a

r
X
i
v
:
1
2
0
9
.
0
1
9
6
v
3


[
p
h
y
s
i
c
s
.
g
e
o
-
p
h
]


9

J
a
n

2
0
1
3
Short-time homomorphic wavelet estimation
Roberto Henry Herrera and Mirko van der Baan
Department of Physics, University of Alberta, Edmonton T6G 2E1, CA
E-mail: [email protected], [email protected]
Abstract. Successful wavelet estimation is an essential step for seismic methods like
impedance inversion, analysis of amplitude variations with oset and full waveform
inversion. Homomorphic deconvolution has long intrigued as a potentially elegant
solution to the wavelet estimation problem. Yet a successful implementation has proven
dicult. Associated disadvantages like phase unwrapping and restrictions of sparsity
in the reectivity function limit its application. We explore short-time homomorphic
wavelet estimation as a combination of the classical homomorphic analysis and log-
spectral averaging. The introduced method of log-spectral averaging using a short-term
Fourier transform increases the number of sample points, thus reducing estimation
variances. We apply the developed method on synthetic and real data examples and
demonstrate good performance.
PACS numbers: 91.30.Bi, 93.85.Rt
Keywords: homomorphic analysis, seismic deconvolution, wavelet estimation
Submitted to: J. Geophys. Eng.
Short-time homomorphic wavelet estimation 2
1. Introduction
Homomorphic wavelet estimation has been popular since its rst application to seismic
deconvolution (Ulrych 1971). In part this is due to the conceptual simplicity of the
method requiring only a forward Fourier transform, then logarithm, followed by an
inverse Fourier transform. In addition, no minimum phase assumptions are made on the
wavelet, nor is the reectivity assumed white in the original formulation (Ulrych 1971).
Despite its promise, the method has often produced mixed results (Jin &
Eisner 1984). This is partially due to an often overlooked condition, namely that the
reectivity series must be suciently sparse (Ulrych 1971). Also the method forces the
resulting wavelet to replicate the strongest arrival in the data, often the rst arrival in
data not subjected to an automatic gain control to equalize amplitudes (Tribolet 1978).
In this paper, we combine elements of Ulrychs (1971) classic application of homomorphic
analysis, Tribolets (1978) short-time cepstral deconvolution with homomorphic wavelet
estimation in the log-spectral domain (Otis & Smith 1977). This relaxes some of the
original assumptions; for instance, the assumption of a suciently sparse reectivity
is replaced by a white, random one. The minimum reectivity assumption is replaced
by short mixed phase segments by windowing. This produces better wavelet estimates
while extending the ability to handle nonminimum phases in both the wavelet and the
reectivity.
We rst describe the classical cepstral liftering method and the stacking procedure
in the cepstral domain. Next we describe our approach. The new proposal is based
on short-time homomorphic analysis by averaging in the log-spectral domain. We
use constant length windows in the log-spectrum whereas Tribolets (1978) short-
time approach was based on averaging dierent window sizes in the cepstral domain.
Finally we test and validate the performance of the short-time homomorphic method
on synthetic and real datasets.
2. Method
2.1. The classical homomorphic analysis
The homomorphic analysis is related to signals which are the outcomes of convolution,
thus (Ulrych 1971):
s
i
(t) = w(t) r
i
(t), (1)
where r
i
(t) is the reectivity of the i
th
trace in a seismic prole, w(t) is the stationary
wavelet, stands for the convolution operator and s
i
(t) is the resulting i
th
seismic trace.
Following the classical homomorphic analysis (Oppenheim et al. 1968, Oppenheim
& Schafer 2010), we apply sequentially a Fourier transform (FT), complex natural
logarithm (ln) and inverse Fourier transform (FT
1
) to get a new signal in the cepstral
domain. In the cepstral domain convolution is mapped into an addition:
s
i
(t) = w(t) + r
i
(t), (2)
Short-time homomorphic wavelet estimation 3
where s
i
(t) = FT
1
{ln[FT{s
i
(t)}]} is the complex cepstrum of s
i
(t) and w(t) and r
i
(t)
are the complex cepstra of the wavelet and the reectivity respectively.
The homomorphism comes from transforming convolved signals into additive
signals. This additive space is known as the quefrency domain, where all the terms
are named by reversing the rst syllable of their spectral domain analogues (Otis &
Smith 1977).
Ulrych (1971) noticed that if the wavelet is time-invariant and has a smooth
spectrum its contribution to the complex cepstrum will be located at the low quefrencies
of the cepstral representation. On the other hand, a rapidly varying log-amplitude and
phase spectra, associated to the reectivity will have contributions at higher quefrency
rhamonics. Thus, by short-pass liftering or high-pass liftering we can extract either the
wavelet or the reectivity function. These lters act as windowing in the quefrency
domain. In addition if the reectivity is minimum phase its complex cepstrum will be
right-sided, making the separation of the additive terms easier. The third condition
Ulrych (1971) observed is related to the sparsity: if the interval time is larger than the
wavelet cepstrum, two reections overlapping in the time domain can be separated in
the cepstral domain. In this latter condition the logarithm of the spectrum plays the
role of whitening the wavelet spectrum to shorten its cepstral representation.
After windowing, the resulting signal is transformed to the time domain, that is:
s
i
(t) = FT
1
{exp(FT{ s(t)

