See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/318817855

5D interpolation with wavefront attributes

Article in Geophysical Journal International · August 2017

DOI: 10.1093/gji/ggx334

CITATIONS READS

0 200

2 authors:

Yujiang Xie Dirk Gajewski

University of Hamburg University of Hamburg
15 PUBLICATIONS 9 CITATIONS 217 PUBLICATIONS 2,305 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

5D interpolation and regularization View project

Unifying aspects of stacking / beamforming View project

All content following this page was uploaded by Yujiang Xie on 15 September 2017.

The user has requested enhancement of the downloaded file.

 Geophysical Journal International
Geophys. J. Int. (2017) 211, 919–941 doi: 10.1093/gji/ggx334
Advance Access publication 2017 August 2
GJI Seismology

5-D interpolation with wave-front attributes

Yujiang Xie and Dirk Gajewski

Institute of Geophysics, University of Hamburg, D-20146 Hamburg, Germany. E-mail: yujiang.xie@studium.uni-hamburg.de

Accepted 2017 August 1. Received 2017 June 7; in original form 2017 January 23

SUMMARY
Most 5-D interpolation and regularization techniques reconstruct the missing data in the
frequency domain by using mathematical transforms. An alternative type of interpolation
methods uses wave-front attributes, that is, quantities with a specific physical meaning like
the angle of emergence and wave-front curvatures. In these attributes structural information of
subsurface features like dip and strike of a reflector are included. These wave-front attributes
work on 5-D data space (e.g. common-midpoint coordinates in x and y, offset, azimuth and
time), leading to a 5-D interpolation technique. Since the process is based on stacking next
to the interpolation a pre-stack data enhancement is achieved, improving the signal-to-noise
ratio (S/N) of interpolated and recorded traces. The wave-front attributes are determined in
a data-driven fashion, for example, with the Common Reflection Surface (CRS method). As
one of the wave-front-attribute-based interpolation techniques, the 3-D partial CRS method
was proposed to enhance the quality of 3-D pre-stack data with low S/N. In the past work
on 3-D partial stacks, two potential problems were still unsolved. For high-quality wave-front
attributes, we suggest a global optimization strategy instead of the so far used pragmatic search
approach. In previous works, the interpolation of 3-D data was performed along a specific
azimuth which is acceptable for narrow azimuth acquisition but does not exploit the potential
of wide-, rich- or full-azimuth acquisitions. The conventional 3-D partial CRS method is
improved in this work and we call it as a wave-front-attribute-based 5-D interpolation (5-D
WABI) as the two problems mentioned above are addressed. Data examples demonstrate the
improved performance by the 5-D WABI method when compared with the conventional 3-D
partial CRS approach. A comparison of the rank-reduction-based 5-D seismic interpolation
technique with the proposed 5-D WABI method is given. The comparison reveals that there are
significant advantages for steep dipping events using the 5-D WABI method when compared
to the rank-reduction-based 5-D interpolation technique. Diffraction tails substantially benefit
from this improved performance of the partial CRS stacking approach while the CPU time is
comparable to the CPU time consumed by the rank-reduction-based method.
Key words: Body waves; Theoretical seismology; Wave propagation; Wave scattering and
diffraction.

1 I N T RO D U C T I O N
3-D pre-stack seismic data are recorded in the 5-D data space:
four spatial dimensions and one temporal dimension [e.g. common-
midpoint (CMP) coordinates in x and y, offset, azimuth and time].
As some natural and anthropogenic factors, for example, field obsta-
cles, dead traces, and budgetary constraints, 3-D pre-stack seismic
data may be irregularly and sparsely sampled during data acquisi-
tion, which would affect the image quality of further applications.
In order to resolve this problem, a simple and straightforward strat-
egy is to introduce interpolated traces into these data gaps. In the
literature, there are about five categories of interpolation methods
reported. The first category is based on sparse transforms, such as
Radon transform (e.g. Kabir & Verschuur 1995; Trad et al. 2009;
Zhang & Lu 2014), Fourier transform (e.g. Liu & Sacchi 2004;
Zwartjes & Sacchi 2007; Trad 2009; Curry 2010; Naghizadeh
& Innanen 2011), and Curvelet transform (e.g. Naghizadeh &
Sacchi 2010). With the sparse transforms, one can gradually at-
tenuate the artefacts and recover the missing data information in
the sparse domain by iteratively thresholding the transformed do-
main of the incomplete seismic data (Chen et al. 2016b). The sec-
ond category of seismic interpolation methods is the prediction-
filtering based interpolation methods (e.g. Spitz 1991; Naghizadeh
& Sacchi 2009; Liu & Fomel 2011), which interpolate high-
frequency aliased data using prediction-error filters derived from
low-frequency non-aliased data. This method works well with


C The Authors 2017. Published by Oxford University Press on behalf of The Royal Astronomical Society. 919
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 920 Y. Xie and D. Gajewski

Figure 1. The initial search apertures used in this work. (a) A given 3-D CRS gather consisting of several 3-D CMP gathers, where the black dots denote
locations of the 3-D CMP gathers. (b) A considered 3-D CMP gather and its corresponding 3-D partial CRS offset aperture (blue) in this 3-D CMP gather.
Note: the blue ellipse here is only a subset of whole 3-D partial CRS aperture in the offset direction. (c) The considered time slice at t0 . (d) A grid-based
full-azimuth regularization performed at t0 , where the black grid points are the trace location regularized along different azimuth directions. Here only the
positive offsets are shown, and the constant angle between the dashed lines is the azimuthal interval. More details see Fig. 3.

Figure 2. Calculating the t0, p and construct the 3-D partial CRS stacking operator for a sample A on trace k located at (m0, x , m0, y , offx, offy). (a) Calculating the
t0, p by a traveltime fitting process. For the sake of simplicity, only one direction (OF) is shown. The black curve is an observed seismic event at this direction.
In order to find t0, p , the traveltime operator (Eq. 2) is used. We start from the first ZO sample (see the top blue dashed curve), then the traveltime surface
(eq. 2) is moved downward until the condition satisfies: (t − t22 )2 ≤ fmin2 , where we assume t0, p = t11 . For better understanding, the traveltime difference
between t0, p and t11 shown in this figure is exaggerated. In practice, the difference between t0, p and t11 is very small. Usually, we can refine t0, p using the
simulated sample A if the wave-front attributes around t0, p are smooth. The coefficient fmin can be automatically calculated by the wave-front attributes of two
neighbouring ZO samples around t0, p . The right green line is a boundary of the offset aperture in the OF direction. (b) An intersection (red) of the 3-D partial
CRS traveltime surface with the vertical OF plane, where its top horizontal short line (blue) is the partial CRS aperture for sample A. Other parts of the 3-D
partial CRS traveltime surface at current CMP and its neighbouring CMPs are not shown here. The whole 3-D partial CRS traveltime surface is expressed as
eq. (3).

regularly sampled data. The third category of methods includes, for
example, the Cadzow rank-reduction method (Trickett 2008; Trick-
ett et al. 2010; Chen et al. 2016a,b), or called the multichannel singu-
lar spectrum analysis (MSSA) method (e.g. Oropeza & Sacchi 2011;
Huang et al. 2015). The rank-reduction-based interpolation meth-
ods assume that missing traces and random noise increase the rank
of the constructed Hankel/Toeplitz matrices, and one can intuitively
reduce the negative effects caused by the missing traces and random
noise by applying rank-reduction operators (Chen et al. 2016a,b).
They intuitively satisfy the local plane-wave assumption. The fourth
seismic interpolation methods is the wave-equation-based methods
(e.g. Ronen 1987; Stolt 2002; Fomel 2003; Kaplan et al. 2010).
They generally assume that the subsurface velocity is known and
are computationally expensive solving the wave equation. The last
kind of interpolation methods is the wave-front-attribute-based in-
terpolation (WABI; e.g. Höcht et al. 2009; Baykulov & Gajew-
ski 2009, 2010; Xie & Gajewski 2016b; Xie 2017), which is a
data-driven, velocity-independent interpolation technique extended
from, for example, the Common-Reflection-Surface (CRS) method
(e.g. Jäger et al. 2001; Mann 2002; Höcht 2002; Müller 2003). This
kind of interpolation methods utilizes wave-front attributes derived
from moveout of events for several neighbouring CMP gathers.

