You are on page 1of 11

Journal of Geophysics and Engineering

J. Geophys. Eng. 12 (2015) 199–209 doi:10.1088/1742-2132/12/2/199

Automated diffraction delineation using an


apex-shifted Radon transform
Sadegh Karimpouli1, Alireza Malehmir2, Hossein Hassani3,
Hossein Khoshdel4 and Majid Nabi-Bidhendi5
1
  Mining Engineering Group, Faculty of Engineering, University of Zanjan, Zanjan, Iran
2
  Department of Earth Sciences, Uppsala University, SE 75236 Uppsala, Sweden
3
  Faculty of Mining and Metallurgy Engineering, Amirkabir University of Technology, Tehran, Iran
4
  Exploration Directorate, NIOC, Tehran, Iran
5
  Institute of Geophysics, University of Tehran, Tehran, Iran

E-mail: s.karimpouli@znu.ac.ir

Received 27 June 2014, revised 27 October 2014


Accepted for publication 12 January 2015
Published 10 March 2015

Abstract
Diffraction arrivals are important data that have increasingly been used to delineate the
sources of diffractors and to explore subsurface discontinuities. In prestack data, diffractions
are both zero- and non-zero offset hyperbolas while reflections are only zero-offset
hyperbolas. An iterative algorithm using an apex-shifted Radon transform (ASRT) approach
is presented in this paper that uses the diffraction hyperbolic trajectory similar to that of
prestack time migration in order to locate diffractors and to estimate their corresponding
background velocities. Because diffraction energy is generally weak in seismic data and
particularly in prestack data, noise attenuation and edge enhancement methods are applied
on the instantaneous phase of the seismic data instead of the amplitude data. This means that
the phase data are input to the ASRT algorithm. The method is then tested on two synthetic
datasets (a point-diffraction model with randomly distributed diffractors and the 2D BP/SEG
salt model) and one real data example. Results show that this method can locate the diffractors
reasonably well on the rough surfaces of the salt dome and the discontinuities associated
with structures such as paleo-channels and faults. Our analysis of the estimated velocities
suggests that they are generally valid for diffraction delineation; however, the accuracy of the
estimation decreases as background velocity and depth increase.

Keywords: diffractions, apex-shifted Radon transform, discontinuities, velocity estimation

(Some figures may appear in colour only in the online journal)

1. Introduction In summary, the following approaches have been studied: (1)


kinematic travel times of the diffracted arrivals (e.g. Landa
Diffracted seismic waves provide valuable data for detecting et al 1987, Kanasewich and Phadke 1988, Tsingas et al
and distinguishing important subsurface structures such as frac- 2011), (2) separation of the weak energy of diffractions from
ture zones, faults, pinch-outs, reef edges, sub-volcanic mounds, the reflected waves (e.g. Taner et al 2006, Fomel et al 2007,
small intrusive bodies, rough surfaces of salt domes and other Berkovitch et al 2009, Al-Dajani and Fomel 2010, Burnett
geological discontinuities. These seismic structures can play and Fomel 2011, Dell and Gajewski 2011, Klokov and Fomel
an important role in, for example, mineral exploration (e.g. 2012), (3) modification of the conventional Kirchhoff migra-
Malehmir and Bellefleur 2009, Malehmir et al 2011, 2012, tion algorithm to include diffracting points (e.g. Schmelzbach
Malinowski et al 2012) and hydrocarbon exploration and/or et al 2008, Malehmir et al 2009, 2010, Zhu and Wu 2010;
production (e.g. Tsingas et al 2011, Karimpouli et al 2013). Bellefleur et al 2012) and (4) application of image processing
Over the past three decades, different properties of dif- techniques usually using pattern recognition techniques such
fractions have been studied to identify their point of origin. as Radon and Hough transforms (Windsor et al 2005, Song

1742-2132/15/020199+11$33.00 199 © 2015 Sinopec Geophysical Research Institute  Printed in the UK