f(t)})}, (3)
with

f the windowing lifter and s
i
(t) is the estimated signal after the liftering process.
Theoretically a low-cut time lifter will produce s
i
(t) = w(t) and a high-cut time lifter
will give s
i
(t) = r
i
(t). Judicious windowing can extract then the propagating wavelet
if the reectivity is sucient sparse. Unfortunately this is rarely the case in recorded
seismic data.
To avoid cepstral windowing, Otis & Smith (1977) proposed a technique based on
the averaging of log-spectral components in an array of seismic traces. They considered
the convolution of a constant source wavelet with a suite of non-stationary reectivity
series. In the cepstral domain the wavelet will be located at low quefrencies while the
reectivity will vary from trace to trace. By averaging the complex cepstra of a seismic
prole, the constant wavelet can be estimated since the variable reectivity with zero
mean will tend to zero:
s(t) =
1
N
N

i=1
s
i
(t) = w(t) +
1
N
N

i=1
r
i
(t), (4)
where i represents the i
th
trace in a seismic prole of N traces.
Averaging in the cepstral domain is equivalent to averaging in the log-spectral
domain, since the Fourier transform is a linear process. Thus,

S(f) =

W(f) +
1
N
N

i=1

R
i
(f), (5)
Short-time homomorphic wavelet estimation 4
where

S(f) = ln S(f) and S(f) is the Fourier transform of s(t), likewise for the wavelet
and the reectivity.
Cepstral stacking will produce successful results when the wavelet is spatially
stationary and the reectivity series are white. The latter condition implies that the
geological structure changes at each shot point (Tribolet 1979). This can be best
realized by combining traces from dierent parts of the 3D volume. The minimum
phase reectivity assumption is maintained in the log-spectral averaging method.
2.2. Our approach
Spectral estimation can be improved by using the Welch approach, that is, by assuming
the signal of interest is stationary, and then averaging the individual spectra obtained
from partially overlapping segments taken from the total observed data (Tribolet 1979).
The complex cepstrum of the i
th
trace is then:
s
ik
(t) = w(t) + r
ik
(t), 0 t L, (6)
where k represents the segment of length L used to compute the complex cepstrum.
If the cepstral structure r
ik
(t) is independent between segments, the estimated
wavelet in the log-spectrum is:

W
e
(f) =
1
NM
N

i=1
M

k=1

S
ik
(f) =

W(f) +
1
NM
N

i=1
M

k=1

R
ik
(f), (7)
where

S
ik
(f) = ln S
ik
(f) and S
ik
(f) is the Fourier transform of s
ik
(t).
The estimated wavelet