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 5-D interpolation with wave-front attributes 921

Figure 3. Different azimuth-based regularization strategies. (a) An azimuth-fixed regularization (e.g. azimuth = 0). (b) Narrow azimuth regularization. (c)
Regularization of one trace k (offx, offy) on the azimuth direction of iaz∗δλ, where δλ is the azimuth interval in different directions, the sign ∗ indicates a
multiplication, and iaz is an index of different azimuth directions. (d) A wide- or full-azimuth regularization. For each azimuthal direction iaz∗δλ in the 3-D
CMP gather, we have hx = hy /tan(0 + iaz∗δλ) and hy = j∗δhy , where j = 0, 1, 2, ..., max(hy )/δhy , and δhy is the offset increment in the hy direction. The
azimuth interval δλ and offset increment δhy should be set by the user in different data. In the 3-D SEG case, we use δλ = 15∗π /180 and δhy = 30 m, which
need to be tested from the data, for example, several δλ and δhy are tried until the one with the best result is gotten. For the 3-D SEG data, the azimuth range
would be better set between [π /4 π /2], for example, azimuth λ = π /4, π /3, 5π /12, π /2. This is because the 3-D SEG data is not a real full-azimuth data.
We do the same calculation in the 3-D simple data but there the offset interval is set in the x direction, that is, δhx = 30 m.

The missing traces are predicted by the so-called partial stack. Gajewski 2010) was successfully applied to denoise and reconstruct
Examples are the partial CRS stacking methods (Baykulov & 3-D pre-stack seismic data with low S/N. However, two potential
Gajewski 2009, 2010), which can significantly increase the signal- problems need to be addressed in this method and are considered in
to-noise ratio (S/N) of every trace and fill the data gaps after this paper:
the interpolation. In this fashion a regularized 3-D data volume
(i) To obtain high-quality 3-D wave-front attributes their deter-
with improved S/N is generated. Since the wave-front attributes
mination should be performed with global optimization. The above-
are derived from moveout which contains information on reflector
mentioned publication adopted a pragmatic search strategy in sub-
dip, strike, and curvature, we consider this kind of interpolation
volumes of the data which may lead to a poor or insufficient fit of
techniques as physics-based interpolation methods to distinguish
the adapted traveltime surface to the full data volume. As any other
them from pure mathematics-based interpolation approaches. The
processing step using wavefield attributes also the 3-D partial CRS
WABI methods are performed within the first Fresnel zone and
benefits of high-quality attributes.
therefore they use traces in the interpolation process, which re-
(ii) The second problem considered in this work is the azimuth-
solve the same structural details. Since the wave-front attributes
based regularization for wide-, rich- or full-azimuth acquisition. In
are determined by kinematic features of the wave field, that is,
previous works the regularization was performed along a specific
moveout, aliasing issues are relaxed. Next to the CRS operator any
azimuth which does not exploit the potential of different azimuth
other operator utilizing wave-front attributes like i-CRS (Schwarz
acquisition.
et al. 2014), non-hyperbolic CRS (Fomel & Kazinnik 2013) or
multifocusing (Gelchinsky et al. 1999) can be used. These oper- An effective solution to resolve the first problem in the con-
ators determine wave-front attributes from pre-stack data equally ventional 3-D partial CRS method is to develop a global search
well (Walda et al. 2017). The 3-D partial CRS method (Baykulov & strategy which determines the wave-front attributes simultaneously

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 922 Y. Xie and D. Gajewski

(a) (b)
0 0.55 0
0.895
0.2 0.2
0.548

Semblance

Semblance
0.8945
0.4 0.4
CR

CR
0.546 0.894
0.6 0.6
0.8935
0.8 0.544 0.8
0.893
1 0.542 1 0.8925
0 1 2 0 1 2
F F
(c) (d)
0 0.324 0
0.0764
0.2 0.322 0.2
0.0762

Semblance

Semblance
0.4 0.32 0.4 0.076
CR

CR
0.6 0.318 0.6 0.0758

0.8 0.8 0.0756

0.316
0.0754
1 0.314 1
0 1 2 0 1 2
F F
Figure 4. Parameter sensitivity analysis for the F and CR in the DE algorithm. (a), (b), (c) and (d) are corresponding to the four ZO samples mentioned
in Fig. 5. In these tests, the parameter F is set between [0 2] with 0.2 increasing interval, and CR is set between [0 1] with 0.1 increasing interval. After a
brute-force search, we use F = 1.2 and CR = 0.9.

(a) (b)
0.55 0.9

0.5
Semblance

Semblance

0.85
0.45
0.8
0.4

0.35 0.75
10 20 30 40 50 10 20 30 40 50
Number of iteration Number of iteration
(c) (d)
0.35 0.08

0.07
Semblance

Semblance

0.3
0.06
0.25
0.05

0.2 0.04
10 20 30 40 50 10 20 30 40 50
Number of iteration Number of iteration
Figure 5. Iteration tests of four zero-offset (ZO) samples picked from the 3D SEG data. (a) The first ZO sample is chosen from CMP 181934, sample 232.
(b) The second ZO sample is taken from CMP 182203, sample 108. (c) The third ZO sample is CMP 182125, sample 205. (d) The fourth ZO sample is CMP
182115, sample 550.

from the entire 5-D data space by assuming that the wave kinemat-
ics are properly described by the 3-D CRS traveltime formula. We
recently introduced two well-known global optimization algorithms
to simultaneously determine the 3-D CRS wave-front attributes (see
Xie & Gajewski 2016a,c). These are the particle swarm optimization
(PSO; Kennedy & Eberhart 1995; Shi & Eberhart 1998) and genetic
algorithm (GA; Holland 1975). Both algorithms provide improved
wave-front attributes when compared to the conventional pragmatic
approach or the general Powell conjugate direction (PCD) algorithm
(Powell 1964). We found that the differential evolution (DE; Storn
& Price 1997) is the most stable algorithm to globally determine
the 3-D CRS wave-front attributes when compared to the PSO, GA,

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017

140
Computational costs [%]
120
100
80
60
40
20
0
DE Pragmatic approach
Figure 6. Computational costs of the pragmatic approach and the DE algo-
rithm, where DE is referenced as 100 per cent.
and the conventional pragmatic approach. In this work, we use the
DE algorithm as the global search method to replace the conven-
tional pragmatic approach. We call the improved 3-D partial CRS
interpolation method as a 5-D WABI technique, where the 3-D CRS
wave-front attributes are obtained by global optimization using the
DE algorithm and the azimuth-based regularization is applied.
After this introduction we briefly describe the 3-D CRS theory
and the search strategy for wave-front attributes then we intro-
duce our concept for regularization and interpolation of data with
wide, rich or full azimuth. Data examples and comparisons with
the conventional 3-D partial CRS approach using the 3-D SEG salt
model complement the paper. Finally we present a comparison of
the proposed 5-D WABI method with a 5-D interpolation approach,
for example, based on the damped rank-reduction method (Chen
et al. 2016b), using a simple 3-D synthetic data set of a diffraction
event and the complex 3-D SEG data set.
2 T H E O RY A N D M E T H O D S
2.1 3-D CRS theory and wave-front attribute search
The 3-D CRS traveltime formula can be derived from the
well-known two-point paraxial travetime approximations (e.g.
Červený 2001; Červený & Moser 2007; Moser & Červený 2007;
Bortfeld 1989; Hubral et al. 1992; Schleicher et al. 1993b). In 3-D
seismic systems, the CRS traveltime parameters are generally ex-
plained with physical meanings, for example, kinematic wavefield
attributes or wave-front attributes. The classical 3-D zero-offset
(ZO) CRS traveltime formula is based on the hyperbolic approx-
imation (e.g. Schleicher et al. 1993b; Jäger 1999; Müller 2003;
Bergler 2004) or geometrical considerations (Höcht 2002). It reads
2
thyp = (t0 + wz · md )2 + mdT Nmd + hT Mh, (1)
where the vector wz includes the near-surface velocity (v0 ), the dip
(α) and azimuth (λ) of the reference ray. The midpoint parameter
satisfies: md = m − m0 , where the vector m0 is the considered
3-D CMP position, and the m indicates any neighbouring CMP
location of the considered 3-D CMP. Half offset of a trace in the
considered 3-D CRS gather is given by h. The 2 × 2 symmetrical
matrices M and N are related to the curvatures of the normal-
incidence-point (NIP) wave front and the normal (N) wave front,
respectively (Hubral 1983; Jäger 1999; Müller 2003; Bergler 2004).
The superscript T denotes the transpose of a matrix.
In general, the 3-D ZO CRS wave-front attributes are searched
by using the semblance (Neidell & Taner 1971) as an objective
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
5-D interpolation with wave-front attributes 923
function. In practice, a large amount of wave-front attributes are
tested by the search algorithms, and then the set of wave-front
attributes that provides the highest semblance is kept as the best
fitting one. In order to calculate the maximum semblance for each
ZO sample, two types of methods are often applied in the CRS
domain. The first one is a brute-force search or called exhaustive
linear search that is used to find a target value within a list. It
sequentially tests each element of the list for the objective function
until all elements are searched. The brute-force search could be
stable and effective in a low-dimensional search, such as the 2-
D velocity analysis that only requires a velocity list to find the
best stacking velocity for each ZO sample. However, in a high-
dimensional search, for example, in the 3-D ZO CRS case, eight
wave-front attributes are required if the data are gotten from a
wide- or full azimuth acquisition. A simultaneous brute-force search
for the eight wave-front attributes is extremely expensive, where
we may have to test 1008 sets of wave-front attributes for each
ZO sample if 100 search intervals of each of the eight wave-front
attributes are considered. It will take about 2 months for one ZO
sample if a CPU with 2.5 GHz and 1 GB RAM is used. Usually,
there are about 105 to 108 ZO samples need to be processed in a
general 3-D seismic data set in the 3-D ZO CRS case, and such large
number of ZO samples are practically impossible to implement in
most computing platforms.
To reduce the high computational costs in the 3-D ZO CRS case
(see Müller 2003), the historical CRS parameter search method,
that is, the well-known, three-step-based pragmatic approach (Mann
et al. 1999) was introduced into the 3-D ZO CRS. It suggests to de-
termine the eight wave-front attributes by three steps. In each step,
only two or three of the eight wave-front attributes are determined si-
multaneously by the brute-force search, which significantly reduces
the computational costs when compared to the simultaneous one-
step brute-force search, that is, (1003 + 1002 + 1003 )  1008 . The
1008 on the right-hand side means the number of all possible tested
wave-front attributes used in the simultaneous one-step brute-force
search. On the left-hand side, the first and the last 1003 represent the
number of NIP and N wave-front attributes to be searched for the
three elements of matrices M and N, respectively. The 1002 indi-
cates the number of dips and azimuths for the two elements of vector
wz . Alternative solution to reduce the high computational costs in
a simultaneous search is to introduce a suitable global optimiza-
tion algorithm, which is iteratively trying to find or move toward to
the best solution of the objective function (e.g. Weise 2009), such
as the above DE algorithm that improves the candidate solution to
get close to the best solution of the semblance function by a mass
of iterations. It could be cheaper than the conventional pragmatic
approach when we apply 100 intervals for each of the eight wave-
front attributes, that is, 2000  (1003 + 1002 + 1003 ), where the
2000 here means that 40 agents and 50 iterations are applied in the
DE algorithm. More importantly, the maximum semblance gotten
from the DE algorithm is higher than the one obtained from the
conventional pragmatic approach (see a detailed discussion below).
2.2 5-D WABI method
In this work, the 5-D WABI method is given as follows:
Firstly, we compute the midpoint locations m = (mx , my ) for each
3-D CMPn gather of the 3-D data. In each 3-D CMP gather, we have
m = ( i=1 mi )/n, where mi is the midpoint location of a shot s and
a receiver g, i indicates the shot-receiver pair, and n is the number
of pairs in this 3-D CMP gather. For each receiver trace in this 3-D
 924 Y. Xie and D. Gajewski