J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

et al 2010, Figueiredo et al 2012). An issue here is that in point, implying that diffraction delineation and a simple
all of these studies, even those separating different sources of velocity analysis can be done simultaneously.
seismic energy, the user or interpreter has to make a decision One issue that should be dealt with is the weak energy of
about the location of the diffraction at some stage of the work. diffracted waves compared with reflected ones. The maximum
This procedure is often inaccurate. amplitude of a diffraction (which occurs at its apex) is one-
In this study, we propose a procedure that is based on the half that of the associated reflection (Trorey 1970) making
delineation of diffracted energy using an apex-shifted Radon diffractions difficult to detect in a seismic section, especially
transform (ASRT) method (Trad et al 2003). This method in shot records. Energy also decays quickly as the diffraction
uses the kinematics of diffraction travel times in the pre- tails off, and as such, sometimes recognizing it is difficult.
stack domain (similar to prestack time migration). In seismic To overcome this problem and enhance diffractions in the
studies, linear or slant stacking (Yilmaz 2001), hyperbolic, presence of reflections, we propose to use the instantaneous
parabolic, hybrid and apex-shifted Radon transforms have phase instead of the amplitude of the data mainly because
often been used, for example for velocity analysis (Thorson coherent events are more easily detectable in phase sections.
and Claerbout 1985), multiple attenuation (Foster and Mosher The instantaneous phase of seismic data can be extracted after
1992, Wang 2003, Trad et al 2003) and diffraction enhance- applying field and elevation statics, surface-consistent decon-
ment and/or attenuation (Nowak 2004, Bansal and Imhof volution and any other processing steps necessary to improve
2005, Klokov and Fomel 2012). In shot gather data, diffrac- the continuity of diffractions. A second step that we propose
tion points appear at any offsets, which means conventional for enhancing seismic events is to apply a Gaussian derivative
Radon transforms cannot locate them, except ASRT. Applying operator, a well-known edge detector function (Canny 1986).
ASRT in a similar procedure to prestack time migration allows In the following sub-sections, these steps are first explained
us to find the locations of diffractions in shot gather data and and then the procedure is thoroughly described.
their background velocities as well. Therefore, shot gather dif-
fraction locations can be normal moveout (NMO) corrected
using the predicted background velocities and then stacked 2.1.  Pre-processing of seismic data
to produce a diffraction location section (output) instead of a
Since in our approach the Radon transform projects seismic
seismic section.
events along a hyperbolic trajectory, it is important that all
Since diffraction signals are often weaker than reflection
coherent events are comparable in quality. In the phase
signals (Berryhill 1977, Klem-Musatov 1994) we use the
domain, it does not matter how strong the recorded energy
instantaneous phase of the data rather than their amplitudes.
is; phase changes are the only important issue. The phase is
To enhance diffraction events, we apply both noise reduction
independent of the amplitude and this makes it more viable to
and edge detection methods. Compared with poststack data,
image coherent but weak events such as diffractions (Taner
this is particularly important since in prestack data diffrac-
et al 1979).
tions are much weaker. Therefore, in this paper we present
Let D(x, t ) be a 2D section of seismic data and D*(x, t ) be
and apply the proposed method to two synthetic datasets
the quadrature section derived by a 2D Hilbert transform of
and one real data example that demonstrate its capability for
this section. A complex or analytic section  Z (x, t ) is defined
locating geological irregularities. The method provides first-
by (Barnes 1996):
hand information on the background velocities, which can
further be used in other applications such as for building an Z (x, t ) = D (x, t ) + iD* (x, t ) = A  exp(iθ )
(1)
initial velocity model that can be used for prestack time or
depth migrations. We show how the method provides reliable A = Z (x, t ) = (D 2 (x, t ) + D*2 (x, t ) )1/2
(2)
results, especially at shallow depths or in low-velocity envi-
ronments such as in sedimentary settings. We also discuss its θ = arg (Z (x, t ) ) = arctan (D(x, t )/ D*(x, t ))
(3)
limitations in heterogeneous and anisotropic environments.
where A is the 2D instantaneous amplitude or envelope, i is
the imaginary unit and θ is the 2D instantaneous phase.
2.  Theory and methodology Figure 1 shows a comparison between an amplitude sec-
tion  and the corresponding 2D instantaneous phase section.
We use the hyperbolic Radon transform in non-zero offsets to Strong events in the amplitude section (figure 1(a)) are weak-
separate diffractions from other coherent events in prestack ened as time increases. Some events are not even recogniz-
seismic data. In shot gathers, semi-hyperbolic reflections (and/ able because of their weak amplitudes. In the instantaneous
or multiples), hyperbolic diffractions and noise (both ambient phase section (figure 1(b)), however, most events are recog-
and source-generated) can be observed. Noise generally has nizable. The phase section  appears complicated since every
no preferred shape, but it can disturb or mask other events. weak event is now apparent. This demonstrates why the phase
Hyperbolic events in seismic data can be focused using the section is superior to the amplitude section for the purpose of
ASRT to a parameter domain. Detection of maximum values diffraction imaging.
in the parameter domain (Shapiro and Stockman 2001) leads In case of low signal-to-noise ratio data, computation of
to identifying the diffraction location in time and space (t – x) the instantaneous phase is problematic and unreliable, which
and the rms-velocity of the medium around the diffraction must also be considered. Even low-level noise can degrade the

200
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

(a) (b)
Offset (m) Offset (m)
-600 -300 0 300 600 -600 -300 0 300 600 Phase (Deg.)

1300 1300
150

1400 1400
100

1500 1500
50

Time (ms)
Time (ms)

1600 1600
0

1700 1700
−50

1800 1800
-100

1900 1900
-150

2000 2000

Figure 1. (a) Amplitude and (b) corresponding instantaneous phase of a synthetic shot gather data. Note that coherent events are better
observed in the phase section even when their amplitudes are much weaker.

actual phase section, and thus can also be a problematic issue. Ds Dr


If noise is not coherent then its effect is negligible during the S D R
focusing procedure of the Radon transform (this will be illus-
trated later). For coherent noise, reduction methods such as
median and/or Wiener filters can be used. Generally, coherent
source-generated noise should be attenuated as much as
Depth

possible.
After the noise reduction, a Gaussian derivative operator
(Canny 1986) is further applied to enhance the seismic events.
The total derivative of the instantaneous phase in two direc-
tions (horizontal and vertical) with a small 2D Gaussian
kernel is used. Trorey (1977) showed that an edge diffractor
generates diffracted waves with opposite phases on the two
tails. The magnitude of the derivatives removes the effect of
the phase change on the edge diffractions. Phase changes thus
Time

have no disturbing effect on the algorithm’s capability for


even detecting edge-diffractions.

2.2.  Apex-shifted Radon transform

The Radon transform was established by Radon (1986) and is


an operator that integrates some physical property of a medium Receivers (Diffracted waves)
along a particular path. As mentioned earlier, numerous appli-
cations of the Radon transform have been applied in explora- Figure 2.  A sketch showing diffraction hyperbola in prestack data.
tion seismology. The ASRT is an extension of the standard The travel time of the hyperbola is obtained using equation (5)
hyperbolic Radon transform with hyperbolic basis functions shown in the text. The apex occurs on the trace located at
diffraction location (D) in space (modified from Berryhill 1977 and
located at every point of a data gather. The mathematical
Kanasewich and Phadke 1988).
description of such an operator is similar to the kinematic
prestack time migration equation, with the horizontal coordi- where t0 is the diffraction location in time and S, R and D
nates being offsets instead of the midpoints (Trad et al 2003). are shot, receiver and diffractor locations along the seismic
Kanasewich and Phadke (1988) presented the travel time of profile, respectively. vrms is the rms-velocity of the medium
diffracted waves in prestack data as the following (figure 2): at the diffractor location. As illustrated in figure 2, the max-
imum amplitude of the diffraction occurs at the receiver cor-
t 2 (S − D )2 t02 (D − R )2 responding to the midpoint coincident with the diffractor
t= 0 +
(4) 2
+ + 2
4 vrms 4 vrms point (Dr = Ds ). Note that, in shot gather data, a phase change

