Evaluation of the Sensitivity of Seismic Inversion Algorithms to Different Statistically Estimated Wavelets

V. M. Gomes et al. Int. Journal of Engineering Research and Application www.ijera.com
ISSN : 2248-9622, Vol. 6, Issue 11, ( Part -5) October 2016, pp.59-73
www.ijera.com 59 | P a g e
Evaluation of the Sensitivity of Seismic Inversion Algorithms to
Different Statistically Estimated Wavelets
V. M. Gomes*, M. A. C. Santos*, R. B. Burgos**, D. M. S. Filho***
*(Department of Geology and Geophysics, Federal Fluminense University, Brazil
** (Department of Structures and Foundations, State University of Rio de Janeiro, Brazil)
***(CENPES/Petrobras, Brazil)
ABSTRACT
Seismic wavelet estimation is an important step in processing and analysis of seismic data. Inversion methods as
Narrow-Band and theConstrained Sparse-Spike ones require information about it so that the inversion solution,
once it is not a unique problem, may be restricted by comparing the real seismic trace with the synthetic
generated by convolution of the estimated reflectivity and wavelet. Besides helping in seismic inversion, a good
estimate of the wavelet enables an inverse filter with less uncertainty to be computed in the deconvolution step
and while tying well logs, a better correlation between the seismic trace and well log can be achieved.
Depending on the use or not of well log information, the methods of wavelet estimation can be divided into two
classes: statistical and deterministic. This work aimed to test the sensitivity of acoustic post-stack seismic
inversion algorithms to wavelets statistically estimated by two distinct methods.
Keywords-Narrow-band, Seismic Inversion, Seismic Wavelet Estimation, CSSI
I. INTRODUCTION
Geophysical techniques are used to
investigate indirectly the structures and properties
associated with the geology of the subsurface.
Depending on the features of the desired target, a set
of techniques will be chosen. Among the existing
methods, the seismic is the most used, considering it
provides good resolution even for great depths and
allows for estimation of the acousticaland elastic
properties of a medium.Estimating these properties
involves an approach through Inverse Problem
Theory to define a suitable method for the problem
at hand. Considering the Earth surface as an acoustic
medium a group of methods can be classified as
acoustic and post-stack (following classification
proposed in [1]) . The latter term refers to the fact
the data is supposed noiseless and generated through
a zero-offset geometry (source and receiver in the
same spatial position) a limitation that can be fairly
approximated using standard processing flows. The
main purpose of these methods is to generate a
consistent impedance (i.e. product of density and
velocity) model. Under the scope of this work the
methods discussed are: Narrow-Band and Sparse-
Spike (CSSI). Narrow-band, occasionally referred as
SEISLOG or VERILOG, is the simplest method of
seismic inversion being common routine in the 90's
[2],[1].CSSI on the other hand, includes more robust
methods which allow fora prioriinformation and
admits more consistent assumptions about
reflectivity. Under a deterministic approach,
nowadays even though model-based methods hold as
most used, there are still applications of CSSI in
light of its robustness as in [3].There is a wide
literature on CSSI methods and often they differ
only in the deconvolution technique employed to
reflectivity estimation. As examples of these, there
are the ones that perform deconvolution using
maximum likelihood estimation [4],[5],[6]; the L1
norm [7],[8],[9]; and methods developed under a
Bayesian point of view [10].The source signature
(signal) as a parameter of the seismic experiment is a
required input to inversion algorithms. The
combination of this signature and the effects on
wave propagation is called seismic wavelet. Thus, its
estimate from the treated data is a fundamental step
to ensure a reliable acoustic model is achieved.
There exists a vast range of works dealing with
wavelet estimation techniques and a possible
classification is to divide them in deterministic and
statistical, their difference being related to the
introduction of well log information. Much of the
work in the area until 1996 can be found in [11]. In
[11] a solution isproposed using the Kolmogorov
method, based on the assumption that the wavelet
has minimum phase and its phase spectrum can be
obtained by the Hilbert transform of the natural
logarithm of the estimated amplitude spectrum.
In[13], authorsseek an estimate using the theory of
homomorphic systems presented in [14]. The work
of[15] seeks a solution to this problem through a
strategy based on a simulated annealing algorithm.