Figure 7. Semblance sections of inline 190 and crossline 300 of the 3-D semblance volume. (a) Inline 190 with pragmatic approach. (b) Inline 190 with
DE-based global optimization. (c) Semblance difference between (b) and (a), where about 88.7 per cent ZO samples have a positive semblance. (d) Crossline
300 with pragmatic approach. (e) Crossline 300 with DE-based global optimization. (f) Semblance difference between (e) and (d), where about 88.1 per cent
ZO samples have a positive semblance. The semblance difference is computed by using the semblance of DE minus the semblance of pragmatic approach.

CMP gather, we have mi = (g + s)/2 and hi = (g − s)/2, where
s = (sx , sy ), g = (gx , gy ), mi = (mi, x , mi, y ), and hi = (hi, x , hi, y ).
Secondly, the location of the given 3-D CRS gather is deter-
mined by using a 3-D CMP gather located at m0 = (m0, x , m0, y ).
We often apply two steps to read traces into a given 3-D CRS
gather:
(i) reading traces into the given 3-D CRS gather if all traces in
the 3-D data satisfy: (mx − m0, x )2 /amx 2 + (my − m0, y )2 /amy 2 ≤ 1,
where the denominator amx and amy are the midpoint aperture (see
Fig. 1a), and mx and my indicate any 3-D CMP position located
within the red ellipse centred at the given 3-D CMP location, that
is, the m0 = (m0, x , m0, y ). The number of 3-D CMP gathers used

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 5-D interpolation with wave-front attributes 925

Figure 8. Time slice 1.2 s of the 3-D semblance volume. (a) Semblance difference between (d) and (c), where about 89.7 per cent ZO samples have a positive
semblance. (b) Interpolation operators with low and high semblance. (c) Time slice of 1.2 s with pragmatic approach. (d) Time slice of 1.2 s with DE-based
global optimization. The semblance difference is computed by using the semblance of DE minus the semblance of pragmatic approach.

inator hx, i , hy, i are the offset aperture centred at t0 (see Fig. 1b).
The offset aperture at each ZO traveltime can be computed from
the pre-determined offset apertures: hx, 0 , hy, 0 , hx, n , and hy, n (see
Appendix). The elliptical aperture used in this work is not the only
way to select traces into a given 3-D CRS gather, but it is a good so-
lution preferable to a rectangular surface. This conclusion is based
on our tests performed for the open 3-D SEG C3WA data.

In the next step we determine the traces which contribute to

Figure 9. Semblance of the four ZO samples mentioned in Fig. 5. The
global optimization here is based on the DE algorithm.
a given 3-D partial CRS gather. The given 3-D CRS gather is
already loaded into the computer memory. If traces within the
given 3-D CRS gather satisfy the relation (hi, x − offx)2 /px, i 2 +
(hi, y − offy)2 /py, i 2 ≤ 1, they are selected into the given 3-D par-
tial CRS gather centred at (offx, offy), where px, i and py, i are the
semi-major and semi-minor axes of the horizontal ellipse (blue) (see
Fig. 1b). Both the px, i and py, i are estimated by the pre-determined
apertures: a, b, c, and d (see Appendix A).
After this step we compute the ZO traveltime t0, p for the 3-D
partial CRS operator. For a sample A of trace k with h = (offx, offy),
the t0, p can be fitted with
2
t22 (o f f x, o f f y) = t11
2
+ hT Mh. (2)

We assume the ZO traveltime t0, p = t11 if the calculated t22 satisfy:

in the given 3-D CRS gather depends on the midpoint aperture that
can be estimated by the Fresnel zone (Hubral et al. 1993a) by means
of the 3-D CRS wave-front attributes.
(ii) Selecting traces into the offset dimension if all traces obtained
in step (i) satisfy: hi, x 2 /hx, i 2 + hi, y 2 /hy, i 2 ≤ 1, where the denom-
(t − t22 )2 ≤ fmin2 , where t11 is the trial ZO traveltime, and t22 is the
calculated hyperbolic traveltime (see Fig. 2a). A refinement of the
t0, p by using Equation 2 at the sample A would be necessary if we
get the ZO wave-front attributes obtained from t0, p . The eight wave-
front attributes at t0, p are read from the attribute files (a byproduct

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 926 Y. Xie and D. Gajewski

Trace number Trace number

Trace number Trace number

Figure 10. A 3-D CMP gather (CMP 181903) picked from the 3-D SEG data. (a) Raw 3-D CMP gather with low S/N. (b) Raw 3-D CMP gather with gaps.
(c) Interpolation with conventional 3-D partial CRS stacking. (d) Interpolation with 5-D WABI method. For the sake of simplicity, only the azimuth λ = π /2
is shown in this 3-D CMP gather after the 5-D interpolation and regularization.

of the DE algorithm) for each t0, p . Now we have all attributes to
apply the 3-D partial CRS operator using
p = (t0, p + wz · md ) + md Nmd + h Mh.
2 2 T T
thyp,
Eq. (3) is a 5-D interpolation operator of the acquisition system
as well as the wave-front attributes in the time domain. With this
operator, the sample A is interpolated at t on trace k located at (m0, x ,
m0, y , offx, offy). The same process is performed for each sample
of trace k, and for all other traces in the given 3-D CMP gather, as
well as for all other 3-D CMP gathers in the whole 3-D data. For
each 3-D CMP gather of the 3-D data, the trace locations are regu-
larized by an azimuth-based regularization in the offset dimension
(see Figs 1 d and 3), where different azimuthal directions are con-
(3)
sidered. In practice, the 3-D partial CRS aperture and the azimuthal
intervals need to be tested from the 3-D data. We often perform a
series of tests to find the optimized 3-D partial CRS apertures and
azimuthal intervals. Amplitude variation in the midpoint dimension
can be estimated from the modulus of the geometrical spreading
factor (e.g. Hubral at al. 1993b; Schleicher et al. 1993a). Shuey’s
approximation (Shuey 1985) can also be applied for an analysis of