W
e
(f) converges to the true wavelet

W(f) if the reectivity
series is white, and the propagating wavelet is stationary. We refer to van der Baan
et al. (2010) for wavelet terminology. Averaging zero mean random reectivities over
the ensemble of windows makes the real-valued part of the last term converge to a
constant, and the imaginary part to zero. Hence the resultant process converges to the
seismic wavelet. The complex log-spectrum of the averaged reectivity is:

R(f) = ln R(f) = ln |R(f)| + j

R
(f), (8)
where the term ln |R(f)| converges to ln
r
which approaches zero as reectivity series
have assumed unitary variance
r
. The reectivity phase
R
tends to zero after
deramping and phase unwrapping if we assume that (1)
R
is uniformly distributed
between and and (2) the reectivity series is dominated by a few large reectors.
The recommended window length should be from three to ve times the wavelet
length, in accordance with Buttkus (1975).
Our approach resembles the Welch transform in the way the seismic trace is divided
into M overlapping segments (Welch 1967). We use overlapping windows to increase
the number of traces thereby improving the estimation variance (Angeleri 1983). The
reduction in the estimation variance is of the order of MO compared to the use of the
entire trace in one go (Appendix A). For a single seismic trace of length T, a window
Short-time homomorphic wavelet estimation 5
length L and a fraction of overlap between windows O, the number of segments per
trace will be:
M =
T LO
L LO
. (9)
In Eq. (7) we use the complex logarithm. This means that phase information is
included and we have to deal with the phase unwrapping problem (Herrera & van der
Baan 2011). Expanding Eq. (7) in terms of amplitude and phase:

W
e
(f) =
1
NM
N

i=1
M

j=1
ln |S
ik
(f)| +
j
NM
N

i=1
M

k=1
arg{S
ik
(f)}, (10)
where j =

1 and the function arg refers to the continuous unwrapped phase

arg{S(f)} =
S
(f) + 2n with n some integer. We need a continuous function
to guarantee the uniqueness of the solution. That is why at every short trace we
estimate the unwrapped and deramped phase. Removing the linear trend in the phase
suppresses the wavelet timing prior to the averaging process, thereby improving the
wavelet estimate.
We tested the performance of various phase unwrapping algorithms (Herrera &
van der Baan 2011) where we concluded that polynomial factorization (Sitton et al. 2005)
is exact but computational costly. On the other hand the -plane method (Kaplan
& Ulrych 2007) and the simple correction of 2 jumps in phase (Oppenheim &
Schafer 2010) have similar performance. Thus to reduce the computational load we
remove all jumps of 2 in the wrapped phase.
The estimated wavelet in the time-domain is nally given by:
w
e
(t) = FT
1
{exp[

W
e
(f)]}. (11)
3. Results
To evaluate the performance of Short-Time Homomorphic Wavelet Estimation
(STHWE), we compare its results with the log-spectral averaging (LSA) (Otis &
Smith 1977, Tria et al. 2007) and with the true wavelet in the synthetic example and
with the rst arrival in the eld dataset. We also include a kurtosis-based method
to estimate a constant-phase wavelet (van der Baan 2008), as it gives an independent
comparison based on an alternative approach using dierent statistical assumptions.
3.1. Synthetic dataset
The dataset consists of a synthetic section of 400 traces of 560 samples with 4 ms time
step, see Figure 1. These data have been produced with a mixed-phase and narrow-
bandwidth Ricker wavelet. The dominant frequency range is between 9 and 37 Hz with
a peak frequency of 23 Hz (i.e., 2 octaves and a ratio bandwidth to peak frequency of
1.18). Kurtosis-based methods require that the bandwidth exceeds the peak frequency
(Longbottom et al. 1988, White 1988). The true wavelet has a complex frequency
Short-time homomorphic wavelet estimation 6
dependent phase with a mean value of 58 degrees. Inspection of the waveform of the
strong rst arrival shows that the wavelet is not zero phase as it is asymmetric. It has
a large positive (black) and large negative (white) lobes.
Both statistical methods, KPE and STHWE are based on the Central Limit
Theorem, thus averaging of traces is needed. The entire section is used to extract a
single wavelet. The wavelet length was xed to wl = 220 ms for both methods.
Synthetic dataset
CDP
T
i
m
e

