GEOPHYSICS, VOL. 83, NO. 2 (MARCH-APRIL 2018); P. F9–F20, 19 FIGS.

10.1190/GEO2016-0603.1
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

An overview of reproducible 3D seismic data processing and imaging

using Madagascar

Can Oren1 and Robert L. Nowack2

Madagascar is one of the most extensive open-source packages for

ABSTRACT the processing of 2D/3D seismic data sets. The Madagascar soft-
ware package is first briefly introduced. The software is then ap-
We present an overview of reproducible 3D seismic data plied to the observed 3D Teapot Dome seismic data set, which is
processing and imaging using the Madagascar open-source now publicly available. Basic information on the data set, such as
software package. So far, there has been a limited number of location, acquisition parameters, and data files, is provided. Next,
studies on the processing of real 3D data sets using open-source several initial processing steps, such as amplitude gaining, ground
software packages. Madagascar with its wide range of individ- roll attenuation, muting, deconvolution, static corrections, and spike-
ual programs and tools available provides the capability to fully like random noise elimination, are applied to the data. The effects of
process 3D seismic data sets. The goal is to provide a stream- these steps are illustrated using individual gathers, as well as an inline
lined illustration of the approach for the implementation of 3D stacked record section. Velocity analysis is then carried out on se-
seismic data processing and imaging using the Madagascar lected common-midpoint (CMP) gathers in which the normal move-
open-sourcesoftwarepackage.Abriefintroductionisfirstgiven out (NMO) velocities are manually picked from a velocity scan. This
to the Madagascar open-source software package and the pub- is followed by an NMO correction and stack to obtain zero-offset
licly available 3D Teapot Dome seismic data set. Several data. A 3D velocity model is obtained by the application of the
processing steps are applied to the data set, including amplitude following steps: interpolation of the rms velocities, rms-to-interval
gaining, ground roll attenuation, muting, deconvolution, static velocity conversion, and time-to-depth scaling. Then, 3D extended
corrections,spike-likerandomnoiseelimination,normalmove- split-step poststack depth migration is performed to produce migrated
out (NMO) velocity analysis, NMO correction, stacking, and seismic image of the Teapot Dome data set. Finally, f-x deconvolu-
band-pass filtering. A 3D velocity model in depth is created us-
tion and structure-enhancing filtering are carried out to reduce
ing Dix conversion and time-to-depth scaling. Three-dimen-
remaining processing and other noise in the migration image. An on-
sional poststack depth migration is then performed followed
line seismic processing tutorial using Madagascar and other materials
by f-x deconvolution and structure-enhancing filtering of the
for the Teapot Dome data set are also provided by Schleicher (2017).
migrated image to suppress random noise and enhance the use-
In the overview given here, our goal is to provide a complementary
ful signal.We show that Madagascar, as a powerful open-source
and streamlined illustration of the approach.
environment, can be used to construct a basic workflow to proc-
ess and image 3D seismic data in a reproducible manner.
Madagascar software package
Madagascar is designed for computational data analysis on vari-
ous platforms (e.g., Linux, Solaris, macOS, and Windows under the
INTRODUCTION Cygwin environment), and it provides a reproducible research envi-
ronment for researchers (Fomel et al., 2013). The Madagascar soft-
In this study, an overview of 3D seismic processing and imaging ware library consists mostly of modules written in C, but it also
using the Madagascar open-source software package is presented. includes modules written in C++, Python, Fortran-77, Fortran-90,

Peer-reviewed code related to this article can be found at http://software.seg.org/2018/0002.

Manuscript received by the Editor 21 November 2016; revised manuscript received 14 November 2017; published ahead of production 21 November 2017;
published online 10 January 2018.
1
Formerly Purdue University, Department of Earth, Atmospheric, and Planetary Sciences, West Lafayette, Indiana, USA; presently Turkish Petroleum,
Ankara, Turkey. E-mail: [email protected].
2
Purdue University, Department of Earth, Atmospheric, and Planetary Sciences, West Lafayette, Indiana, USA. E-mail: [email protected].
© 2018 Society of Exploration Geophysicists. All rights reserved.

F9
 F10 Oren and Nowack

MATLAB, and Java. The main scripts, referred to as SConstruct 3D Teapot Dome data set
scripts, are written in Python syntax. SCons, a Python-based
make-like utility, is then used to run the SConstruct scripts. We use the 3D Teapot Dome seismic data set to illustrate the seis-
There are four main commands that are used in SConstruct mic processing and imaging results using a selection of programs
scripts. “Fetch” is used for downloading data files either from a available in Madagascar. The Teapot Dome oil field is located ap-
local computer or from a server. “Flow” is used for creating an out- proximately 25 mi north of Casper, Wyoming, where a 3D seismic
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

