3 Coherence

45
Introduction
Coherence is a measure of similarity between wave
forms or traces. When seen on a processed section, the seis
mic waveform is a response of the seismic wavelet con
volved with the geology of the subsurface. That response
changes in terms of amplitude, frequency, and phase, de
pending on the acousticimpedance contrast and thickness
of the layers above and below the reflecting boundary. In
turn, acoustic impedance is affected by the lithology, porosi
ty, density, and fluid type of the subsurface layers. Conse
quently, the seismic waveforms that we see on a processed
section differ in lateral character that is, strong lateral
changes in impedance contrasts give rise to strong lateral
changes in waveform character.
Figure 1a shows a laterally stable waveform indicating
a coherent event. Figure 1b shows a synclinal but laterally
invariant waveform. In contrast, Figure 1c and 1d shows vari
ations in waveform that are the result of channels. Geologi
cally, highly coherent seismic waveforms indicate laterally
continuous lithologies. Abrupt changes in waveform can
indicate faults and fractures in the sediments. In this chap
ter, we will demonstrate that lateral changes in coherence
provide interpretation insights.
Figure 2a shows a segment of a seismic section, and
Figure 2b shows its equivalent coherence section. Notice
that there are no sharp breaks within the highlighting hex
agon; however, a close examination does reveal changes in
the seismic waveforms. The coherence section shows these
changes as lowcoherence features. It is that sensitivity to
waveform changes that makes coherence a useful tool for
extracting subtle information from seismic data. Similarly,
Figure 3b demonstrates the ease with which faults can be
seen on the vertical coherence section and easily can be put
on a seismic section (Figure 3a).
In Figure 4, we redisplay one of the earliest published
applications of coherence: application of the original three
trace crosscorrelation coherence algorithm to a large 3D
survey acquired over South Marsh Island, Louisiana, U.S.A.
In Figure 4a we see a time slice through seismic data, and
in Figure 4b we see the crosscorrelationbased coherence
volume. The extremes of the high and low values of coher
ence were indicated by yellow and red at the time. Unless
otherwise stated, other examples in the book use a more
consistent blackgraywhite color scheme.
3D seismic interpretation
A 3D volume of seismic data allows us to visualize the
spatial evolution of structural or stratigraphic features. Such
a continuous evolution is basic to our understanding of sed
imentary deposition and of the eventual tectonic folding or
faulting, and to our study of the configuration and fluid
content of a reservoir. The first step toward 3D interpreta
tion is to use time slices along with the vertical sections
pulled out of the volume. Whereas vertical seismic sections
describe inline dips, the crossline dip can be investigated
either by animating through successive inline seismic sec
Chapter 3
Coherence
Chapter Objectives
After reading this chapter, you will be able to
summarize the physical and mathematical basis of currently available seismic coherence algorithms
evaluate the impact of spatial and temporal analysis window size on the resolution of geologic features
recognize artifacts that result from structural leakage and seismic zero crossings
apply best practices for structural and stratigraphic interpretation
Chopra_CH03_v7.indd 45 6/5/2007 5:56:21 PM
Downloaded 10 Dec 2011 to 198.3.68.20. Redistribution subject to SEG license or copyright; Terms of Use: http://segdl.org/
46 Seismic Attributes for Prospect Identification and Reservoir Characterization
must be interpreted. Most of the channels and faults (barring
the obvious ones) could go undetected if only vertical sec
tions are interpreted. Time slices are a great help for such an
interpretation. Obviously, the interpretation depends on the
objective at hand and the quality of the data. Interpretation
of subtle details could be a nightmare if the data are of poor
quality.
Interactive workstations
Interactive seismicinterpretation workstations are valu
able tools for interpreting large volumes of 3D seismic data
efficiently. Inlines, crosslines, time slices, and arbitrary
profiles can be accessed readily from a single data volume,
and that facilitates convenient review and fast editing of the
data. Folded displays such as that shown in Figure 5 are a
great help in our understanding of subsurface geology.
Rapid displays of horizon amplitude maps are possible that
inject detail into our understanding of the field. The work
stations increased dynamic range of color displays and color
enhancement allows interpreters to readily detect subtle fea
tures for stratigraphic interpretation. Because stratigraphy
is best studied and displayed with vertical exaggeration, zones
of interest can be conveniently zoomed and color can be
enhanced interactively.
In areas with high dip or with stratigraphic features that
cut through different stratigraphic horizons, flattening con
sistent stratigraphic horizons or surfaces helps us obtain a
greater understanding of the stratigraphy. Investigation of
amplitudes along a seismic horizon on which sequences
prograde can give the direction of progradation, when the seis
mic volume is flattened on the horizon and sliced through.
We may conduct similar analysis when we study an uncon
formity surface and the layers subcropping below it, to
image a channel system or to determine the areal extent of
a reefal buildup. An important shortcoming exists, however
interpretive bias usually enters the data set when we use
horizon slices in tracing stratigraphic features, because the
Figure 1. Examples of lateral variations in seismic wave
forms: (a) a fat, laterally invariant, or coherent, waveform,
(b) a synclinal, but otherwise laterally invariant, or coherent,
waveform, (c) a laterally variable waveform indicative of
lateral changes in impedance or thickness, and (d) a rapidly
varying waveform associated with three channels.
a)
b)
c)
d)
Figure 2. (a) Lateral changes seen on a vertical slice through
a seismic data volume. (b) Such lateral changes show up as
lowcoherence features in the corresponding coherence slice.
a) b)
Figure 3. Vertical slices through (a) a seismic data volume
and (b) the corresponding coherence volume, indicating the
clarity with which the faults appear on coherence displays.
a) b)
tions or by taking a perpendicular crossline slice through
the seismic data volume. In contrast, time slices show reflec
tor strike. Changes in strike can be tracked by animating
through successive time slices.
Depositional systems often show up better on time
slices than on vertical sections. For example, river channels
usually cut their neighboring geologic strata in characteris
tically meandering patterns. Such patterns may be obvious
on a time slice. The areal disposition of such channels or
fault planes is not apparent on one vertical section, so to get
a feel for such patterns, several vertical inlines or crosslines
Coherence 47
difficult, timeconsuming and subjective process of picking
the horizon has to be followed.
3D data for reservoir description
In addition to providing structural and stratigraphic de
tails, 3D seismic surface measurements have come into
prominence as an essential element for reservoir descrip
tion. Their use is particularly widespread in those parts of
the world where new reserves are generated by infill and
extension drilling that is based on detailed knowledge of
the reservoir characteristics.
Reservoirs exhibit a range of physical properties that
can be detected as changes in response over appropriate
time intervals (e.g., by reservoir monitoring during en
hanced oil recovery). Timelapse seismic analysis is evolv
ing rapidly to identify bypassed hydrocarbons in a produc
ing reservoir. Threedimensional seismic surveys are carried
out over the reservoir area at different times during the life
of the field. The first survey generally is done before produc
tion is begun, and the second and subsequent surveys are
conducted after significant production has started. Changes
Figure 4. One of the frst applications of
coherence, showing a time slice at t = 1.200
s through (a) the seismic data volume and (b)
the crosscorrelationbased coherence volume.
The color scale is red (for lowest coherence),
black, gray, white, and yellow (for high
est coherence). After Bahorich and Farmer
(1995).
a)
b)
between preproduction and postproduction surveys are com
puted and attributed to changes in fluid properties, such as
saturation, density, and the like that have taken place be
tween the times of the surveys. Thus, 3D seismic data are
used to identify portions of the reservoir that have been de
pleted and other regions that still have commercially viable
hydrocarbon accumulations that could be targeted by an in
filldrilling program. Because 3D seismic measurements
are being integrated with other information technologies,
the potential of reservoir descriptions also is expanding.
Combined with welllog data, 3D seismic information is
used routinely to map trends in reservoir heterogeneity. Such
mapping was impossible previously with 2D seismic data.
With 2D seismic data, reservoir heterogeneity resulting from
smallscale faults often was ignored in reservoircharacter
ization projects because of the difficulty in correlating the
faults between coarsely spaced seismic lines. Threedimen
sional seismic data now permit us to do detailed geologic
modeling, and powerful graphics give us stronger visual evi
dence. Those advances allow us to define and extend reser
voirs and locate development wells optimally for draining
off reservoirs effectively and efficiently.
Coherence helps us accurately portray stratigraphic
plays and the associated faults. Sometimes, we must use
the gauged information to carefully model the structure and
stratigraphy and fault blocks so we can visualize the fault
planes that intersect, clip, or extend into 3D space. Usually,
the results of such exercises provide valuable input to geo
cellular and reservoirmodeling programs aiming to pro
duce moreaccurate volumetric and hydraulic models. Thus,
using coherence not only reduces exploration risk but, by
giving us a better understanding of the reservoir, ultimately
leads to optimum reservoir management.
In this chapter, we present the three common methods
for calculating coherence (crosscorrelation, semblance, and
eigenstructure). Then we present comparative examples that
highlight the differences among three coherence methods.
Higherorder statistics (Lu et al., 2003) or entropy measures
(Cohen and Coifman, 2002) also have been used for com
puting coherence but are not discussed here.
Crosscorrelation-based Coherence
All coherence measures operate on a spatial window of
neighboring traces. The simplest crosscorrelation algorithm
operates on three neighboring traces (Figure 6a), whereas
more computationally intensive algorithms based on sem
blance and eigenstructure operate on five, nine, or more
neighboring traces (Figure 6b).
Let us begin by examining the crosscorrelation be
tween the magenta master trace (trace 0) and the orange
target trace (trace 1) in the inline, or x direction, shown in
Figure 6a. We display the crosscorrelation operation graph
ically in Figure 7. If we slide the target trace by an amount
x
and define our vertical analysis window to range between
K samples above and below our analysis point at time t,
Inline
Inline
C
r
o
s
s
l
i
n
e
C
r
o
s
s
l
i
n
e
Figure 6. Spatial (or multitrace) analysis windows commonly used in coherence calculations for (a) the crosscorrelation algo
rithm and (b) the semblance and eigenstructure algorithms. In the crosscorrelation algorithm, we frst crosscorrelate the target
trace (in magenta) with the inline trace (in orange) over a suite of temporal lags. Then we repeat that process between the target
trace and the crossline trace (in cyan). The coherence estimate is obtained using equation 3.4. In the semblance and eigenstruc
ture algorithms, we frst estimate dip and azimuth, as discussed in Chapter 2, and then calculate either the semblance (given by
equation 3.13) or a covariance matrix (given by equation 3.17) between the target trace (in cyan) and its nearest neighbors. Here
we show four nearest neighbors in green and eight nearest neighbors in green and gray. These sets of nearest neighbors give rise
to either fve or ninetrace coherence algorithms, respectively.
a) b)
Figure 5. A folded or edge display showing a time slice and
vertical section. Courtesy of ONGC, India.
Coherence 49
we can write the normalized crosscorrelation coefficient
x

