You are on page 1of 5

Predicting bed thickness with cepstral decomposition

MATTHEW HALL, Red Brick Consulting, Calgary, Canada

Downloaded 11/08/14 to 31.7.79.46. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

onventional seismic data do not resolve thin beds. Below


a thickness of one quarter of a wavelength, the top and base
of a thin bed cannot be interpreted accurately. Even at thicknesses greater than but close to this limit, resolution is compromised. Cepstral decomposition is a new approach to
measuring bed thickness even when the bed itself cannot
be interpreted. The technique builds on the widely used
process of spectral decomposition.
Spectral decomposition has emerged recently as an
enlightening seismic attribute, producing very informative
maps of thin beds, especially in clastic successions with
sharp impedance contrasts (Partyka et al., 1999). These maps
are typically interpreted qualitatively, using geomorphologic pattern-recognition, or semi-quantitatively, to infer relative thickness variation. A number of commercial and
noncommercial implementations of spectral decomposition
are now in use, and the technique is de riguer in the analysis of subtle stratigraphic plays and fractured reservoirs.
There is also considerable scope for applying spectral
decomposition quantitatively. As shown in Figure 1, the
spectrum of a thin bed has characteristic periodic notches
resulting from interference between the signals reflected
from the top and base of the thin bed. The spacing of the
notches is the reciprocal of the two-way time thickness of
the thin bed. Unfortunately, estimating the notch spacing in
real data is difficult, because the notches are commonly
somewhat cryptic. To solve this problem, I propose using
Fourier analysis to elicit the notch spacing. This transforms
the spectrum into the cepstrum, and the seismic trace itself
beyond the frequency domain into the realm of quefrency.
The cepstrum is unfamiliar to most seismic interpreters but
well known in several other branches of signal processing.
A brief history of quefrency. During the Cold War, the United
States government was quite concerned with knowing about
when and where nuclear test detonations were happening.
One technique they employed was seismic monitoring. In
order to discriminate between nuclear detonations and earthquakes, a group of mathematicians from Bell Telephone
Laboratories proposed detecting and timing echoes in the
seismic recordings. These echoes gave rise to periodic but
cryptic notches in the spectrum, the spacing of which was
inversely proportional to the timing of the echoes. This is
exactly analogous to the seismic response of a thin-bed.
To measure notch spacing, Bogert et al. (1963) invented
the cepstrum (an anagram of spectrum and therefore usually
pronounced kepstrum). The cepstrum is defined as the Fourier
transform of the natural logarithm of the Fourier transform
of the signal: in essence, the spectrum of the spectrum. To
distinguish this new domain from time, to which it is dimensionally equivalent, they coined several new terms. For example, frequency is transformed to quefrency, magnitude to
gamnitude, phase to saphe, filtering to liftering, even analysis to alanysis. Only cepstrum and quefrency are widespread
today, though liftering is popular in some fields.
Today, cepstral analysis is employed extensively in linguistic analysis, especially in connection with voice synthesis. This is because, as shown in Figure 2, voiced human
speech (consisting of vowel-type sounds that use the vocal
chords) has a very different signature in time and frequency
from unvoiced speech. These differences are most easily

Figure 1. Nonband-limited spectra of equal and opposite reflection coefficient pairs. The reciprocal of the notch spacing is equal to the bed thickness. The spectra of two pairs are shown: one with a thickness of 20 ms,
resulting in 50 Hz notch spacing, and the other with a thickness of 50 ms,
giving 20 Hz notches.

