A new approach to post-stack diffraction separation

Endrias G. Asgedom∗ University of Oslo, Leiv -J. Gelius University of Oslo, Andreas Austeng University of Oslo, and
Martin Tygel State University of Campinas

SUMMARY CRS DIFFRACTION TRAVELTIME

Downloaded 10/09/16 to 181.223.137.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Recently, separation and imaging of diffracted waves have The CRS hyperbolic moveout equation for a given reflection
shown the capability of providing highly resolved images of event, w, is specified by the ZO traveltime, τ0w , and a (refer-
discontinuities associated with complex geological areas. In ence) trace location, x0w :
this paper we propose a novel and highly robust technique for
separating diffractions from reflections. The main idea is that [τwθ (xm , h)]2 = [τ0w + Aw (xm − x0w )]2
diffracted waves (either from seismic or Ground Penetrating + Bw (xm − x0w )2 +Cw h2 , (1)
Radar (GPR)) can be described by a modified version of the
where (xm , h) are the midpoint and half-offset coordinate of a
moveout equation normally employed by the Common Reflec-
source-receiver pair in the vicinity of the reference location.
tion Surface (CRS) technique. To ensure an optimal selection
Moreover,
of the diffracted energy a coherency measure has to be applied.
θ = {Aw , Bw ,Cw } , (2)
Two different measures were tested in this study: Semblance
and MUltiple SIgnal Classification (MUSIC). The potential is the CRS parameter vector, with three parameters, Aw , Bw
of the new separation technique has been demonstrated using and Cw , to be estimated from the data.
both Marmousi data and a multi-offset GPR field dataset. In
In case the recorded data stems from a diffraction, the condi-
both cases well separated diffractions were obtained.
tion Bw = Cw holds. This is because, as the reflector shrinks to
a point, the N-wave turns out to be identical to the NIP-wave
(Zhang et al., 2001). As a consequence, the hyperbolic move-
INTRODUCTION out of diffractions (or diffraction traveltime), reduces to,

Seismic or GPR diffracted waves may carry high resolution [τwθ (xm , h)]2 = [τ0w + Aw (xm − x0w )]2
structural information associated with complex geological struc- + Bw [(xm − x0w )2 + h2 ] . (3)
tures like faults, pinchouts, wedgeouts or sudden changes in
Direct comparison between Eqs. (1) and (3) demonstrates that
facies that are often associated with hydrocarbon traps
the reflections and diffractions follow different traveltime curves.
(Kanasewich and Phadke, 1988; Moser and Howard, 2008).
Hence, in order to enhance the diffracted contributions relative
Accurate imaging of such structures requires optimal use of the
to those being reflected, the CRS stacking should be carried out
diffracted wave energy (Neidell, 1997). However, diffracted
based on Eq. (3). However, to ensure that optimal CRS param-
waves in seismic or GPR usually manifest themselves as sig-
eters are determined corresponding to the best fitting between
nificantly weaker signals compared to the reflections. As a
the CRS traveltime moveout and the actual diffraction event(s)
consequence, traditional imaging schemes, tailored for reflec-
present in the measurements, a coherency measure has to be
tions and treating diffractions as noise, generally suffer in qual-
employed.
ity and resolution (Neidell, 1997). Recently, however, many
attempts have been made to overcome this problem by separat-
ing diffractions from reflections followed by higher resolution CRS PARAMETER ESTIMATION BY COHERENCY
imaging. MEASURES
To enhance diffractions and suppress (or attenuate) reflections
present in a common offset section, Landa et al. (1987) pro- Standard parameter searching algorithms used in seismic or
posed the use of a specialized double-square-root traveltime GPR are often based on Semblance as coherency measure.
moveout. On the other hand, Fomel (2002) and Fomel et al. Even though Semblance in general is a good measure of co-
(2007) introduced plane-wave destruction filters to separate herency, it will often fail in case of interfering events (Biondi
diffractions in a zero-offset (ZO) section for later use within and Kostov, 1989; Kirlin, 1992). As recognized within sonar
migration velocity analysis. Moser and Howard (2008) used and radar applications, methods exploiting the properties of
the concept of anti-stationary filtering to perform depth imag- the eigenstructure of the data covariance matrix (i.e. subspace
ing of diffractions. In this paper we follow the basic approach methods) can lead to far better resolution than Semblance. Here
as in Landa et al (1987), but propose to use the Common Re- we briefly review both Semblance and one of the subspace
flection Surface (CRS) technique for describing the diffraction techniques named MUSIC. Both these coherency measures will
moveout in a ZO (stacked) section. To be able to select the be tested out for selecting the optimum diffraction stacking
optimal CRS diffraction-stack parameters a proper coherency surface.
measure has to be included. The work presented here focuses
Semblance
on the separation issue. However, the practical use of the
diffracted waves can be in residual velocity analysis, as well Following Du and Kirlin (1993), Semblance can be formulated
as in local high-resolution imaging. in terms of the covariance matrix of the data. Within a selected