More recent studies have evolved to more efficient
methods at the cost of more complexity as[16],
[17]and [18].Sensitivity of the inversion algorithms
was analyzed for wavelets estimated using the
RESEARCH ARTICLE OPEN ACCESS

methods: Kolmogorov (or Hilbert Transform)as in
[12]and a combination of smoothing the
seismogram's amplitude spectrum and setting a
constant linear model to the phase spectrum, similar
to the methodology used in [19]. To reduce
uncertainty in the phase estimate, after estimation a
correction in the phase spectrum based on[20] was
applied. In this work it was possible to analyze the
behavior of the seismic inversion methods: Narrow-
band and Bayesian, regarding wavelets estimated by
the aforementioned methods and noise. In general
the second got far superior results, admitting more
robust estimates of the impedance in all tests. For
wavelet estimation methods, the second (smoothing)
obtained better results, although for both, the phase
estimates were not good.
II. THEORETICAL BACKGROUND
A. Convolutional Model
The convolutional model in its simplest
form was originally described in [12] and is a valid
approximation for several steps of seismic
processing. Its validity depends on some
assumptions two of which are: (1) the reflectivity is
a random process and (2) the wavelet has minimum
phase. Considering a reflectivity series r and a
wavelet w, the seismic trace or seismogram s can be
mathematically represented as:
𝑠𝑡 = 𝑤𝑘 𝑟𝑡−𝑘 𝑜𝑟 𝑠(𝑡)
∞
𝑘=0
= 𝑤(𝑡) ∗ 𝑟(𝑡)
(1)
It can be viewed as an approximation to an
ideal zero-offset acquisition where the wavelet is an
idealization of a waveform that could account both
for the one created by the source and all the wave-
field propagation phenomena changing it along time
(i.e. dispersive effects). As long as a suitable
processing workflow is applied to the recorded data,
the final post-stack section or volume reasonably
satisfies the convolutional model assumptions thus
admitting inversion with the latter discussed
methods.
B. Wavelet estimation
The two methods discussed here are:
Hilbert transform method and Smooth spectra
method. Before going further the fact they are a
good approximation to wavelet amplitude spectrum
while producing a poor estimate of phase spectra
must be emphasized. To overcome this drawback of
both methods a correction based on the Automatic
Phase Correction (APC as referred herein)
introduced in [20] was adopted. In the original APC
the wavelet phase spectrum is rotated by a constant
angle and a synthetic calculated for each rotation
until an entropy norm is maximized. Here the best
solution was found considering the minimum error
norm seeking an increase in correlation.
1) The Hilbert Transformor Kolmogorov
factorization method
This method estimates the wavelet
amplitude spectrum of its minimum phase version
considering the data auto-correlation. However the
actual phase might not be minimum thus phase
spectrum must also be estimated. Using the
convolutional model and assuming reflectivity as a
random (white Gaussian) series, the wavelet auto-
correlation can be fairly approximated by the
seismogram's one, differing only by a scale factor
equal to the total energy of reflectivity, as shown in
[21].
As stated in [22], calculation of spectral density of a
signal can be accomplished by the Fourier transform
(FT) of its auto-correlation (ɸ) and by the square of
its magnitude spectrum ( 𝑊(𝜔) ) which are related
through:
𝑊(𝜔) = ɸ(𝜔) (2)
Therefore using the above relation an estimate of a
minimum phase wavelet amplitude spectrum can be
found from the square root of the seismogram's
spectral density.
The phase spectrum is obtained using the equation
derived in [12]:
∡𝑊(𝜔) =
−2 𝑠𝑖𝑛(𝜔𝑡)
1
2𝜋
𝑐𝑜𝑠(𝜔𝑡) ln 𝑊(𝜔) 𝑑𝜔
𝜋
0
∞
1 (3)
Derivation of the above equation can be done using
the Hilbert transform of the logarithm of the
wavelet's amplitude spectrum as stated in [23] thus
the method's name.