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 5-D interpolation with wave-front attributes 927

Trace number Trace number

Trace number Trace number

Figure 11. A 3-D CMP gather (CMP 182113) picked from the 3-D SEG data. (a) Raw 3-D CMP gather with low S/N. (b) Raw 3-D CMP gather with gaps.
(c) Interpolation with conventional 3-D partial CRS stacking. (d) Interpolation with 5-D WABI method. For the sake of simplicity, only the azimuth λ = π /2
is shown in this 3-D CMP gather after the 5-D interpolation and regularization.

the amplitude versus offset (AVO) in the offset dimension in each 3 R E S U LT S A N D D I S C U S S I O N S

3-D CMP gather. The amplitude of the sample A is a summation of
all amplitudes along the 3-D partial CRS traveltime surface within 3.1 3-D SEG C3WA data and CRS midpoint
the 3-D partial CRS aperture, that is, an improvement of the S/N and offset apertures
ratio is obtained. Considering the AVO may be omitted since we
The 3-D SEG C3WA data is an open pre-stack seismic data built
stack/interpolate locally along the offset direction. However, AVO
by the SEG research committee. Some important parameters of the
analysis is necessary in the case of data with large gaps since in
data need to be pointed out. Each trace of the data has 625 samples
such case a large 3-D CRS aperture is applied to each trace of the
with a sampling interval of 8 ms. The maximum CDP fold is 17,
considered 3-D CMP gather.
and an effective offset is ranged from 40 to 2695 m. The midpoint

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 928 Y. Xie and D. Gajewski

Figure 12. Inline 190 of the first 3-D CO volume. (a) Raw CO section. (b) Raw CO section with gaps. (c) Interpolation with conventional 3-D partial CRS
stacking. (d) Interpolation with the 5-D WABI method.

distance between inlines is 40 m, twice the density of the crosslines, 3.2 Usage of the pragmatic approach and the DE
leading to double ZO traces in the inline section. An example of pre- algorithm
processing the SEG/EAGE data is given by (Xu et al. 2004). To see
Before we present a comparison of the conventional 3-D partial
the performance of the conventional 3-D partial CRS stacking and
CRS stacking and the 5-D WABI method, the 3-D CRS wave-
the 5-D WABI method in this work, some traces in the 3-D SEG data
front attributes require to be determined from the gapped SEG data
were randomly killed and some random Gaussian noise with an S/N
by means of the pragmatic approach and the DE algorithm. Re-
= 10 was added to the seismograms, that is, the RMS magnitude
lated to the pragmatic approach, one can reference Müller’s work
of the signal amplitude is 10 times the RMS magnitude of the
(Müller 2003). Here we briefly introduce how to use the DE algo-
random noise. Fig. 1(a) shows the midpoint aperture (red ellipse),
rithm in the 3-D CRS case. Four parameters need to be determined
being initially set to 200 m in both x- and y-direction, that is, amx
in the DE algorithm. The first two are the differential weight F and
= amy = 200 m. After the initial search, the midpoint aperture is
the crossover probability CR, a priori suggested values are given
estimated by the Fresnel zone (Hubral at al. 1993a), which in turn
by Pedersen (2010). However, we found that the F and CR may be
can be computed from the determined wave-front attributes. The
slightly different in different data sets. Fig. 4 shows a test of the F
offset aperture is set manually, for example, the observed seismic
and CR from the 3-D SEG data, which illustrates that the maximum
events in the offset dimension are totally included within the offset
semblance is stable as long as the F is large than 1. CR is not sensi-
aperture at the beginning, then some offset apertures are tried until
tive to the semblance function in Fig. 4 when F is larger than 1. In
the one with the possible best results is found. The criterion of the
this work, we use F = 1.2 and CR = 0.9. The last two parameters
best results is driven by the determined wave-front attributes as well
are the population size NP and the number of iterations IT, which
as the semblance. A detailed setting of the CRS offset apertures are
also need to be tested from the data. In our tests to the 3-D SEG
given in Appendix.
data, we found that a safe suggestion of the NP should be larger

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017

Figure 13. Crossline 300 of the first 3-D CO volume. (a) Raw CO section. (b) Raw CO section with gaps. (c) Interpolation with conventional 3-D partial CRS
stacking. (d) Interpolation with the 5-D WABI method.
than 40, and the IT would be better larger than 50 (see a test out
of ten shown in Fig. 5). In order to determine the F and CR in the
above, a very large IT and NP are used at the beginning (making
sure the results only sensitive to the F and CR), then the IT and NP
are tested with the determined F and CR.
3.3 Semblance difference between the pragmatic approach
and the DE algorithm
In our tests, we found that the DE algorithm is computationally
more efficient than the traditional pragmatic approach to determine
the 3-D CRS wave-front attributes (see Fig. 6). To estimate the
semblance difference obtained from the pragmatic approach and
the DE algorithm, three sections of the 3-D semblance volume
are shown. These are inline 190, crossline 300, and a time slice
at 1.2 s (Figs 7 and 8). The semblance difference of the three
sections indicates that, for close to 90 per cent of the ZO samples
the semblance provided by the DE algorithm is higher than the one
for the pragmatic approach. A detailed semblance comparison for
four ZO samples of the 3-D semblance volume is given in Fig. 9.
The semblance provided by the DE algorithm turns out to be higher.
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
5-D interpolation with wave-front attributes 929
3.4 Conventional 3-D partial CRS stacking and the 5-D
WABI method
3.4.1 CMP gathers after interpolation
The 5-D WABI technique presented in this work includes two as-
pects: (i) we interpolate a sample A with traveltime t on trace k
located at (m0, x , m0, y , offx, offy); (ii) the trace location relies on the
azimuth-based regularization in each 3-D CMP gather (see Fig. 3).
To show the conventional 3-D partial CRS stacking and the 5-D
WABI method performed on individual 3-D CMP gather, two 3-D
CMPs were randomly taken from the 3-D SEG C3WA data. The first
one is CMP 181903 (see Fig. 10, displayed in a 2-D section, that is,
all traces are sorted with increasing offsets). In this gather, seismic
events on the raw data are visible but unclear. Fig. 10(b) shows the
same data, where some traces were deleted to simulate data gaps.
After determination of the 3-D CRS wave-front attributes obtained
from the reduced 3-D data, the conventional 3-D partial CRS stack-
ing and the 5-D WABI method are applied separately. Fig. 10(c)
displays the 3-D CMP gather processed by the conventional 3-D
partial CRS stacking, where the reflection events are clearly visible
and continuous with less noise, compared to the input 3-D CMP
 930 Y. Xie and D. Gajewski
Figure 14. Time slice 1.2 s of the first 3-D CO volume. (a) Raw data. (b) Raw data with gaps. (c) Interpolation with conventional 3-D partial CRS stacking.
(d) Interpolation with the 5-D WABI method.
gather (Fig. 10b). Fig. 10(d) shows the same 3-D CMP gather but
processed by the 5-D WABI method, which also displays improved
results compared to the input 3-D CMP gather. In this 3-D CMP
gather with CMP 181903, we do not see a big difference between
the conventional 3-D partial CRS stacking and the 5-D WABI. The
second example shows data for CMP 182113 of the 3-D SEG data,
which is noisier and the seismic events are difficult to identify. Af-
ter processing with the conventional 3-D partial CRS stacking and
the 5-D WABI method the gaps are filled and the reflection events
are clearly visible (see Figs 11c and d). However, the result of 5-
D WABI displays an improved result. The semblance is used to
confirm this visual observation.
3.4.2 Common-offset (CO) sections after interpolation
In order to see the 5-D interpolation results on more 3-D CMP
gathers simultaneously, two 3-D CO volumes are shown, of which
each 3-D CMP gather contains one trace only, and all traces in each
3-D CO volume have the same half offset. The first 3-D CO volume
is chosen from a constant half offset of h = (0 m, 100 m). The
second 3-D CO volume is taken from a far constant half offset with
h = (0 m, 1000 m). In each of the two 3-D CO volumes, three
sections comprising the salt body are considered. These are inline
190, crossline 300, and a time slice at 1.2 s. In the Common-offset
(CO) sections after interpolation, all figures have the same colour
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
bar with a minimum (blue) and a maximum (red) amplitude ranged
from −0.6 to 0.6.
Fig. 12 shows inline 190 of the first 3-D CO volume, where the
image quality of the raw data is poor, and many deep reflection
events and diffraction patterns are almost invisible (see Fig. 12a).
Fig. 12(b) displays the section with removed traces. We can see a
better continuity of horizons at different time levels and the S/N
is increased significantly after the conventional 3-D partial CRS
stacking and the 5-D WABI method were applied. Some differences
between the conventional 3-D partial CRS stacking and the 5-D
WABI method are marked by red arrows (see Figs 12c and d). Fig. 13
shows crossline 300 of the first 3-D CO volume, where most of the
seismic events of the raw data are difficult to identify except the top
horizontal layers (Fig. 13a). However, we observe many reflection
events and diffraction patterns after the conventional 3-D partial
CRS stacking is applied. It can provide continuous seismic events
at all time levels with a better S/N (see Fig. 13c). Even improved
results are obtained when the 5-D WABI method is applied (see
Fig. 13d). It displays more continuous events than the conventional
3-D partial CRS. We note that the resolution of the crossline section
is worse than that of the inline section. This is due to a sparse data
acquisition along the crossline direction. Fig. 14 shows the time
slice at 1.2 s of the first 3-D CO volume, where the structure of the
salt body is hardly visible both on the raw time slice and the time
slice where traces were deleted. However, a significantly improved

