Instantaneous spectral analysis: Detection of low-frequency

shadows associated with hydrocarbons

JOHN P. CASTAGNA and SHENGJIE SUN, University of Oklahoma, Norman, U.S.
ROBERT W. SIEGFRIED, Gas Technology Institute, Des Plaines, Illinois, U.S.

I nstantaneous spectral analysis (ISA) is a continuous time-

frequency analysis technique that provides a frequency spec-
trum for each time sample of a seismic trace. ISAachieves both
excellent time and frequency localization utilizing wavelet
transforms to avoid windowing problems that complicate
conventional Fourier analysis. Applications of the method
include enhanced resolution, improved visualization of strati-
graphic features, thickness estimation for thin beds, noise sup-
pression, improved spectral balancing, and direct hydrocarbon
indication. We have seen four distinct ways in which ISA can
help in the detection of hydrocarbons: (1) anomalously high
attenuation in thick or very unconsolidated gas reservoirs, (2)
low-frequency shadows in reservoirs where the thickness is
not sufficient to result in significant attenuation, (3) preferen-
tial illumination at the “tuning” frequency which can be dif-
ferent for gas or brine-saturated rocks, and (4)
frequency-dependent AVO. In this paper, we describe the ISA
technique, compare it to other spectral decomposition meth-
ods, and show some examples of the use of ISA to detect low-
frequency shadows beneath gas reservoirs.
The ISA method involves the following steps: (1) decom- Figure 1. Synthetic waveform with transient arrivals (black seismogram),
pose the seismogram into constituent wavelets using wavelet constituent wavelets (color coded by center frequency), and time-
transform methods such as Mallat’s matching pursuit decom- frequency analysis (red represents high amplitude). At any given time, the
position, (2) sum the Fourier spectra of the individual wavelets spectrum of the seismogram is the weighted superposition of the wavelet
spectra. Since the exact wavelets summing to form the seismogram are
in the time-frequency domain to produce “frequency gathers,” known, exact spectra versus time may be computed.
and (3) sort the frequency gathers to produce common (con-
stant) frequency cubes, sections, time slices, and horizon slices.
The results can be viewed using animation techniques avail-
able in commercial interpretation and visualization packages.
Figure 1 shows a synthetic seismic trace and the corre-
sponding ISAtime-frequency analysis. The time-frequency plot
shows amplitude spectra for each time sample. We refer to
this kind of plot as a “frequency gather.” The first arrival on
the synthetic seismogram results from an isolated reflector. The
frequency spectrum is the spectrum of the wavelet. Note that
the duration of the spectrum is identical to the duration of the
arrival in the time domain as opposed to Fourier-based meth-
ods in which the time duration is equal to the window length.
The second event is a composite of two events of differing cen-
ter frequency arriving precisely at the same time. The fre-
quency spectrum indicates a low-frequency arrival spread
over time and a higher-frequency arrival that is more local-
ized in time. The third event is caused by two interfering Figure 2. Comparison of true spectra and ISA for a synthetic waveform.
arrivals of the same frequency. Although the presence of two
arrivals is not immediately apparent on the seismogram, the (DFT) have been applied (DFT having the advantages of
time-frequency decomposition clearly shows two distinct greater speed and not mandating transform lengths that are
arrivals. The fourth event is a composite of four waveforms a power of 2 as required by FFT). Both techniques have lim-
arriving at two distinct times evident on the time-frequency ited vertical resolution because the seismogram must be win-
analysis. The final event consists of three arrivals of the same dowed. The spectral energy is distributed in time over the
frequency that are very closely spaced in time. The three dis- length of the window, thereby limiting resolution. If the time
tinct arrivals cannot be resolved at low frequencies, but the window is too short, the spectrum is convolved with the trans-
separation is clearly evident on the time-frequency analysis fer function of the window, and frequency localization is lost
at high frequency. It is apparent that ISA provides a useful (i.e., the frequency spectrum is smeared). This can be mitigated
representation of the information contained in a seismic trace. to some extent by tapering the window, but it is obviously
preferable to avoid windowing altogether. Another disad-
Spectral decomposition methods. Various techniques have vantage of a short window is that side lobes of arrivals appear
been utilized in time-frequency analysis. Traditionally, the as distinct events in the time-frequency analysis. If the time
Fast Fourier transform (FFT) and discrete Fourier transform window is lengthened to improve frequency resolution, mul-