(
s
)
50 100 150 200 250 300 350 400
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Figure 1. Seismic dataset under study. Both statistical methods extract a global
wavelet for the entire dataset.
The LSA method averages the entire trace while the short-time method uses small
overlapping windows. In the STHWE method the critical parameter is the window
length which guarantees the reduction of the estimation variance in the wavelet phase
spectrum. We use an analysis window of 660ms (i.e., three times the wavelet length),
tapered with a Hamming window, and set a 50 % overlap between windows. Two cycles
of averaging are implicit in our implementation, rst one wavelet is estimated for each
trace and the nal estimated wavelet is the result of their average. For each trace we
have M = 19 segments, obtained by the Welsh relation in Eq. 9 and the total set is
NM = 7600 segments.
The true wavelet (black trace) is shown in Figure 2 along with the rst arrival
(FA in green), the KPE-wavelet(red), the LSA-wavelet (magenta) and the STHWE-
wavelet (blue). The LSA estimated wavelet is unstable compared to the other three
wavelets which all have similar time-domain waveforms (left panel) and spectral content.
LSA does not produce a very accurate result since it involves many more degrees of
freedom (i.e., individual frequencies) compared with the STHWE method, combined
with less averaging due to the longer windows. Finally we present the unwrapped and
deramped phases for all the estimated wavelets. Their average phases are respectively

KPE
= 65.30,
STHWE
= 68.14,
FA
= 58.42 and
LSA
= 49.06, which are all close to
Short-time homomorphic wavelet estimation 7
100 50 0 50 100
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Time [ms]
N
o
r
m
a
l
i
z
e
d

a
m
p
l
i
t
u
d
e
Estimated Wavelets


Truewav
FA
KPEwav
LSAwav
STHWEwav
0 50 100
0
2
4
6
8
10
Frequency [Hz]
A
m
p
l
i
t
u
d
e
Amplitude Spectrum


Truewav
FA
KPEwav
LSAwav
STHWEwav
10 20 30 40
50
0
50
Frequency [Hz]
P
h
a
s
e

[
d
e
g
r
e
e
s
]
Phase Spectrum


Truewav
FA
KPEwav
LSAwav
STHWEwav
Figure 2. Estimated wavelets. In the left in black is the true wavelet (True-wav),
rst arrival (FA) is plotted in green, the KPE method (red) and the short-time
homomorphic method (blue). In the upper right panel the four amplitude spectrums
are plotted and in right bottom their corresponding phase spectra. Average unwrapped
phases are
FA
= 58.42 degrees,
KPE
= 65.30 degrees,
LSA
= 49.06 degrees and

STHWE
= 68.14 degrees.
58 degrees, the average phase of the true wavelet.
3.2. Noise Stability Test
The noiseless synthetic example represent the optimum conditions for the proposed
algorithm. In this section we explore the robustness of our approach in the presence
of noise. We use the same dataset shown in Figure 1, but contaminated with white
Gaussian noise in a wide range of signal to noise ratios. Both methods, the kurtosis
based estimation and the homomorphic wavelet estimation are tested using 50 Monte
Carlo simulations per signal-to-noise ratio. All signal-to-noise ratios are dened as the
standard deviation of the signal over that of the noise.
For fairness of comparison, we also include a frequency dependent wavelet
estimation method, called robust blind deconvolution (RBD) of van der Baan &
Pham (2008). This method is based on a modied mutual information criterion and
operates on the amplitude and phase spectrum of the wavelet separately. The mutual
information rate is a general-purpose criterion to measure whiteness using statistics of
all orders. Its principal parameter is the wavelet length which is set to 220 ms. The
three wavelet estimates are correlated with the true wavelet, and the average of the
correlation coecients and their standard deviations are calculated. In each Monte
Short-time homomorphic wavelet estimation 8
Carlo simulation, dierent noise realizations are added to all traces. Here we follow the
experimental setup by van der Baan & Pham (2008).
Correlation coecients are bounded by unity and therefore not normally
distributed, which biases the estimated standard deviations. We use the Fishers
transformation to translate the correlation values to an almost normal distributed space.
Then we compute the average and the upper and lower standard deviations. Finally,
by inverse transformation we recover these values. See VanDecar & Crosson (1990) for
further details.
0 1 2 3 4 5 6 7
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
signaltonoise ratio
C
o
r
r
e
l
a
t
i
o
n
Noise stability test