201
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

(a) (b)
1

100 100 1500


0.75

Sample number
Sample number

200 200
1000
0.5
300 300
500
0.25
400 400

500 0 500 0
100 200 300 400 500 100 200 300 400 500
Trace number Trace number
(c) (d)

300
100 100 150
250
Sample number

Sample number

200 200 200


100
150
300 300
100 50
50
400 400
0 0

500 500
100 200 300 400 500 100 200 300 400 500
Trace number Trace number

Figure 3. (a) Synthetic hyperbola with 20% added random noise relative to the maximum amplitude of the data with i0 = 250, j0 = 250,
jS = 250 dx = 4 m, dt = 2 ms and vrms = 2000 m s−1 (see equation (5)). A 3D array of the parameter domain corresponding to the Radon
transform is shown in 2D arrays from (b)–(d). (b) Parameters domain for vrms = 2000 m s−1; (c) for vrms = 4000 m s−1; and (d) for
vrms = 1000 m s−1. The maximum value of the parameter domain (located in (b)) indicates that the unknown parameters of the hyperbola
are i0 = 250, j0 = 250, and vrms = 2000 m s−1, i.e. the same values as the real model.

occurs at this trace (R in figure 2). After discretization, t = i dt, represents the seismic data in the t – x domain (or in the i – j
t0 = i0 dt,  S = jS dx, R = jr dx, D = j0 dx equation  (4) can be discrete domain) and dm contains dt, dx and dvrms for time,
written as (still in shot gather domain): trace and rms-velocity intervals, respectively (di, dj and dv in
the discrete domain). This relation can be written as:
i 2 (j − j )2 dx 2 i 2 (j − j )2 dx 2
i= 0 + S 20 2
(5) + 0 + 0 2r 2 +∞
4 vrms dt 4 vrms dt
(7)
Rf (i0, j0 , vrms ) = ∬∫−∞
f (i, j ) δ (i − ξ )  di  dj  dv
where dx and dt are receiver (or trace) interval and sampling where δ(.) is the Dirac delta function which is zero where i ≠ ξ.
rate, respectively; (i0, j0 ) defines the apex of the hyperbola in The parameters domain is informative since diffractions are
discrete space and jS and jr are shot and receiver locations, converted to points. Therefore, finding voxels with the max-
respectively. imum value in this domain allows us to detect the coordinates
The general form of the Radon transform can be written as of the apex (i0, j0 ), and the rms-velocity (vrms ).
(Helgason 1999): Figure 3 shows how the ASRT is capable of estimating the
unknown parameters of equation (5). Figure 3(a) is a 500 by
Rf (ξ ) =
(6) f (s ) dm(s ) ∫ 500 sample synthetic dataset containing a diffraction hyperbola
contaminated by 20% random noise based on the maximum
ξ
amplitude of the data. The diffraction hyperbola according
where Rf (.) is the Radon transform of f (.) and dm is the to equation  (5) is defined as i0 = 250, j0 = 250, js = 250,
Euclidean measure on the hyperplane ξ in Euclidean space. dx = 4 m, dt = 2 ms and vrms = 2000 m s−1. A 3D array was used
In our case, ξ is the hyperbolic path (equation (5)) with for the parameters domain in which i0 and j0 varied from the
three parameters, i0, j0 and vrms, where Rf (.) is the ASRT, f (.) minimum to maximum values of the data size (from 1 to 500)

202
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

(a) (b)

(c) (d) App

Hhyp

Figure 4.  A schematic illustration of iterative procedure used in the method. In the first iteration (a), the maximum value of the parameter
domain (b) is selected and the corresponding hyperbola is removed from the data (c) and iterations continue until N hyperbolas are found.
After the first iteration ((a)–(c)), to speed up the algorithm, Radon transform is calculated just for the shaded window (d) for the remaining
iterations. ‘App’ is an optimum aperture for diffraction hyperbola which is selected by the user and ‘Hhyp’ is the height of the hyperbola.

and di = dj = 1, implying that each data point could poten- anisotropic fabrics, but will have two different characters (still
tially be a diffraction location; vrms was varied from 1000 to asymmetric in 3D). In the case of vertical transverse anisot-
4000 m s−1 with an increment of 1000 m s−1 (dv = 1000 m s−1). ropy diffractions will also be asymmetric and distorted from
Three values of the velocities are shown in figure 3. One of a perfect hyperbolic shape. The velocities estimated in all
them is the real velocity (2000 m s−1) and the two others are these three scenarios are likely to be incorrect and reflect the
1000 m s−1 and 4000 m s−1, which are lower and higher than anisotropic velocities that are faster parallel to the fabric and
the real velocity, respectively. Resolution of the velocity is slower perpendicular to the anisotropic fabric. Note that an
discussed later in this paper. Figures 3(b)–(d) shows the values NMO correction ignoring the anisotropy effects also leads to
of the parameters domain calculated using the ASRT of the incorrect results. These observations may still help to iden-
diffraction model (figure 3(a)). The location of the data point tify anisotropy in the data (for example by assuming there is
with the maximum value in the parameter domain is found at one dominant direction and then search for the orientation and
i0 = 250, j0 = 250 and vrms = 2000 m s−1 (figure 3(b)), which velocities) by investigating the different behavior of the dif-
are the same as the values for the real model. fracted energy. Such an analysis is, however, beyond the scope
of our paper and will be investigated in the near future. For
2.2.1. Point-like diffractor assumption.  The assumption of a simplicity, we do assume that diffractions are generated from
point-like diffractor as the source of the diffracted energy in point sources but are aware of the issues with this assump-
a homogeneous and isotropic media is an issue. If, however, tion. There are still cases of point-like diffractors that occur in
the point diffraction assumption does not hold (i.e. target fea- nature; for example diffractions from small intrusive bodies,
tures have a dimension about one wavelength: Mie scattering sub-volcanic mounds, and cavities can be considered as gen-
regime), the diffracted arrivals, especially those from dipping erating from point sources.
structures, will be asymmetric around the diffractor (Bohlen
et al 2003, Malehmir et al 2009, 2013, Dehghannejad et al 2.2.2. Iterations.  There are some effective methods to
2012, Bellefleur et al 2012, Cheraghi et al 2013), leading to determine maximum values in the parameter domain. Local
incorrect results. The same can be said where the media is maxima (Shapiro and Stockman 2001) and thresholding
strongly anisotropic. In the case of tilted transverse isotropy (Cooper 2006) are common processes when using the Radon
diffractions will be asymmetric in almost all directions. For transform; however, an iterative form of this method can be
horizontal transverse isotropy diffractions may be symmetric more effective (figure 4). In the first iteration (figure 4(a)),
in both the perpendicular and horizontal orientations to the the location of the maximum value in the parameter domain