quantified in the cepstrum, in which prominent peaks appear


at characteristic quefrencies for voiced speech. These peaks
are completely absent for unvoiced speech, which is noisebased and contains no resonant components. Cepstral methods help speech synthesizers switch seamlessly between
two different filters to make convincing human speech.
What is the cepstrum? To describe the key properties of the
cepstrum, we must look at two fundamental consequences
of Fourier theory. Firstly, convolution in time is equivalent
to multiplication in frequency. Secondly, the spectrum of an
echo contains periodic peaks. Let us look at these in turn.
A noise-free seismic trace s can be represented in the time
t domain by the convolution of a wavelet w and reflectivity series r
Then, in the frequency f domain
In other words, convolution in time becomes multiplication in frequency. The cepstrum is defined as the Fourier
transform of the log of the spectrum. Thus, taking logs of
the complex moduli
Since the Fourier transform F is a linear operation, the
cepstrum is
We can see that the spectra of the wavelet and reflectivity series are additively combined in the cepstrum. This
property has given rise to its use in deconvolution and signal estimation. It could also allow the removal of wavelet
effects from the cepstrum for more accurate bed thickness
estimation. I am currently working on this possibility.
How does the cepstrum respond to a thin bed? To study
this, the reflectivity series r can be represented as the sum
of two equal and opposite signals, one of which is delayed
by time :

The Fourier transform of a delayed signal is obtained


by multiplication by 2cos(2f), so
FEBRUARY 2006

THE LEADING EDGE

199

Downloaded 11/08/14 to 31.7.79.46. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 2. Time series (top), spectra (middle), and cepstra (bottom) for the sounds [ss] and [ee], as spoken by the author. The unvoiced consonant is
random noise with a whitish spectrum and flat cepstrum. In contrast, the voiced vowel sound, which uses the vocal chords, has a notched spectrum,
giving rise to a strong cepstral peak at 11.9 ms. This reflects the pitch of the author's voice.

Calculating the cepstrum as before gives


So adding an echo to the original signal has the effect of
adding a cosinusoidal pulse with period to the cepstrum
(Bogert et al. 1963). Cepstral analysis is concerned with
detecting this pulse and measuring its period.
In practice, I have found that taking the log of the spectrum sometimes has an adverse effect on the results. It is
not a necessary step for simply finding a periodic signal
in the spectrum. Also, there is no reason why any other transformation into the frequency domain could not be used
for example, the wavelet transform, S-transform, or
Wigner-Ville transform may offer better temporal resolution
than the Fourier transform (FT). In this paper, however, I
have used the true cepstrum, that is FT(ln(FT(signal))),
taking the complex modulus of the spectrum at each step.
Variants, where another transform is used, or where logs
are not taken, could be called pseudo-cepstra.
Why go to all this bother? The cepstrum offers unprecedented resolving power. Kallweit and Wood (1982) showed
that the thinnest bed min that can be interpreted is approximately

Figure 3. A synthetic seismic trace (top), its band-limited spectrum (middle), and its cepstrum (bottom). It is possible to measure the position of
the first spectral peak, then to double it, to get the notch spacing (middle).
Alternatively, the Fourier transform of the spectrum gives the cepstrum,
which has a peak at the bed thickness (bottom). The strong peak at the
short quefrencies is an effect of the limited bandwidth. Note that the cepstral peak is the first in a family of n harmonics of magnitude n-1.

bandwidth. In principle, the Fourier transform is capable of


resolving periods in the spectrum as long as 2 fmax, therefore
where fmax is the highest frequency present in the data.
Experience indicates that this is an optimistic estimate, and
the actual interpretable thickness is greater than this, possibly as thick as 1/0.5 fmax.
In contrast, the chief limit on the cepstral method is
200

THE LEADING EDGE

FEBRUARY 2006

In other words, the cepstrum is capable of resolving


beds that are only 70% the thickness of an interpretable bed,

Downloaded 11/08/14 to 31.7.79.46. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 4. Nonequal reflection coefficients. Time series (top), spectra (middle), and cepstra (bottom) for three opposite reflection pairs with different
reflection coefficient ratios. The positions of the cepstral peaks are unaffected, but their gamnitudes are reduced significantly if reflection coefficients are not equal.

