
di recti ons- common
illustrate the effectiveness of one of. the methods, aleconvolution. Referring to Figure 4.•, notice that a
scheme involves the convolutional proauct of four
must therefore average over four different geometry
source, common receiver, common depth point (CDP), and
con, non offset (COS). The averaging must be performed iteratively and there
are several different ways to perform it. The example in Figures 4.4 ana 4.5
shows an actual surface-consi stent case study which was aone in the following
way'
(a) Compute the autocorrelations of each trace,
(b) average the autocorrelations in each geometry eirection to get four
average autocorrel ati OhS,
(c) derive and apply the minimum-phase inverse of each waveform, and
(•) iterate through this procedure to get an optimum result.
Two points to note when you are looking at the case study are the
consistent definition of the waveform
in the surface-consistent
approach an•
the subsequent improvement of the stratigraphic interpretability of the stack.
We can compare all of the above techniques using Table 4-1 on the next
page. The two major facets of the techniques which will be compared are the
wavelet estimation procedure and the wavelet shaping procedure.
Introduction to Seismic Inversion Methods Brian Russell
Table 4-1 Comparison of Deconvol ution MethoUs
m m ß ß m METHOD Spiking Deconvol ution Predi cti ve Deconvol uti on Zero Phase Deconvol utton Surface-cons. Deconvolution Stratigraphic Deconvol ution Maximum- L ik el i hood deconvol ution WAVELET ESTIMATION
Min.imum phase assumption
Random refl ecti vi ty
assumptions.
No assumptions about
wavelet•
Zero phase assumption.
Random refl ectt vi ty
assumption.
Minimum or zero phase.
Random reflecti vi ty
assumption.
No phase assumption.
However, well must match
sei smi c.
No phase assumption.
Sparse-spike assumption.
WAVELET SHAPING
Ideally shaped
to spike.
In practice, shaped
to minimum
phase, higher frequency
output.
Does not whiten data well.
Removes
short and long period
multiples. Does not affect
phase
of wayel
et for long lags.
..1_, mPhase is not altered.
Amplitude spectrum i$
whi tened.
Can shape
to desired output.
Phase
character i s improved.
Ampl i rude spectrum i s
whitened
less than in single
trace methods.
Phase of wavelet is zeroed.
Amplitude
spectrum
not
whi tened.
Phase of wavelet is zeroed•
Amp 1 i rude spectrum i s
whi tened.