203
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

Seismic data
The number of iterations is defined by the user and the proce-
Prestack data dure stops when all iterations are complete.
Now, we show the implementation of the method on shot
gather data from two synthetic models and on a real dataset.
Pre-processing of the seismic data To compare the results with the models, diffraction locations
1. General pre-processing steps are corrected for their NMO in each shot gather using the
2. Extracting instantaneous phase estimated velocities and are stacked together to form a 2D
3. Noise reduction (median filter) section of diffraction locations. No additional velocity infor-
4. Edge enhancement (Gaussian derivative operator) mation is used to produce the sections.

ASRT (iteration #1) 3.  Synthetic data


1. Transforming the data to the parameters domain
2. Extracting the most plausible hyperbola 3.1.  Point-diffraction model

We built a simple 2D model (figure 6(a)) comprised of sev-


Updating the input data eral point diffractors and generated synthetic shots using an
Removing extracted hyperbola
For i = 2 to N

elastic finite-difference code (Juhlin 1995) in Seismic Unix


(Stockwell and Cohen 2008). Here, we only consider the ver-
ASRT (iteration #i) tical component of the data. The background synthetic model
1. Transforming part of the data affected by consists of three layers with average properties of shallow
removed hyperbola to the parameters domain sedimentary rocks. All parameters of the model are shown in
2. Extracting the most plausible hyperbola table 1. The model consists of 2000 by 2000 cells discretized
by 2 m cells. We generated 56 synthetic shots across the model
Figure 5.  Proposed workflow of the 3-parametric hyperbolic Radon with 25 m shot spacing and 1000 receivers spaced at every
transform for diffraction delineating used in this study. 4 m. Synthetic seismograms were generated using a Ricker
wavelet with a 50 Hz central frequency. Thirty small (4 by
(figure 4(a)) specifies the most plausible unknown parameters 4 m) diffractors were randomly positioned within the model to
of equation (5) (i.e. i0, j0 and vrms). Then, the corresponding be delineated using the ASRT method.
hyperbola in the input data is removed from the data using After implementing our method to each individual shot
the same equation (equation (5)) (figure 4(c)). This can sim- gather, diffraction points found in each shot gather were cor-
ply be done by setting all values of the data along the esti- rected for NMO with the corresponding estimated velocities
mated hyperbolic path to zero. A conservative manner of this and then stacked to obtain a 2D section of the point diffrac-
destructive procedure is to select a number ranging from 0 tors. The desired window was then used through the section to
to 0.5. In the second iteration, the ASRT is calculated for the calculate the number of points within it. The final result can
updated data (figure 4(c)). The maximum value is determined be shown as a probability or redundancy section  of the dif-
in this iteration and then the second hyperbola is removed fraction occurrences (figure 6(b)). Cells with high redundancy
from the data. Iterations continue until a number N of dif- match well the locations of the diffractors and are shown by
fractions are found. Removing the hyperbolic-shaped events red circles for display purposes (figure 6(b)). Note the effect
changes the values of a limited part of the input data while of the seismic fold on the results. Diffractions are better delin-
the other parts of the data remain untouched. This means that eated in the middle of the model where the seismic fold is
the ASRT can be calculated for a small part of the parameter much higher compared to the margins of the model. Since the
domain corresponding to the removed part of the input data NMO correction is done using the estimated velocities, the
(shaded data in figure 4(d)) in a subsequent iteration. Accord- overlap of the final results with the diffraction locations con-
ing to figure 4(d), the affected data are dark shaded but we firms the accuracy of the estimated velocities. However, some
consider both dark and light shaded data areas as the target solutions do not accurately match the locations of diffractors,
area since this is easier to calculate in a rectangular window. which is due to errors in the estimated velocities. Figure  7
Because the central point of the window is an apex point, the shows the velocities estimated from the method and the rms-
length and width of the window is equal to a search aper- velocity of the model. The small error in the velocity estima-
ture (App; see figure  4), which is selected by the user, and tion is due to the coarse velocity intervals used in the method.
the height of the hyperbola (Hhyp) on each side of the central A relatively coarse velocity interval (50 m s−1) was selected
point (figure 4(d)). compared with the gradual increase of the real rms-velocity
Figure 5 shows the workflow of the proposed algorithm for (less than 5 m s−1).
automated diffraction delineation using the iterative format of However, the method is not able to predict estimated
the ASRT. The entire procedure involves two main parts: (1) velocity better than 50 m s−1 using a finer velocity interval
pre-processing of the input data containing the general steps since only small change in the aperture of the diffraction
mentioned above, extracting the instantaneous phase of the hyperbola is produced by such a small velocity change. We
seismic data, noise reduction and edge enhancement, and (2) scrutinized this using a simple model with two (low and high)
extracting hyperbolic features using the iterative approach. velocity layers, each of them containing two diffraction points