2) Smoothing Spectra method
The theory behind this method is simpler
than the previous one. Under the assumption
reflectivity is a white Gaussian series and the
wavelet has a band-limited smooth varying
amplitude spectrum, the smooth "trend" of the
seismogram's amplitude spectrum would be
associated with the wavelet. Though simple it holds
for most of the cases hence leading to good wavelet
estimates. Numerical implementation involves the
choice of a smoothing operator to the amplitude
spectrum of the seismogram and a minimum , mixed
or maximum phase guess for the spectra, where the
most common is the selection of a constant value.
The parameter that most influences the final result is
the operator's size, which can be a simple moving
average filter.
C. Seismic inversion methods
For seismic data inversion Narrow-band and a
Bayesian method were used, both discussed below.

1) Narrow-band inversion
Considering a wavelet, an estimated earth
reflectivity and an initial value for most superficial
layer a recursive equation can be used to calculate an
impedance model iteratively using the index i:
𝑟𝑖 =
𝑍𝑖+1 − 𝑍𝑖
𝑍𝑖+1 + 𝑍𝑖
(4)
From algebraic manipulation of (4) we achieve:
𝑍𝑖+1 = 𝑍𝑖
1 + 𝑟𝑖
1 − 𝑟𝑖
(5)
(5)represents a recursive process for obtaining the
acoustic impedances of subsurface and is the base
for several seismic inversion methods.
The big difference between the Narrow-
band method other recursive involves the choice
ofdeconvolution technique used for reflectivity
estimation. In this specific case, techniques
following the theory of optimum Wiener filters[24],
are adopted. The deconvolution filers are
constructed in an adaptive way, being however
highly sensitive to noise. Seismic deconvolution
comes from image processing theory, citing
Yilmaz[21], "it is responsible for estimating the
reflectivity by compressing the wavelet and
attenuating reverberations and short period
multiples, hence providing greater temporal
resolution".Optimum Wiener deconvolution in the
broad sense of Narrow-band inversion has a very
low computational cost. However for seismic
inversion purposes it represents the worst alternative
due to the limited bandwidth property, characteristic
of the estimated reflectivity. Low frequency loss
represents absence of information regarding the
main trend of the geological model of subsurface,
while the loss of high frequency components leads to
reduction in temporal resolution. This problem is
extensively discussed in literature, e.g.[1], and as it
is not the focus of this paper, a workaround would
be to use a wavelet with small central frequency to
ensure the low frequency part of the spectrum is
present.
2) A Bayesian approach (CSSI)
Seeking to overcome the limited bandwidth
problem discussed above, the sparse-spike methods
emerged. Considering the deconvolution algorithms
used, they will try to estimate a reflectivity
comprised of sparse deltas, ensuring an increase in
the estimated frequency band [10]. Sparsity will be
provided by constraining the estimated model
through minimization of a some sparse norm (e.g.
L1, Huber,Cauchy). Considering the inversion
methods, a bigger band will lead to less uncertainty
in the results. Additionally toincreasing the
bandwidth through the sparsity assumption,a set of
constraints are usually input asa priori information,
commonly low frequency information (a smooth
model) calculated from well logs. Nevertheless
improvement is associated with higher complexity
algorithms and consequently a rise in computational
cost, which in this case become irrelevant compared
to the robustness of results. In a Bayesian framework
both sparsity regularization norms and constraint
dependence can be introduced in the objective
function describing the problem. As in [10] a blocky
impedance model can be found from minimization
of the functional:
𝐽 = 𝜅𝐽𝑐
1
2
1
𝜎
𝑊𝑟 − 𝑠
2
+
1
2
𝑁−1
𝐶𝑟 − Β 2
(6)
In (6)𝜅 is a hyperparameter that ponders
sparsity in reflectivity, 𝐽𝑐 is the sparsity norm, 𝜎 is
an estimate of the noise level, 𝑊is the convolution
matrix associated with the wavelet, 𝑟 is the
reflectivity, 𝑠 the observed seismic data, 𝐶 is an
integration operator and Β is the natural logarithm of
the normalized impedance or double the cumulative
sum of reflectivity. The term 𝑁 is the diagonal
matrix 𝑁𝑘,𝑘 = 𝜗 𝑘 where 𝜗 𝑘 can be seen as a vector
of the uncertainties associated with a priori
information, thus 𝑁 imposes some restrictions on the
estimated model.