put file using the library modules that are applied to an input file. survey was acquired by WesternGeco for the Rocky Mountain Oil
“Plot” is analogous to Flow, but the output file is now an output Test Center (RMOTC) in January 2001. A location map within the
plot. “Result” is analogous to Plot, but the output plot is saved state of Wyoming of the Teapot Dome oil field is shown in Figure 1.
in a “Fig” directory that is automatically created by the Result com- A historical presentation of past oil operations at Teapot Dome is
mand. The output plot can then be used within a publication docu- given by Anderson (2013). The data set is now publicly available,
ment. The software is open source and is described by Fomel et al. and it is provided by RMOTC, a facility of the U.S. Department
(2013) and made available at Madagascar (2017). A list of programs of Energy.
within Madagascar can also be found there including their docu- Unprocessed 3D shot gathers, preprocessed CMP gathers, 3D fil-
mentation and a list of reproducible papers in which they have been tered migrated data, 3D postdip moveout velocity text file, and well
previously used. All reproducible papers and their computational logs are available through the SEG Wiki website, SEG Wiki (2017).
recipes (SConstruct files) are also made available to Madagas- The original processing of the data set was performed by EXCEL
car users. Geophysical Services Company. The data set downloaded from the
In addition to Madagascar, Seismic Unix (SU) maintained by website is in SEGY format, and it first needs to be converted to the
the Colorado School of Mines and SEPlib maintained by the Stan- regularly sampled format (RSF) using the program “sfsegyread” in
ford Exploration Project are among the other popular open-source Madagascar in which RSF is the data format used by Madagascar
software packages for reflection seismology. Although SU is a programs.
strong processing platform for smaller-scale 2D seismic data sets, For the original acquisition of the 3D Teapot Dome seismic data
it provides only a limited range of processing and imaging tools set by WesternGeco, a 1200 channel I/O System II recording system
for larger scale 3D seismic data sets compared with Madagascar. and four AVH III392 vibrators were used. The geophones were
SEPlib also provides a convenient environment for 3D data process- deployed with a group interval of 67 m along each line and a line
ing and imaging, but it is not as accessible as Madagascar in terms spacing of 268 m. The sources were deployed with an interval of
of reproducibility. Therefore, from a reproducible 3D seismic 67 m and a line spacing of 670 m. The bin size was determined to be
processing vantage point, Madagascar is a modern environment 33.5 × 33.5 m (110 × 110 ft).
and allows for dynamic development.
PROCESSING OF TEAPOT DOME
DATA SET
The shot and group coordinates are shown in
Figure 2, and the geometry information was
taken from the header file of the shot gathers. In
addition, a specification sheet of the recording
parameters is provided with the data set. The lo-
cation of the active receivers for two different
shots is displayed in Figure 3, in which the shot
and active receiver locations are shown for shot
indices 214 and 825. Figure 4 shows the fold
map with a maximum fold of 57. To reproduce
the plots in Figures 2–4, type the commands
below:
scons geom/sxsygxgycoord.view
scons geom/sxsygxgycoord12.view
scons geom/fold.view.
The raw shot gathers are used to show the
features of several programs in Madagascar, in-
cluding amplitude gain applications, ground roll
attenuation, muting, deconvolution, and static
corrections. Figure 5 shows a portion of shot
gather 214 without any processing. Three of the
17 receiver lines indicated in Figure 3a are plot-
ted. To reproduce the plot in Figure 5, type the
command below:
Figure 1. Location map of the Teapot Dome oil field, Wyoming. The Teapot Dome oil
field is also known as Naval Petroleum Reserve No. 3 (NPR-3) (adapted from Li, 2014). scons ampcor/fldr_214.view.
 3D seismic processing using Madagascar F11

The data are then sorted into CMP gathers to perform spike-like roll is shown. For the LTF decomposition, we have selected the
noise elimination, NMO velocity analysis, NMO correction, stack- smoothing radius (rect) to be 20 samples in time, the number of
ing, band-pass filtering, migration, f-x deconvolution, and structure- frequencies (nw) to be 200, and the frequency step (dw) to be
enhancing filtering. To use in the migration step, a velocity model is 0.5 Hz. After the decomposition, a 2D soft thresholding with a clip
generated using Dix conversion and time-to-depth scaling. This of 25% is applied to estimate the ground roll energy (Figure 8b).
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

processing sequence takes approximately 16 h

on a Mac computer with 32 GB of memory
and a 4 GHz Intel i7 processor. a) b)

Amplitude gain applications

To improve the amplitude visibility of the later
arrivals, different gain operators can be applied to
the shot gathers, including a time-power ampli-
tude-gain correction (tpow ) and automatic gain
control (AGC). Receiver line 2 from Figure 5
is shown in Figure 6a, and it is selected to
compare different amplitude gain functions in
Madagascar. Figure 6b shows the result of a
time-power correction on receiver line 2 on a
common-shot profile, and it applies a gain func-
tion in time to increase the amplitudes of the later
arrivals. In this example, the program “sfpow” is
used with a power of two.
AGC can also be used to increase the ampli-
tude levels of weak reflections, and it is applied
to single traces to equalize the amplitudes along
each trace by a sliding window of fixed length
(Yilmaz, 2001). The selection of the window size
is important and needs to be chosen with care Figure 2. (a) Shot and (b) group receiver geometry for the Teapot Dome data set. The
because relatively shorter windows tend to intro- distance on the horizontal axis (x) is approximately 6.27 km, and the distance on the
vertical axis (y) is 11.53 km. This figure can be created by typing scons geom/sxsygx-
duce unwanted artifacts, particularly when used gycoord.view.
for visualization purposes. In Figure 6c, an AGC
with a smoothing radius of 500 samples in time is
applied using the program “sfagc.” The artifacts at a) b)
earlier times (above first breaks) introduced by the
AGC can be removed by muting the data. To
reproduce the plots in Figure 6, type the command
below:

scons ampcor/fldr214_all.view.