204
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

(a) Velocity (m/s)


(b) Redundancy
3000 20
100
400 2500 200 15
300

Time (ms)
Depth (m)

800
2000 400 10
1200 500
1500 600 5
1600
700
2000 1000 800 0
400 800 1200 1600 2000 400 800 1200 1600 2000
Distance along the model (m) Distance along the model (m)

Figure 6. (a) 2D point-like diffractor synthetic model used to generate synthetic data. The properties of each layer are found in table 1.
Fifty-six shots spreading from 300 to 1700 mm with 25 m spacing and 1000 receivers with 4 m spacing spreading all along the line were
generated. (b) 2D diffraction section representing the locations of the diffraction points found using the proposed method. The results are
color-coded according to their redundancy in the solution and shown together with the real diffraction points (after converting the depth
section to time) as circles.

Table 1.  Main properties of the point-like diffractor model shown 3.2.  2D BP model-2004 EAGE/SEG dataset
in figure 6.
The 2D BP model has a streamer configuration, comprising a
Layer Thickness (m) Vp (m s−1) Vs (m s−1) Rho (g cm−3) 15 km long array of receivers spaced at every 12.5 m and 50 m
1 500 1800 990 2.3 shot interval (Billette and Brandsberg-Dahl 2005). A total of
2 700 2500 1375 2.4 1340 shots were generated each with 1201 receivers. There
3 800 3000 1650 2.5 were two salt domes in the model, of which, one was selected
in this study to test the method; 250 shot gathers (from shot
at different depths (figure 8). Deeper diffractions have hyper- number 101–350) with 6 s record length were chosen.
bolas with wider apertures/tails than the shallower ones in a The target structure is the top of the salt dome, which is
constant velocity model. This effect is also considered in the expected to generate diffractions in both prestack and unmi-
model. Table 2 shows the properties of the model. The same grated stacked data. The high velocity contrast between the
acquisition geometry as the point diffraction model was used salt and the background, and the roughness of the salt surface,
to generate the seismic data, but just a single shot gather with generate lots of diffractions, implying that it is possible to
the source located on the top of the diffractors was generated find the top of the salt by locating diffractions generated by it.
as shown in figure 9. If the shot depth and dimensions of the Figure 11 shows the locations of diffraction apexes and their
diffractors are ignored, the hyperbolic events at approximately corresponding velocities estimated in a pre-processed phase
400, 600, 1180 and 1270 ms correspond to the point diffrac- image of shot number 200. In this shot gather, 50 hyperbolas
tions. A reflection from the second layer is observed at about (the number of iterations was 100) were detected, of which,
1000 ms and an event at about 800 ms is a multiple of the first only 19 were accepted by the criteria introduced later in the
diffraction. Our method was implemented on these data, with paper. Most of the rejected hyperbolas were reflectors (i.e.
the corresponding results shown in figure  10. As expected, geological layers). Diffraction apexes and velocities for all the
diffraction locations areaccurately predicted. To evaluate shots were obtained and after NMO correction and stacking
the accuracy of the estimated velocities, different meshes were converted from time to depth. High redundant apexes
from fine to coarse intervals of velocity (10, 25, 50, 100 and (>50%) are shown in figure 12, which is the velocity model
200 m s−1) were tested. We observed that, in a low velocity used to generate the BP seismic data (Billette and Brandsberg-
background, small changes in the velocity had a large effect Dahl 2005). In almost all places, our method has been suc-
on the extent of the aperture of the diffraction hyperbola while cessful in locating the rough surfaces on the top of the salt
for a high velocity background large changes in the veloci- body. Since diffraction points are located in a linear pattern
ties had no important impact on the extent of the diffraction and are not well distributed in the model, displaying a velocity
aperture. We conclude from these results that the accuracy of section  was not considered. However, diffraction locations
this method to estimate the velocities at shallow depths with match well with the location of the salt surface, which is an
a low background velocity can be as good as ±50 m s−1, while indicator for the goodness of the estimated velocities.
at greater depths and high background velocity the resolu-
tion decreases dramatically. This decrease in the accuracy 4.  Real data
will obviously have negative effects on the NMO correction
and migration. Therefore, velocity information obtained by The real data used to demonstrate the capability of the method
this method is reliable for diffraction delineation in rather are from a 2D land seismic survey, which was acquired for
low velocity media such as in sedimentary environments or at hydrocarbon exploration. Eighty-five shot gathers with 488
exploration depths and certainly for shallow and near-surface receivers in each shot were selected for this study. Shot
applications such as for ground penetrating radar (GPR) data. and receiver intervals were 45 m and 22.5 mm, respectively,

205
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

(a) (b)
Velocity (m/s) Velocity (m/s)

2400 100 2400


100
200 2300 200 2300
300 300

Time (ms)
Time (ms)

2200 2200
400 400
2100 2100
500 500
2000 2000
600 600
700 1900 700 1900

800 1800 800 1800


400 800 1200 1600 2000 400 800 1200 1600 2000
Distance along the model (m) Distance along the model (m)

Figure 7.  Comparison between the rms-velocity of the point-diffraction model (a) and the estimated velocity section using the method (b).
Note that the estimated velocities are related to just the diffractor points in the model but have been extended for better comparison.

Offset (m)
-600 -200 200 600
0
Depth (m)

200

400

600
Time (ms)

Distance along the model (m)

Figure 8.  A 2D model containing four point-like diffractors used


800
to generate synthetic data to explore the accuracy of the estimated
velocities. The properties of each layer are found in table 2. The
geometry of the seismic survey is the same as the point-diffraction
model shown in figure 6, but this time just a single shot gather data 1000
located on top of the diffraction locations is generated.

