Sie sind auf Seite 1von 6

Spectral decomposition of 3D ground-penetrating radar

data from an alluvial environment


ISABEL C. GEERDES, Chevron Energy Technology Company, Houston, USA
ROGER A. YOUNG, University of Oklahoma, Norman, USA

T his study uses spectral decomposition of GPR data to rec-


Downloaded 07/03/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ognize tuning in a wedge of unconsolidated alluvial sedi-


ments. It also compares the spectral decomposition analysis
to the results of more traditional attributes such as instanta-
neous frequency. The study area is on the northern bank of
the Canadian River, south of Norman, Oklahoma, USA. (Figure
1). The alluvium ranges in thickness from approximately 10
to 15 m and is mostly sand, with varying amounts of silt and
clay and rare occurrences of gravel. Underlying the alluvium
is the Hennessey Shale.
The data set consists of 29, common-offset, single-channel
profiles, each separated by 2 ft. The offset of the transmitter
and receiver was 4 ft and the trace spacing was 1 ft. The high
temporal resolution of GPR data (nominal frequency of 100
Hz and sampling rate of 800 ps), and the close spacing of the
3D survey allow imaging of fine-scale sedimentary changes.
Spectral decomposition has proven very valuable in oil Figure 1. Location of Norman Landfill relative to Norman, USA, and
the Canadian River drainage basin.
exploration, particularly for imaging and mapping bed thick-
nesses and geologic discontinuities. In addition, the spectral
analysis method has been successfully used in the detection
of hydrocarbons in the presence of anomalous attenuation
effects.
The hypothesis tested here is that spectral decomposition
would help to identify the anomalously high attenuation of
the GPR reflections associated with clay layers, and that it
would highlight preferential illumination of specific features
caused by tuning effects. The goal of this research was to inte-
grate the spectral decomposition results with complex-trace
attributes and “hard” data (cores with grain size analysis and
conductivity data) in order to draw some conclusions about
the application of this analysis tool to GPR data.

Data preparation. The first part of the processing was per-


formed with the WinEkko Pro software by Sensors and
Software. The second part of the processing, merging the 29
profiles into a 3D data set, was performed using the Seismic
Processing Workshop (SPW) software by Parallel Geoscience
Corporation (PGC). The complete processing flow is shown
in Figure 2.
A final step in processing was needed to prepare the data
for spectral analysis. The matched pursuit decomposition
(MPD) algorithm that performs the spectral analysis relies on
a wavelet dictionary to map the time signal into the frequency
domain. The radar wavelet is minimum-phase, whereas the
wavelet dictionary is based on a zero-phase Ricker wavelet.
Through trial and error, it was found that the MPD result was
unstable for the radar data because its wavelet is minimum-
phase. Applying a 90° phase rotation on the radar data approx-
imated a zero-phase wavelet while leaving the frequency
spectrum unaltered.

Spectral decomposition. The concept behind spectral decom-


position is that a reflection from a thin bed has a characteris-
tic expression in the frequency domain and this is indicative
of the temporal bed thickness. The interference pattern in the
amplitude spectrum from a tuned reflection characterizes the
relationship between acoustic properties of the beds that gen- Figure 2. Processing flow used on the landfill data. The left column
erate the reflection. The method used to map a time series into contains the processing steps performed in WinEkko Pro. The right
the frequency domain determines the amount and quality of column is the conventional flow performed in SPW.

1024 THE LEADING EDGE AUGUST 2007


Downloaded 07/03/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 3. Comparison of the true spectra of the synthetic seismogram


and the MPD decomposition. The time-frequency plots show the ampli-
tude spectra for each sample. The MPD result matches the true spectra
very well. This method provides excellent time and frequency localiza-
tion (from Castagna et al., 2003).