as
(3.1)
where
(3.2)
denotes the runningwindow mean of the nth trace. Because
we know that the true mean of properly processed seismic
data is 0, we often make the computational simplification
that <u>(t) = 0 as well. That is a reasonable approximation
if the analysis window is greater than a seismic wavelet.
We continue the process for a range of lags,
x
, and plot
the results in Figure 7b. Recalling Chapter 2, we note that
the lag
x
that has the maximum positivevalued normal
ized crosscorrelation coefficient is a firstorder approxima
tion of the inline apparent dip (because the pair of traces
displayed in Figure 7 are in the inline direction).
Next, we calculate the normalized crosscorrelation co
efficient between the magenta master trace (trace 0) and the
neighboring crossline cyan trace (trace 2) shown in Figure
6a, to obtain the crosscorrelation coefficient
y
:

. (3.3)
As before, the lag
y
that has the maximum positive
valued normalized crosscorrelation coefficient is a first
order estimate of crossline apparent dip. Bahorich and
Farmer (1995) defined the 3D crosscorrelation coherence
estimate, c
xc
, by combining the inline and the crossline cor
relation coefficients defined above using the formula
, (3.4)
where
t x y
x x i i
x
max ( , , , )

and
t x y
y y i i
y
max ( , , , )

denote
crosscorrelation values at lags
x
and
y
, respectively, for
which
x
and

y
are maxima. By taking the maximum val
ues of the crosscorrelations with respect to their respective
lags, this definition for coherence automatically accommo
dates for local dip. This analysis is continued for all sam
ples down the master trace and for all the traces in the vol
ume. Thus, a new 3D data volume is generated that contains
measurements of the degree of tracetotrace similarity.
Typical window lengths vary from 40 to 100 ms.
Channels often are associated with gascharged sand in
the Gulf of Mexico. In Figure 8a, we show a 40ms aver
ageabsoluteamplitude (AAA) slice and the crosscorrela
tion coherence through the same volume, showing such a
channel (Figure 8b). The AAA is a common measure of re
flectivity within an analysis window and is available on al
most all commercial workstation systems. We note that the
channel edges are difficult to see in the AAA slice but are
quite visible in the coherence slice. However, these two at
tributes are complementary, and when we use them together
we interpret the highamplitude anomaly as a gascharged
sand lying within the channel. (We will discuss defining
such complementary displays in Chapter 9).
In Figure 9a, we show a vertical cross section through
the seismic data for the same volume as in Figure 8. Al
though the channels are obvious to an experienced inter
preter, we do not know from this lone vertical slice whether
Trace #1
Shifted windows of
Trace #2
Crosscorrelation 20 10 0 +10 +20 Lag, :
40 ms
Maximum
coherence

Figure 7. A schematic diagram showing crosscorrelation
between two traces. Trace number 1 is held fxed while a
window of trace number 2 (here, 40 ms) is slid along at a
suite of time lags and is crosscorrelated. The lag having the
maximum signed crosscorrelation is a crude measure of
inline (or crossline) dip. The crosscorrelation value that cor
responds to this peak is then used in equation 3.4 to generate
a 3D estimate coherence.
a) b)
Figure 8. (a) A horizontal time slice of average absolute amplitude (AAA) calculated over a 40ms vertical analysis window.
(b) The corresponding threetrace 40ms crosscorrelation coherence slice through the same volume. The bright spot seen in (a)
corresponds to the highest coherence value displayed in yellow in (b), lying within the channel. Because bright spots generally
have a high signaltonoise ratio, it is common for them to show up with very high coherence. After Bahorich et al. (1995).
a) b)
Figure 9. (a) Channels (indicated by arrows) seen on a conventional vertical seismic slice through the same volume as that in
the previous fgure. (b) Channels in map view generated by an 11trace 40ms semblancebased coherence algorithm. Note the
higher signaltonoise ratio of the image here compared with the threetrace crosscorrelation coherence shown in Figure 8b.
The map view helps us differentiate between separate channels (perhaps having been formed at different geologic times) and a
single meandering channel. After Haskell et al. (1998).
a) b)
we have three channels or a single meandering channel that
repeatedly crosses the vertical section. Examination of the
coherence slice (Figure 9b) shows clearly that we have
three distinct channels, although of course they could be
either the same channel occupying different spatial posi
tions at different geologic times, or three simultaneous
channels.
Faults may or may not be easy to see on time slices.
Faults that cut perpendicularly to reflector strike, such as
those indicated by the white arrows in Figure 10b, usually
are easy to identify. However, faults that cut parallel to strike,
such as those indicated by black arrows in Figure 10b, can be
quite difficult to identify. Coherence allows us to identify
faults regardless of their orientation (Figure 10a).
Semblance-based Coherence
In the semblance approach to computing coherence, we
define a space and a time aperture, or a 3D analysis win
dow, for the data. We also need to define a dip and azimuth
for each point in the data volume. This dip and azimuth
may be userdefined (implicitly) through flattening, or de
Coherence 51
fined explicitly through a separate dip calculation, most
typically using the semblancedriven dip scan described by
Marfurt et al. (1999) and summarized in Chapter 2.
Let us define an analysis window (either elliptical, as
in Figure 11a, or rectangular, as in Figure 11b) containing J
traces centered about the analysis point, as shown in Figure
11. We define the semblance (t,p,q) to be the ratio of the
energy of the average trace to the average energy of all the
traces along a specified dip:
(3.5) ( , , )
( )
( )
t p q
J
u t px qy
J
u t px qy
j j j
j
J
j j j

{ }
=
1
1
1
2
2
j
J
=
1
where the subscript j denotes the jth trace falling within the
analysis window, x
j
and y
j
are the x and y distances of the
jth trace from the center of the analysis window, and the ap
parent dips pandqare measured in milliseconds per meter
or per foot and define a local planar event at time t. This
semblance is a measure of the degree of similarity to each
other of all of the traces along the selected dip within the
selected square or elliptical outline.
The semblance estimate given by equation 3.5 will be
contaminated most severely by background noise if we
compute the semblance of strong, coherent events near or
at their zero crossings. To avoid that problem, we adopt
conventionalvelocityanalysis semblance scans and calcu
N
3 km
Figure 10. Time slices at t = 2.600 s
through (a) the coherence volume and
(b) the corresponding seismic volume,
for a survey acquired in northwestern
Louisiana, U.S.A. White arrows indi
cate faults that cut perpendicularly to
strike and are easy to see on the seis
mic time slice. Black arrows indicate
faults that cut nearly parallel to strike
and are diffcult to differentiate from
normal waveform changes seen on the
time slice. Small black rectangles in
(a) and corresponding gray rectangles
in (b) indicate acreage that was not
permitted for exploration. Data are
courtesy of Seitel.
a) b)
Figure 11. (a) Elliptical analysis window and (b) rectangular analysis window, each centered about an analysis point defned
by the length of major axis a, the length of minor axis b, and the azimuth of major axis
a
. The term
0
indicates the azimuth of
the inline seismic axis from north. After Marfurt et al. (1998).
a) b)
late an average semblance over a vertical analysis window:

(3.6)
c t p q
J
u t k t px qy
J
u t k t
s
j j j
j
J
k K
K
j
( , , )
( )
(
=
+
+
= =

1
1
1
2
ppx qy
j j
j
J
k K
K
= =

)
2
1
where x
j
and y
j
denote the x and y distances of the jth trace
from the master trace (Figure 6b) and we sum over (2K+ 1)
samples. Graphically, we display the input to the analysis
as the windowed five traces shown in Figure 12a. In Figure
12b, we display the average trace. In Figure 12c, we replace
each original trace by the average of all the traces (this
extra step will help us compare the semblance algorithm to
the eigenstructure algorithm presented later in Figure 17).
The semblance is then simply the ratio of the energy of the
data shown in Figure 12c to the energy of the data shown in
Figure 12a.
Increasing the number of traces in the 11trace sem
blance estimate of coherence shown in Figure 9b reduces
the random speckles seen in the threetrace crosscorrelation
estimate of coherence shown in Figure 8b.
Variance-based Coherence
Another popular measure of waveform similarity is the
variance. As implemented, we will show that the variance
estimate of coherence is identical to the semblance estimate
discussed above. Modifying the formal definition of vari
ance, var(t,p,q), to be computed along samples lying on a
dipping reflector, we write
, (3.7)

where the mean, <u(t,p,q)>, is defined as
. (3.8)
Note here that <u>is calculated for each plane parallel to
the reflector with the analysis window. Whereas equations
3.7 and 3.8 provide the formal definition of variance, most
statisticians use the much more efficient (and mathemati
cally equivalent) computational form
var( , , )
( , , )
( ,
t p q
J
u t px qy x y
J
u t px qy x
j j j j
j
J
j j j
=
1
1
1
2
,, ) y
j
j
J
=
1
2
. (3.9)
To estimate coherence, we sum the variance over a ver
tical analysis window of 2K+ 1 samples and normalize by
the energy of all the traces to obtain
J
u t k t px qy
j j j
j
J
( )
c t p q
v
( , , )
+
1
1
2
k K
K
=
+
J
u t k t px qy
j j j
( +
1
))
j
J
k K
K
= =
+

1
2
( )
j j j
j
J
k
J
u t k t px qy
=
2
1
1
=
+
K
K
, (3.10)
or
c t p q
J
u t k t px qy
J
u t k
v
j j j
j
J
k K
K
j
( , , )
( )
(
=
+
+
= =
+

1
1
1
1
2
t px qy
j j
j
J
k K
K
=
= =
+

)
2
1
c t p q
s
( , , ) 1 . (3.11)
A
n
a
l
y
s
i
s
w
i
n
d
o
w
tKt
t+Kt
Dip
A
n
a
l
y
s
i
s
w
i
n
d
o
w
tKt
Dip
t+Kt
Figure 12. A schematic diagram showing the steps used in
semblance estimation of coherence. (a) First, we calculate
the energy of the fve input traces within an analysis win
dow, (b) then we calculate the average trace, and (c) fnally,
we replace each trace by the average trace and calculate the
energy of the fve average traces. The semblance is the ratio
of the energy of (c) to the energy of (a). If each windowed
trace in (a) has the identical waveform and amplitude, the
semblance is 1.0; otherwise, it is less than 1.0.
a)
b)
c)
Coherence 53
Thus, the variance estimate of coherence is identical nu
merically to one minus the semblance estimate of coherence.
The variance (and therefore also the semblance) is a measure
of how well each trace fits the mean trace. If all traces are
equal, the variance estimate of coherence, c
v
, is 0.0 (and the
semblance estimate of coherence, c
s
, is 1.0). However, even
if all the traces have the exact same waveform, if their ampli
tudes are different, c
v
is greater than 0.0 and c
s
is less than
1.0. In contrast, the crosscorrelation estimate of coherence,
c
xc
,
is 1.0 as long as the waveform is the same, regardless of

the amplitude of each trace. (We will show later in this chap
ter that the eigenstructure estimate of coherence has this
same desirable sensitivity only to waveform).
Artifacts resulting from ignoring
reflector dip and azimuth in
coherence calculations
Coherence calculations based on crosscorrelation be
tween the single adjacent trace and the center trace automati
cally accommodate for local dip through the selection of the
maximum value of the crosscorrelation as a function of the
time lag, . By contrast, because semblancebased coherence
calculations typically include many traces about the desired
output location, the local dip and azimuth calculation should
be completed as a first step. Both semblance and variance
estimates of coherence of seismic data currently are provided
as an option in a popular interpretive workstation software
product. Unfortunately, for computational efficiency, in sev
eral implementations no search over dip exists, such that the
values of p and q in the preceding equations are set identi
cally to zero. A finite dip will misalign the seismic wave
forms, thereby causing a variance or semblance anomaly as
sociated with structure, as seen in Figure 13ac. Such
structureinduced anomalies should not be confused with
faults or fractures, although it is not unreasonable to expect
fractures where we experience a change in dip.
In general, two good methods exist for calculating co
herence volumes in the presence of structural dip. The first
workflow, which involves a direct search of volumetric dip
and azimuth like that discussed in Chapter 2, prior to or as
part of the coherence calculation, is 50200 times more in
tensive computationally than a coherence calculation with
out a dip search. Except for small surveys, such calcula
tions are performed more appropriately on a computer
server by a processing center or by an oil company technol
ogist. In such a scenario, the interpreter loads a precomput
ed coherence volume that includes the advantage of an ex
plicit volumetric dip and azimuth search and then extracts
either time slices or horizon slices (Figure 14a and b).
The second workflow is more appropriate to a stand
alone workstation. The interpreter first flattens a slab of the
seismic volume about an interpreted horizon, implicitly de
fining a dip/azimuth for each trace. The interpreter then
Figure 13. Artifacts generated by not calculating coherence along the dip and azimuth of the refector: (a) coherence calcu
lated over an entire volume without a dip search and then sliced along a picked horizon; (b) the coherence image shown in
(a) overlain by structural contours; and (c) coherence calculated along the dip and azimuth of the refector. Features seen on
the horizon slice in (a) indicated by white arrows initially were thought to show collapse features and fractures of interest.
However, when the structural contours were overplotted in (b), it was clear that most of the features were simply the result of
structure. After Simon (2005).
a) b) c)
calculates coherence without the timeconsuming dip search
on the flattened slab of data (Figure 14c).
The Manhattan Distance and
Other Norms That Measure
Waveform Similarity
The semblance estimate given by equation 3.6 is based
on notions of similar trace energy, whereas the classical
variance estimate given by equation 3.7 is based on notions
of squared error about the mean. When they are normalized
by the total energy of the data in the analysis window, these
two measures become identical. Geophysicists have long
used other error measures, in particular when we wish to
assign a lower weight to outlier measurements. Equation
3.6 can be generalized to

, (3.12)

c t p q
J
u t k t px qy
J
u t k t px qy
r
j j j
j
J
r
k K
K
j j
( , , )
( )
(
=
+
+
= =