Table 2.  Main properties of the four-point diffractor model shown 1200
in figure 10.
Layer Thickness (m) Vp (m s−1) Vs (m s−1) Rho (g cm−3)
1400
1 1000 2000 1100 2.4
2 1000 4500 2475 2.5 Figure 9.  Seismic response and results of the method (red squares)
for the shot generated using the model shown in figure 10. The
hyperbolic events at 400, 600, 1178 and 1267 ms are from the point
diffractions (P1–P4) shown in figure 10. The reflection from the
resulting in a 2D section 3825 m long (figure 13). The ampli- second layer can be seen at 1000 ms and an event at 800 ms seems
tude section  shows a horizontally layered medium gently to be a multiple from the first diffraction. Results suggest that the
dipping to the right-hand side of the section. The section can accuracy of the estimated velocities decreases by increasing depth
and velocity.
be divided into three parts: (1) an upper part (down to about
2000 ms)—horizontal reflections with strong energy; (2) a
features cannot be explained by the overlaid structures and
middle part (from about 2000–2200–2300 ms)—discontin-
require further checking. Note that these results were obtained
uous structures including a channel and some faults and (3)
with no velocity information available.
a lower part (from 2200–2300 ms to the end)—horizontal
reflections with weak energy. Our targets here were the
channel, faults and other discontinuities in the middle part of 5. Discussion
the seismic section.
The proposed method was applied on the shot gather data There are some important issues about our proposed method
and then they were NMO corrected and stacked using the esti- that should be discussed. One particular issue is how the
mated velocities. Final results are also shown in figure 13, a method distinguishes between a reflection and a diffraction.
2D section with a high number of redundant diffraction points Since reflections have a semi-hyperbolic shape and wider
(>50%) superimposed on the corresponding seismic section. aperture, even for dipping reflections, estimated velocities
The channel and some fault locations are detected using the are higher than the rms-velocities of the medium (figure 10).
corresponding diffracted signals. However, some of the other Therefore, velocity estimation can be one method to distinguish

206
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

Offset (m) Velocity (m/s)


-600 -200 200 600 4500
0
4000
1200
200
3500
2400

Depth (m)
400 3000
3600
2500
600
Time (ms)

4800
2000
800
6000 1500
7500 10000 12500 15000 17500
Distance along the model (m)
1000
Figure 12.  A portion of the BP velocity model. The red circles are
highly redundant (>50%) time-to-depth converted apexes found
1200 on the prestack shot gathers but are superimposed on the velocity
model for comparison.

1400
1800
Figure 10.  The results of the proposed ASRT method to find both DC DC F
2000
Time (ms)

diffraction location (solid rectangles) and velocity for the four-point


diffractor model. Note that the accuracy of the estimated velocities
decreases as depth and velocity increase. 2200

CH
DC
2400

F
600
2600
1200 0 1000 2000 3000
Distance along the profile (m)
1800
Figure 13.  2D section of a real seismic data overlaid with
diffraction points (solid circles) obtained by the ASRT method.
2400
(1500) The results match well with target structures such as channel (CH),
Time (ms)

faults (F) and other discontinuities (DC).


3000
NMO correction, no reflection remains in the final results. As
3600 (1700)
a consequence, some diffractions may also be missed by this
(1600) removal. However, since there are many shots (prestack data),
4200 (1700)
the excluded diffractions in one shot can hopefully be cap-
(1700)
4800
tured in other shots. This is also an advantage when prestack
data are used for diffraction delineation using the described
(1900)
5400 (1800) (2000) Radon transform method. It is obvious from figures 6(b), 12
(1900)
(2000) and 13 that the method is capable of distinguishing reflections
(1900) (2100)
6000 from diffraction points.
200 400 600 800 1000 1200 Another issue with the method is the choice of the
Trace number
number of iterations (N). The number of iterations depends
Figure 11.  Pre-processed phase image of shot number 200 of the on the complexity of the media and is determined by the
BP synthetic data and diffraction apexes (solid circles) detected by user based on the target(s) and perhaps some preconceptions
the method. 50 apexes were detected, of which, 19 were accepted about the media. A high N leads to more accurate results,
as the most plausible solutions. The estimated velocity is shown for
but the run time of the method also increases. Therefore, an
each apex.
optimum value can be obtained for N using trial and error.
between diffractions and reflections, but we prefer a simpler As the user defines the number of iterations, the method
way. In shot gather data, points associated with reflections (or generates N number of results, which are the most plausible
multiples) are found at different times, but around the shot hyperbolic shapes. This can become a concern when there
location ( js in equation  (5)). If the points located within a are no diffractions (or as many) in the data, but N solutions
certain range around the shot trace are removed before the are found. In a complex medium with a mixture of events, it