new information that will be obtained (Chakraborty and


Okaya, 1995).
Traditionally, windowed Fourier methods have been
employed for spectral decomposition. They assume that the
time series is nonstationary and then decompose the seismic
signal in a nonstationary sense. This approach runs into some
resolution limitations, and the true spectra may be distorted
(Burnett et al., 2003). If the time window is too short, then the
frequency spectrum is smeared and this compromises fre-
quency discrimination. On the other hand, if the window is
too long, the result contains the spectral contribution of mul-
tiple events, making resolution of high-frequency geologic
responses unstable or impossible. The uncertainty principle,
or Heisenberg inequality, applies to windowed methods in the
sense that once a window function has been determined, the
time-frequency resolution is fixed over the entire time-fre-
quency plane (Chakraborty and Okaya, 1995). Tapering the
data helps to maximize the localization of the Fourier trans-
form, thus alleviating the resolution problem to a certain
extent.
The matched pursuit decomposition (MPD) method was
applied to the GPR data. By utilizing wavelet transforms to
map a seismogram into frequency-time space, this continu-
ous time-frequency analysis technique avoids windowing
requirements and therefore achieves excellent time and fre-
quency localization.The resulting decomposition provides a
frequency spectrum for each time sample of a time series
(Castagna et al., 2003).
This spectral analysis method consists of three steps: (1)
decomposition of the seismogram into constituent wavelets
using a matching pursuit algorithm, (2) summing the spectra
of the constituent wavelets in the time-frequency domain to
produce frequency gathers, and (3) sorting the gathers to pro-
duce the final products to be interpreted. The algorithm used
is based on Mallat and Zhang’s (1993) matching-pursuit
decomposition. Figure 3 shows an example of spectral decom-
position on a modeled trace.
Two common spectral decomposition products are the
“tuning cube” and “discrete-frequency energy cubes.” The tun- Figure 4. MPD discrete-frequency cube results for line 22. The panels
ing cube permits the study of a local area of interest by trans- show the following frequencies: 40 MHz (top), 60 MHz, 80 MHz, and
forming a short temporal window around the area into the 90 MHz.
frequency domain. The resulting cube is made up of three ele-
ments: thin bed interference, noise, and the wavelet overprint. quency slices, which aid in the visualization and identifica-
Each frequency slice is equalized by its average amplitude. tion of geologic textures and patterns.
The tuning cube allows the interpreter to animate through fre- The discrete frequency energy cubes are computed via run-

AUGUST 2007 THE LEADING EDGE 1025


frequency, and reflection strength (or envelope). Because spec-
tral decomposition is nonunique, attribute information can be
used to make an interpretation more robust.