Following the options for the sparsity norm in [10]
the Cauchy norm was chosen,
𝐽𝐶𝑎𝑢𝑐 ℎ 𝑦 =
1
2
ln 1 +
𝑟𝑘
2
𝜗 𝑘
2
𝑘
(7)
Finally, it is noteworthy to comment that the choice
of parameters 𝜅 and 𝜎 as well as the initial model
must be carefully made to ensure convergence to the
solution or to the global minimum of the objective
function.
III. NUMERICAL EXPERIMENTS
The methodology used for sensitivity
testing, in general, involved modeling of a synthetic
seismogram from an input model, wavelet
estimationand finally the inversion.The acoustic
model assumed here is the Marmousi, shown in
Fig.1. It was created in 1988 by the French
Petroleum Institute, based on a geological profile in
the Kwanza basin in North Quenguela. It represents
a complex geological environment and is one of the
most widespread models for seismic analysis,
modelling and inversion, hence its choice.For
modelling of the reference (observed) seismogram
the convolution of reflectivity with a Ricker wavelet
with center frequency of 10 Hz was used. The choice
of this value is a way to overcome the problem of
absence of low frequencies as previously discussed.
The results can be divided into two stages; the first
was analysis of the wavelet estimation methods the
second of inversion techniques. Results of the
wavelet estimation considered synthetic models
generated with minimum and zero phase wavelets

(Fig.2), both tests admittingalso the existence of
noise. The inversion analysis used a seismogram
modeled with a zero phase wavelets as reference and
also accounted for noise influence. The noise used in
both analyses is white Gaussianwith standard
deviation 𝜎 = 10−3
.
D. Wavelet estimation analysis
As stated above, the tests to estimate the
wavelet were performed assuming data modeled
with minimum and zero phase Ricker wavelets.
Fig.2show modeled traces without noise. For the
Hilbert transform method, the results are shown in
figures Fig.3andFig.4. As expected, better results are
reached for the minimum phase situation. For the
zero phase case even after APC the estimated
wavelet does not have a shape similar to the true. It
was discussed in the subsection B that this method
depends on the assumption that the wavelet is
minimum phase, which explains the results.
Moreover noise addition to the trace worsens the
algorithm's performance, creating artifacts on the
shape of the estimated wavelet and for the zero
phase case satisfactory results cannot be achieved
showing the method is highly sensitive to noise.
Fig.1 -Marmousi model and trace 276 in two-way-travel-time.
Fig.2 - Ricker wavelets with zero and minimum phase, their respective amplitude spectrum and modeled traces

Fig.3 -Minimum phase ricker wavelet estimation with Hilbert transform method. (a) without and (b) with noise
Fig.4 -Zero phase ricker wavelet estimation with Hilbert transform methods. (a) without and (b) with noise.
Analyzing the results of estimation with the
Smoothing spectra method, shown in Fig.5 and
Fig.6, it can be noted that the presence of noise
doesn't have as much influence on the results as for
the previous method. Also, as there is not a
dependency of results with assumptions about the
phase, the APC correction proved more effective for
this technique. However, it can be seen that the best
estimates for the shape and phase spectrum
correspond to the zero phase case.It was observed
that better performance of the method is achieved
when guessing a linear phase spectrum for the
estimated wavelet before APC. The best results for
the zero phase case were obtained when considering
small values inside the interval [1,2] for the phase
spectrum; for the minimum phase case, larger values
achieved more success. Adopting a moving average
filter for spectral smoothing is largely responsible
for attenuation of noise effects as the shape artifacts
as seen for the previous method.
Fig.5 -Minimum phase ricker wavelet estimation using Smooth spectra method. (a) without and (b) with noise.
Fig.6 -Zero phase ricker wavelet estimation using Smooth spectra method. (a) without and (b) with noise.

In general the second method can achieve
an estimate with less uncertainty for the wavelet
when analyzed with respect to noise and phase. An
option for improvement of the results of both
techniques would be to use a more suitable
algorithm for correction and estimation of the phase
spectrum considering the estimated phase spectra are
far from the real (Fig.7,Fig.8,Fig.9 andFig.10) while
good estimates were obtained for the amplitude
spectrum, as shown by Fig.11, Fig.12, Fig.13 and
Fig.14.