120 THE LEADING EDGE FEBRUARY 2003 FEBRUARY 2003 THE LEADING EDGE 0000
 trace and wavelets

Figure 3. Comparison of ISA and time-frequency decomposition obtained Figure 5. Comparison of ISA and time-frequency analysis obtained with
with FFT using a 200-ms window. Notice the poor vertical resolution and MEM using a long time window. Although MEM at times has superb
spectral distortion (ribs and notches) caused by FFT windowing. frequency resolution, vertical resolution is limited by the window length
necessary for reliable spectral determination.

Figure 4. Comparison of ISA and time-frequency analysis obtained with

DFT using a short time window. Although DFT with a short window has Figure 6. A frequency gather (left) and a common frequency section
excellent vertical resolution, the frequency spectrum has been smoothed by (right) obtained by sorting many frequency gathers according to
convolution with the spectrum of the window and false events are associ- frequency. The common frequency gathers can be thought of as instanta-
ated with side lobes of transient arrivals. neous amplitude at a given frequency.

tiple events in the window will introduce notches that dom- frequency over a long time interval, thereby superimposing
inate the spectrum. Long windows thus make it very difficult a cross-hatch pattern on the time-frequency analysis. This is
to ascertain the spectral properties of individual events. This a consequence of computational shortcuts designed specifi-
is one reason why Q-attenuation over short intervals is diffi- cally for rapid computation made possible by selecting a par-
cult to measure quantitatively using Fourier-based techniques. ticular wavelet dictionary. We prefer to select the wavelet
Recently, the maximum entropy method (MEM) has been dictionary to better capture the features of the seismogram
used for spectral decomposition. This technique can achieve while selecting parameters judiciously and avoiding as many
excellent frequency resolution but can be unreliable if the sig- cross-correlation operations as possible to achieve reasonable
nal violates the assumptions of the method or if the window computation time. Thus, wavelet-transform based instanta-
is too short. The major disadvantage, in our experience, is that neous spectral analysis (ISA) can be done accurately with
it seems to be unstable especially for less-than-expert users. acceptable speed while simultaneously achieving excellent
Note in the example presented in this article that, even for the time and frequency resolution.
first arrival, there are two peak frequencies.
Wavelet transforms decompose a seismogram into con- Comparison of methods. When comparing different spectral
stituent wavelets. The superposition principle tells us that the decomposition techniques, it is useful to have “the right
frequency spectrum of a seismogram is the sum of the fre- answer” as a guide. When dealing with synthetic data, con-
quency spectra of the wavelets that sum to produce that seis- structed as a superposition of wavelets, the “true” time-fre-
mogram. At any given time, the frequency spectrum is the quency analysis is readily calculated as the sum of the spectra
superposition of weighted wavelet spectra in the vicinity of of the known wavelets. As shown in Figure 2, the ISA tech-
that time sample. In a somewhat overlooked paper published nique does not yield the true spectrum precisely because
in GEOPHYSICS in 1995, Chakraborty and Okaya showed how wavelet decomposition is not unique in very much the same
wavelet transforms can be used in time-frequency analysis. way that seismic inversion is not unique. This fact can be used
They utilized matching pursuit decomposition to produce to prove that time-frequency analysis itself is not unique.
high-resolution time-frequency analyses. However, our expe- Many different time-frequency decompositions can result
rience with this method suggests that it introduces artifacts from the same seismogram and, conversely, can be inverse
into the time-frequency analysis manifested as high-amplitude transformed to produce the same seismogram. This then leads
bursts at a given time over a wide frequency band or at a given to the question: How can a particular frequency-decomposi-