KPEconstant phase
STHWE frequency dependent phase
RBD frequency dependent phase
Figure 3. Noise stability test. Signal-to-noise ratio versus correlation coecient. The
wavelet estimated by the constant-phase KPE and the frequency-dependent phase
methods STHWE and RBD are compared to the true wavelet. The homomorphic
method outperforms the other frequency-dependent technique; yet KPE is most
appropriate here as this true wavelet can be accurately described by a constant-phase
approximation.
Figure 3 shows the correlation coecients between the estimated wavelets and the
true wavelet. All three wavelet estimation techniques perform well for high signal-
to-noise ratios. The KPE method performs best since the compact nature of the
true wavelet favors the constant phase estimation. KPE achieves, in this case, better
reconstructions as seen by the higher correlation coecients, but van der Baan & Pham
(2008) show that complex wavelets such as dispersive ones cannot be described by a
constant phase approximation. Therefore, the two frequency-dependent methods are
most useful for more band-limited data e.g., old legacy data and situations where long
wavelets with complex phases are anticipated.
The homomorphic method outperforms the RBD method, giving satisfactory result
for signal-to-noise ratios over 1.5. The three methods have low standard deviations in
their estimates at all noise levels.
Short-time homomorphic wavelet estimation 9
3.3. Real dataset
Finally, we show the performance of the new wavelet estimation method on a real seismic
dataset. A stacked section of marine data is shown in Figure 4. The original section
consists of 300 traces and 1074 samples, with a sampling frequency of 250 Hz. The
rst arrival occurs around 0.55 s, the ocean bottom. There is a strong reector at
approximately 1.6 s, below which few arrivals are observed. In real data the wavelet
length can often be selected from observation of the length of the rst arrival, in this
case wl = 140 ms. The analysis window is set to be three times the wavelet length,
using a 50 % overlap and a Hamming lter for tapering.
Real dataset
CDP
T
i
m
e

(
s
)
50 100 150 200 250 300
0.5
1
1.5
2
2.5
Figure 4. Real stacked seismic section. The rst arrival corresponds to the ocean-
bottom reection.
The estimated wavelets using the constant-phase rotation approach (KPE) and
the homomorphic method (STHWE) are compared with the rst arrival (FA) at 0.5 s.
Figure 5 shows in the left plot the estimated wavelets by the constant-phase rotation
approach (red) and the new short-time homomorphic method (blue). Both estimates
lead to similar wavelets and reproduce the FA waveform, as expected in this marine
example with a strong ocean-bottom reection.
The mean value of the frequency dependent phase for the STHWE method and
the rst arrival are respectively
STHWE
= 45.13 and
FA
= 45.97 degrees. The
KPE method delivers a constant phase of
KPE
= 45 degrees. Thus, both statistical
methods produce similar results comparable to the rst arrival.
4. Discussion
The minimum phase constraint on the reectivity function is a fundamental assumption
in the technique of Ulrych (1971). It is also a necessary condition for full trace log-
Short-time homomorphic wavelet estimation 10
50 0 50
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
1.2
Time [ms]
N
o
r
m
a
l
i
z
e
d

a
m
p
l
i
t
u
d
e
Estimated Wavelets


FA
KPEwav
STHWEwav
0 50 100
1
2
3
4
5
Frequency [Hz]
A
m
p
l
i
t
u
d
e
Amplitude Spectrum