Fig.7 -Phase spectra of minimum phase ricker wavelet estimated with Hilbert transform method. (a) without and
(b) with noise.
Fig.8 -Phase spectra of zero phase ricker wavelet estimated with Hilbert transform method. (a) without and (b)
with noise.
Fig.9 -Phase spectra of minimum phase ricker wavelet estimated with Smooth spectra method. (a) without and
(b) with noise.
Fig.10 -Phase spectra of zero phase ricker wavelet estimated with Smooth spectra method. (a) without and (b)
with noise.

Fig.11 -Amplitude spectra of minimum phase ricker wavelet estimated with Hilbert transform method. (a)
without and (b) with noise.
Fig.12 -Amplitude spectra of zero phase ricker wavelet estimated with Hilbert transform method. (a) without
and (b) with noise.
Fig.13 -Amplitude spectra of minimum phase ricker wavelet estimated with Smooth spectra method. (a) without
and (b) with noise.
Fig.14 -Amplitude spectra of zero phase ricker wavelet estimated with Smooth spectra method. (a) without and
(b) with noise.

E. Seismic inversion analysis
As previously stated synthetics used as the
observed (reference) data for the inversion analysis
were modeled with a zero phase wavelet. For
simplicity, density was assumed to be constant so
that the estimates, even when referred to as
impedance, are the velocities of P-wave propagation.
Results are presented and discussed below. First the
performance of the methods for situations when
noise is present is analyzed, then their sensitivity to
the estimated wavelets is observed. Considering the
Narrow-band method for the case where the real
wavelet is used, Fig.15 and 16show that problem is
solved even in the presence of noise, although the
results are not as good as in the absence of it. It can
be concluded that if the wavelet is well estimated,
even in the presence of noise good results are
obtained. In
Fig.16 this idea is well represented,
showing the similarity between the estimated and
synthetic models. It is necessary to discuss here the
parameter µ which represents knowledge about the
noise. A good estimate will lead to results with less
uncertainty. One can think of it as a regularization
parameter to the inversion since the results are
highly influenced by its choice. This constant will
also act on situations with estimated wavelets, even
in the absence of noise, as will be seen ahead, once
these wavelets carry an amount of uncertainty
associated with the estimation process which will be
interpreted as noise by the inversion algorithm. It
was observed that for the tests with noise addition
𝜎 = 10−3
, a value that achieve good results is
𝜇 = 0.1.
Fig.15 -Narrow-band inversion with true wavelet and without noise addition.
Fig.16 -Narrow-band inversion with true wavelet and with noise (𝜎 = 10−3
) addition for trace 276.

The use of an estimated wavelet, contrary
to the above results, will not produce such reliable
results as previously discussed. Figs. 17, 18, 19 and
20 show the results to the sensitivity of the Narrow-
band inversion to the estimated wavelets. The
striking effect in all cases is the discrepancy between
amplitudes, although interfaces have been well
delimitated. This effect on the other hand is related
to the previously mentioned problem of lack of low
frequencies, in fact, during the tests it was observed
that lowering the central frequency of the wavelet
to5Hz produces better results. It must be noted that
the method has not failed; it succeeded in identifying
the most significant subsurface
interfaces.Nevertheless since this method does not
use an initial model this behavior is something one
should expect in absence of low frequencies,
therefore a solution would be to add the low
frequency impedance model to the final estimate
from Narrow-band inversion. For results with noise
and an estimated wavelet, the noise does not show
significant influence when a good estimate for µ was
used. This confirms the earlier statement that the
method can solve, to some extent, effects related to
the presence of noise. It could also be noticed that,
the results using a wavelet estimated by the Hilbert
method are more sensitive to noise, requiring larger
values ofµ in comparison with the Smooth spectra
ones..
Comparison of the results using different estimation
techniques again shows that the results of Smooth
are superior (
Fig.17 and
Fig.19). As a final statement regarding Narrow-band
inversion results, it must be pointed out the need to
estimate a value for µ even in the absence of noise.