122 THE LEADING EDGE FEBRUARY 2003 FEBRUARY 2003 THE LEADING EDGE 0000
 Figure 10. Broad-band seismic section from NW Shelf of Australia. Gas
Figure 7. Broad-band migrated stacked section for offshore Tertiary clastic sands are pink and brine sands are blue.
section. Troughs are blue, and peaks are red. The reservoir (arrow) is a
classic bright spot (low-impedance gas sands with a characteristic leading 2) The sum of the time-frequency analysis over time should
trough). No shadowing beneath the reservoir is apparent. Timing lines
represent 20 ms.
approximate the spectrum of the seismic trace.
3) Distinct seismic events should appear as distinct events on
the time-frequency analysis. In other words, the vertical res-
olution of the time frequency analysis should be compara-
ble to the seismogram. The time duration of an event on
the time-frequency analysis should not differ from the time
duration on the seismogram.
4) Side lobes of events on the seismogram should not appear
as separate events on the time-frequency analysis.
5) The amplitude spectrum of an isolated event should be
undistorted. The spectrum should not be convolved with
the spectrum of the window function.
6) There should be no spectral notches related to the time sep-
aration of resolvable events.

A sure sign of improper spectral decomposition is a large

dc (zero frequency) component for given events on data that
have had low-cut filters applied and should have little low-
Figure 8. 10-Hz common frequency section corresponding to the broad- frequency energy. This is usually caused by spectral smear due
band section in Figure 7. Significant low-frequency energy occurs beneath to windowing.
the reservoir but is absent elsewhere. Timing lines are 20 ms. By definition, the ISA technique is designed to mathe-
matically conform to criteria (1) and (2). As the ISA method
involves no windowing of the seismogram it also does well
at meeting criteria (3) to (6). As can be seen in Figure 2, ISA
does a good job of capturing the essential features of the true
time-frequency spectra. One caveat: the more appropriate the
selection of the wavelet dictionary, the better the time-fre-
quency representation. An inappropriate wavelet dictionary
will cause the ISA method to fail in the sense that criteria (3)
and (4) will not be achieved.
Figure 3 compares ISA and FFT spectra for the same syn-
thetic trace discussed previously. The FFT was calculated with
a 200-ms window. Notice that the FFT spectral energy is spread
out over the window length. When multiple arrivals occur in
the same window, severe spectral notches appear as well as
“ribs” between events. The location of the notches and peri-
Figure 9. 30-Hz common frequency section corresponding to the broad- odicity of the ribs depend entirely on the location of events in
band section in Figure 7. The low-frequency shadow in Figure 8 has time and tell little about the spectral characteristics of indi-
disappeared. Events immediately below the reservoir appear somewhat
attenuated. Timing lines are 20 ms. vidual reflectors. Certainly, closely spaced arrivals cannot be
resolved in time on the FFT time-frequency analysis. Their
tion be judged as better than another if there is no unique spacing in time can only be inferred from location of spectral
answer? In our opinion, the important question is not which notches—this being a strongly model dependent inference. The
decomposition is right or wrong, but which captures the essen- temporal resolution of the FFT can be improved by reducing
tial features important in interpretation. We believe it is impor- the window length, the cost being loss of frequency resolu-
tant that the following criteria are met: tion and smearing of the spectrum (Figure 4). In this case, DFT
was used rather than FFT because the window length was not
1) The sum of the time-frequency analysis over frequency a power of 2. At first glance, DFT appears to have resolution
should approximate the instantaneous amplitude of the comparable to ISA. Part of this improvement is illusory, how-
seismic trace. ever, as false events (that can be associated with side lobes of

124 THE LEADING EDGE FEBRUARY 2003 FEBRUARY 2003 THE LEADING EDGE 123
 a)

Figure 12. Broad-band seismic section for offshore Gulf of Mexico bright
spot. The top of reservoir is the trough at 2500 ms at trace 190. Low-
b) frequency shadowing is not particularly apparent on the broad-band data.

hydrocarbon indicator. These shadows are often attributed by

explorationists to abnormally high attenuation in gas-filled
reservoirs. However, it is often difficult to explain observed
shadows under thin reservoirs where there is insufficient
travel path through absorbing gas reservoir to justify the
observed shift of spectral energy from high to low frequen-
cies. At the 1996 SEG/EAGE Summer Research Workshop,
Dan Ebrom summarized at least 10 mechanisms that can
explain these low-frequency shadows. Other than intrinsic
attenuation, we think that one or more of the following mech-
anisms for introducing low-frequency shadows may be at
work at any given time: stacking in of locally converted shear-
waves and peg-leg multiples; NMO stretch of far-offset infor-
c) mation; improper moveout correction and consequent loss of