Results. Matched-pursuit decomposition (MPD) was applied


to the final processed version of the Norman Landfill 3D data.
In addition, a discrete Fourier transform (DFT) algorithm was
tested for comparison with the MPD results. The following
Downloaded 07/03/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 5. Spectral decomposition problem for radar data. general results were observed for the entire data set. (1)Tuning
effects could be detected by a gradual increase of spectral
amplitude in the thinning section as frequency increases. (2)
Multiples of diffractions show up as vertical columns of anom-
alous spectral amplitudes. (3) The clay onset marks where
amplitudes die off. (4) The clay onset has anomalous bright
amplitudes between 80 and 90 MHz.
Figure 4 shows a frequency progression for line 22 where
results one through four can be identified. The frequencies
shown on the figure are 40, 60, 80, and 90 MHz. At 40 MHz,
the airwave (gray arrow) dominates the radar section. Another
dominating event is the vertical column of bright amplitude
(oval). This corresponds to a diffractor and its multiples that
have been collapsed to points through migration. The sand-
Figure 6. Base map with cores and log locations. The cores coincide clay onset reflector (red arrow) is not anomalous when com-
with the CMP locations, while the logs are designated ECO (from pared to the rest of the section. Everything below the thick
Faust, 2004). white line is very attenuated, so this has been identified as the
clay onset. At 60 MHz, the airwave starts to dissipate and the
ning window-spectral analysis (DFT) for each sample in a seis- rest of the section brightens. This indicates that the airwave
mic volume; then the components are sorted into multiple has a lower peak frequency than the rest of the section, which
monofrequency volumes. The same can be generated using a was also noticed in the dispersion analysis. The clay onset is
MPD algorithm. This allows the interpreter to characterize very bright at this frequency (red arrow), but it doesn’t seem
larger seismic volumes using specific frequencies of interest. to be brighter than other events. At 80 MHz, there is no sig-
nificant airwave energy (gray arrow). The clay onset is now
Spectral decomposition and signature: Seismic versus GPR. anomalous when compared to the rest of the section (red
In seismic exploration and production, it is essential to be able arrow). The diffraction amplitudes are gone (oval), and we
to link an attribute response or signature to some physical have anomalous high amplitude where a fine sand package
property or properties. In the present GPR study, we have wedges against the water table reflection (purple arrow). There
adopted the same goal. is also another bright event (brown arrow) which is probably
According to Dilay and Eastwood (1995), spatial changes the thin-bed interference effect from another shallow clay
in seismic data spectra are caused by either lithologic or petro- stringer. At 90 MHz, the clay onset is still anomalous (red
physical factors. Changes in lithology are best observed directly arrow), the shallow interference event is still anomalous
in the zone of interest, and the typical spectral signatures are (brown arrow), and the wedge effect (purple arrow) is begin-
caused by: (1) tuning effects producing wavelet interference ning to dim. An important observation for the MPD results is
which alters the spectra, (2) time sagging associated with low that the spectral amplitudes jump from peaks to lobes for the
velocity intervals, or a real thickening of the sediments, (3) radar data. The MPD algorithm attempts to match the data
amplitude changes from lateral impedance changes, and (4) to a Ricker wavelet. Although the radar data have been phase
geologic scattering of seismic energy leading to small static shifted by 90° to try to mitigate this problem, some events still
errors and a consequent loss of higher frequencies. look very much like a sine wave instead of a zero-phase Ricker
Petrophysical spectral signatures are generally caused by wavelet. Figure 5 illustrates this problem. For the trace shown
intrinsic attenuation associated with variations in rock-fluid in Figure 5, there are two possible decompositions: two pos-
properties. They can be identified by windowing either at the itive polarity Ricker wavelets with the side-lobes interfering,
zone of interest or directly below. or two negative Ricker wavelets with side-lobes interfering.
The lithologic factors that can be expected to affect GPR The decomposition is done trace by trace; so, for side-by-side
data are: tuning effects, amplitude changes caused by electri- traces, the resulting compositions will look choppy or broken
cal impedance changes, and geologic scattering. Since petro- up because there are two perfectly suitable decompositions.
physical spectral signatures are linked to attenuation, it is From the discrete frequency cube (MPD results) analysis,
expected that for GPR data they will be related to both fluid it was determined that at 80 MHz features of interest, such as
saturation below the water table and clay content. the clay onset, and thin-bed interference effects had anomalous
high spectral amplitude. In order to draw some conclusions
Complex trace attributes. Attribute analysis has been applied about whether the spectral decomposition signatures are
to GPR data to predict fluid permeability and mudstone dis- caused by lithologic or petrophysical factors, we tied the spec-
tributions, and to aid in geologic interpretation. The compu- tral analysis to our “hard data.” For the Norman Landfill data,
tation of complex-trace attributes is basically a transformation we have at our disposal nine cores with grain-size analysis
that separates the amplitude and phase information in a time and five conductivity logs. Figure 6 shows the location of the
signal. cores and logs.
Although an exhaustive attribute extraction and inter- Figure 7 shows an example of the tie of the radar data with
pretation was outside the scope of this project, three attrib- the core and conductivity log data, and the corresponding fre-
utes were examined: instantaneous phase, instantaneous quency gathers. These gathers show the amplitude spectra for