1
1
1
jj
r
j
J
k K
K
)
= =

1
where the symbol
f (x)
indicates the absolute value of the

function f (x), and where now we measure what statisticians
call the distance using the more general rnorm rather than
using the Pythagorean squared norm (the sum of the squares,
which provides a measure of energy). One major vendor
estimates coherence using the Manhattan distance with r =
1, rather than the Pythagorean distance with r = 2such as is
used in equation 3.6. Although inhabitants of New Jersey
and other uncivilized parts of the world west of the Hudson
River measure distances as the crow flies, where the shortest
distance between two points (x
2
,y
2
) and (x
1
,y
1
) is given by a
straight line of length
( ) ( ) x x y y
2 1
2
2 1
2
+
, the inhabit
ants of densely built Manhattan measure distances as so
many blocks uptown or downtown, by so many blocks
crosstown, giving a distance of
x
2
x
1
+
y
2
y
1
.
Such an r = 1 norm as a measure of error can provide supe
rior results when we are fitting noisy outlier samples. How
ever, we find that by the time most seismic data have been
migrated and find their way to a workstation, the strong
spikes have been removed, such that equations (3.12) with
r = 1 and (3.6) with r = 2 provide comparable images.
Using Analytic Traces in
Coherence Computations
If we use a small verticalanalysis window in equation
3.6, say, K = 1, we still may encounter artifacts in our co
herence image near zero crossings in the data. In general,
we can assume that a constant background level of incoher
ent seismic noise exists. Once the amplitude of our signal
goes below that background level, our coherence algorithm
will think the data are incoherent, thereby giving rise to
lowcoherence artifacts that follow structure, as we see in
the example shown in Figure 15. We circumvent such a
problem by applying the semblance computation to the an
Figure 14. Alternative ways to compute and visualize coher
ence volumes. Computation of a coherence volume requires a
value of dip and azimuth at every input sample. (a)(b): The
simplest but computationally most intensive approach is to
calculate coherence for the entire seismic volume, searching
for dip and azimuth as part of the calculation. The interpreter
then uses a conventional work fow of viewing and interpret
ing the coherence volume on (a) time slices or (b) horizon
slices. (c) A computationally more effcient but in general less
accurate approach is frst to fatten a window of data along a
picked horizon. The dip magnitude of this fattened volume is
implicitly equal to zero, such that no dip and azimuth search
is needed. The resulting coherence subvolume is a suite of
phantom horizons parallel to the original picked horizon. This
second approach inherits the interpreter bias and errors made
in the original picks. Furthermore, many geologic features,
such as diapirs, karst, fuvial deltaic systems, and masstrans
port complexes may not be associated with an overlying or
underlying horizon that can be picked accurately. Figure cour
tesy of Sue Nissen.
a) b) c)
Coherence 55
of the real trace using equation 3.6 (Figure 16a). Note that
Figure 16b is significantly whiter, or more coherent, than Fig
ure 16a, with greatly reduced lowcoherence artifacts re
sulting from zero crossings.
Eigenstructure-based Coherence
One drawback of the multitrace semblance and vari
ancebased estimates of coherence is that they are sensitive
not only to waveform but also to lateral changes in seismic
amplitude. In contrast, the threetrace crosscorrelation al
gorithm, although it is more sensitive to noise, is sensitive
only to changes in waveform, not in amplitude. Kalkomey
(1997), Barnes and Laughlin (2002), and others recommend
that when we use multiple seismic attributes for reservoir
characterization, those attributes should (1) have a firm basis
on physics or geology and (2) be mathematically indepen
dent of each other. Our philosophy is that, if possible, each
attribute should measure only one property of the seismic
response. Then multiple attributes can be combined later
either graphically (see Chapter 9) or numerically (using ei
ther geostatistics or neural networks). In Chapter 5, we will
discuss coherent amplitude gradients, which are a direct
measure of lateral changes in reflector amplitude. We will
show that although changes in waveform and changes in
amplitude often are coupled through the underlying geolo
gy, by using eigenstructure analysis, we can separate them
into two mathematically independent measures that are
amenable to subsequent analysis.
We begin the eigenstructure method with estimates of
dip and azimuth, which we discussed in Chapter 2. Not sur
prisingly, the eigenstructure method can be used as a mea
sure for discretely scanning dip and azimuth. Graphically,
the eigenstructure method analyzes a window of traces
(Figure 17a) and determines which (yet to be scaled) wave
let best represents the waveform variability (Figure 17b).
Then this wavelet is scaled to fit to each input trace, provid
ing what we call the coherent component of the data within
alytic trace (discussed in Chapter 1), rather than simply to
the real trace, as given by
c t p q
J
u t k t px qy
s
j j j
j
J
( , , )
( )
+
,
,
]
]
]
+
1
1
2
J
u t k t px qy
k K
K
j j j
( )

+ ,
]
]
+
J
u t k t px
j
H
j
( +
1
qqy
j
j
J
)
,
]
]
]
1
2
2
1
uu t k t px qy
j
H
j j
j
J
k K
K
( ) + ,
]
]
2
1
, (3.13)
where the superscript H denotes the Hilbert transform (or
quadrature, or 90 phase rotation) of the input seismic data.
When the original trace goes to zero, the magnitude of the
quadrature trace is at a maximum, thus circumventing the
previously mentioned problem of accurately calculating the
semblance in a smallamplitude portion of a trace in the
presence of noise.
Conversely, when the quadrature goes to zero the mag
nitude of the original trace is at a maximum, again provid
ing a robust semblancebased coherency calculation in the
presence of noise. With both real and imaginary parts of the
analytic trace included in equation 3.13, the signaltonoise
ratio is good if the envelope (discussed in Chapter 1) of the
trace, u, is significantly larger than the level of background
noise. Not surprisingly, if the envelope (also known as the
reflection strength) falls below the backgroundnoise level,
such as we sometimes see in shaleonshale reflections, we
still will see bands of low coherence following structure. We
also will see low coherence associated with situations of
geologic interest such as turbidites, along angular and ero
sional unconformities, and within salt and shale diapirs.
Figure 16 shows the improved results obtained by estimating
the semblance of the analytic trace using equation 3.13 (Fig
ure 16b) versus those obtained by estimating the semblance
8 ms
Figure 15. The effect of zero crossings
in the signal, using the traditional sem
blance algorithm given by equation 3.6.
(a) The signaltonoise ratio is low when
a short analysis window (here, 8 ms, cor
responding to three samples) straddles
a zero crossing. (b) The resulting coher
ence image seen on a time slice at t =
2.164 s. White arrows indicate artifacts
associated with low signaltonoise about
zero crossings. Black arrows indicate
geologic faults of interest.
a) b)
the analysis window (Figure 17c). Eigenstructure coher
ence is simply the ratio of the energy of the coherent com
ponent of the data (Figure 17c) to the energy of the original
traces (Figure 17a) within the analysis window. For 3D
data, the analysis window includes a suite of traces cen
tered around the analysis point (Figure 11).
To help the reader understand the wealth of published
information on eigenstructure (also called principalcom
ponent) analysis, we will introduce some terminology used
by statisticians and show step by step how to calculate the
covariance matrix (Figure 18). In step 1, we extract a suite
of sample vectors from the data (Figure 12). If our analysis
window includes nine traces, the datas sample vector will
be of length 9. There will be one sample vector that corre
sponds to a suite of timeinterpolated data samples, for each
value of k ranging between +K and K, that falls within our
verticalanalysis window and is extracted parallel to a plane
defined by apparent dips p and q. Each sample vector forms
a row in the data matrix. In step 2, we take each column of
the data matrix (which is simply a shifted window of each
seismic trace) and crosscorrelate that column with itself
and with each other column, forming what we call a covari-
Figure 17. A schematic diagram showing the steps used in eigenstructure estimation of coherence. (a) First, we calculate the
energy of the input traces within an analysis window. (b) Next, we calculate the seismic waveform that best approximates the
waveform of each input trace, (c ) Finally, we replace each trace by a scaled version of (b) that best fts the input trace. The
eigenstructure coherence is the ratio of the energy of (c) to the energy of (a). If each windowed trace in (a) has exactly the same
waveform (but perhaps a different amplitude), the coherence is 1.0; otherwise, it is less than 1.0.
A
n
a
l
y
s
i
s
w
i
n
d
o
w
Dip
A
n
a
l
y
s
i
s
w
i
n
d
o
w
tKt
t+Kt
Dip
tKt
t+Kt
c)
b)
a)
Figure 16. Semblance estimates of coherence generated using (a) the input data trace and (b) the analytic trace. Vertical win
dow length = 20 ms. At a fxed level of noise, the signaltonoise ratio can become low near refector zero crossings, thereby
resulting in lowcoherence artifacts that follow the structure (arrows). Using the analytic trace (equation 3.13) avoids this prob
lem. When the magnitude of the real input trace is low, the magnitude of the quadrature component is high. Likewise, when the
magnitude of the quadrature component is low, the magnitude of the real input trace is high, thereby maintaining a good signal
tonoise ratio in the presence of strong refectors. It still is common to see lowcoherence trends following structure when we
have lowrefectivity (and hence low signaltonoise ratio) shaleonshale events, and for truly incoherent geology such as that
encountered with erosional and angular unconformities, or with karst, masstransport complexes, and turbidites.
a) b)
Coherence 57
ance matrix. The covariance matrix contains all possible
unnormalized crosscorrelations and autocorrelations of all
nine windowed traces with each other. Those correlations
then are ordered in a systematic fashion, as shown in Figure
18. Specifically, crosscorrelating (without normalization) a
window of trace 2 (column 2 of the data matrix) with the
corresponding window of trace 8 (column 8 of the data ma
trix) gives us the C
28
element of the 9 9 covariance ma
trix (step 3). Mathematically, the C
ij
component of the co
variance matrix is given as
c t p q
[u t k t px qy
ij
i i i
( , , )
( ) <u(t,p,q)>]
=
+
[u t k t px qy
j j j
( ) <u(t,p,q)>] +
k K
+ K
=
, (3.14)
Figure 18. Steps for generating the data covariance
matrix used in calculating coherence, coherent amplitude
gradients (discussed in Chapter 5), and structurally ori
ented principalcomponent fltering (discussed in Chapter
8). In this illustration, K = 1, resulting in three sample
data vectors. See text and equation 3.14 for details.
where, unlike in equation 3.8, <u(t,p,q)> is now the mean
of each windowed trace. Statisticians often approximate the
(unknown) true mean by the sample mean. Fortunately, for
properly recorded seismic data, we can assume that the true
mean is zero.
When we have computed our covariance matrix, C, the
next step is to decompose it into eigenvectors, each of
which has a corresponding eigenvalue. Interpreters who
work with attributes recognize that eigenvectors (also called
principal components) and eigenvalues are used routinely
in multiattribute analysis for reservoir characterization. If
we assume that the mean amplitude value, , is zero, then
the first eigenvalue,
1
, tells us what portion of the energy
of all the traces in the analysis window can be expressed by
a single waveform, with a different scaling parameter for
each trace. The corresponding eigenvector, v
1
, is a list of
those scaling coefficients, one for each trace, that when
multiplied by the waveform, best fit the data. We will dis
cuss eigenvectors in Chapter 5 on coherent amplitude gra
dients and in Chapter 8 on structureoriented filtering. For
now, all we need to know is that any matrix can be decom
posed into eigenvectors and eigenvalues. For any diagonal
ly symmetric, realvalued covariance matrix, we can write