Ground roll attenuation

Ground roll is a type of coherent noise that
mostly appears in land seismic surveys. In
general, it has a dispersive low-frequency and
high-amplitude character that degrades the data
quality. Here, a local time-frequency (LTF) de-
composition method (Liu and Fomel, 2013) is
applied to the shot gathers to attenuate its dis-
ruptive effects. In this method, a forward LTF
decomposition is applied to each trace of a gather
to form a 3D time-frequency cube.
Note that the unprocessed shot gather in Fig-
ure 6a is shown in Figure 7a with a clip percent-
age of pclip = 98 to better illustrate the ground Figure 3. (a) Shot (in red) and active receiver (in blue) locations for shot index 214.
Specific receiver lines in (a) are noted by 1, 2, and 3. (b) Shot (in red) and active receiver
roll contamination on the near-offset traces. In (in blue) locations for shot index 825. The distance on the horizontal axis (x) is approx-
Figure 8a, the LTF decomposition of the data in imately 6.27 km, and the distance on the vertical axis (y) is 11.53 km. This figure can be
Figure 7a with the dominant effect of the ground created by typing scons geom/sxsygxgycoord12.view.
 F12 Oren and Nowack

An inverse transform from the time-frequency domain back to the linear airwave arrival in the unprocessed shot gather (Figure 7a) is
time domain is then performed to subtract the estimated ground roll also estimated with the application of soft thresholding (Figure 7b).
in Figure 7b from the raw shot gather in Figure 7a. Note that a t2 To reproduce the plots in Figures 7 and 8, type the commands
gain correction is also applied to the unprocessed data prior to sub- below:
traction. The shot gather after ground roll attenuation and its LTF
scons groll/ltftalltd.view
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

decomposition are shown in Figures 7c and 8c, respectively. Note

scons groll/ltftallfd.view.
that the reflections become more obvious after reducing the effects
of the ground roll energy. In addition to the ground roll, a coherent For the application to the whole data set, several of the parameters
for the ground roll elimination are changed (nw is reduced from 200
to 50 and dw is increased from 0.5 to 2) to reduce the computation
time. However, the results are not significantly affected by these
changes. The procedure is first applied on each shot gather to es-
timate the ground roll. The estimated ground roll is then subtracted
from the entire data set.

Muting
In reflection seismology, muting can be applied to zero out the
undesirable part of the data, such as direct arrivals and wide-angle
reflections. Muting can be performed by first picking the time and
offset pairs on the prestack data. The selected time and offset pairs
can then be used either in the program “sftahmute” or in the pro-
gram “sfmutter” in Madagascar. In Figure 9a, a shot gather is shown
after applying an AGC with a window of 1000 ms to the shot gather
shown in Figure 7c and a mute function to remove the unwanted
part of the prearrival data and direct wave. In this example, the time
(tmute = 0, 1.20, 1.82 s) and offset (xmute = 0, 8560, 18,000 ft)
pairs are used with an 80 point taper. As shown in Figure 9a, the
early portion of the data is eliminated using the mute.

Deconvolution (prediction error filtering)

Predictive deconvolution removes the unknown source wavelet
and the predictive component of the signal, such as multiples and
Figure 4. The fold map of the Teapot Dome data set in which the reverberations in the seismograms. It also yields an output seismo-
horizontal axis is the crossline number and the vertical axis is the gram with a higher temporal resolution. Here, deconvolution is per-
inline number. This figure can be created by typing scons geom/ formed by applying a prediction error filter to the shot gathers using
fold.view. the program “sftahpef” in Madagascar. The prediction error filter is
designed from the autocorrelation of traces. A unit prediction lag
converts the predictive deconvolution into spiking deconvolution
(Yilmaz, 2001). The maximum lag of the filter and the autocorre-
lation window are other parameters that need to be specified for the
deconvolution. Spiking deconvolution is based on the assumption
of a minimum phase wavelet, and the wavelet is statistically esti-
mated and then removed from the seismograms. The shot gather
in Figure 9a after spiking deconvolution using the program “sftah-
pef” is shown in Figure 9b in which the temporal resolution is
shown to be increased. The program “sftahpef” applies a single-
trace deconvolution rather than a surface-consistent one. In this
example, a prediction gap of 0.002 s (the sampling interval), a maxi-
mum lag of 0.14 s, and a total autocorrelation window of 3 s are
used. Note that the total autocorrelation window is longer than the
selected maximum lag of the prediction error filter.