Figure 11. (a) 10-Hz common frequency section corresponding to the

broad-band seismic section in Figure 10. The low-frequency shadow
beneath the lower gas sand is the strongest event at 10 Hz. (b) 20-Hz
common frequency section corresponding to the broad-band seismic sec-
tion in Figure 10. The low-frequency shadow beneath the lower gas sand
is now weaker than the overlying gas sands. (c) 30-Hz common frequency
section corresponding to the broad-band seismic section in Figure 10. The
low-frequency shadow beneath the lower gas sand is now gone and the
gas sands are the strongest events on the section.

arrivals) are evident at low frequencies. Furthermore, it can

be seen that frequency resolution has been lost and the spec-
tra have been spread out over a broader frequency band than
is actually present in the data.
Superb frequency resolution can be achieved with MEM
(Figure 5); however, the method becomes unstable when short
windows are used. Thus, in practice, the temporal resolution
that can be achieved is relatively poor.

Low-frequency shadows. Since the inception of bright spot

technology in the 1960s, low-frequency shadows beneath
amplitude anomalies have been used as a substantiating

124 THE LEADING EDGE FEBRUARY 2003 FEBRUARY 2003 THE LEADING EDGE 125
 a)
a)

b)

b)

c)

Figure 14. (a) 6-Hz common frequency horizon slice (left) on the top of a
reservoir (dotted black line) and for a 50-ms window immediately below
the reservoir base (right). At 6 Hz, two strong events are apparent—one
associated with the reservoir dimensions (a shadow) and another indeter-
c) minate elongate feature toward the lower left. (b) 14-Hz common fre-
quency horizon slice on the top of a reservoir (left) and for a 50-ms
window immediately below the reservoir base. At 14 Hz, the reservoir is
now bright, the low-frequency shadow is gone, and the indeterminate
elongate feature toward the lower left remains. (c) 21-Hz common fre-
quency horizon slice on the top of a reservoir (left) and for a 50-ms win-
dow immediately below the reservoir base. At 21 Hz, the reservoir
remains bright, the low-frequency shadow is gone and the indeterminate
elongate feature is starting to look like a channel.

strong low-frequency energy beneath the reservoir. At 30 Hz

(Figure 9) the reservoir is clearly defined, though less anom-
alous in amplitude, and the energy under the reservoir appar-
ent at 10 Hz is gone.
The next example (Figure 10) from the NW Shelf of
Figure 13. (a) 8-Hz common frequency section corresponding to the
broad-band seismic section in Figure 12. The low-frequency shadow just Australia exhibits two distinct gas reservoirs. At 10 Hz (Figure
beneath the reservoir is the strongest event on the section. The reservoir 11a) the brightest event on the section is beneath the deeper
extends from xline 165 to 200. (b) 12-Hz common frequency section gas pay. We believe this to be a low-frequency shadow. At 20
corresponding to the broad-band seismic section in Figure 12. The low- Hz, the gas reservoirs are brighter than the shadow which still
frequency shadow and the reservoir have comparable amplitude, but the persists. The shadow has completely disappeared at 30 Hz
extent of the shadow better defines the reservoir dimensions (from xline (Figure 11c).
165 to 200). (c) 20-Hz common frequency section corresponding to the
broad-band seismic section in Figure 12. The low-frequency shadow is Returning to the Gulf of Mexico, a weak amplitude anom-
completely attenuated. aly can be seen at the crest of the structure in Figure 12. Figure
13 shows the relative variations in event amplitudes as fre-
high frequencies upon stacking; and time-varying deconvo- quency increases. At 8 Hz the strongest event is a low-fre-
lution where the gas bright spot is in the design window. quency shadow under the reservoir which extends to the
In the examples which follow, the frequency gathers are reservoir limits (Figure 13a). At 12 Hz, the top reservoir-sand
sorted into common frequency cubes, sections, and horizon reflection is now strong and the shadow persists (Figure 13b).
slices (Figure 6). Each common frequency display is thus the At 20 Hz the shadow is gone (Figure 13c).
spectral amplitude for that frequency versus time. Viewing frequency-dependent effects in map view is also
The first example is a bright spot from the Gulf of Mexico very revealing. Figure 14 shows frequency-dependent hori-
(Figure 7). The reservoir has a characteristic leading-trough zon slices at the top of the reservoir (left) and for a 50-ms time
(blue) on the broadband seismic data and is slightly brighter window immediately beneath the reservoir. The reservoir
than nearby events. Figure 8 shows the corresponding ISAsec- dimensions are outlined by the dashed contour. At 6 Hz
tion at 10 Hz. The reservoir is anomalously bright at this fre- (Figure 14a) the reservoir amplitude is not particularly bright.
quency, but what is most intriguing is the zone of abnormally The deeper window shows a strong shadow under the reser-