207
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

6. Conclusions
c1
Diffractions contain valuable information, which can be used
a c2 to detect fracture zones, faults, pinch-outs, mineral deposits,
c3 cavities, reef edges and other geological discontinuities.
During the last three decades, various aspects of diffractions
b have been studied to locate their corresponding sources. We
b have presented a procedure for automated diffraction delin-
eation and velocity analysis based on an apex-shifted Radon
transform algorithm. After pre-processing the seismic data,
consisting mainly of extracting the instantaneous phase, noise
reduction and edge enhancement, the method is applied in
Figure 14.  Schematic model demonstrating potential issues that order to characterize the diffraction hyperbolas in an itera-
maybe occurred during the diffraction hyperbola detection using tive manner. We built a synthetic model with randomly dis-
a hypothetical shot gather data. Events marked by (a), (b), and tributed point diffractors and then generated shot gather data.
(c1–c3) are from a reflector, diffraction and predicted hyperbolas,
respectively. All the three predicted diffractions (c1–c3) are false The implementation of our method presented in a redundancy
solutions; (c1) occurs because of high value cells located at the apex section of the diffraction locations showed a good correlation
of the semi-hyperbola shape of the reflector; (c2) occurs because of with the model. The estimated velocities also showed a good
high value cells located on the tail of the reflector hyperbola; correlation with the rms-velocities of the model. We also used
(c3) occurs on the cross-points of the two hyperbolas. the 2D BP salt model to find the location of the salt body using
is inevitable that some diffraction solutions are not true. To this method. Results showed that the rough surfaces of the salt
avoid erroneous interpretations, we suggest some ‘accept or dome, considered as the main source of the diffractions, were
reject’ criteria on the generated results after every iteration well detected. Finally, the method was tested on a real data
to allow exclusion of artifacts. Figure  14 shows problem- example. In this case, channel, faults and some other disconti-
atic situations in a hypothetical shot gather also relevant for nuities were the target structures, most of which were success-
some geological scenarios. Situation ‘(c1)’ in figure  14 is fully detected using this method.
when the apex of a reflection can be wrongly replaced by a While the method has some limitations, it also has some
high velocity diffraction hyperbola. As mentioned before, a potential to be further exploited. Future work may include:
muting procedure with a reasonable range around the shot (1) further improvement of the input data; (2) the design of
trace can be helpful to remove these solutions. Situation some more effective criteria, which can help to exclude incor-
‘(c2)’ in figure 14 happens when high pixel values are pre- rect identification of diffractors in complex geological media;
sent close to the tails of a reflection hyperbola. An operator (3) applying some antialiasing algorithms such as delay-time
that compares the two sides of a predicted hyperbola is useful Radon transform (Wang 2002) and (4) further studies to detect
to reject such solutions. Situation ‘(c3)’ is the cross-point other important subsurface structures such as fracture zones in
of two hyperbolas. This situation is automatically rejected hydrocarbon reservoirs and in anisotropic media.
during the iterative procedure, because all points along the
hyperbola (‘(b)’ in figure  14) turn to zero. However, with Acknowledgments
a reasonable number of iterations and/or by limiting the
velocity search (a priori information), the detection chance SK visited Uppsala University where this research was ini-
of these solutions is reduced. tiated and conducted. He acknowledges the kind technical
The proposed method is one of the first attempts at using support provided by the geophysical staff at the Geophysics
the ASRT to locate diffractions and their corresponding Program, especially those by S Cheraghi and C Juhlin. The
velocities. It is easy to run and does not require any com- authors thank G Kylberg for his valuable comments. We also
plicated computation. Input data (even prestack data) do not thank BP and Frederic Billette for preparing the 2D BP model
need to be heavily processed, except for some routine steps available on the SEG website. An earlier version of this paper
of pre-processing such as static corrections, deconvolution was critically commented on by several anonymous review-
and attenuation of source-generated noise and, more impor- ers, which helped to improve its readability and quality.
tantly, no velocity information is needed. The method took
only 15 s to generate results for a shot gather containing 1000
by 1000 samples; 100 iterations and 100 traces for aperture References
of the hyperbola were used. An Intel® Core™ i7 CPU with
4 GB random access memory (RAM) was used. On the other Al-Dajani A and Fomel S 2010 Fractures detection using multi-
hand, since all the steps are identical for each shot gather, the azimuth diffractions focusing measure: is it feasible?: SEG
algorithm can be parallelized for massive data computing Expanded Abstracts pp 287–91
Bansal R and Imhof M G 2005 Diffraction enhancement in prestack
and testing of parameters. Further studies are recommended seismic data Geophysics 70 V73–9
to exploit the application of the method in anisotropic and Barnes A E 1996 Theory of 2D complex seismic trace analysis
strongly heterogeneous settings. Geophysics 61 264–72