Static corrections
One of the crucial problems for seismic land data is static time
Figure 5. Part of shot gather 214 is shown without any processing. shifts. For land-based data, the complex near-surface velocity var-
See Figure 3 for the location of receiver lines 1, 2, and 3 for this dis- iations or rapid changes in shot and receiver elevations tend to
play. This figure can be created by typing scons ampcor/fldr_214.view. cause reflection traveltime distortions. Because of these static shifts,
 3D seismic processing using Madagascar F13

prestack reflections often do not exhibit a hyperbolic behavior. is also windowed to include 26 traces to test the TVMF method.
Therefore, corrections for near-surface statics need to be made prior In Figure 10a, the reflections from 0.9 to 1.2 s are obscured by
to any further processing steps. The static corrections can be deter- the presence of high-frequency random noise. To attenuate the
mined from common-shot and common-receiver gathers using high-frequency random noise, the reference median window length
first-arrival refraction analysis (Yilmaz, 2001). In Madagascar, (nfw) is selected to be seven points, and the default values are used
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

the program “sftahstatic” is then used to apply the static corrections. for the rest of the parameters (α ¼ 2, β ¼ 0, γ ¼ 2, and δ ¼ 4). We
The source and receiver refraction statics (in milliseconds) found these values to be sufficient to attenuate the random, spike-
have been previously computed by the con-
tractor and included in the header files of the shot
a) b) c)
gathers. To adjust for the misalignments, static
information is obtained from the header files of
the data and used in the following equation:
tout ¼ tin − sign  ðsstat þ gstat Þ, where tout is the
time of the output trace, tin is the time of the input
trace, sstat is the source static time correction, gstat
is the group static time correction, and sign is a
scale factor (1) that controls the option of shift-
ing the traces up or down. The shot gather shown
in Figure 9a after the application of spiking de-
convolution and static corrections is shown in
Figure 9c. The hyperbolas can be seen to become
more coherent with repositioning based on the
static time shifts. Applying ground roll attenua-
tion, deconvolution, and statics corrections takes
approximately 12 h on a Mac computer with
32 GB of memory and a 4 GHz Intel i7 processor.
Most of the runtime is in the ground roll attenu-
ation. To reproduce the plots in Figure 9, type the
command below:
Figure 6. A comparison between various amplitude gain methods. (a) The unprocessed
scons deconstat/agcmutepefstaticshot.view. receiver line 2 from Figure 5. (b) The application of tpow with a power of two. (c) The
application of AGC with a window of 500 samples in time. This figure can be created by
typing scons ampcor/fldr214_all.view.
Random spike-like noise attenuation
using time-varying median filter
Spike-like noise present in seismic data tends to a) b) c)
lower the signal-to-noise ratio (S/N). High-cut fil-
ters can be applied to eliminate this type of noise,
but there is also the possibility of removing the
useful signal from data at high frequencies. To
overcome this issue, a 1D time-varying median
filter (TVMF) aims to suppress random, spike-like
noise in the CMP domain (Liu et al., 2009). In this
method, a threshold value is first calculated based
on the reference median window length given by
the user. To better distinguish between the useful
signal and noise, time-varying filter windows are
then designated using four constant even numbers
(α, β, γ, and δ). In essence, time-varying filter win-
dows are combinations of the reference filter win-
dow and these four constant numbers. The main
principle of this method is that the time-varying
filter windows need to be small to preserve the
useful signal, whereas the filter windows need
to be large to eliminate the noise. Here, our goal
is to apply this method to the CMP gathers prior to Figure 7. (a) The shot gather shown in Figure 6a with clip percentile of pclip = 98 to
velocity picking to obtain more accurate sem- better show the low-frequency, high-amplitude ground roll energy in the data. (b) The
estimated ground roll from using the LTF decomposition for the gather shown in (a).
blance panels. (c) The shot gather after ground roll attenuation obtained by subtracting (b) from
A CMP gather is extracted for the midpoint (a) where the direct waves and refractions are also muted. This figure can be created
location of crossline 120 and inline 160, and it by typing scons groll/ltftalltd.view.
 F14 Oren and Nowack

like noise. Once we apply a larger window length using the default NMO velocity analysis
values for the rest of the parameters, we start to lose some use-
ful signal. In Figure 10b, the reflections starting from 0.9 to It is important to build an accurate subsurface velocity model for
1.2 s become more evident and the eliminated noise can be seen stacking and migration imaging to be able to obtain flat CMP and
in Figure 10c. To reproduce the plots in Figure 10, type the com- image gathers. One of the ways to build an initial velocity model is
to pick maximum coherency peaks as velocity-time pairs from a
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

mand below:
velocity scan, which are derived from selected CMP gathers (Taner
scons tvmf/gather_all.view. and Koehler, 1969). These velocities are called root-mean-squared
(rms) velocities and are used for NMO correction
of reflection traveltimes prior to stacking. In Ma-
a) b) c) dagascar, rms velocities can be picked either man-
ually or automatically from the velocity scan.
Interactive picking can be performed using the
program “sfipick,” whereas automatic picking
can be performed using the program “sfpick.”
Although automatic picking would be useful in
terms of reproducibility, the velocities have been
picked manually from the velocity scans every
25th CMP gather along the inline and crossline
directions. Figure 11 shows 51 CMP locations
where the velocity picking has been performed.
To reproduce the plot in Figure 11, type the com-
mand below:
scons vel/xlil.view.