© 2011 SEG
SEG San Antonio 2011 Annual Meeting 3861
 A new approach to post-stack diffraction separation

time window along the chosen trial moveout, Semblance can be frequency independent. This allows us to handle wideband
be mathematically written as seismic or GPR data. Windowing the event can be interpreted
as steering the covariance matrix before eigendecomposition
uT R(θ )u and using unity steering vectors for generating the MUSIC
Sc = , (4)
Mtr(R(θ )) pseudo-spectra (Kirlin, 1992).
where u is a column vector of ones, which can be referred to as
the unitary steering vector, and R(θ ) is the steered covariance
Downloaded 10/09/16 to 181.223.137.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

matrix (or the covariance computed witin a window follow- RESULTS

ing the defined traveltime trajectory). Here, θ is the vector
The concept of diffraction separation based on CRS is tested
containing all the parameters of the traveltime (cf. Eq. 2). Fol-
using both Marmousi and a multi-offset Gound Penetrating
lowing conventional notation, T and tr() represent transpose
Radar (GPR) dataset (Perroud and Tygel, 2005). The overall
and matrix trace, respectively. Equation (4) provides the in-
diffraction separation process is divided into two steps:
terpretation that Semblance can be regarded as a normalized
output/input energy ratio within the selected time window. (i) Sort the data in CMP gathers, where xm = x0w in Eq. (3),
and search for parameter Cw . Since Cw = 4/Vrms2 , where V
rms
MUSIC
is the root mean square velocity (in practice we assume that
The basic idea behind subspace methods like MUSIC is to de- the stacking velocities resemble the rms-velocities well) a ZO
compose the data covariance matrix into two orthogonal sub- section can be generated using these rms velocities.
spaces: (i) the signal subspace spanned by the eigenvectors
(ii) Next apply the CRS diffraction condition Bw = Cw and use
associated with large eigenvalues and (ii) the noise subspace
a wide aperture to avoid reflection contributions from within
spanned by the eigenvectors corresponding to small or zero
the Fresnel zone and search for parameter Aw in the ZO sec-
eigenvalues. Mathematically, this eigendecomposition of the
tion. Finally, perform CRS stacking based on Eq. (3) and gen-
covariance matrix can be written as
erate a diffraction-only ZO section.
R(θ ) = Vs (θ )Σs (θ )VH
s (θ ) + Vn (θ )Σn (θ )Vn (θ ) ,
H
(5)
Marmousi data result
where Vs (θ ) and Vn (θ ) are the signal and noise subspace
As a first step in the diffraction separation process we gen-
eigenvector matrices while Σs (θ ) and Σn (θ ) are the correspond-
erated a stacked section of the Marmousi dataset based on
ing eigenvalue matrices, respectively.
smoothed rms velocities obtained from the known interval ve-
MUSIC exploit the fact that the “correct” moveout, represented locites in depth. Only a selected part of the Marmousi model
as a steering vector, must lie in the signal subspace and there- was considered for demonstrating the diffraction separation as
fore is orthogonal to the noise subspace eigenvectors. As a indicated on the Marmousi interval velocity model shown in
consequence, the projection of the steering vector onto the Fig. 1(b). The corresponding ZO stack is shown in Fig 1(a).
noise subspace provides a nearly vanishing value. The inverse Using the rms velocites to define the values of parameter Cw
of such a projection (namely the sum of the dot products of and introducing the CRS diffraction condition Bw = Cw we
the steering vector with the noise eigenvectors) should peak performed parameter Aw search within an aperture of 100 traces.
when the steering vector represents a correct moveout. From Eventhough this size of the aperture is too wide for the shallow
this consideration the MUSIC measure of coherency (referred part of the data and will cause the CRS technique to loose its
to as MUSIC pseudo-spectrum) is given by power of describing a seismic event, our main interest is in the
relatively deeper part of the selected section where we need
uuH a wider aperture. Based on this choice, the coherency maps
PMU SIC (θ ) = , (6)
uPn (θ )uH computed by Semblance (cf. Fig. 2) and MUSIC (cf. Fig. 3)
show the possible locations of diffraction energy. The part of
where Pn (θ ) = Vn (θ )Vn (θ )H is the steered noise subspace the coherency maps in Figs. 2 and 3 that is above 0.4s shows
projection matrix. a very low coherence value due to the selection of the wide
The MUSIC algorithm is originally designed for narrowband aperture. The MUSIC coherency map represents the MUSIC
uncorrelated signals. However, seismic or GPR signals are pseudo-spectrum after being normalized by the energy of Sem-
both wideband and highly correlated. As a consequence, the blance. The Semblance balancing is applied to condition the
original MUSIC algorithm need to be modified accordingly pseudo-spectrum since MUSIC gives unconstrained coherency
before being applicable to such data. In order to handle cor- values. Finally utilizing the determined values of parameter Aw
related signals, spatial smoothing over the covariance matrix and also the condition Bw = Cw , we performed a CRS stack
can be employed (Biondi and Kostov, 1989; Kirlin, 1992). To based on Eq. (3). The results obtained from Semblance and
be able to implement such spatial smoothing in seismic or MUSIC are shown respectively in Figs. 4 and 5. Both co-
GPR, one has to taper the data within a window following the herency measures seem to work well, but with MUSIC giving
event. This tapering is to make the delay times of the event lin- an overall higher resolution (compare areas marked by black
ear (which is the basic requirement behind spatial smoothing) rectangles).
(Biondi and Kostov, 1989). The other advantage of perform- GPR dataset example
ing the analysis in a given window is to make the steering vec-
tors, required for generating the MUSIC pseudo-spectrum, to We also tested the diffraction separation process on a GPR