This relates to numerical noise introduced by the use
of the Discrete Fourier Transform algorithm
(calculation of the deconvolution step is performed
in the frequency domain). Looking now at the results
of the Bayesian method, Fig.21, Fig.22 and
Fig.22Fig.23 show the inversion results using the
true wavelet. One can see that the method can obtain
good estimates for impedance even if noise is
present. In general, the algorithm converged well to
all methods.The first discussion regarding the
Bayesian inversion must be about the parameters
chosen for the purpose of ensuring better results. In
(6) and (7) all parameters involved are shown,
however the ones actually used as input to the
algorithm are: an estimate of the standard noise
deviation (𝜎) and the uncertainty vector associated
with the a priori information (𝜗). Nonetheless, for
each test performed a new set of parameters should
be selected once each setting possesses different
levels of uncertainty. Complementing the discussion
in the paragraph above, it is necessary to talk about
the use of an initial impedance model. The one used
here was generated by smoothing the true model. It
is responsible to introduce a priori information and
consequently increase bandwidth, therefore ensuring
better convergence. As previously discussed, the low
frequency components carry relevant information
about trends of subsurface geology and thereby, use
of an initial model allowed incorporation of these
components and hence of the actual impedance
trend, constraining the results as observed in the
tests.
Fig.17 - Narrow-band inversion with wavelet estimated using Hilbert transform method for trace 276.

Fig.18 -Narrow-band inversion with wavelet estimated using Hilbert transform method and noise (𝝈 = 𝟏𝟎−𝟑
)
addition for trace 276
Fig.19 -Narrow-band inversion with wavelet estimated using Smooth spectra method for trace 276.
Fig.20 -Narrow-band inversion with wavelet estimated using Smooth spectra method and noise (𝜎 = 10−3
)
addition for trace 276
By observing the influence of noise, Fig.23,
Fig.24 and Fig.25 prove the method as more
efficient than the previous to deal with these
situations, even when using estimated wavelets. A
characteristic effect of noise are the artifacts with
oscillatory behavior in the calculated impedance,
easily noticeable in the last two above-mentioned
figures what however, does not prevent the
identification of the main interfaces. Finally it is
reiterated that the good performance of the method is
highly dependent on the above mentioned
parameters so that, good results can only be
achieved through a good estimate of these.Regarding
the performance tests with estimated wavelets, again
this method outperforms the Narrow Band,
achieving good results for the two estimation
techniques. In tests with the noise, however, the
presence of the artifacts discussed is significant,
especially when using wavelets estimated by Hilbert
transform method (Fig.24). By observing the results,
once again the Smooth appeared as the best
alternative, although in the tests discrepancies were
not as evident, showing that the Bayesian can solve
for uncertainties associated with the estimation of
the wavelet. Fig.26 and Fig.27 illustrate the
differences for results with both estimated wavelets.
As a final discussion about this method, the
artifacts in Fig. Fig.21 will be treated. They emerge
as a consequence of numerical instability during the
Conjugate Gradient optimization. As the inversion is
performed trace by trace, situations where the
method cannot converge well or when errors
associated with numerical approximation appear
might happen, justifying the artifacts in a few traces.
Nonetheless, the Conjugate Gradient method
minimizes the misfit with a step size per iteration,
the search for this step size is referred as line search
and was implemented here in a somehow empirical
way, where from an initial value the next was found
through fitting a polynomial between the previous
and values closer to it. Therefore the selected step
size could go over the global minimum hence
increasing the error and leading to the discussed
discontinuities.Comparison of the above results
show that Bayesian inversion has guaranteed less
estimation errors, proving superior to Narrow-band
for all tests. However, it is necessary to mention that
its computational cost is higher, in some cases
requiring more than 100 iterations for convergence
to an acceptable result. In any case, its sensitivity to
the uncertainties introduced by noise and estimated
wavelet is smaller, what can justify its choice over
the Narrow-band.