1026 THE LEADING EDGE AUGUST 2007


opposed to 1 ft in the line direction).
The turquoise rectangle over the
time section (Figure 7, row 1) indicates
the airwave and the refracted airwave
events. The depth section was corre-
lated with the core and conductivity
log data by recognizing major sediment
subdivisions. The first subdivision,
Downloaded 07/03/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

enclosed in red, is consistent in charac-


ter throughout the entire data set, and
is interpreted to be the fine-sand section
shown on the core and log data section
(Figure 7, row 2, enclosed in red). The
base of this fine-sand package dips
toward the east, and has a wedge effect
toward the west where the reflector
merges with the water-table reflection.
The depth of the base of the fine-sand
varies from 4 ft (west) to 7 ft (east) in
the depth section (row 1, right) and the
core and conductivity log section (row
2), so the estimated depth conversion
correlates well with the control data.
The second subdivision (Figure 7,
row 1, enclosed in yellow), covers the
areas of the data where there is signif-
icant amplitude decay. This amplitude
drop-off is interpreted as the onset of
clay shown in the core and log section
(Figure 7, row 2, thick blue line). The
response of the log shows high con-
ductivity areas that correlate with the
clay in the cores, so the sand-clay
boundary is well recognized. The clay
onset seems to correlate with the deep-
Figure 7. Time and depth data (row 1), core and conductivity log data (row 2), and frequency
gathers (row 3) for line 10. The time section (left, row 1) has four interpreted horizons, and the est reflector interpreted (Figure 7, row
turquoise rectangle indicates the airwave and a critical refraction. The depth section has some 1, green horizon). The depth of the clay
interpretations that are tied with the core and log data. The core and log section with grain-size onset in the depth section (Figure 6,
analysis is displayed in the middle. The frequency gathers are taken at the core and conductivity row 1, left) varies between 10 and 12 ft,
log locations. Higher attenuation coincides with clay content. The clay onset is indicated in the so this again correlates with the core
depth section, the core/log section and the frequency gathers (rows 1, 2, and 3). and conductivity log data.
A study of the frequency gathers
each time sample. The line shown is line 10, in both time and reveals the following trends: higher attenuation coincides with
depth (Figure 7, row 1). The depth conversion of the radar higher clay content, and higher amplitudes and frequencies
data was done using a single velocity function (the same one correspond to sandier sections. The spectral amplitude on the
used for the migration but converted to interval velocity). frequency gathers quickly decays below the clay onset (Figure
This doesn’t account for lateral velocity variation; however, 7, row 3, yellow line). The presence of reflectors within this
the purpose of the depth conversion was to give an estimate attenuated zone correlates with the core and log data where
of the event depths. Although a more detailed depth conver- we can see that there are thin sand layers below the clay onset.
sion using spatial-velocity variation was attempted with all The frequency gathers that correspond to cores 7 and 8, as well
nine of the velocity functions, it suffered the same problems as conductivity log ECO507, show anomalous amplitude
as the migration. The variability of the near-surface velocities decay when compared to the other two. This correlates with
created distortions in the depth-converted data that rendered a higher clay content as indicated in the core and conductiv-
it more difficult to interpret. The locations of the cores and ity log section. The gathers corresponding to core 9 and log
logs are indicated on the radar data with yellow broken lines ECO510 have overall higher amplitudes, and the events
and labeled (Figure 7, row 1). The frequency gathers shown between the airwave and the clay onset have higher peak fre-
are extracted from the core and conductivity log locations. quencies. This effect correlates with higher sand content as
The time section (Figure 7, row 1) has four horizons inter- can be seen in the core and conductivity log section.
preted on it: the water-table in turquoise, a prominent lithol- Further study of line 10 confirms that there is an anom-
ogy boundary in yellow, and two bright reflectors in purple alous spectral signature for the clay onset and for thin-bed
and green. Although it may seem that the horizons are con- interference effects. Figure 8 shows a discrete frequency sec-
tinuous when looking at one line in particular, tracking them tion for line 10 at 80 MHz, the reflection strength (or enve-
in the crossline direction was more difficult. The only hori- lope) attribute and the instantaneous frequency attribute. The
zons that could be tracked with certainty were the water table attributes add value to the interpretation by representing the
and the yellow horizon (horizon A). There is an appreciable radar data in the different ways shown. It was established from
amount of small-scale variability that is not imaged properly the tuning cube and discrete frequency volume analysis that
in the crossline direction because the trace spacing is 2 ft (as the thin-bed interference effects (purple arrows), the wedge