Figure 15. Inline 190 of the second 3-D CO volume. (a) Raw CO section. (b) Raw CO section with gaps. (c) Interpolation with conventional 3-D partial CRS
stacking. (d) Interpolation with the 5-D WABI method.
result is obtained by both, the conventional 3-D partial CRS stacking
and the 5-D WABI method (see Figs 14c and d). Again, the latter
performs better leading to smoother images than the conventional
3-D partial CRS for most of the ZO sample locations.
Results obtained from the second 3-D CO volume are signif-
icantly improved with the 5-D WABI method when compared to
the 5-D WABI method applied to the first 3-D CO volume. See
the corresponding inline 190 and crossline 300 of the second 3-D
CO volume (Figs 15 and 16). Fig. 17 shows the time slice at 1.2 s,
where the 5-D WABI method provides a smoother image with less
noise, compared to the conventional 3-D partial CRS. In the central
part of the time slice, the structure of the salt body imaged by the
5-D WABI method is better visible than the one obtained for the
conventional 3-D partial CRS stacking. For about 90 per cent of the
ZO samples, the results obtained for the 5-D WABI method are
better than for the conventional 3-D partial CRS (as an example
where this is not the case, see the right top corner of Figs 17c and
d). For these cases these areas, the semblance calculated from the
DE-based wave-front attributes is lower than the semblance for the
conventional 3-D partial CRS stacking (see Figs 8c and d). With
low-semblance traveltime interpolation operator, the data sample
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
5-D interpolation with wave-front attributes 931
may be interpolated at a wrong (W) position, instead of the right
(R) position (see Fig. 8b). Usually, one may use a local optimiza-
tion algorithm to refine the DE-based wave-front attributes in these
low-semblance areas. However, this option failed in our tests, where
two local optimization algorithms, namely modified simulated an-
nealing (MSA) and Nelder–Mead method were tested. We assume
that a local optimization may not guarantee to find or move forward
to the global best if the initial guess is too far away from the global
best. We conclude, that for most ZO samples, nearly 90 per cent in
this case, the results obtained from the 5-D WABI are obtained in
better quality using less computational time.
Using a wave-front-attribute-based interpolation operator and
considering several events for the same ZO sample (conflicting
dips) may be cumbersome in the interpolation process if we have
complex data and large data gaps. This conclusion, however, is ap-
plicable to trace interpolation in general since the projected Fresnel
zone set physical limits with respect to size of handleable data gaps.
For some ZO samples, the number of seismic events at the con-
sidered ZO sample can be counted and searched by the algorithm.
In such case, a well-interpolated result can be obtained. However,
at some ZO samples with sparse data comprising large data gaps,
 932 Y. Xie and D. Gajewski
Figure 16. Crossline 300 of the second 3-D CO volume. (a) Raw CO section. (b) Raw CO section with gaps. (c) Interpolation with conventional 3-D partial
CRS stacking. (d) Interpolation with the 5-D WABI method.
the 3-D hyperbolic or non-hyperbolic CRS traveltime operator may
fail in the process of determining wave-front attributes which be-
long to different events. In this work, we consider only the dom-
inant event at each ZO sample. Obviously, events not considered
are not interpolated and therefore missing after the interpolation
process. The conventional 3-D partial CRS stacking has already
demonstrated good performance for reflection data (see Baykulov
& Gajewski 2009, 2010). This work has shown that the 5-D WABI
method provides reliable results for reflections and diffractions.
Limited by the offset range of the 3-D SEG C3WA data, we use the
hyperbolic 3-D CRS operator in this work.
3.5 The 5-D WABI and the rank-reduction-based 5-D
interpolation
The proposed 5-D WABI method, as an extension of the conven-
tional 3-D partial CRS approach, has shown its potential to recon-
struct reflection and diffraction data. However, related to the 5-D
interpolation, there are several other types of 5-D seismic interpo-
lation approaches proposed recently. In order to check the potential
usage of the proposed 5-D WABI, we also compare it with the rank-
reduction-based 5-D interpolation method, for example, the damped
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
rank-reduction method (Chen et al. 2016b), which is worked quite
well for reflection data. The rank-reduction codes used in our work
are modified from the open-source Matlab code package (Chen
et al. 2016a). The compared results are given in two 3-D data sets
discussed below.
3.5.1 Model with three layers and a spherical diffractor
We first use a simple 3-D model to generate pre-stack synthetic
seismic data which are close to the benchmark 3-D data used in
many published interpolation works (e.g. Trad 2009; Oropeza &
Sacchi 2011; Chen et al. 2016a,b). In addition we consider a spher-
ical diffractor in the model since most recently published 5-D in-
terpolation results mainly focus on data dominated by reflections
(e.g. Trad 2009; Chopra & Marfurt 2013; Kreimer et al. 2013;
Gao et al. 2015; Ely et al. 2015; Chen et al. 2016b). Diffraction
is more and more gaining interest in both academic and industry
applications since it images small-scale heterogeneities and struc-
tures, for example, fractures, pinch-outs, thin lenses etc. (e.g. Dell
& Gajewski 2011a,b; Rad et al. 2015). Preserving diffractions (i.e.
fractures) in 5-D interpolation is a discussed topic and leaves space
for discussions and investigations (Trad 2014).
 5-D interpolation with wave-front attributes 933

Figure 17. Time slice 1.2 s of the second 3-D CO volume. (a) Raw data. (b) Raw data with gaps. (c) Interpolation with conventional 3-D partial CRS stacking.
(d) Interpolation with the 5-D WABI method.

120 Ricker-wavelet with a prevailing frequency of 20 Hz is applied. The

3-D synthetic pre-stack seismic data has a CMP spacing of 12.5 m
in the x direction, and 25 m in the y direction. The maximum CMP
100
fold is 20, and the sampling interval is 4 ms. We add 20 per cent
random noise to the data, that is, the maximum RMS amplitude of
the signal is 5 times than that of the random noise. In the reduced
data set, 50 per cent traces is randomly removed. In the 3-D CRS
Computational costs [%]

80
60
40
20
0
Figure 18. Computational costs of the 5-D WABI method and the 5-D
damped rank-reduction method (5-D DRRM), where the 5-D WABI is ref-
erenced as 100 per cent. We mention that the computational costs of the
5-D WABI tested here include a global determination of the wave-front
attributes.
The simple 3-D model is separated by three layers with the veloc-
ity of 1500, 1800 and 2000 m s−1 , respectively. A spherical diffractor
with a velocity of 4000 m s−1 and with a lateral extension of 100 m
is buried in the bottom layer. In the forward modelling, a zero-phase
wave-front attribute search, the midpoint aperture is 75 m in the x
direction and 50 m in the y direction. The 3-D CRS offset apertures
and 3-D partial CRS offset apertures are estimated as discussed in
Appendix.
In the following, we will compare the two interpolation meth-
ods with respect to computational efficiency and image quality for
reflections and diffractions. Fig. 18 shows the computational costs
between the two 5-D interpolation methods tested from the simple
3-D data, which indicates that the damped rank-reduction method
has about the same magnitude of computational costs compared
with the 5-D WABI method when we use the same computational
setting, for example, the number of CPUs. The comparison may be
slightly different with different implementations or different param-
eter setting, but it indicates that both methods could be comparable
in the computational costs. The parameter setting is crucial in the
damped rank-reduction method which needs some testing for the
data under consideration. We use rank K = 3, which is a parameter
accounting for the reconstructed events. The damping factor N is
set to 2. The greater the N, the weaker is the damping, that is, we
may get higher S/N data with smaller N. The number of iterations
for each frequency slice is set to 10. The minimum and maximum
frequency to reconstruct are between 0 and 200 Hz. These parame-
ters are obtained after some testing from the data. For this test, the

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 934 Y. Xie and D. Gajewski

Figure 19. Inline 10 of the 3-D CO volume (hx = 25 m, hy = 0 m). (a) Original data section. (b) Reduced section after adding 20 per cent random noise and
killing 50 per cent traces randomly. (c) Denoising and data reconstruction with the damped rank-reduction method. (d) Denoising and data reconstruction with
the 5-D WABI method.