Cv
(m)
=
m
v
(m)

, (3.15)
where v
(m)
indicates the mth eigenvector and
m
indicates
the corresponding mth eigenvalue. In general, for a J J
covariance matrix, there are J mathematically independent
eigenvectors, where J is the number of traces in the spatial
analysis window. By convention, the eigenvectors and ei
genvalues are ordered, with the one best expressing the en
ergy of data within the analysis window being first, the one
best expressing the remaining data variation being second,
and so on.
In general, only a few eigenvectors and eigenvalues are
needed to represent 95% of the data. Indeed, the eigenstruc
ture coherence estimate (Gersztenkorn and Marfurt, 1999)
uses only the first eigenvalue:

c
e

j
J
=
1
C
jj
1

, (3.16)
where J is the number of traces used in the analysis win
dow, the denominator represents the energy of all the traces,
and the numerator
1
represents the energy that can be rep
resented by the first eigenvector, v
1
. Thus, if all the traces
in the same analysis window have the exact same shape
(but perhaps with different amplitudes, including even trac
es with reversed polarity), the eigenstructure coherence, c
e
,
will be 1.0. By contrast, if the waveforms are consistent but
the amplitudes are not identical, the semblance coherence,
c
s
, will not be equal to 1.0 and the variance estimate of co
herence, c
e
, will not be equal to 0.0. Thus, the eigenstruc
ture coherence has the advantage of being blind to traceto
trace amplitude scaling differences.
Like the crosscorrelation, semblance, and variance es
timates of coherence, the eigenstructure coherence estimate
is applied to a window centered on each sample of each
trace in the seismic data volume. At each point we extract a
data matrix, form the corresponding covariance matrix, cal
culate the first eigenvalue, calculate the energy of the traces
that compose the data matrix, take their ratio, and store the
solution as c
e
. The analysis cube is moved throughout the
3D seismic volume, and the output is a 3D data set consist
ing of coherence coefficients defined for each amplitude
point of the input 3D data volume.
In Figure 19, we show one of the first applications of
the eigenstructure coherence estimate and compare it with
the crosscorrelation and semblance estimates of coherence
discussed above. In this early application, the eigenstruc
ture algorithm assumed that the apparent dips, p and q,
were identically zero, which led to some structural artifacts
or leakage such as that discussed earlier in Figure 13. Such
artifacts are particularly evident in structurally complex ter
rains such as those shown next in Figure 20a. By first cal
culating volumetric dip and azimuth as discussed in Chap
ter 2 and then performing the calculation using equations
(3.14) through (3.16), we obtain the artifactfree result
shown in Figure 20b.
In Figure 21, we display a vertical seismic section and
a set of seismic time slices at 0.2s intervals. Note a central
syncline bounded by two anticlines. In Figure 22, we dis
play corresponding slices through a coherence volume that
was generated using equations 3.14 through 3.16 with the
assumption that the dip components p and q were identi
cally zero. Arrows indicate two bands of low coherence
that appear on the vertical section exactly where we might
anticipate fractures. These features are the result of struc
tural leakage, not fractures, as was verified by a more care
ful (and more expensive) calculation of coherence along
the estimated structural dip and shown in Figure 23.
Typically, dip scans sample the data discretely at finite
intervals, such as 5 or 0.001ms/m. We can improve
our calculations by interpolating the dip as we discussed in
Chapter 2. In Figure 24, we display time slices through co
herence volumes generated using a threetrace crosscorre
lation and a fivetrace semblance, an eigenstructure without
a dip scan, an eigenstructure with a discrete dip scan, and
an eigenstructure with interpolated dips. Each algorithmic
approach provides a small but incremental improvement in
our image quality.
A final improvement to the eigenstructure calculation
is to use the analytic trace rather than the real input trace
when we compute the covariance matrix. Marfurt (2006)
evaluated two alternatives forming a complex, symmet
ric covariance matrix directly with the analytic trace, and
using the quadrature component to form additional sample
vectors in equation 3.13. The first approach allows for com
plex eigenvectors and hence some amount of wavelet rota
tion within the analysis window, which does not fit our pre
conception of a coherent reflection. In contrast, the second
(preferred) method does not. Thus, in the second approach,
the equation for our covariance matrix becomes
C t p q
u t k t px qy t p q
ij
i i i
( , , )
( ) ( , , )
=
+
[ ]

k K
K
u t k t px qy
j j j
( ) + (( , , ) t p q
=
+
( ) ( , , ) u t k t px qy t p q
H
i i
H
i
+ +

k K
K
=
+
( u t
H
j
+ kk t px qy t p q
j j
H

) ( , , )
(3.17)
Coherence 59
where the superscript H denotes the Hilbert transform (or
quadrature) of the input data.
In Figure 25, we apply the different generations of co
herence algorithms to a data volume from offshore eastern
Canada. We can see a progressive improvement in the reso
lution of discrete faults as we move from the crosscorrela
tion algorithm, through the fivetrace semblance and the
fivetrace eigenstructure (without dip scan), to the fivetrace
eigenstructure with the dip scan, using the analytic trace.
Gradient Structure
Tensor-based Coherence
Bakker (2003) developed and analyzed a very com
plete suite of interpretation and datafiltering algorithms
based on the gradient structure tensor (GST). The GST was
described in Chapter 2 as a robust means of estimating re
flector dip and azimuth. In contrast to the JJ covariance
matrix given by equation 3.14, whose size is determined by
Figure 19. Comparison of alternative coherence algorithms used on data from South Marsh Island, Gulf of Mexico, U.S.A.
The volume shown contains both structural and stratigraphic features associated with deposition over a terrane infuenced by
salt tectonism. (a) A time slice through the timemigrated seismic data at 1.8 s. S denotes a salt dome, and F indicates sev
eral radial faults. Corresponding slices through coherence volumes were generated using the (b) threetrace crosscorrelation
algorithm, (c) a fvetrace semblance algorithm, and (d) a fvetrace eigenstructure algorithm. All coherence computations used
the same 80ms vertical analysis window. The circular rings seen in (a) correspond to sediments that are cut by radial faults
(indicated by F) and that are dipping against a salt dome. The disorganized feature indicated by C in the northeast is inter
preted to be a canyon. The salt dome and faults appear to be incoherent (black) in (b) through (d). Note that there is consider
ably less speckle noise in the fvetrace semblance algorithm than in the threetrace crosscorrelation algorithm. An even greater
improvement in signal to noise and lateral resolution accompanies the fvetrace eigenstructure algorithm. The structural arti
facts (leakage) about the salt dome indicated by L and the overall grayer level of the image in (d) are the result of a failure, in
this early work, to search over structural dip. After Gersztenkorn and Marfurt (1996).
a) b)
c) d)
Figure. 20. Comparison of coherence volumes generated
(a) without and (b) with a dip search using an identical
eigenstructure estimate of coherence. The computational
effort needed to generate the coherence volume in (a) is only
10% of that needed to generate the coherence volume in
(b). However, failure to estimate waveform similarity along
refector dip results in artifacts that some call structuralleak-
age, thereby giving rise to contourlike features indicated by
arrows in (a). The lowcoherence contours result from the
structural dip being such that the lateral analysis window
spans events from different geologic horizons, each of which
has its own waveform. The contours in this image mask the
fault that is seen more clearly in (b). After Chopra (2002).
a) b)
Figure 21. A vertical seismic slice AA and time slices at
0.200s intervals through a timemigrated seismic volume
from the western U.S.A. We note a central syncline bounded
by two anticlines.
Figure 22. A vertical slice AA and time slices through the
eigenstructure coherence volume corresponding to the slices
through the seismic volume shown in Figure 21. Coherence
was calculated assuming a fat dip. Arrows indicate structural
leakage artifacts that might be misinterpreted as lowcoher
ence fracture zones.
the number of traces in the data volume, the covariance ma
trix given by equation 2.11 is 3 3, corresponding to the
threedimensionality of the directional derivative sample
vectors. Bakker describes several shape attributes, based on
the three eigenvectors and eigenvalues of the GST covari
ance matrix. The shape attribute that is most similar to the
coherence images discussed so far in this chapter is his
measure of how planar the features falling within the analy
sis window are (i.e., his measure of their planarity). For
discussion, we will call this calculation the GSTbased esti
Coherence 61
mate of coherence:

c
GST

+

1 2
1 2

, (3.18)
where
1
,and
2
are the first two eigenvalues of the GST. If
1
>>
2
, we have a coherent (and in general, dipping) pla
nar reflector. If
1
=
2
=
3 ,
we have totally incoherent
energy, and c
GST
approaches zero. If
1
=
2
>>
3
, we have
a lineament, and c
GST
again approaches zero. In Figure 26,
we display Bakkers (2003) comparison of the GSTbased
coherence versus the dipscanbased coherence. The GST
based coherence has lower lateral resolution because it uses
a larger analysis window in calculating the derivatives used
in equation 2.10. Because the GST matrix is based on de
rivatives of seismic amplitude, it will join semblance and
variance as an attribute that is sensitive both to changes in
waveform and to lateral changes in amplitude.
Randen and his colleagues (Randen et al., 2000) have
also worked with the GST, defining a measure they call
chaos, which we will denote by c
:

c

=
+

2
1
2
1 3

. (3.19)
Note that if
1
>>
2
, the coherence is high and c
goes
to 1. If
1
3
,
goes to 0. Finally, if
1
2
, but
3
0 [Bakkers (2003) lineament attribute], c
goes to +1.
At the time of writing this book, we have not found
enough published examples to compare quantitatively how
the more recently developed GSTbased coherence and
chaos algorithms compare with the more mature, dipscan
based coherence algorithms. However, we expect that a
similar suite of incremental algorithmic improvements will
make these newer techniques competitive and perhaps su
perior to dipscanbased coherence.
Least-squares-based Coherence
Our final algorithm, presented by Bednar (1998), is
based on a leastsquares fit to a plane through a seismic
data window. Bednars paper does not provide explicit de
tails, so we will provide our own interpretation of the meth
od. Clearly, one can fit a plane to a 3D analysis window of
seismic data using either a bruteforce leastsquares fit or a
more elegant predictionerror filter. The objective of either
approach is to minimize the squared error given by the
equation
E t p q u k t px qy r
k j j j j
j
J
k K
K
2
2
1
( , , ) ( ) =
{ }
= =
+

, (3.20)
with the constraint that

r
j
j
2
1
=

. (3.21)
In equation 3.20, the unknown coefficients
k
are in
versely proportional to the wavelet that best fits the data
within the window, and r
j
is proportional to the amplitude
of the coherent part of the trace. As we did with the eigen
structure algorithm, we solve these equations using an iter
ative technique, beginning with the mean trace as an esti
mate of an initial wavelet that best fits the data within the
analysis window. Many of the concepts associated with
leastsquares fitting are closely linked to eigenvector analy
sis. We present a result of Bednars (1998) leastsquares fit
as Figure 27.
Sensitivity of Coherence to
Analysis-window Size
One beauty of seismic coherence is that no matter what
the implementation, only a limited number of parameters
exists the maximum dip to be searched, a dipscanning
increment, the size of the vertical analysis window, and the
Figure 23. A vertical slice AA and time slices through the
eigenstructure coherence volume corresponding to those in
Figures 21 and 22. Coherence was calculated along refec
tor dip. Note that the structural leakage has disappeared,
although we do see bands of lowcoherence shales following
structure.
size of the spatial analysis window. Even though there are
only a few parameters to determine, the choice of parame
ters has a bearing on computation of the final coherence
coefficient and hence on the display. We already have noted
that an accurate estimate of structural dip is critical. In
practice, the interpreter scans through the seismic data and
determines the maximum reflector dip by inspection and
enters it as an input parameter. Searching for dips that are
unreasonably large not only will increase the computation
cost, it may also allow the algorithm to misidentify steeply
dipping backscattered ground roll as coherent reflectors. In
and of itself, the dipazimuthscan component of coherence
computations is a very powerful filter.
Sensitivity to spatial window size
Increasing the size of our analysis window increases
the run time and the angular resolution, and at a fixed depth
level it decreases the lateral resolution. Because frequency
content decreases with increasing depth, it often makes
sense to calculate coherence using a small window for a
shallow objective and using a larger window for a deeper
objective.
We begin our analysis with coherence computation of
a channel system using the following algorithms shown in
Figure 28: (b) a threetrace crosscorrelation, (c) a fivetrace
( r = 12.5m radius) semblance, and (d) a fivetrace (r =
12.5m radius) eigenstructure. (Figure 28a shows a time
slice for the same channel system.) The vertical analysis
window is 40 ms for all three algorithm images.
Of the three algorithm images, we prefer the eigen
structure image and will use that for our subsequent analy
sis. The improvement in lateral resolution in Figure 28d re
sults from the amplitude insensitivity of the eigenstructure
algorithm. In Figure 29, we vary the size of the spatial anal
ysis window from r= 12.5 m(shown earlier as Figure 28d)
Eigenstructure
(without
dip scan)
Eigenstructure
(discrete dip
scan)
Cross-
correlation
Semblance
(with dip
scan)
A
Eigenstructure
(interpolated
dip)
Vertical
seismic section
2.0
1.5
2.5
1.0
3.0
A
T
i
m
e

(
s
)
Figure 24. Evolution of seismic coher
ence algorithms, illustrated by time
slices at t = 1.500 s through coherence
volumes generated by crosscorrelation,
semblance with dip search, eigenstructure
without dip search, eigenstructure with
a discrete dip search, and eigenstructure
with an interpolated dip search. Black
arrows correlate the discontinuities seen
on the coherence slice to the vertical
seismic section. Failure to interpolate the
semblanceestimated dip results in the
structural leakage indicated by the white
arrows on the simple dipscancoherence
calculations.
Coherence 63
Figure 25. Evolution of coherence algo
rithms, applied here to faults and fractures
off the east coast of Canada. A time slice
through the seismic data, a threetrace
crosscorrelation coherence, a fvetrace
semblance coherence, a fvetrace eigen
structure coherence, and a modifed fve
trace eigenstructure coherence. The tem
poral aperture is 40 ms for all coherence
calculations. After Chopra (2002).
I
n
l
i
n
e

s
l
i
c
e
C
r
o
s
s
l
i
n
e

s
l
i
c
e
T
i
m
e

s
l
i
c
e
Seismic GST coherence Dip-scan coherence
Figure. 26. Orthogonal vertical slices and
time slices through a seismic data volume,
an eigenstructure coherence generated
using the gradient structure tensor (GST),
and an eigenstructure coherence generated
using a discrete dip scan. In these images
we see greater lateral resolution in the
eigenstructure dipscan coherence. After
Bakker (2003).
2