1028 THE LEADING EDGE AUGUST 2007


Figure 8.
(a) Discrete-
frequency
section (80
MHz); (b)
envelope
attribute;
(c) instanta-
neous fre-
Downloaded 07/03/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

quency for
line 10. Figure 9. Envelope and instantaneous frequency responses from a low-
Thin-bed impedance wedge in a homogeneous half-space (from Robertson and
interference Fisher, 1988). The envelope response is anomalous only where the
and clay wavelets are tuned (black arrow). The instantaneous frequency-
onset have response peaks well before that (black arrow) and then drops as the
very charac- wedge continues to thin (changes from green back to white). The red
teristic re- arrow indicates the frequency and envelope response for a thin-bed that
sponses in has not reached maximum-amplitude tuning.
all three
displays.

Figure 10. Isochron between water table and horizon A. Time separa-
tion between the two horizons decreases toward the west. Bright colors
(red-yellow white ~6 ns) are thin sections and dark purple indicate
thick sections (~32 ns).

the red arrow on Figure 9.


A second spectral decomposition was performed with a
DFT algorithm. Our goal in using a windowed spectral decom-
position was to compare the results with the MPD method.
Two products were generated, the tuning cube and the dis-
crete frequency volumes. The tuning cube, in spite of suffer-
ing from limitations, provided some interesting results.
effect (oval), and clay onset (yellow arrow) appear anomalous Although several parameters were tested, the discrete fre-
at 80 MHz. These anomalies correlate with the envelope (b) quency volumes were lacking in vertical resolution and had
and instantaneous frequency (c) attribute responses. The enve- artifact energy, so the results were unsatisfactory.
lope attribute shows that the most prominent event is the air- The biggest limitation of the windowed spectral decom-
wave. The two thin-bed interference effects (red arrows) position is defining the window length. Windowing does not
corresponding to the events in the frequency section above allow one to focus on a single event because generally win-
have anomalous low amplitudes. The wedge (white oval) has dows encompass several events. One of the tuning effects
a high amplitude anomaly where the wavelets tune. The clay studied with the tuning cube was the wedge created when
onset (yellow arrow) has a very localized high-energy the fine-sand interface (horizon A) merged with the water-table
response. The thin bed interference effects (red arrows) have reflection. Figure 6 showed line 10 with the interpretation of
the characteristic high instantaneous frequency response. The the water-table (turquoise) and horizon A (yellow). The two
clay onset has only a slightly higher frequency than the back- horizons wedge toward the west. Figure 10 shows an isochron
ground (black arrow). Wedge (oval) also presents anomalous map between the two horizons, where again we confirm that
frequency response. the event separation decreases toward the west.
Figure 9 illustrates the classic attribute response to a wedge Figure 11 shows the tuning cube frequency progression
model. These responses can be seen in Figure 8. The thin-bed (60, 80, 90, 100, and 120 MHz) that was created using the DFT
response (ovals in Figure 8) where the fine-sand reflection on a window length of 40 ns centered about horizon A (yel-
wedges against the water reflection shows exactly the same low horizon on Figure 6). This spectral signature is directly
attribute response as Figure 9: The instantaneous frequency linked to the thinning of the wedge as shown in the isochron
increases as the wedge thins until it reaches a maximum, and map. The shape of the tuned event at 90 MHz has exactly the
then it drops anomalously low; the envelope response is max- same geometry as the red to white zone in the isochron map
imum where the wavelets are tuned (λ/4). The other two thin (Figure 10). Thinner events will tune at higher frequencies, so
bed effects show an anomalous frequency response, yet there the spectral amplitude migrates toward the west (thinner sec-
is no envelope response. This is the case that corresponds to tion) as the frequency increases.