Hankel matrix and the three block Hankel matrices being set close
to square matrices. For a detailed discussion on parameters we refer
to Chen et al. (2016b).
In order to show the 5-D interpolation results in 2-D planes,
we first extract a 3-D CO volume, for example, setting the con-
stant half offset h = (25 m, 0 m), from the 5-D interpolated vol-
ume, then the 3-D CO volume is cut into two sections (inline 10
and crossline 100) and two time slices (1.66 and 2.4 s). The two
time slices are used to show the reflection (1.66 s) and the bot-
tom diffraction (2.4 s). The colour bar of all figures shown be-
low means the amplitude after the interpolation. Fig. 19 shows
inline 10 of the 3-D CO volume. We note that the damped rank-
reduction method displays reflections slightly stronger, while the
5-D WABI method is superior for diffractions. In the top part
with reflection-only data, the damped rank-reduction method re-
constructs the reflection to the boundary even if the reflection ter-
minates prior to the boundary (see the left-hand white arrows).
In the bottom part of the data, the diffraction pattern is recov-
ered quite well with the 5-D WABI method. The damped rank-
reduction method preserves only parts of the diffraction. Fig. 20
is the crossline 100 of the 3-D CO volume, where both interpo-
lation methods are comparable when the crossline is taken at the
centre of the salt body, that is, closer to the spherical diffractor.
Fig. 21 displays the time slice at 1.66 s cutting through the second
reflector. Here the damped rank-reduction method shows a better
performance than the 5-D WABI to reconstruct the reflection. How-
ever, in the bottom part at time slice 2.4 s which cuts through
the diffraction pattern, the 5-D WABI is considerably better in
preserving the diffraction even when compared with the damped
rank-reduction method (see Fig. 22). For a more detailed investiga-
tion of the damped rank-reduction method on diffraction-only data,
the bottom part of the whole 3-D data is used. The comparison
is shown in Fig. 23. In this comparison we see that the damped
rank-reduction method can recover parts of the diffraction but still
fail to preserve the diffraction tails. If the frequency component
of the random noise is close to that of the weak diffraction tails,
we may have a challenge to predict or reconstruct the diffraction
tails in the frequency domain without a data enhancement facil-
ity which is included in the 5-D WABI procedure. Perhaps we
need to design adaptive sizes for the Hankel matrix and for the

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 5-D interpolation with wave-front attributes 935

(a) (b)
0 0
0.8 0.8

1 0.6 1 0.6
Time [s]

Time [s]
0.4 0.4
2 2

0.2 0.2

3 3
0 0

4 -0.2 4 -0.2
5 10 15 20 5 10 15 20
Inline number Inline number

(c) (d)
0 0
0.8 0.8

1 0.6 1 0.6
Time [s]

Time [s]
0.4 0.4
2 2

0.2 0.2

3 3
0 0

4 -0.2 4 -0.2
5 10 15 20 5 10 15 20
Inline number Inline number
Figure 20. Crossline 100 of the 3-D CO volume (hx = 25 m, hy = 0 m). (a) Original data. (b) Reduced section after adding 20 per cent random noise and
killing 50 per cent traces randomly. (c) Denoising and data reconstruction with the damped rank-reduction method. (d) Denoising and data reconstruction with
the 5-D WABI method.

three block Hankel matrices to account for the diffraction in the
future.
3.5.2 Applications to the 3-D SEG C3WA data
In this section, we apply the two 5-D interpolation methods to the
open 3-D SEG data, with S/N = 5, and 50 per cent traces randomly
removed in the data. We use K = 15, N = 6, and 10 iterations in
the damped rank-reduction method. The minimum and maximum
reconstructed frequencies are set between 0 to 250 Hz. These pa-
rameters are also tested from the 3-D SEG data in which we get
almost the same result if the maximum reconstructed frequency
is larger than 50 Hz. The 3-D CRS apertures and the 3-D partial
CRS apertures used here are set the same as in Section 3.4. Fig. 24
shows inline 190 of a 3-D volume that is extracted from the 5-D
interpolated result with a constant half offset of (0 m, 20 m). We
note that the damped rank-reduction method is strong to reconstruct
the reflection events, but weak to preserve the diffraction events by
the salt body in the middle of the model. The 5-D WABI is strong
in preserving the diffraction patterns including the weak multiple
diffractions, compared to the damped rank-reduction method. Sim-
ilar conclusions are obtained for crossline 300 shown in Fig. 25. In
the time slice at 1.2 s shown in Fig. 26, the damped rank-reduction
method performs well to reconstruct the strong reflections, but fails
to preserve the diffraction events caused by edgy structures, while
the 5-D WABI keeps both, the reflection and the diffraction, which
are recovered very well when compared with the original data.
4 C O N C LU S I O N S
We have extended the conventional 3-D partial CRS stacking ap-
proach for 5-D trace interpolation. The global optimization strat-
egy to determine wave-front attributes and the azimuthal interpo-
lation process better account for the potential of wide-, rich- and
full-azimuth data. The 5-D WABI method comprises next to the
usual 5-D data space (e.g. CMP coordinates in x and y, offset, az-
imuth and time) the wave-front attributes which contain surface
expressions of subsurface features. In the 5-D WABI method the

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 936 Y. Xie and D. Gajewski

(a) (b)

0.8 0.8

Inline number 50 0.6 50 0.6

Inline number
0.4 0.4
100 100
0.2 0.2

150 0
150 0

-0.2 -0.2
5 10 15 20 5 10 15 20
Crossline number Crossline number

(c) (d)

0.8 0.8

50 0.6 50 0.6
Inline number

Inline number

0.4 0.4
100 100
0.2 0.2

150 0
150 0

-0.2 -0.2
5 10 15 20 5 10 15 20
Crossline number Crossline number
Figure 21. Time slice of 1.66 s of the 3-D CO volume (hx = 25 m, hy = 0 m). (a) Original data section. (b) Reduced section after adding 20 per cent
random noise and killing 50 per cent traces randomly. (c) Denoising and data reconstruction with the damped rank-reduction method. (d) Denoising and data
reconstruction with the 5-D WABI method.

wave-front attributes are determined by a 3-D CRS approach using
global optimization. An algorithm based on differential evolution is
used for this purpose. The global optimization provides improved
wave-front attributes when compared to the conventional method
using the pragmatic optimization strategy. The improvements are
particularly visible at far offset traces. The comparison of the 5-
D WABI with a 5-D interpolation technique, for example, based
on the damped rank-reduction-method, revealed a better perfor-
mance in the interpolation process for steeply dipping events like,
for example, diffraction tails. Because of this feature, diffractions
are much better preserved by the 5-D WABI than by the rank-
reduction-based interpolation. Since diffraction imaging is develop-
ing into a powerful add on to plain reflection processing, this feature
is substantial for all kind of processing steps utilizing diffractions.
Since the wave-front attributes are determined from kinematic prop-
erties of the wavefield, namely moveout, aliasing issues are relaxed.
The computational effort of the 5-D WABI approach is comparable
to the CPU time consumed by the damped rank-reduction method
where the comparison includes the CPU time for the determina-
tion of wave-front attributes which may be used for several other
processing purposes like velocity model building, time and depth
migration, diffraction separation and tomography, just to name a
few. The 5-D WABI provides a powerful alternative to other 5-D
interpolation methods with improved handling of diffractions and
relaxed aliasing issues. Next to the interpolation capability itself the
process also provides a data enhancement facility.
AC K N OW L E D G E M E N T S
We thank the Applied Seismics Group in Hamburg for continuous
discussion. We thank the sponsors of the Wave Inversion Technol-
ogy (WIT) Consortium for technical support and SEG for provid-
ing the data. Y. Xie would like to thank Ransheng Chen (China
JK Institute of Engineering Investigation and Design) to discuss
the DE algorithm in actual implementation and also thanks the
China Scholarship Council (CSC) for partially funding this work.
We are grateful to the editors and reviewers for their comments and
suggestions.

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 5-D interpolation with wave-front attributes 937

(a) (b)
0.1 0.1

50 0.05 50 0.05

Inline number

Inline number
100 0 100 0

-0.05 -0.05
150 150

-0.1 -0.1
5 10 15 20 5 10 15 20
Crossline number Crossline number

(c) (d)
0.1 0.1

50 0.05 50 0.05
Inline number

Inline number
100 0 100 0

-0.05 -0.05
150 150

-0.1 -0.1
5 10 15 20 5 10 15 20
Crossline number Crossline number

Figure 22. Time slice of 2.4 s of the 3-D CO volume (hx = 25 m, hy = 0 m). (a) Original data section. (b) Reduced section after adding 20 per cent random noise
and killing 50 per cent traces randomly. (c) Denoising and data reconstruction with the damped rank-reduction method. (d) Denoising and data reconstruction
with 5-D WABI method.