Figure 6. Time series (top), spectra (middle), and cepstra (bottom) for a
wavelet (left), a reflectivity series containing three 10-ms thin beds (middle), and the corresponding synthetic trace (right). The band-limited
wavelet has a featureless cepstrum, whereas the reflectivity series clearly
shows two sets of harmonic peaks, corresponding to the thin beds (each 10
ms thick) and the thicker composite package.

spectrum, caused by more complex stratigraphy. I will return


to this point later.
I have tested this idea on 1D and 2D synthetic data. My
results show that it is indeed possible to use cepstra as an
accurate measure of the thickness of thin beds from seismic.
In addition, more complex stratigraphic patterns can also
be qualitatively characterized.

Figure 5. Nonopposite reflection coefficients. Time series (top), spectra


(middle), and cepstra (bottom) for three nonopposite pairs with different
reflection coefficient ratios. As for opposite reflections, the positions of the
cepstral peaks are unaffected, but their gamnitudes are reduced significantly if reflection coefficients are not equal.

and may be able to resolve the thickness of a bed as thin as


25% of the interpretable thickness. I will show this later in
the paper with a simple wedge model.
Partyka et al. (1999) also described a quantitative thickness estimation technique in their landmark paper. Rather than looking at the
whole spectrum to detect periodicity, as
I propose, the frequency of the first local
maximum in the amplitude spectrum is
doubled to estimate the position of the
first notch (as shown in Figure 3).
Unfortunately, the first spectral peak
frequency is prone to perturbation by
noise and interference effects, and in
my experience, the result is often difficult to reconcile with bed thicknesses
measured from wells without considerable support from modeling. Also,
because it is a single number, it cannot
represent more complex patterns in the

The cepstral peak measures bed thickness. Figure 3 shows


a single seismic trace representation of a 20-ms thin bed. The
notches in the spectrum are obvious, and have a spacing of
50 Hz. The cepstrum reflects this notching with a strong peak
at 20 ms, plus harmonics at integer multiples of the fundamental frequency. This simple example shows that the concept is valid.
What if reflections are not equal? Figure 4 shows the time
series, spectra, and cepstra for three cases, two of which have
unequal reflections. Clearly, the quefrencies of the peaks in
the cepstrum are not affected, but their gamnitudes are. In
fact, the gamnitude of the nth peak is proportional to the
ratio of reflection coefficients to the nth power. This means
that the cepstral peaks are markedly weaker if the thin-bed
does not have equal reflection coefficients. This is a good
thing, because the more unequal pairings of reflection coefficients are, the less likely they are to be related to the same
stratal unit.
What if reflections are not opposite? Figure 5 shows the

FEBRUARY 2006

THE LEADING EDGE

201

Downloaded 11/08/14 to 31.7.79.46. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 7. Filtered time series (top two panels),


spectrum (third panel), and cepstrum (bottom
panel) for a wedge model. The reflectivity model
was filtered with an Ormsby 10-60 Hz wavelet.
Interpretation (shown in green) in the time
domain poorly represents the model. Automatic
detection of the first spectral peak (orange) also
provides a poor estimate (see Figure 8). Manual
interpretation of the first peak (blue) would be an
improvement, but this is impossible in most real
data sets. In contrast, the cepstral peak (red) gives
a good representation of thickness down to
around 4 ms. The Fourier transform windows are
shown on the time series and frequency panels,
along with their respective Gaussian and cosine
tapers (purple).

effect of nonopposite reflection coefficients on the spectrum and the cepstrum.