The program “sfvscan” is used to generate

velocity scans, and the first scanned velocity
(v0), the step in velocity (dv), the number of
scanned velocities (nv), and the semblance aver-
Figure 8. (a) The LTF spectra of the raw shot gather in Figure 7a. (b) LTF spectra of the aging (nb) are among the main parameters. In this
estimated ground roll in Figure 7b. (c) LTF spectra of the shot gather after ground roll application, these parameters have been selected
attenuation in Figure 7c. For these flat cube plots, the top, side, and front frame numbers to be 9000 ft∕s, 100 ft∕s, 101, and 10, respec-
are selected to be 0.75 s, 540 Hz, and 20 Hz, respectively. This figure can be created by tively. Full-offset CMP gathers are used as input
typing scons groll/ltftallfd.view.
to generate semblance panels. An auxiliary offset
file is required to define the irregular trace off-
sets. Using Python for loop in the SConstruct
a) b) c) script, 1D velocity functions have been picked
at 51 different CMP locations. For each CMP,
the program generates a text file that includes
corresponding picked time-velocity pairs. Using
a Python code, velfile.py, all the separate text
files related to the time-velocity pairs along with
their CMP numbers are then concatenated to be
used as an input to the interpolation code. The
Python code, velfile.py, and the concatenated
velocity text file, rms_vel.txt, are provided with
the SConstruct script.
The CMP gather with TVMF applied in Fig-
ure 10b is used to illustrate the application of the
velocity scan and NMO correction (Figure 12a).
Figure 12b shows the velocity scan obtained us-
ing the program “sfvscan” for the CMP gather
shown in Figure 12a and the manually picked
rms velocities shown by the solid line. The
NMO correction is then carried out using the
program “sftahnmo,” and Figure 12c shows
Figure 9. (a) The shot gather shown in Figure 7c after AGC (1000 ms) and mute are
applied, (b) the shot gather in (a) after spiking deconvolution is applied, and (c) the shot the NMO-corrected CMP gather in which the
gather in (b) after static corrections are applied. This figure can be created by typing scons nonzero offset traveltimes are mapped onto
deconstat/agcmutepefstaticshot.view. the zero-offset traveltimes after NMO correc-
 3D seismic processing using Madagascar F15

tion. To reproduce the plots in Figure 12, type the com- program “sfspline.” Then, several 2D interpolation programs such as
mand below: “sfextract,” “sftri2reg,” “sfenoint2,” and “sfinvbin” can be used to
interpolate the data points on each time slice and to obtain a 3D
scons nmostack/gathervscanpicknmo.view. velocity cube.
The 1D velocity-time functions are then interpolated to build a In many of the migration algorithms in Madagascar, an interval
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

3D rms velocity cube that will be used for further processing and velocity model is needed as an input. Interval velocities represent
imaging steps, such as NMO correction and poststack migration. the velocities within each layer and can be estimated from the rms
To illustrate the use of an external code into the processing flow, velocities using Dix’s formula (Dix, 1955). Figure 13 demonstrates
we use a Python code, interpvel.py, which is located in the a 3D interval velocity model for the Teapot Dome data set in which
$RSFSRC/book/data/teapotdome/vels directory for the interpola- the interval velocities are estimated using the program “sfdix.” This
tion of the velocities. We have slightly modified the code to program converts rms velocities to interval velocities using least-
change the time axis from milliseconds to seconds and the regular squares and shaping regularization. For the parameters, a maximum
expression pattern for reading the time-velocity pairs from the of 100 iterations are performed during the inversion part and a
given velocity file. The modified version of the interpolation code smoothing radius is selected to be 15 samples on the time axis. To
is provided with the SConstruct script. The time axis is first lin- reproduce the 3D interval velocity cube in Figure 13, type the com-
early interpolated. The interpolation is then carried out on each mand below:
time slice using a radial basis function interpolator in Python. scons vel/vint3d.view.
To be able to use both interpolation functions, Python’s scipy li-
brary should be installed. Also note that the provided velocity text A depth representation of the interval velocity model that is
file and interpolation code need to be copied to the “vel” directory needed for depth migration can be obtained using the time-to-depth
before running the command below. Although the interpolation is conversion program “sftime2depth.” The transformation is con-
performed using an external Python code, there are several 1D/2D trolled by depth sampling (dz ¼ 30 ft) and number of depth sam-
interpolation programs in Madagascar. For instance, 1D velocity ples (nz ¼ 1200). The depth model is then converted to slowness to
functions can be first interpolated using the cubic spline interpolation be used as an input for the zero-offset depth migration.

a) b) c)

Figure 10. (a) Original CMP gather extracted for the midpoint of crossline 120 and inline 160, (b) denoised gather using seven-point TVMF,
and (c) eliminated noise. This figure can be created by typing scons tvmf/gather_all.view.
 F16 Oren and Nowack