208
J. Geophys. Eng. 12 (2015) 199 S Karimpouli et al

Bellefleur G, Malehmir A and Müller C 2012 Elastic finite- Malehmir A, Bellefleur G and Müller C 2010 3D diffraction and
difference modeling of volcanic-hosted massive sulfide modeconverted scattering signatures of base-metal deposits
deposits: a case study from Halfmile Lake, New Brunswick, Bathurst Mining Camp, Canada First Break 28 41–5
Canada Geophysics 77 WC25–36 Malehmir A and Bellefleur G 2009 3D seismic reflection imaging
Berryhill J R 1977 Diffraction response for non-zero separation of of VHMS deposits, Insights from reprocessing of the
source and receiver Geophysics 42 1158–76 Halfmile Lake data, New Brunswick, Canada Geophysics
Berkovitch A, Belfer I, Hassin Y and Landa E 2009 Diffraction 74 B209–19
imaging by multifocusing Geophysics 74 WCA75–81 Malehmir A, Dahlin P, Lundberg E, Juhlin C, Sjöström H and
Billette F and Brandsberg-Dahl S 2005 The 2004 BP velocity Högdahl K 2011 Reflection seismic investigations in
benchmark EAGE 67th Conference & Exhibition B035 the Dannemora area, central Sweden: insights into the
Bohlen T, Müller C and Milkereit B 2003 Elastic wave scattering geometry of poly-phase deformation zones and magnetite-
from massive sulfide orebodies: on the role of composition skarn deposits J. Geophys. Res. 116 B11307
and shape Hard Rock Seismic Exploration ed B Milkereit Malehmir A, Schmelzbach C, Bongajum E, Bellefleur G, Juhlin J
et al (Tulsa, OK: SEG) pp 70–89 and Tryggvason A 2009 3D constraints on a possible
Burnett W and Fomel S 2011 Azimuthally anisotropic 3D velocity deep >2.5 km massive sulphide mineralization from 2D
continuation Int. J. Geophys. 2011 484653 crooked-line seismic reflection data in the Kristineberg
Canny J A 1986 Computational approach to edge detection IEEE mining area, northern Sweden Tectonophysics
Trans. Pattern Anal. Mach. Intell. 8 679–98 479 223–40
Cheraghi S, Malehmir A, Bellefleur G, Bongajum E and Bastani M Malinowski M, Schetselaar E and White D 2012 3D seismic
2013 Scaling behavior and the effects of heterogeneity on imaging of volcanogenic massive sulfide deposits in the
shallow seismic imaging of mineral deposits: a case study Flin Flon mining camp, Canada: part 2—forward modeling
from Brunswick No. 6 mining area Can. J. Appl. Geophys. Geophysics 77 WC81–93
90 1–18 Nowak E J 2004 Application of Radon transform, stratigrapgic
Cooper G R J 2006 Geophysical applications of the Hough filering and object base stochastic reservoir modeling PhD
Transform South African J. Geol. 109 555–60 thesis Virgina Polytechnic Institute and State University
Dehghannejad M, Malehmir A, Juhlin C and Skyttä P 2012 3D p 66
constraints and finite-difference modeling of massive Radon J 1986 On the determination of functions from their integral
sulfide deposits: the Kristineberg seismic lines revisited, values along certain manifolds, IEEE Trans. on Medical
northern Sweden Geophysics 77 WC69–79 Imaging 5 170–6
Dell S and Gajewski D 2011 Common-reflection-surface-based Shapiro L and Stockman G 2001 Computer Vision (Englewood
workflow for diffraction imaging Geophysics 76 S187–95 Cliffs, NJ: Prentice-Hall)
de Figueiredo J J S, Oliveira F, Esmi E, Freitas L, Schleicher J, Schmelzbach C, Simancas J F, Juhlin C and Carbonell R 2008
Novais A, Sussner P and Green S 2012 Automatic detection Seismic-reflection imaging over the South Portuguese Zone
and imaging of diffraction points using pattern recognition fold-and-thrust belt, SW Iberia J. Geophys. Res.
Geophys. Prospecting 61 368–79 113 B08301
Fomel S, Landa E and Taner M T 2007 Poststack velocity Song X, Wu R and Liu J 2010 Underground diseases identification
analysis by separation and imaging of seismic diffractions of airport runway using GPR PIERS Proc. (Xi'an, China,
Geophysics 72 U89–94 22–26 March 2010) pp 92–6
Foster D J and Mosher C C 1992 Suppression of multiple Stockwell J W Jr and Cohen J K 2008 The New SU User’s Manual
reflections using the Radon transform Geophysics (Golden, CO: Colorado School of Mines)
57 386–95 Taner M T, Koehler F and Sheriff R E 1979 Complex seismic trace
Helgason S 1999 Radon Transform (Boston, MA: Birkhauser) p 190 analyesis Geophysics 44 1041–63
Juhlin C 1995 Finite difference elastic wave propagation in 2D Taner M T, Fomel S and Landa E 2006 Separation and imaging of
heterogeneous transversely isotropic media Geophys. seismic diffractions using plane-wave decomposition SEG
Prospecting 43 843−58 Expanded Abstracts pp 2401–5
Kanasewich E R and Phadke S M 1988 Imaging discontinuities on Thorson J R and Claerbout J F 1985 Velocity-stack and slant-stack
seismic sections Geophysics 53 334–45 stochastic inversion Geophysics 50 2727–41
Karimpouli S, Hassani H, Malehmir A, Nabi-Bidhendi M and Trad D, Ulrych T and Sacchi M 2003 Latest view of sparse Radon
Khoshdel H 2013 Understanding the fracture role on transform Geophysics 68 386–99
hydrocarbon accumulation and distribution using seismic Trorey A W 1970 A simple theory for seismic diffractions
data: a case study on a carbonate reservoir from Iran Geophysics 35 762–84
J. Appl. Geophys. 96 98–106 Trorey A W 1977 Diffractions for arbitrary source-receiver
Klem-Musatov K 1994 Theory of Seismic Diffractions (Tulsa, OK: locations Geophysics 42 1177–82
SEG) Tsingas C, Marhfoul B E, Satti S and Dajani A 2011 Diffraction
Klokov A and Fomel S 2012 Separation and imaging of seismic imaging as an interpretation tool First Break 29 57–61
diffractions using migrated dip-angle gathers Geophysics Wang Y 2002 Antialiasing condition in the delay-time Radon
77 S131–43 transform Geophys. Prospecting 50 665–72
Landa E, Shtivelman V and Gelchinsky B 1987 A method for Wang Y 2003 Multiple attenuation: coping with truncation effect in
detection of diffracted waves on common-offset sections the Radon transformed effect Geophys. Prospecting
Geophys. Prospect. 35 359–73 51 75–87
Malehmir A, Andersson M, Lebedev M, Urosevic M and Windsor C G, Capineri L and Falorni P 2005 The estimation of
Mikhaltsevitch V 2013 Experimental estimation of buried pipe diameters by generalized Hoough transform of
velocities and anisotropy of a series of Swedish crystalline radar data Progress in Electromagnetics Research Symp.
rocks and ores Geophys. Prospect. 61 153–67 (Hangzhou, China, 22–26 August 2005) pp 345–9
Malehmir A, Durrheim R, Bellefleur G, Urosevic M, Juhlin C, Yilmaz O 2001 Seismic Data Analysis: Processing, Inversion,
White D, Milkereit B and Campbell G 2012 Seismic and Interpretation Of Seismic Data (Tulsa, OK: SEG)
methods in mineral exploration and mine planning: Zhu X and Wu R S 2010 Imaging diffraction points using the local
a general overview of past and present case histories and a image matrices generated in prestack migration Geophysics
look into the future Geophysics 77 WC173–90 75 S1–9

209

You might also like