AUGUST 2007 THE LEADING EDGE 1029


1) The wedge effect between “horizon A” (base of fine-sand
package) and the water table was clearly documented in
all the spectral decomposition results, both MPD and the
DFT tuning cube. It had the classic wedge response in the
complex-trace attributes: high frequency anomaly that
turns into low as the wedge thins out further, and a local-
ized high-amplitude response in the envelope attribute as
the wavelets are tuned.
Downloaded 07/03/16 to 130.194.20.173. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

2) The clay onset’s spectral response was preferentially illu-


minated at 80 MHz, but localized, coinciding with where
the envelope response was high. The instantaneous fre-
quency response for the onset was varied. The frequency
gathers and common frequency sections all showed that
there was a very high signal attenuation below the clay
onset.
3) Other thin-bed effects and multiple diffraction energy
showed up as anomalous in the spectral analysis, with dif-
ferent preferential illumination frequencies.

Finally, both the MPD and DFT methods were compared


and the former showed better resolution in time and fre-
quency. The MPD algorithm’s limitation is that the radar data
wavelet (minimum-phase) is not compatible with the wavelet
dictionary (based on zero-phase Ricker wavelet), which causes
some jumping of the spectral amplitudes from peaks to
troughs.

Suggested reading. “Application of spectral decomposition to


gas basins in Mexico” by Burnett et al. (TLE, 2003). “Instantaneous
spectral analysis: Detection of low-frequency shadows associ-
ated with hydrocarbons” by Castagna et al. ( TLE , 2003).
“Frequency-time decomposition of seismic data using wavelet-
based methods” by Chakraborty and Okaya (GEOPHYSICS, 1995).
“Prediction of 3D fluid permeability and mudstone distributions
from ground-penetrating radar (GPR) attributes: Example from
the Cretaceous Ferron Sandstone Member, east-central Utah” by
Corbeanu et al. (GEOPHYSICS, 2002). “Spectral analysis applied to
seismic monitoring of thermal recovery” by Dilay (TLE, 1995).
“A three-dimensional ground-penetrating radar study at the
Norman landfill, Norman, OK” by Faust (master’s thesis,
University of Oklahoma, 2004). “Processing and interpretational
aspects of spectral decomposition” by Gridley and Partyka (SEG
1997 Expanded Abstracts). “Geomorphology and sedimentology
of the Canadian River alluvium adjacent to the Norman Landfill,
Norman, OK” by Mallat and Zhang (School of Geology, Oklahoma
State University, 1993). “Geomorphology and sedimentology of
the Canadian River alluvium adjacent to the Norman Landfill,
Norman, OK” by Marston et al. (School of Geology, Oklahoma
State University, 2001). “Interpretational applications of spectral
decomposition in reservoir characterization” by Partyka et al.
(TLE, 1999). “Complex-seismic trace attributes” by Robertson
and Fisher (TLE, 1988). “Interpretation of reflection attributes in
a 3D GPR survey at Valle d'Ossau, western Pyrenees, France” by
Figure 11. Tuning-cube frequency progression (60 MHz at top, 80, 90,
100, and 120 MHz at bottom). The DFT was calculated using a win- Sénéchal et al. (GEOPHYSICS, 2000). “Complex-seismic trace analy-
dow length of 40 ns centered about horizon A of Figure 7, row 1. The sis” by Taner et al. (GEOPHYSICS, 1979). TLE
spectral amplitude migrates toward the west as the frequency increases
corresponding to a westward wedge thinning. Greatest brightening Acknowledgments: The authors thank the University of Oklahoma for finan-
occurs at 100 MHz. cial support during this research and the USGS for borehole data. Also, the
authors thank the following companies for providing software and support:
Conclusions. The final MPD results helped to identify the Parallel Geoscience Corporation (SPW); Sensors and Software (WinEkko
anomalously high attenuation associated with clay layers, Pro); Landmark (SeisWorks, SpecDecomp), Fusion Geophysical (Inspect,
and indicated other effects, such as preferential illumination SpecMan), SeismicMicroTechnology (Kingdom Suite).
of thin-bed interference and the clay onset. The spectral decom-
position results were integrated with the other data available Corresponding author: icsa@chevron.com
to us (cores with grain size analysis, conductivity data, and
complex-trace attributes) and the following conclusions could
be drawn:

1030 THE LEADING EDGE AUGUST 2007

Das könnte Ihnen auch gefallen