Stacking Migration
Stacking is an effective tool to improve the S/N and to eliminate Migration yields a seismic image by collapsing diffractions and
the ground roll in the seismic data. It is also used for compressing moving dipping structures to their correct positions. Here, extended
the data in the offset domain, and it is a required step for poststack split-step Fourier migration is implemented (Kessinger, 1992). It is
migration imaging. After applying NMO correction using the 3D a 3D poststack depth migration method that is based on the solution
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

rms velocity cube, the CMP gathers can be stacked using the pro- of the one-way wave equation. This method also has the advantage
gram “sftahstack” to obtain the zero-offset data. A stacked 3D data of handling laterally varying velocity changes. The 3D extended
cube for the Teapot Dome data set after band-pass filtering applied split-step migration is performed using the program “sfzomig3” in
between 12 and 90 Hz is shown in Figure 14. Prominent reflectors Madagascar. Because the migration is carried out in the frequency
can be observed between 0.8 and 1.2 s on the crossline and inline domain, the Fourier transform of the band-pass filtered poststack
sides of the data cube. To reproduce the 3D stacked cube in Fig- data is first computed. The Fourier-transformed data are filtered
ure 14, type the command below: up to 50 Hz before running the migration algorithm. The Fourier-
transformed data and the slowness model are the inputs for the mi-
scons nmostack/npr3_stack.view. gration. Because the output of the migration is given in depth slices,
To further investigate the impact of the processing steps, such as the resulting migrated data cube needs to be transposed back to the
spiking deconvolution and static corrections, a subset of the stacked conventional coordinate system. Figure 17 depicts the 3D poststack
data is examined. Figure 15 shows the inline 225 stacked section after depth-migrated image of Teapot Dome data set. The migrated data
t2 gain correction, muting, and AGC with a window of 1000 ms. in Figure 17 show the expected improved fault resolution when
Figure 16 shows the inline 225 stacked section after t2 gain correc- compared with the unmigrated stack in Figure 14. The 3D poststack
tion, muting, AGC with a window of 1000 ms, spiking deconvolu- depth migration takes approximately 2.5 h on a Mac with 32 GB of
tion, and static corrections. After spiking deconvolution and static memory and a 4 GHz Intel Core i7 processor. To reproduce the mi-
corrections, the reflectors in the stacked section in Figure 16 become grated data in Figure 17, type the command below:
more coherent and have a higher temporal resolution than that of scons mig/ssfmigz.view.
Figure 15. To reproduce the 2D stacked sections in Figures 15
and 16, type the commands below: After examining the migration result shown in Figure 17, it can
be seen that there exists some processing and other noise that dis-
scons deconstat/muteagcstack225.view torts the migration images even though the prominent reflectors are
scons deconstat/mutepefagcstaticstack225.view. clearly imaged by the migration algorithm. A partial remedy for this
is to apply f-x deconvolution to the migrated images to remove any
remaining spurious noise.

Random noise attenuation by f-x deconvolution

The migration results generally include residual noise and
processing artifacts, which could make the interpretation more dif-
ficult. Hence, the noise in the migration images needs to be attenu-
ated to achieve a clearer image. In Madagascar, frequency-space
(f-x) deconvolution, which is a prediction technique applied in
the f-x domain, can be applied to eliminate the noise using the pro-
gram “sffxdecon.” In this technique, small temporal and spatial win-
dows of input data are first Fourier transformed to the f-x domain.
A deconvolution operator is used to predict the coherent part of the
data in the lateral domain. The predicted coherent data are then sub-
tracted to isolate the random noise component. Finally, the random
noise is inverse Fourier transformed and then subtracted from the
data. This procedure is repeated with overlapping windows of the
data (Gulunay, 1986).
Figure 18 shows the extended split-step migration image after f-x
deconvolution was applied three times with default parameters to
the image shown in Figure 17 in which the random noise is attenu-
ated and the prominent reflectors are preserved. The fault structures
in the extended split-step image are well-migrated and more evident
in the f-x deconvolution result. We have found the default param-
eters (number of traces for filter lenf = 4 and number of traces in
Figure 11. The blue plus signs indicate the CMP locations where the window n2w = 10) to be sufficient in the application. To reproduce
rms velocities have been manually picked. Picking has been per- the f-x deconvolution result in Figure 18, type the command below:
formed every 25th CMP location along the inline and crossline di-
rections. This figure can be created by typing scons vel/xlil.view. scons mig/ssfmigz_fx.view.
 3D seismic processing using Madagascar F17