k
m
3.0
3.4
3.2
2.8
2.6
T
i
m
e

(
s
)
3.0
3.4
3.2
2.8
2.6
T
i
m
e

(
s
)
Figure 27. (a) A seismic data
volume and the (b) corre
sponding coherence volume
calculated by leastsquares
ftting of a plane to the seis
mic data within a dataanaly
sis window using an effcient
prediction error flter algo
rithm. After Bednar (1998).
a) b)
a)
c)
Figure 28. Analysis of a channel system showing the evolu
tion of coherence algorithms. as applied to a channel system.
Time slices through (a) a seismic data volume and (b)(d)
through its corresponding (b) threetrace crosscorrelation
coherence, (c) fvetrace semblance, and (d) fvetrace eigen
structure coherence volumes. Note the improvement in lat
eral resolution as we progress from (b) through (d). The ver
tical analysis window is 40 ms for all coherence calculations.
b)
d)
through r= 25.0 m, r=37.5 m, and ending with r= 50 m.
The vertical analysis window is fixed at 16 ms. Not surpris
ingly, a narrow channel indicated by the white arrow is best
delineated on the image generated with the smallest lateral
window (Figure 29a). Less certain is why the edges of a
larger channel indicated by the gray arrow also show up
best on this image. The largerwindow calculations com
pute not only coherence but also dip and azimuth using
more traces, such that the local estimates of reflector dip
and azimuth upon which coherence is calculated may be
different. An intermediatesize channel, indicated by the
striped arrow, appears only when the aperture is greater
than or equal to 25 m (Figure 29b, c, and d). Although most
features become progressively more blurred in the 37.5m
and 50mradius images (Figure 29c and d), we can see
considerably more detail inside one of the channels in the
37.5m window, as indicated by the dotted arrow (Figure
29c). We will discuss the impact of larger window analysis
of longer wavelength features in Chapters 4 and 5.
Sensitivity to vertical window size
Because the data are of good quality, in Figure 30 we
will use our fivetrace (r = 12.5 m) eigenstructure image
and vary the temporal (vertical) analysis window. The thin
channel indicated by the white arrow is best resolved using
the 8ms and 16ms analysis windows (Figure 30a and 30b,
respectively), as is the feature shown by the gray arrow. We
expect narrow channels also to be thin. As we increase the
vertical analysis window first to 24 ms and then to 32 ms
(Figure 30c and 30d, respectively), these two channels
bleed away, implying that they are low in amplitude and are
confined temporarily. Stacking more (stratigraphically un
correlated) data into the analysis does not improve the
image. In contrast, the channel indicated by the striped
arrow appears when we use the larger vertical windows.
Most likely, that channel is slightly below the time slice
Coherence 65
displayed and falls outside the smaller analysis windows.
The fact that we are able to detect it using the smaller, 16
ms vertical window and the larger, r = 25 m spatial window
implies that above this channel, smooth differential com
paction may exist that is detected only by the larger spatial
window.
In general, larger temporal (vertical) analysis windows
tend to blur stratigraphic features, as Figure 31a and b show.
Blumentritt et al. (2003) determined that for thin channels,
the optimum vertical analysis window is the reciprocal of
the dominant frequency. They noted that shorter windows
were more contaminated by noise, whereas longer windows
blurred the images and mixed stratigraphic features.
Figure 29. The effect of varying the spatial radius on the
channel system shown in Figure 28d, using the eigenstruc
ture algorithm with various spatial windows: (a) r = 12.5 m
including fve traces, (b) r = 25 m including 13 traces, (c) r
= 37.5 m including 25 traces, and (d) r = 50 m including 49
traces. The temporal analysis window = 16 ms or 5 samples
for all images. Note that the very thin channel feature (white
arrow) and the edges of a broad channel (gray arrow) are
illuminated by the 12.5m window but are washed out in
the other images from larger analysis windows. In contrast,
a somewhat wider channel (vertically striped arrow) is bet
ter illuminated when r > 12.5 m. Likewise, the horizontally
striped arrow indicates greater detail within a meandering
channel complex.
Spatial radius = 12.5 m Spatial radius = 25 m
Spatial radius = 37.5 m Spatial radius = 50 m
a)
c)
b)
d)
Figure 30. The effect of varying the temporal analysis
window on the channel system shown in Figures 28 and
29, here using the eigenstructure algorithm with (a) 8 ms or
three samples, (b) 16 ms or fve samples, (c) 24 ms or seven
samples, and (d) 32 ms or nine samples. The spatial analysis
window is r = 12.5 m, including fve traces. Note that the
channel features indicated by the white and gray arrows are
resolved better for shorter temporal analysis windows, imply
ing that they are confned to this time slice. In contrast, the
channel indicated by the striped arrow is imaged better by
the larger temporal analysis windows, indicating that it may
lie above or below the time slice displayed. In general, we
recommend using small analysis windows for stratigraphic
analysis to avoid mixing the information content of different
levels.
Temporal aperture = 8 ms Temporal aperture = 16 ms
Temporal aperture = 24 ms Temporal aperture = 32 ms
a)
c)
b)
d)
In contrast, larger temporal (vertical) analysis windows
tend to improve vertical faults, as shown in Figure 31c and
d. If the fault is vertical, discontinuities align in the stack
that is implicit in our crosscorrelation, semblance, and eigen
structure algorithms, and at the same time, noise, including
the stratigraphic features of potential interest, is attenuated.
Unfortunately, not all faults are vertical. Current implemen
tations of coherence integrate the discontinuities over the
vertical axis. Listric faults and other structural features of
interest that are not vertical become progressively blurred
in coherence displays. When a listric fault cuts a strong re
flector and we have a large vertical analysis window, we
may even see the fault twice once when it cuts the re
flector above the time slice of interest on the footwall, and
a second time when it cuts the reflector below the time slice
of interest on the hanging wall. We will point out such arti
facts in Chapter 11.
Next, we return to the data volume from South Marsh
Island, Louisiana, U.S.A., that we showed initially in Fig
ure 4. In Figure 32, we display time slices at 1.2 s through
the original seismic volume (Figure 32a) and a coherence
volume (Figure 32b). Although we can see parts of chan
nels and faults on the seismic time slice (indicated by white
and gray arrows in Figure 32a), many of the other linea
ments, such as the one indicated by the black arrow, are
quite ambiguous. Conventional interpretation would re
quire slicing the data vertically at various angles to deter
mine the cause of the feature on the time slice.
The corresponding coherence slice (Figure 32b) clearly
shows this lineament to be the edge of a channel. Indeed,
many structural and stratigraphic features now appear on
the data.
The reader may be somewhat suspicious about such a
dramatic difference in image quality. If you are wondering
Figure 31. A larger temporal analysis window tends to smear stratigraphic features such as channels [compare the circled
features in (a) and (b)] that are confned to a given geologic horizon. In contrast, larger temporal analysis windows improve
the appearance of vertical faults [compare the faults in (c) and (d) that are identifed by white and gray arrows]. The latter is
because the coherence computation stacks similar discontinuities over a larger vertical window.
a) b)
c) d)
Coherence 67
about the origin of the additional information portrayed in
Figure 32b, we remind you that the coherence time slice
reveals information that is in the displayed time slice and
information in eight additional amplitude time slices that
resided in the analysis window. If we were able to animate
through these nine time slices and memorize key features,
we could convince ourselves that coherence is not creating
something out of nothing. To quote Adam Gersztenkorns
concluding remarks in his 1996 SEG presentation (Gersz
tenkorn and Marfurt, 1996), Given a sufficiently large sal
ary and sufficient time, a good geoscientist doesnt need
coherence to interpret these features.
Benefits of Coherence Extraction
along Time Slices versus
along Horizon Slices
Now we will compare the appearance of amplitude and
coherence data like those we just saw from along a constant
time slice, with amplitude and coherence data from the same
volumes but taken from along a selected horizon. In Figure
32, we displayed a time slice through the seismic and coher
ence data volumes at a selected constant time. In Figure 33,
we show corresponding horizon slices at approximately the
same level; these slices follow a picked horizon of interest.
Figure 32. Time slices, at t = 1.200 s, through (a) a seismic data volume and (b) the corresponding coherence volume calcu
lated using an 11trace semblance algorithm. Although channels (white arrows) and faults (gray arrows) can be seen on the
seismic time slice, lineaments such as that indicated by the black arrow are more ambiguous. The coherence slice allows us to
interpret our data more confdently in the timeslice mode. The black arrow clearly indicates the edge of a channel. Previously
unrecognizable faults and channels now appear. After Marfurt et al. (1998).
a) b)
Figure 33. Horizon slices through (a) a seismic data volume and (b) a coherence volume extracted along an interpreted
Pleistocene horizon corresponding approximately to the time slice shown in Figure 32. The area displayed is smaller than that
seen in Figure 32, because the horizon did not exist over a salt diapir in the southwest and was truncated in the northeast. In
general, stratigraphic features are seen best on horizon slices, whereas structural features are seen best on time slices. See text
for discussion. Semblancebased coherence was calculated using 11 traces and a 64ms vertical analysis window. After Marfurt
et al. (1998).
a) b)
Figure 34. The impact of using a smaller analysis window on resolving stratigraphic features. Both images use a threesam
ple, 16ms analysis window that is shifted 24 ms higher in (a) than in the image shown in (b), which is centered about the