Because I have used the complex modulus to compute gamnitude, it is not
apparent in the diagram that the cepstral peaks are of alternating sign. So the
technique as I have applied it does not
allow one to discriminate against nonopposite pairs of reflections, which
arguably are unlikely to represent thin
beds anyway, but could allow for it with
some modification.
What if there are many thin beds?
Figure 6 shows that the spectrum of a
compound thin-bed model synthetic
trace reveals two sets of overlapping
notches. The period of the notches with
respect to frequency Pf = 1/ = 20 Hz and
100 Hz, where is the two-way time bed
thickness. In a typical quantitative application of spectral decomposition, the
interpreter measures the first spectral
peak frequency, in this case either around
8.5 Hz or 50 Hz (depending on how the
first peak is selected). This is doubled to
estimate Pf. This approach fails to separate the two scales of bedding, resolving
either the thin beds themselves (the 50
Hz peak) or the package (the 8.5 Hz
peak). Notice that the thicker package is
estimated to be about 1/17 60 ms
thick, as opposed to the actual 50 ms.
The cepstrum shows unambiguous
peaks at 10 ms and 60 ms, representing
the thin beds themselves and the threebed package respectively. The cepstrum
accurately reflects the notches in the
spectrum, but like the first spectral peak
method, it also gives the wrong answer
for the total package thickness. The erroneous 60 ms cepstral peak is the result
of interference between the numerous
pairs of reflection coefficients present in
the time seriesthere are 15 ways to
combine the six reflectivities.
How thin is a thin bed really? I have also
applied the technique to a two-dimensional synthetic model. For this work, I
used Landmarks Spectral Decomposition
software, which shares its heritage with
202

THE LEADING EDGE

FEBRUARY 2006

Figure 8. Comparison of thickness estimates by interpretation (green), first spectral peak


frequency (orange), and cepstral peak (red). The cepstral peak gives the best estimate for 80% of
cases below about twice the tuning thickness (as shown by the red indicator line at the top of the
graph). The cepstral estimate has less than 20% error down to about 4 ms.

Downloaded 11/08/14 to 31.7.79.46. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

FreeUSP.orgs tune3d tool. The software automatically normalizes the spectral slices, effectively whitening the spectrum
and making the otherwise very weak high frequencies resolvable.
A simple Widess wedge model (Figure 7) was decomposed using the discrete Fourier transform to give the spectrum. The natural log of the spectrum was calculated and the
result was then run through the spectral decomposition software a second time to give the cepstrum.
Again, the cepstrum is considerably easier to interpret
than the spectrum, both qualitatively and quantitatively. At
thicknesses below about 18 ms, it is clear that the interpreted horizons are no longer representative of the reflectivity model. This is much thicker than the theoretical tuning
thickness, which is 1/(1.4  60 Hz) 12 ms for this data
set. In contrast, the cepstral peak faithfully reproduces the
reflectivity model, even well below the tuning thickness.
In summary, the thickness of the wedge can be calculated in three ways: from the time difference of the interpreted seismic events, from the first spectral peak frequency,
or from the cepstral peak. A comparison of the thickness estimates from these three methods is shown in Figure 8. The
systematic errors in the time-difference and first spectral
peak estimates are very clear, but the cepstral peak gives
good estimates down to around 4 ms.

position has the potential to significantly improve the accuracy of bed thickness estimation from seismic. In the oneand two-dimensional models presented in this paper, estimates are much more accurate than those from horizon
interpretation alone, or from traditional spectral decomposition workflows. In addition, the technique is able to characterize multiple beds of various thicknesses from the same
window, the inability to do this easily being one of the limitations of spectral decomposition. As with any seismic
attribute, indiscriminate and offhand application could result
in errors. In particular, likely pitfalls include having insufficient bandwidth to represent the target beds, a laterally
variable wavelet, or a window of inappropriate length. But
as a companion to these other techniques, cepstral decomposition promises to be a valuable stratigraphic analysis tool.
Suggested reading. The quefrency alanysis of time series for
echoes: cepstrum, pseudo-autocovariance, cross-cepstrum, and
saphe-cracking by Bogert et al. (Proceedings of the Symposium
on Time Series Analysis, Wiley, 1963). Interpretational applications of spectral decomposition in reservoir characterization
by Partyka et al. (TLE, 1999). The limits of resolution of zerophase wavelets by Kallweit and Wood (GEOPHYSICS, 1982).
How thin is a thin bed? by Widess (GEOPHYSICS, 1973). TLE
Acknowledgment: SpecDecomp software is a trademark of Landmark.

Conclusion. All of my work so far has been on synthetic


data. The next step in this research is naturally a move to
real seismic data. But it is already clear that cepstral decom-

204

THE LEADING EDGE

FEBRUARY 2006

Corresponding author: matt.hall@red-brick.com

You might also like