126 THE LEADING EDGE FEBRUARY 2003 FEBRUARY 2003 THE LEADING EDGE 125
 voir but also other strong energy to the lower left of the reser- when their prospect may in fact be more anomalous at some
voir. At 14 Hz (Figure 14b) the reservoir is a clear bright spot, other frequency. In our opinion, such broad-band amplitude
the shadow is gone, and the high-frequency energy to the lower anomaly mapping will become obsolete in the not too distant
left persists (indicating that this energy has another origin). future.
At 21 Hz, the energy to the lower left has developed a crisp It should also be apparent that the old idea of “tuning thick-
channel-like character showing that it is a stratigraphically ness” is also obsolete. Because we can investigate the data at
older geologic feature unrelated to the reservoir. any frequency, there is no single tuning thickness.
Explorationists need to think in terms of the “tuning fre-
Discussion and conclusions. In this article we have shown quency” for a given reservoir, not the tuning thickness for a
that the ISA method has a much better combination of tem- given seismic data set. This will be discussed in a future arti-
poral and frequency resolution than conventional spectral cle.
decomposition methods. This enables the use of ISAas a direct
hydrocarbon indicator. Low-frequency shadows are much Suggested reading. As a general introduction to spectral decom-
more apparent on spectrally decomposed data than on broad- position we recommend various papers by Greg Partyka or Kurt
band seismic sections. Marfurt in GEOPHYSICS. For an introduction to wavelet-transform
For every example shown, the shadow was stronger than based spectral decomposition, Chakraborty and Okaya’s 1995
the reservoir reflection at lower frequencies, suggesting that GEOPHYSICS paper “Frequency-time decomposition of seismic data
shadows are not necessarily a simple attenuation phenome- using wavelet-based methods” is a must read. “The low-fre-
non because low-frequency energy must have been added or quency gas shadow on seismic section” by Ebrom is available in
amplified by some physical or numerical process. Attenuation the proceedings for the 1996 SEG/EAGE Summer Workshop on
alone should simply attenuate higher frequencies, not boost Wave Propagation in Rocks. TLE
lower frequencies. Furthermore, attenuation should be less
localized in time than what we are observing. We believe that Acknowledgments: Thanks to the Gas Technology Institute and Fusion
these shadows are caused by one or more of the mechanisms Geophysical for financial backing in the development and evaluation of ISA.
described by Ebrom. Seismic data were provided by ChevronTexaco. Thanks are also owed to Sven
Conventional amplitude analysis is made at the arbitrary Treitel who doesn’t even know that he gave me the clue I needed to do this
and largely accidental dominant frequency of the seismic data in a conversation relating the Fourier transform to correlation with sine waves.
resulting from a complex interaction of acquisition parame- Special thanks are due to my late father John F. Castagna. I didn’t believe
ters, earth filtering, and data processing. When one sees how him when he told me years ago that geophysicists were doing things all wrong
amplitudes change with frequency, the inadequacy of “broad- and that we needed to look at things a single frequency at a time. Thanks,
band” amplitude analysis becomes immediately apparent. Dad, for being so arrogant in suggesting the ridiculous and for teaching me
Explorationists who construct conventional horizon amplitude to do the same.
maps need to ask themselves if they should continue to gen-
erate maps at the accidental dominant frequency of the data Corresponding author: castagna@ou.edu

126 THE LEADING EDGE FEBRUARY 2003 FEBRUARY 2003 THE LEADING EDGE 127