Structure-enhancing filtering need to be designated for the application of the filter. In this exam-
ple, smoothing (smnclip) and sharpening (shnclip) parameters are
Attenuation of random noise while preserving the structural in- chosen to be 13 samples. According to the default relationship be-
formation in migration images is very crucial. Structure-enhancing
filtering can be used to enhance the quality of the seismic images
and is a nonstationary smoothing operator. For this purpose, we first
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

construct a structural prediction of the seismic events and then apply

lower-upper-middle (LUM) filtering (Liu et al., 2010). Even though
we applied f-x deconvolution to the poststack depth-migrated im-
age, there still exists some remaining random noise in the image in
Figure 18.
Structure prediction comprises two main steps. First, a local dip
(slope) estimate is obtained by plane-wave destruction (Fomel,
2002). Second, each seismic trace is predicted from its neighbor
traces based on the local slope by adding an extra prediction dimen-
sion. A nonlinear structure-enhancing LUM filter is then applied
along the prediction direction to obtain the denoised image. Hence,
we initially use the program “sfdip” for 3D dip estimation along the
inline and crossline directions. A dip smoothness (rect#) of 10 sam-
ples is used for all dimensions. Then, we apply a 3D plane-wave
spray (“sfpwspray2”) to form the structure prediction volume by
choosing a total of four prediction steps (two on the left and right
sides) for parameters ns2 and ns3 around the reference trace. Figure 13. A 3D interval velocity model in time for the Teapot
The LUM filter is defined as a very efficient nonlinear filter, Dome data set. This plot displays selected sections as the faces of
the cube. For this cube plot, the top, side, and front frame numbers
which can be used to attenuate random noise in seismic images are selected to be 0.65 s, 225, and 125, respectively. This figure can
(Liu et al., 2010). Two parameters for smoothing and sharpening be created by typing scons vel/vint3d.view.

a) b) c)

Figure 12. (a) The CMP gather with TVMF applied in Figure 10b, (b) the velocity scan derived from the CMP gather in (a) and the solid line
shows the manually picked velocities, and (c) the CMP gather in (a) after NMO correction. This figure can be created by typing scons nmostack/
gathervscanpicknmo.view.
 F18 Oren and Nowack

tween these two parameters and the filter length (smnclip = shnclip The SConstruct script creates 10 separate directories (fetch, geom,
= (nfw + 1)/2) defined in the program “sflum” in Madagascar, a ampcor, groll, deconstat, tvmf, vel, nmostack, mig, and Fig). The
filter window length (nfw) of 25 samples is selected. In Figure 19, “fetch” directory contains raw prestack SEGY files downloaded
it can be seen that these parameters work well to attenuate most of from the Internet and their internal RSF versions with the header
the random noise and protect structures compared with the image in files. The “geom” directory has all shot and receiver coordinates,
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 18. To reproduce the image after structure-enhancing filter- single shots 214–825 and the corresponding receiver locations as
ing in Figure 19, type the command below: well as fold information. The “ampcor” directory comprises a por-
tion of shot 214 and the amplitude gain comparisons (tpow and
scons mig/ssfmigz_fx_lum.view.
AGC) applied to receiver line 2 from shot 214. The “groll” directory
contains the same receiver line and its LTF spectra before and after
ground roll elimination. The full shot gathers after ground roll elimi-
PROCESSING SCRIPTS
A single SConstruct script along with two Python codes (velfile.
py and interpvel.py) and a velocity file (rms_vel.txt) are provided.

Figure 16. Inline 225 stacked section after t2 gain correction, mut-
ing, AGC with a window of 1000 ms, spiking deconvolution, and
static corrections. This figure is compared with Figure 15 to observe
Figure 14. The stacked 3D cube for the Teapot Dome data after the improvements due to deconvolution and static corrections. This
applying a band-pass filter between 12 and 90 Hz. This plot displays figure can be created by typing scons deconstat/mutepefagcstatic-
selected sections as the faces of the cube. For this cube plot, the stack225.view.
top, side, and front frame numbers are selected to be 0.65 s, 225,
and 125, respectively. This figure can be created by typing scons
nmostack/npr3_stack.view.

Figure 17. A 3D poststack depth-migrated image of the Teapot

Dome data set using the extended split-step method. This plot dis-
plays selected sections as the faces of the cube. For this cube plot, the
Figure 15. Inline 225 stacked section after t2 gain correction, mut- top, side, and front frame numbers are selected to be 3510 ft, 225, and
ing, and AGC with a window of 1000 ms. This figure can be created 125, respectively. Compare to Figure 14 to see the effect of migration
by typing scons deconstat/muteagcstack225.view. process. This figure can be created by typing scons mig/ssfmigz.view.
 3D seismic processing using Madagascar F19

nation will also be in this directory. The “deconstat” directory com- interpvel.py) the velocity text file (rms_vel.txt) downloaded with
prises deconvolution and refraction statics examples applied to the this manuscript should be located in the vel directory. The “nmo-
same receiver line along with the inline 225 stacked section. It also stack” directory includes the same CMP gather, its velocity scan,
has the CMP gathers after deconvolution and statics applied at NMO corrected gather, and the 3D stacked cube for input to depth
which the velocity picking has been performed. The “tvmf” migration. The “mig” directory contains the 3D depth-migrated data
before/after f-x deconvolution and structure-enhancing filtering. Fi-
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