(a) (b)
2 2
0.03 0.03

0.02 0.02

0.01 0.01
Time [s]

Time [s]

2.5 2.5
0 0

-0.01 -0.01

-0.02 -0.02

100 125 100 125

Crossline number Crossline number

(c) (d)
2 2
0.03 0.03

0.02 0.02

0.01 0.01
Time [s]

Time [s]

2.5 2.5
0 0

-0.01 -0.01

-0.02 -0.02

150 175 150 175

Crossline number Crossline number

Figure 23. 5-D interpolation results with the diffraction-only data, that is, the bottom part of the 3-D CO volume (hx = 25 m, hy = 0 m). The top two columns
are selected from the central part of the diffraction-only data, where (a) is gotten from the damped rank-reduction method and (b) is obtained from the 5-D
WABI method. The bottom two columns selected from the right-hand side of the diffraction-only data, where (c) is gotten from the damped rank-reduction
method and (d) is obtained by the 5-D WABI method.

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 938 Y. Xie and D. Gajewski

Figure 24. Inline 190 of the 3-D SEG CO volume (hx = 0 m, hy = 20 m). (a) Original data section with 20 per cent random noise. (b) Reduced section with
50 per cent randomly deleted traces. (c) Denoising and data reconstruction with the damped rank-reduction method. (d) Denoising and data reconstruction with
5-D WABI method. The four subfigures have the same colour bar with a minimum (blue) and a maximum (red) amplitude ranged from −0.6 to 0.6.

(a) (b)
0 0

1 1

2 2
Time [s]

Time [s]

3 3

4 4

5 5
50 100 150 50 100 150
Inline number Inline number

(c) (d)
0 0

1 1

2 2
Time [s]

Time [s]

3 3

4 4

5 5
50 100 150 50 100 150
Inline number Inline number

Figure 25. Crossline 300 of the 3-D SEG CO volume (hx = 0 m, hy = 20 m). (a) Original data section with 20 per cent random noise. (b) Reduced section
with 50 per cent randomly deleted traces. (c) Denoising and data reconstruction with the damped rank-reduction method. (d) Denoising and data reconstruction
with 5-D WABI method. The four subfigures have the same colour bar with a minimum (blue) and a maximum (red) amplitude ranged from −0.6 to 0.6.

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 5-D interpolation with wave-front attributes 939

(a) (b)

50 50

100 100

150 150
Inline number

Inline number
200 200

250 250

300 300

350 350

50 100 150 50 100 150

Crossline number Crossline number

(c) (d)

50 50

100 100

150 150
Inline number

Inline number

200 200

250 250

300 300

350 350

50 100 150 50 100 150

Crossline number Crossline number
Figure 26. Time slice of 1.2 s of the 3-D SEG CO volume (hx = 0 m, hy = 20 m). (a) Original data section with 20 per cent random noise. (b) Reduced section
with 50 per cent randomly deleted traces. (c) Denoising and data reconstruction with the damped rank-reduction method. (d) Denoising and data reconstruction
with 5-D WABI method. The four subfigures have the same colour bar with a minimum (blue) and a maximum (red) amplitude ranged from −0.6 to 0.6.

REFERENCES Chen, Y., Zhang, D., Jin, Z., Chen, X., Zu, S., Huang, W. & Gan,
S., 2016b. Simultaneous denoising and reconstruction of 5-D seis-
Baykulov, M. & Gajewski, D., 2009. Prestack seismic data enhancement with
mic data via damped rank-reduction method, Geophys. J. Int., 206,
partial common-reflection-surface (CRS) stack, Geophysics, 74, V49–
1695–1717.
V58.
Chopra, S. & Marfurt, K.J., 2013. Preconditioning seismic data with 5D in-
Baykulov, M. & Gajewski, D., 2010. Partial 3D CRS Stack, in 72nd EAGE
terpolation for computing geometric attributes, Leading Edge, 32, 1456–
Conference and Exhibition incorporating SPE EUROPEC 2010.
1460.
Bergler, S., 2004. On The Determination and Use of Kinematic Wavefield
Curry, W., 2010. Interpolation with Fourier-radial adaptive thresholding,
Attributes for 3D Seismic Imaging, Logos.
Geophysics, 75, WB95–WB102.
Bortfeld, R., 1989. Geometrical ray theory: rays and traveltimes in seismic
Dell, S & Gajewski, D., 2011a. 3D CRS-attributes based diffraction imaging,
systems (second-order approximation of the traveltimes), Geophysics, 54,
in 73rd EAGE Conference and Exhibition incorporating SPE EUROPEC
342–349.
2011, doi:10.3997/2214-4609.20148991.
Červený, V., 2001. Seismic Ray Theory, Cambridge Univ. Press.
Dell, S. & Gajewski, D., 2011b. Common-reflection-surface-based workflow
Červený, V. & Moser, T.J., 2007. Ray propagator matrices in three-
for diffraction imaging, Geophysics, 76, S187–S195.
dimensional anisotropic inhomogeneous layered media, Geophys. J. Int.,
Ely, G., Aeron, S., Hao, N. & Kilmer, M.E., 2015. 5D seismic data com-
168, 593–604.
pletion and denoising using a novel class of tensor decompositions,
Chen, Y., Huang, W., Zhang, D. & Chen, W., 2016a. An open-source Matlab
Geophysics, 80, V83–V95.
code package for improved rank-reduction 3D seismic data denoising and
Fomel, S., 2003. Seismic reflection data interpolation with differential offset
reconstruction, Comput. Geosci., 95, 59–66.
and shot continuation, Geophysics, 68, 733–744.

Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes

by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
 940 Y. Xie and D. Gajewski
Fomel, S. & Kazinnik, R., 2013. Non-hyperbolic common reflection surface,
Geophys. Prospect., 61, 21–27.
Gao, J., Stanton, A. & Sacchi, M.D., 2015. Parallel matrix factorization
algorithm and its application to 5D seismic reconstruction and denoising,
Geophysics, 80, V173–V187.
Gelchinsky, B., Berkovitch, A. & Keydar, S., 1999, Multifocusing homeo-
morphic imaging: Part 1. Basic concepts and formulas, J. Appl. Geophys.,
42, 229–242.
Höcht, G., 2002. Traveltime Approximations for 2D and 3D Media and
Kinematic Wavefield Attributes, University of Karlsruhe.
Höcht, G., Ricarte, P., Bergler, S. & Landa, E., 2009. Operator oriented CRS
interpolation, Geophys. Prospect., 57, 957–979.
Holland, J., 1975. Adaptation in Natural and Artificial Systems, University
of Michigan Press.
Huang, W., Wang, R., Zhang, M., Chen, Y. & Yu, J., 2015. Random noise
attenuation for 3D seismic data by modified multichannel singular spec-
trum analysis, in 77th Annual International Conference and Exhibition,
EAGE Extended Abstracts, doi:10.3997/2214-4609.201412830.
Hubral, P., 1983. Computing true amplitude reflections in a laterally inho-
mogeneous earth, Geophysics, 48, 1051–1062.
Hubral, P., Schleicher, J. & Tygel, M., 1992. Three-dimensional paraxial ray
properties. Part I. Basic relations, J. seism. Explor., 1, 265–279 [Part II.
Applications, J. seism. Explor., 1, 347–362].
Hubral, P., Schleicher, J., Tygel, M. & Hanitzsch, C., 1993a. Determination
of Fresnel zones from traveltime measurements, Geophysics, 58, 703–
712.
Hubral, P., Schleicher, J. & Tygel, M., 1993b. Three-dimensional primary
zerooffset reflections, Geophysics, 58, 692–702.
Jäger, R., 1999. The common reflection surface stack-introduction and ap-
plication, Diploma thesis, Karlsruhe University, Germany.
Jäger, R., Mann, J., Höcht, G. & Hubral, P., 2001. Common-reflection-
surface stack: Image and attributes, Geophysics, 66, 97–109.
Kabir, M.M.N. & Verschuur, D.J., 1995. Restoration of missing offsets by
parabolic Radon transform, Geophys. Prospect., 43, 347–368.
Kaplan, S.T., Naghizadeh, M. & Sacchi, M.D., 2010. Data reconstruc-
tion with shot-prole least-squares migration, Geophysics, 75, WB121–
WB136.
Kennedy, J. & Eberhart, R., 1995. Particle swarm optimization, in Proceed-
ings of IEEE International Conference on Neural Networks, pp. 1942–
1948.
Kreimer, N., Stanton, A. & Sacchi, M.D., 2013. Tensor completion based
on nuclear norm minimization for 5D seismic data reconstruction, Geo-
physics, 78, V273–V284.
Liu, B. & Sacchi, M.D., 2004. Minimum weighted norm interpolation of
seismic records, Geophysics, 69, 1560–1568.
Liu, Y. & Fomel, S., 2011. Seismic data interpolation beyond aliasing using
regularized nonstationary autoregression, Geophysics, 76, V69–V77.
Mann, J., 2002. Extensions and applications of the Common-Reflection-
Surface Stack Method, PhD thesis, University of Karlsruhe.
Mann, J., Jager, R., Muller, T., Hocht, G. & Hubral, P., 1999. Common-
reflection-surface stack - a real data example, J. Appl. Geoph., 42, 301–
318.
Moser, T.J. & Červený, V., 2007. Paraxial ray methods for anisotropic inho-
mogeneous media, Geophys. Prospect., 55, 21–37.
Müller, N.A., 2003. The 3D Common-reflection-surface stack—theory and
application, Diploma thesis, University of Karlsruhe.
Naghizadeh, M. & Innanen, K.A., 2011. Seismic data interpolation using a
fast generalized Fourier transform, Geophysics, 76, V1–V10.
Naghizadeh, M. & Sacchi, M.D., 2009. f-x adaptive seismic-trace interpo-
lation, Geophysics, 74, V9–V16.
Naghizadeh, M. & Sacchi, M.D., 2010. Beyond alias hierarchical scale
curvelet interpolation of regularly and irregularly sampled seismic data,
Geophysics, 75, WB189–WB202.
Neidell, N.S. & Taner, M.T., 1971. Semblance and other coherency measures
for multichannel data, Geophysics, 36, 482–497.
Oropeza, V. E. & Sacchi, M.D., 2011. Simultaneous seismic data de-noising
and reconstruction via multichannel singular spectrum analysis, Geo-
physics, 76, V25–V32.
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017
Pedersen, M.E.H., 2010. Good Parameters for Differential Evolution, Hvass
Laboratories Technical Report HL1002.
Powell, M., 1964. An efficient method for finding the minimum of a function
of several variables without calculating derivatives, Comput. J., 7, 155–
162.
Rad, P.B., Vanelle, C. & Gajewski, D., 2015. 3D CRS-based prestack diffrac-
tion separation and imaging, in 77th EAGE Conference and Exhibition.
Ronen, J., 1987. Wave-equation trace interpolation, Geophysics, 52, 973–
984.
Schleicher, J., Tygel, M. & Hubral, P., 1993a. 3D true-amplitude finite offset
migration, Geophysics, 58, 1112–1126.
Schleicher, J., Tygel, M. & Hubral, P., 1993b. Parabolic and hyperbolic
paraxial two-point traveltimes in 3D media, Geophys. Prospect., 41, 495–
513.
Schwarz, B., Vanelle, C., Gajewski, D. & Kashtan, B., 2014. Curvatures and
inhomogeneities: an improved common-reflection approach, Geophysics,
79, S231–S240.
Shi, Y. & Eberhart, R.C., 1998. A modified particle swarm optimizer, in
Proceedings of IEEE International Conference on Evolutionary Compu-
tation, pp. 69–73.
Shuey, R.T., 1985. A simplification of the Zoeppritz equations, Geophysics,
50, 609–614.
Spitz, S., 1991. Seismic trace interpolation in the f-x domain, Geophysics,
56, 785–794.
Stolt, R.H., 2002. Seismic data mapping and reconstruction, Geophysics,
67, 973–984.
Storn, R. & Price, K., 1997. Differential evolution - a simple and efficient
heuristic for global optimization over continuous spaces, J. Glob. Optim.,
11, 341–359.
Trad, D., 2009. Five-dimensional interpolation: recovering from acquisition
constraints, Geophysics, 74, V123–V132.
Trad, D., 2014. Five-dimensional interpolation: new directions and chal-
lenges, in CSEG Recorder, pp. 40–46.
Trad, D., Ulrych, T. & Sacchi, M., 2002. Accurate interpolation with
high-resolution time-variant Radon transforms, Geophysics, 67, 644–
656.
Trickett, S., 2008. F-xy cadzow noise suppression, CSPG CSEG CWLS
Convention, 67, 303–306.
Trickett, S., Burroughs, L., Milton, A., Walton, L. & Dack, R., 2010. Rank-
reduction-based trace interpolation, SEG Technical Program Expanded
Abstracts, 29, 3829–3833.
Walda, J., Schwarz, B. & Gajewski, D., 2017. A competitive com-
parison of multiparameter stacking operators, Geophysics, 82(4),
V275–V283.
Weise, T., 2009. Global Optimization Algorithms - Theory and Application,
http://www.it-weise.de, self-published.
Xie, Y., 2017. 3D prestack data enhancement with a simplified CO CRS
operator, in 79th Annual International Conference and Exhibition, EAGE,
Extended Abstracts.
Xie, Y. & Gajewski, D., 2016a. Automatic estimation of the 3D CRS
attributes by a metaheuristic-based optimization, in 78th Annual In-
ternational Conference and Exhibition, EAGE, Extended Abstracts,
doi:10.3997/2214-4609.201601436.
Xie, Y. & Gajewski, D., 2016b. Interpolation and regularization with the 3D
CRS operator, SEG Technical Program Expanded Abstracts, pp. 4092–
4096.
Xie, Y. & Gajewski, D., 2016c. Simultaneous estimation of the 3D CRS at-
tributes by an evolutionary-based Nelder-Mead algorithm, SEG Technical
Program Expanded Abstracts, pp. 4326–4330.
Xu, S., Lambare, G. & Calandra, H., 2004. Fast migration/inversion with
multivalued ray fields: Part 2—Applications to the 3D SEG/EAGE salt
model, Geophysics, 69, 1320–1328.
Zhang, Y.Q. & Lu, W.K., 2014. 2D and 3D prestack seismic data regu-
larization using an accelerated sparse time-invariant Radon transform,
Geophysics, 79, V165–V177.
Zwartjes, P.M. & Sacchi, M.D., 2007. Fourier reconstruction of
nonuniformly sampled, aliased seismic data, Geophysics, 72,
V21–V32.

APPENDIX
A1 3-D CRS offset apertures
Fig. 1(b) shows an offset aperture of the 3-D CRS stacking used
in this work. The parameters hx, 0 and hy, 0 are the semi-major and
semi-minor axes of the top ellipse, where a selection of the location
(cutting through a sample) and the size of the top ellipse depends
on the used data. In practice, the observed seismic events from the
top to the bottom in the offset dimension are constrained within
the top and the bottom ellipses initially, then different candidate
apertures are probed until the one with the possible best perfor-
mance is found. The location and size of the bottom ellipse is de-
fined in the same way. In this case, the semi-major and semi-minor
axes of any ellipse cutting through a sample on the ZO trace are
computed by
h x,i = h x,0 + i ∗ (h x,n − h x,0 )/n, (A1)
h y,i = h y,0 + i ∗ (h y,n − h y,0 )/n, (A2)
where hx, n and hy, n are the semi-major and semi-minor axes of the
bottom ellipse, and n is the maximum number of ZO samples. In
the 3-D SEG data, we set hx, 0 = 140 m, hy, 0 = 800 m, hx, n = 140 m,
Downloaded from https://academic.oup.com/gji/article-abstract/211/2/919/4060700/5-D-interpolation-with-wave-front-attributes
by Bibliothekssystem Universitaet Hamburg user
on 15 September 2017 stats
View publication
5-D interpolation with wave-front attributes 941
hy, n = 1350 m. In the simple 3-D data, we set hx, 0 = 800 m, hy, 0 =
50 m, hx, n = 1000 m, hy, n = 50 m. The 3-D SEG data has a small
half offset in the x direction, up to 140 m, and the simple 3-D data
has a small half offset in the y direction, up to 50 m.
A2 3-D partial CRS offset apertures
The 3-D partial CRS stacking apertures for the offset are determined
as:
px,i = b + i ∗ (d − b)/n, (A3)
p y,i = a + i ∗ (c − a)/n, (A4)
where a, b, c and d are estimated by applying different possible
apertures and keeping the one with the best image quality. In the
3-D SEG data, we set a = c = 400 m, b = d = 100 m. So for all
samples, we have px, i = 100 m, py, i = 400 m. Similarly, we set
a = c = 50 m, b = d = 500 m in the simple 3-D data case. One
may note that the offset apertures are used to cut the tested data
from the original data. The observed seismic events are included in
the apertures as long as the apertures are large than the maximum
length of the seismic events in the horizontal direction.