FA
KPEwav
STHWEwav
10 20 30 40
50
0
50
Frequency [Hz]
P
h
a
s
e

[
d
e
g
r
e
e
s
]
Phase Spectrum


FA
KPEwav
STHWEwav
Figure 5. Estimated wavelets for the stacked section displayed in Figure 4. The
constant-phase rotation approach (KPE) in red, and the new short-time homomorphic
method (blue) lead to similar wavelet estimates, with high similarity to the rst arrival
(FA) in green. The KPE amplitude spectrum is smoother than the homomorphic
estimates due to its constant-phase nature. The KPE phase estimate is
KPE
= 45
degrees, whereas the average phase for the STHWE and rst arrival are respectively

STHWE
= 45.13 and
FA
= 45.97 degrees.
spectral averaging proposed by Otis & Smith (1977). In the proposed short-time
averaging method it is no longer required as many reections are taken into account.
Phase unwrapping and deramping tends to place the largest reection at low quefrency
values, thus allowing the repeated averaging to estimate the underlying propagating
wavelet.
The sparsity condition, in the conventional homomorphic deconvolution method,
aims to separate the wavelet from the reectivity in the cepstral domain. This means
that the interarrival time should be sparse enough to isolate the wavelet from the
reectivity. As the time separation between arrivals is unpredictable, induced whitening
due to applying a logarithm to the wavelet passband may not be sucient to separate
individual arrivals. In our case a whiteness assumption of the reectivity leads to random
rhamonics in the log-spectrum, which after averaging converge to the smooth wavelet
log-spectrum, i.e., the averaging process acts as a lifter.
The stationary condition imposed on the wavelet, is a limitation in time varying
wavelet estimation. In this case, we suggest to assume that the wavelet is piecewise
stationary and divide the seismic section into two or more separate time windows to
estimate dierent wavelets.
Short-time homomorphic wavelet estimation 11
5. Conclusions
Homomorphic wavelet estimation was rst introduced over 40 years ago and has been
revisited often with its promise of nonminimum-phase wavelet estimation. The original
method of cepstral liftering assumes the wavelet has a smooth spectrum and that the
reectivity series is minimum phase and suciently sparse. However, in most of the
cases, the latter assumption is rarely honored.
Log-spectral averaging mitigates the need for the sparsity constraint, but requires
a large number of independent reectivity series while maintaining the minimum phase
constraint. The method of log-spectral averaging using a short-term Fourier transform
increases the number of traces, thus reducing estimation variances. Furthermore, no
assumptions regarding the phase of the wavelet or the reectivity are required. A
comparison using a synthetic example shows similar results, with regards to constant-
phase wavelet estimation based on kurtosis maximization. The short-time homomorphic
method and the kurtosis-based method produce similar wavelets but the short-time
homomorphic technique allows for a frequency-dependent phase estimation, whereas
the kurtosis-based method assumes a constant phase.
The noise stability test, using Monte Carlo simulations on a synthetic gather,
demonstrates the feasibility of statistical wavelet estimation from noisy seismic traces.
Both statistical methods lead to reliable wavelet estimates that could be employed
as quality control tool for deterministic methods. When dierent well logs in the
same seismic section produce dierent extracted wavelets, the statistical methods could
interpolate the wavelet phase between the non-matching wells. Seismic inversion in
absence of well logs is one the main applications for the statistical methods. The
method also shows promise for estimating wavelets from time-lapse datasets where the
acquisition parameters may have changed between monitoring and baseline surveys or
signicant time-lapse changes are observed around key horizons.
Acknowledgments
The authors thank Chevron for permission to use the synthetic data and BP for
providing the real dataset and the Sponsors of the BLISS project for their nancial
support. We are grateful to the two anonymous reviewers for their valuable comments
and suggestions to improve the original manuscript. R. Herrera thanks Mauricio Sacchi
for his useful comments related to the Homomorphic Deconvolution Method.
Appendix A. Phase variance in the short-time homomorphic analysis
The quality of the estimated wavelet log-spectrum is highly dependent on the number
of averaging windows. Better results are obtained with a larger number of averaged
segments for a stationary wavelet. The objective is to minimize the eect of the
reectivity in the second right term of Eq. 10 for the time-invariant case.
Short-time homomorphic wavelet estimation 12
Angeleri (1983) evaluates the number of windows needed for a successful averaging
process. Denoting the variance of the phase spectrum of the averaged reectivity by
2
R
the original expression is:

2
R
=
BL
12N
, (A.1)
where B is the frequency bandwidth, L is the trace length and N is the total number
of traces.
Assuming statistical independence between non overlapping segments it is
reasonable to assume we can approximate the averaged reectivity as:

2
R
=
BL
12NMO
, (A.2)
with O the fraction of overlap in each segment and M the total number of segments.
Using short-time windows reduces therefore the estimation variance (Ulrych et al. 1995).
Note that the dimensionless value of
2
R
refers to variance of the phase in radians.
References
Angeleri G 1983 A Statistical Approach to the Extraction of the Seismic Propagating Wavelet
Geophysical Prospecting 31(January), 726747.
Buttkus B 1975 Homomorphic ltering theory and practice Geophysical Prospecting 23(4), 712 748.
Herrera R & van der Baan M 2011 Revisiting Homomorphic Wavelet Estimation and Phase
Unwrapping Recovery 2011 CSPG CSEG CWLS Convention .
Jin D J & Eisner E 1984 A review of homomorphic deconvolution Reviews of Geophysics 22(3), 255
263.
Kaplan S T & Ulrych T J 2007 Phase Unwrapping: A Review of Methods and a Novel Technique
CSPG CSEG Convention pp. 534537.
Longbottom J, Walden a T & White R E 1988 Principles and Application of Maximum Kurtosis Phase
Estimation Geophysical Prospecting 36(2), 115138.
Oppenheim A & Schafer R 2010 Discrete-Time Signal Processing (3rd Edition) (Prentice Hall Signal
Processing) Signal Processing 3rd edn Prentice Hall.
Oppenheim A, Schafer R & Stockham T 1968 Nonlinear ltering of multiplied and convolved signals
Proceedings of the IEEE 56(8), 12641291.
Otis R M & Smith R B 1977 Homomorphic Deconvolution by Log Spectral Averaging Geophysics
42(6), 1146 1157.
Sitton G, Burrus C, Fox J & Treitel S 2005 Factoring very-high-degree polynomials Signal Processing
Magazine, IEEE 20(6), 2742.
Tria M, van der Baan M, Larue A & Mars J 2007 Wavelet estimation and blind deconvolution of
realistic synthetic seismic data by log spectral averaging SEG Technical Program Expanded
Abstracts 26(1).
Tribolet J M 1978 Applications of short-time homomorphic signal analysis to seismic wavelet
estimation Geoexploration 16(1-2), 7596.
Tribolet J M 1979 Seismic applications of homomorphic signal processing Prentice-Hall, Englewood
Clis, N.J. :.
Ulrych T 1971 Application of homomorphic deconvolution to seismology Geophysics 36(4), 650 660.
Ulrych T J, Velis D R & Sacchi M D 1995 Wavelet estimation revisited The Leading Edge 14(11), 1139
1142.
van der Baan M 2008 Time-varying wavelet estimation and deconvolution by kurtosis maximization
Geophysics 73(2), V11.
Short-time homomorphic wavelet estimation 13
van der Baan M, Fomel S & Perz M 2010 Nonstationary phase estimation: A tool for seismic
interpretation? The Leading Edge 29(9), 10201026.
van der Baan M & Pham D T 2008 Robust wavelet estimation and blind deconvolution of noisy surface
seismics Geophysics 73(5), V37V46.
VanDecar J C & Crosson R S 1990 Determination of teleseismic relative phase arrival times using multi-
channel cross-correlation and least squares Bulletin of the Seismological Society of America
80(1), 150169.
Welch P 1967 The use of fast fourier transform for the estimation of power spectra: A method
based on time averaging over short, modied periodograms IEEE Transactions on Audio and
Electroacoustics 15(2), 70 73.
White R E 1988 Maximum Kurtosis Phase Correction Geophysical Journal International 95(2), 371
389.