© 2011 SEG
SEG San Antonio 2011 Annual Meeting 3862
 A new approach to post-stack diffraction separation
0 0

5500
0.2 500
5000

4500
0.4 1000
4000

Depth [m]
Time [s]
3500
0.6 1500
3000

2500
0.8 2000
2000
Downloaded 10/09/16 to 181.223.137.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

1500
1 2500

3000
3000 4000 5000 6000 7000 8000 0 1000 2000 3000 4000 5000 6000 7000 8000 9000
Midpoint coordinate [m] Distance [m]

(a) (b)

Figure 1: (a) Zero-offset section of a selected part of the Marmousi data with AGC length of 50ms applied. (b) Interval velocity
model with chosen target area.

0.55

0.2 0.5 0.2

0.45

0.4 0.4 0.4

0.35

Time [s]
Time [s]

0.6 0.3 0.6

0.25

0.8 0.2 0.8

0.15

1 0.1 1

0.05
1.2
3000 4000 5000 6000 7000 8000 3000 4000 5000 6000 7000 8000
Midpoint coordinate [m] Midpoint coordinate [m]

Figure 2: Coherency map of parameter A based on Semblance. Figure 5: CRS diffraction stack based on MUSIC.

field data set. The data is composed of 28 different offsets

0.2 0.5
running every 0.2 m from 0.6 to 6 m. The CMP spacing is 0.1
0.4 0.4 m covering a 55-m-long profile. The stacking velocities re-
quired to generate the ZO section were obtained by means of
Time [s]

0.3
0.6
Semblance computed within a window of 11 samples. Figure 6
0.8 0.2
shows the ZO section obtained where most of the diffractions
are hidden behind the stronger reflection energy. Applying the
1 0.1

0
3000 4000 5000 6000 7000 8000
Midpoint coordinate [m]
0.02

0.04
Figure 3: Coherency map of parameter A based on MUSIC.
0.06
Time (µs)

0.08

0.2
0.1

0.4
0.12
Time [s]

0.6 0.14

0.8 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Midpoint (cm)

1.2 Figure 6: Zero offset section for the GPR data with an AGC of
length 392µ s applied.
3000 4000 5000 6000 7000 8000
Midpoint coordinate [m]

Figure 4: CRS diffraction stack based on Semblance. CRS diffraction condition Bw = Cw and using an aperture con-

© 2011 SEG
SEG San Antonio 2011 Annual Meeting 3863
 A new approach to post-stack diffraction separation

taining 40 traces, we then searched for parameter Aw . The

coherency map corresponding to parameter Aw obtained from 0.02

respectively Semblance and MUSIC are shown in Figures 7

0.04
and 8. Finally, we performed CRS stacking based on Eq. (3)
0.06
using the parameters determined by Semblance (cf. Fig. 9) and

Time (µs)
MUSIC (cf. Fig. 10). 0.08

0.1
Downloaded 10/09/16 to 181.223.137.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

0.12

0.14
0.02 0.6

500 1000 1500 2000 2500 3000 3500 4000 4500 5000
0.04 0.5 Midpoint (cm)

0.06
0.4
Time (µs)

Figure 9: GPR CRS diffraction stack based on Semblance.