directory has the TVMF application implemented to a CMP gather

and also contains the amplitude spectra of a trace from the gather. nally, the “Fig” directory contains eight directories (except fetch)
The “vel” directory comprises the velocity analysis CMP loca- with the same name in which the final reproducible figures used in
tions, 3D rms velocity cube, 3D interval velocity cube in time this document are stored.
and depth, and 3D slowness cube used as an input to depth migra-
tion. It is important to note that two Python codes (velfile.py and
CONCLUSION
In this study, an overview of reproducible 3D seismic data process-
ing and imaging using Madagascar open-source software package
has been presented. The processing and imaging steps are applied
to a real land 3D data set, Teapot Dome, which is publicly available.
To illustrate the use of Madagascar, several processing steps, such as
amplitude gaining, ground roll attenuation, muting, spiking deconvo-
lution, static corrections, spike-like random noise elimination, NMO
velocity analysis, NMO correction, and band-pass filtering are ap-
plied to the data. Dix conversion and time-to-depth scaling are used
to form a 3D velocity model in depth. The zero-offset depth migra-
tion is then computed to obtain a seismic image of the data set. In the
migrated image, random noise is attenuated using f-x deconvolution
and structural continuity is improved using structure-enhancing filter-
ing. The processing script (SConstruct) along with the velocity file
and interpolation code is provided for training purposes in reproduc-
ible research, and readers are also encouraged to improve on the re-
sults. The processing framework will be useful to researchers who
Figure 18. A 3D seismic image after applying f-x deconvolution to wish to test new processing programs.
the image shown in Figure 17. This plot displays selected sections
as the faces of the cube. For this cube plot, the top, side, and front
frame numbers are selected to be 3510 ft, 225, and 125, respec- ACKNOWLEDGMENTS
tively. Compare to Figure 17 to see the effect of f-x deconvolution.
This figure can be created by typing scons mig/ssfmigz_fx.view. We greatly acknowledge the associate editor Y. Liu and reviewer
K. Schleicher for their valuable help in improving the manuscript.
We would like to thank three anonymous reviewers for their con-
structive comments and suggestions. We would also like to thank
the developers of the Madagascar open-source software package
and the organizers of the 2015 SEG 3D Seismic Processing Work-
ing Workshop for inspiring us to write this overview. We also thank
S. Fomel for his encouragement to submit this study. We acknowl-
edge RMOTC and the U.S. Department of Energy for making the
Teapot Dome 3D seismic data available. C. Oren acknowledges
Turkish Petroleum for the financial support during his graduate
studies at Purdue University.

REFERENCES
Anderson, T., 2013, History of geologic investigations and oil operations at
Teapot Dome, Wyoming: AAPG Rocky Mountain Section Meeting.
Dix, C. H., 1955, Seismic velocities from surface measurements: Geophys-
ics, 20, 68–86, doi: 10.1190/1.1438126.
Fomel, S., 2002, Applications of plane-wave destruction filters: Geophysics,
67, 1946–1960, doi: 10.1190/1.1527095.
Figure 19. Structure-enhancing result using LUM filtering. This Fomel, S., P. Sava, I. Vlad, Y. Liu, and V. Bashkardin, 2013, Madagascar:
plot displays selected sections as the faces of the cube. For this cube Open-source software project for multidimensional data analysis and
reproducible computational experiments: Journal of Open Research
plot, the top, side, and front frame numbers are selected to be 3510 Software, 1, e8, doi: 10.5334/jors.ag.
ft, 225, and 125, respectively. Compare to Figure 18 to see how the Gulunay, N., 1986, Fxdecon and complex wiener prediction filter: 56th
remaining random noise is attenuated while preserving the geologic Annual International Meeting, SEG, Expanded Abstracts, 279–281.
structures. This figure can be created by typing scons mig/ Kessinger, W., 1992, Extended split-step Fourier migration: 62nd Annual
ssfmigz_fx_lum.view. International Meeting, SEG, Expanded Abstracts, 917–920.
 F20 Oren and Nowack

Li, Y., 2014, An uncertainty analysis of modeling geologic carbon seques- Madagascar open-source software project, 2017, http://www.ahay.org/, ac-
tration in a naturally fractured reservoir at Teapot Dome, Wyoming: Ph.D. cessed 1 November 2017.
thesis, University of Wyoming. Schleicher, K., 2017, http://www.reproducibility.org/RSF/book/data/teapotdome/
Liu, Y., and S. Fomel, 2013, Seismic data analysis using local time- paper.pdf, accessed 1 November 2017.
frequency decomposition: Geophysical Prospecting, 61, 516–525, doi: SEG Wiki, 2017, https://wiki.seg.org/wiki/Open_data, accessed 1 Novem-
10.1111/j.1365-2478.2012.01062.x. ber 2017.
Liu, Y., S. Fomel, and G. Liu, 2010, Nonlinear structure-enhancing filtering Taner, M. T., and F. Koehler 1969, Velocity spectra — Digital computer der-
Downloaded 07/09/18 to 128.119.168.112. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

using plane-wave prediction: Geophysical Prospecting, 58, 415–427, doi: ivation and applications of velocity functions: Geophysics, 34, 859–881,
10.1111/(ISSN)1365-2478. doi: 10.1190/1.1440058.
Liu, Y., C. Liu, and D. Wang, 2009, A 1-D time-varying median filter for Yilmaz, O., 2001, Seismic data analysis: SEG.
seismic random, spike-like noise elimination: Geophysics, 74, no. 1,
V17–V24, doi: 10.1190/1.3043446.