same horizon shown in Figure 33b. Note the improved lateral resolution that we gain by using a smaller analysis window. Also
note that we now see that the narrow (and probably thin) channel features indicated by gray arrows are shallower than those
indicated by the black arrows. The broader channel indicated by white arrows has cut down from a shallower horizon. After
Marfurt et al. (1998).
a) b)
The lateral extents of these horizonbased images are
not as large as the extent of the previous time slice, because
the selected horizon did not exist in the southwest over a
salt diapir and had been eroded away in the northeast. Such
a limited field of view is one limitation of horizon slices
versus time and depth slices. A second limitation is time. It
took one of the authors (Marfurt) a full three days to pick
the Pleistocene horizon corresponding to Figure 32. (Read
ers who know this author may have serious doubts also
about the quality of his picks, which in polite circles is
called interpreterbias and which constitutes a third limita
tion.) Nevertheless, use of horizon slices (whether they are
explicitly picked horizons, phantom horizons, or slabs of
data flattened about a horizon) is the preferred method for
interpreting stratigraphic features. That preference holds in
the seismic data directly (Figure 33a) and in the corre
sponding attribute volume (Figure 33b). In these images,
we can see clearly the complex distributary channels, in
cluding point and longitudinal bars, of the Pleistocene paleo
Mississippi River. However, the westsouthwest to east
northeasttrending channel indicated by white arrows does
not fit with our distributarychannel model and instead cuts
the major system at right angles. We suspect that it is part
of a later system where the Mississippi River changed its
course.
The coherence volume shown in Figure 33b was run
with a 64ms analysis window, to better map faults associ
ated with salt diapirism. To better unravel our channel ge
ometry, we reran the coherence computation using a shorter
16ms analysis window, and we display the corresponding
horizon slice in Figure 34b and a shallower phantom hori
zon, 24 ms above, in Figure 34a. Because the channel indi
cated by the white arrows is delineated more clearly along
the shallower phantom horizon, we conclude that that chan
nel is a later event that has cut down through the main chan
nel system we see in Figure 34b. Although the images in
Figure 34 are less coherent (blacker) in general, they also
have better lateral resolution than the image did that used
the longer vertical analysis window and was shown in Fig
ure 33b. Longer windows mix stratigraphy, thereby in some
sense adding geologic noise to our image. We will discuss
such issues in greater detail in Chapter 12, in our discussion
of attribute illumination of clastic depositional environ
ments.
We conclude this chapter by examining the coherence
expression of the Delaware Basin survey displayed earlier
in Chapter 2. Figure 35 shows a suite of time slices at 0.8 s,
1.0 s, 1.2 s, 1.4 s, 1.6 s, and 1.8 s through the coherence
volume corresponding to the dip and azimuth seismic over
lays shown in Figure 20 of Chapter 2. The carbonate plat
Coherence 69
form appears to be relatively featureless until we approach
its edge. There we encounter the Brushy Canyon sequence
dumped into the basin from above (yellow arrows). We do
not see the broad fold indicated by the magenta arrow at
1.2 s, but we see it deeper at 1.4 s. We also see the broad
fold indicated by the blue arrows and seen earlier on the dip
and azimuth slices. The tightly folded strata seen in Figure
20e and 20f of Chapter 2 are quite organized here in Figure
35e and 35f.
Because we know that the reflectors are folded, we at
tribute these incoherent patterns to thinning of sediments
on the anticlinal parts of the folds. We note that such pat
terns easily could be misinterpreted as channels from the
coherence time slice alone. Such misinterpretations can be
avoided by testing our hypotheses against conventional
vertical slices through the seismic data. These discontinui
ties appear as vertical discontinuities in Figure 36. Although
their appearance is similar to that seen for faults, we inter
pret these discontinuities to be zones where the rocks have
been squeezed tightly and deformed. We will return to these
images in subsequent chapters when we discuss curvature
and coherent energy gradients.
Figure 35. Time slices at t
= (a) 0.800 s, (b) 1.000 s, (c)
1.200 s, (d) 1.400 s, (e) 1.600
s, and (f) 1.800 s through the
coherence volume correspond
ing to the dip and azimuth
overlays from a survey in the
Delaware Basin, New Mexico,
U.S.A., and shown in Chapter
2, Figure 20. We have retained
the same arrows shown in that
fgure to aid in the comparison.
a) b)
c) d)
e) f)
Figure 36. Vertical slices
through the coherence volume
corresponding to Figure 21 in
Chapter 2. Line locations are
indicated in Figure 35a.
a) b)
c) d)
e) f)
Coherence 71
Chapter Summary
Coherence is a direct measure of waveform similarity.
All coherence computations should be performed along the
local dip and azimuth of the best reflector at the analysis
point. Failure to do so will result in lowcoherence artifacts
that reflect structures overprinting anomalies of interest
that result from discontinuities in the waveform. Crosscor
relation and eigenstructure estimates measure only changes
in reflector waveform, whereas semblance, variance, and
GST estimates of coherence are sensitive to both waveform
and lateral changes of reflector amplitude. Although it is
computationally more intensive, the eigenstructure algo
rithm provides maximum lateral resolution. Furthermore,
by using the analytic trace rather than simply the input
trace, we can improve the fidelity of coherence images at
zero crossings for small vertical analysis windows.
Coherence integrates the information content of adja
cent traces and samples in a nonlinear manner that allows
us to extract information not seen on any individual time
slice. Coherence volumes significantly enhance our ability
to see structural and stratigraphic discontinuities on time
slices, thereby allowing us to more quickly produce a high
quality 3D interpretation that includes accurate fault and
channel geometries. In general, stratigraphic features are
shown best on horizon slices using a vertical analysis win
dow that is approximately equal to the period of the domi
nant frequency. In contrast, structural features that cut
across stratigraphy, such as vertical faults, are seen best on
constanttime (or depth) slices, which lack the interpreter
bias that would be present on horizonbased extractions. In
creasing the temporal (or depth) analysis window over ver
tical faults can improve their clarity by stacking the vertical
discontinuities seen in the seismic section. Increasing the
spatial (lateral) analysis window increases the cost of com
putation and decreases lateral resolution, but it can improve
the signaltonoise ratio, particularly in the deeper parts of
the seismic volume. In surveys with varying stratigraphic
and structural objectives, more than one coherence volume
should be generated, with apertures chosen to help visual
ize the full information content of the seismic data.
References
Bahorich, M. S., and S. L. Farmer, 1995, 3D seismic
coherency for faults and stratigraphic features: The
Leading Edge, 14, 10531058.
Bahorich, M.S., J. A. Lopez, N. L. Haskell, S. E. Nissen,
and A. Poole, 1995, Stratigraphic and structural inter
pretation with 3D coherence: 65th Annual Interna
tional Meeting, SEG, Expanded Abstracts, 97100.
Bakker, P., 2003, Image structure analysis for seismic
interpretation: Ph.D. thesis, Technische Universiteit
Delft.
Barnes, A. E., and K. J. Laughlin, 2002, Comparison of
methods for unsupervised classification of seismic data:
64th Conference and Technical Exhibition, EAGE, Ex
tended Abstracts, P222.
Bednar, J. B., 1998, Leastsquares dip and coherency at
tributes: The Leading Edge, 17, 775776.
Blumentritt, C. H., E. C. Sullivan, and K. J. Marfurt, 2003,
Channel detection using seismic attributes on the Cen
tral Basin Platform, west Texas: 73rd Annual Interna
Chopra, S., 2002, Coherence cube and beyond: First
Break, 20, 2733.
Cohen, I., and R. R. Coifman, 2002, Local discontinu
ity measures for 3D seismic data: Geophysics, 67,
19331945.
Gersztenkorn, A., and K. J. Marfurt, 1996, Eigenstructure
based coherence computations, 66th Annual Interna
, 1999, Eigenstructure based coherence compu
tations as an aid to 3D structural and stratigraphic
mapping: Geophysics, 64, 14681479.
Haskell, N. L., and S. E. Nissen, 1998, 3D seismic co
herency and the imaging of sedimentological fea
tures, in F. M. Gradstein, K. O. Sandvik, and N. J.
Milton, eds., Sequence stratigraphy concepts and ap
plications: Elsevier, 197214.
Haskell, N. L., S. E. Nissen, J. A. Lopez, and M. S. Baho
rich, M. S., 1995, 3D seismic coherency and the imag
ing of sedimentological features, 65th Annual Interna
Kalkomey, C. T., 1997, Potential risks when using seis
mic attributes as predictors of reservoir properties:
The Leading Edge, 16, 247251.
Lu, W., Y. Li, H. Xiao, and S. Zhang, 2003, Higherorder
statisticsbased coherency estimation algorithm: 73rd
Annual International Meeting, SEG, Expanded Ab
stracts, 17321735.
Marfurt, K. J., 2006, Robust estimates of 3D reflector
dip and azimuth: Geophysics, 71, 2940.
Marfurt, K. J., R. L. Kirlin, S. H. Farmer, and M. S. Baho
rich, M. S. 1998, 3D seismic attributes using a run
ning window semblancebased algorithm: Geophys
ics, 63, 11501165.
Marfurt, K. J., V. Sudhakar, A. Gersztenkorn, K. D. Craw
ford, and S. E. Nissen, 1999, Coherency calculations
in the presence of structural dip: Geophysics, 64, 104
111.
Randen, T., E. Monsen, C. Signer, A. Abrahamsen, J.
Hansen, T. Saeter, and J. Schlaf, 2000, Threedimen
sional texture attributes for seismic data analysis,
70th Annual International Meeting, SEG, Expanded
Abstracts, 668671.
Simon, Y. S., 2005, Stress and fracture characterization
in a shale reservoir, north Texas, using correlation
between new seismic attributes and well data: M.S.
thesis, University of Houston.

3 Coherence

Uploaded by

Copyright:

Available Formats

3 Coherence

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

3 Coherence

Uploaded by

Copyright:

Available Formats

45

is 1.0 as long as the waveform is the same, regardless of

indicates the absolute value of the

You might also like