0.08
0.3
0.1
0.2
0.12 0.02

0.1
0.14 0.04

0.06
1000 2000 3000 4000 5000

Time (µs)
Midpoint (cm)
0.08

0.1

Figure 7: GPR Coherency map of parameter A based on Sem- 0.12

blance.
0.14

500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Midpoint (cm)

0.02 0.6
Figure 10: GPR CRS diffraction stack based on MUSIC.
0.04
0.5

0.06 moveout equation requires the determination of optimal CRS

0.4
Time (µs)

0.08
moveout parameters based on the use of a coherency measure.
0.3 In this paper both Semblance and MUSIC were tested out, and
0.1 the latter was found to give better result in terms of resolution.
0.2
0.12
The application of the proposed technique to both Marmousi
0.1
and a multi-offset GPR dataset demonstrated its potential of
0.14
separating diffractions from reflections.
0
1000 2000 3000 4000 5000
Midpoint (cm)

ACKNOWLEDGMENTS
Figure 8: GPR Coherency map of parameter A based on MU-
SIC. The authors would like to thank Dr. Hervè Perroud for provid-
ing the GPR dataset.

In both cases diffractions are successfully separated from re-

flections and can then be used for further high-resolution imag-
ing or velocity model building. In general, the result obtained
using MUSIC is slightly better resolved and with less noise
than that obtained by Semblance. This is especially noticeable
within the two areas marked by black rectangles in Figs. 9 and
10.

CONCLUSIONS

A CRS based diffraction separation scheme has been intro-

duced and tested out. It uses the CRS traveltime moveout
valid for diffractions derived as a special case of the general-
ized hyperbolic CRS equation. Practical use of this diffraction

© 2011 SEG
SEG San Antonio 2011 Annual Meeting 3864
 EDITED REFERENCES
Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2011
SEG Technical Program Expanded Abstracts have been copy edited so that references provided with the online metadata for
each paper will achieve a high degree of linking to cited sources that appear on the Web.
Downloaded 10/09/16 to 181.223.137.170. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

REFERENCES
Biondi, B., and C. Kostov, 1989, High-resolution velocity spectra using Eigenstructure methods:
Geophysics, 54, 832–842, doi:10.1190/1.1442712.
Fomel, S., 2002, Applications of plane-wave destruction filters: Geophysics, 67, 1946–1960,
doi:10.1190/1.1527095.
Fomel, S., E. Landa, and M. Taner, 2007, Poststack velocity analysis by separation and imaging of
seismic diffractions: Geophysics, 72, no. 6, U89–U94, doi:10.1190/1.2781533.
Kanasewich, E., and S. Phadke, 1988, Imaging discontinuities on seismic sections: Geophysics, 53, 334–
345, doi:10.1190/1.1442467.
Kirlin, R. L., 1992, The relationship between semblance and eigenstructure velocity estimators:
Geophysics, 57, 1027–1033, doi:10.1190/1.1443314.
Landa, E., V. Shtivelman, and B. Gelchinsky, 1987, A method for detection of diffracted waves on
common-offset sections: Geophysical Prospecting, 35, no. 4, 359–373, doi:10.1111/j.1365-
2478.1987.tb00823.x.
Moser, T., and C. Howard, 2008, Diffraction imaging in depth: Geophysical Prospecting, 56, no. 5, 627–
641, doi:10.1111/j.1365-2478.2007.00718.x.
Neidell, N., 1997, Perceptions in seismic imaging, Part 2: Reflective and diffractive contributions to
seismic imaging: The Leading Edge, 16, 1121–1123, doi:10.1190/1.1437744.
Perroud, H., and M. Tygel, 2005, Velocity estimation by the common-reflection-surface (CRS) method:
Using ground-penetrating radar data: Geophysics, 70, no. 6, B43–B52, doi:10.1190/1.2106047.
Zhang, Y., S. Bergler, and P. Hubral, 2001, Common-reflection-surface (CRS) stack for common offset:
Geophysical Prospecting, 49, no. 6, 709–718, doi:10.1046/j.1365-2478.2001.00292.x.

© 2011 SEG
SEG San Antonio 2011 Annual Meeting 3865