IV. CONCLUSION
Arising from the need to obtain estimates of
the parameters associated with subsurface geology
mainly with the purpose of exploration, geophysical
methods have increasingly gained space once it is an
indirect way and therefore more economical. In the
case of the seismic method, probably the most
popular, information on acoustic properties can be
obtained using seismic inversion algorithms, among
which two were discussed here, Narrow band and
Bayesian.The proposed methodology aimed to test
the sensitivity of the inversion methods considering

estimated wavelets and noise. Comparison of the
wavelet estimation methods showed that Smooth
produces better results. However, the amplitude
spectrum of the estimated wavelets showed that both
can solve the problem. Phase spectrum analysis
indicates otherwise, so that even after APC phase
correction the spectrum was not well approximated.
Techniques that are able to produce better estimates
to the phase spectrum would obtain in this way,
superior results.For inversion, the tests indicated
Bayesian as the most efficient once it converged to
the real model in all tests, even in the most complex
situation with wavelet estimated and noise, though
artifacts appeared in the inverted data. Anyway this
method produces good estimates of the impedance
model, delimiting all the interfaces and solving the
uncertainties related with both the noise and an
estimated wavelet. Summarizing, through the tests
carried out, the inversions methods performed well
under a low noise level situation (which can be
accomplished through robust pre-conditioning
techniques) and showed uncertainty in the wavelet
estimation can affect drastically inversion results.
Moreover, even though no well information was
input to the experiments, statistical wavelets
provided good results. Finally It can also be inferred
that by using a consistent low frequency model these
wavelets are valid though a comparison between
them and deterministic wavelets in the seismic
inversion context should be performed to confirm
this.
ACKNOWLEDGEMENTS
The authors thank Petrobras for the
financial support to the infrastructure of the GISIS
(Group of Seismic Inversion and Imaging) Research
Group. We are grateful to the members of the Group
for the comments and discussions. We also thankful
to the Department of Geology and Geophysics of
Universidade Federal Fluminense for allowing this
work to be completed.
REFERENCES
[1]. B.H. Russell, Introduction to seismic
inversion methods (Society of Exploration
Geophysicists, 1988).
[2]. D.A. Cooke, W.A. Schneider, Generalized
linear inversion of reflection seismic data,
Geophysics, 48(6), 1983, 665-676.
[3]. T.J. Campbell, F.W.B. Richards, R.L.
Silva, G. Wach, L. Eliuk, Interpretation of
the Penobscot 3D seismic volume using
constrained sparse spike inversion, Sable
sub-Basin, offshore Nova Scotia, Marine
and Petroleum Geology, 68(Part A), 2015,
73-93.
[4]. J.J. Kormylo, J. Mendel, Maximum
likelihood detection and estimation of
bernoulli-gaussian processes, Information
Theory, IEEE Transactions on, 28(3), 1982,
482-488.
[5]. C. Chi, J. Goutsias, J. Mendel, A fast
maximum-likelihood estimation and
detection algorithm for bernoulli-gaussian
processes, Acoustics, Speech, and Signal
Processing , IEEE International
Conference on ICASSP'85, 10, 1985, 1297-
1300.
[6]. B. Ursin, O. Holberg, Maximum-likelihood
estimation of seismic impulse responses,
Geophysical Prospecting, 33(2), 1985, 233-
251.
[7]. D. Oldenburg, T. Scheuer, S. Levy,
Recovery of the acoustic impedance from
reflection seismograms, Geophysics,
48(10), 1983, 1318-1337.
[8]. M.D. Sacchi, Reweighting strategies in
seismic deconvolution, Geophysical
Journal International, 129(3), 1997, 651-
656.
[9]. Y. Wang, Seismic impedance inversion
using l1-norm regularization and gradient
descent methods, Journal of Inverse and
Ill-posed Problems, 18(7), 2011, 823-838.
[10]. T.J. Ulrych, M.D. Sacchi, Information-
based inversion and processing with
applications, 36 (Elsevier, 2005).
[11]. O. Osman, E. Robinson, Seismic source
signature estimation and measurement,
Geophysics Reprint Series (Society of
Exploration Geophysicists, 1996).
[12]. E.A. Robinson, Predictive decomposition of
time series with applications to seismic
exploration, PhD thesis, Massachusetts
Institute of Technology, Cambridge, MA,
1954.
[13]. T. Ulrych, Application of
homomorphicdeconvolution to seismology,
Geophysics, 36(4), 1971, 650-660.
[14]. A.V. Oppenheim, Superposition in a class
of nonlinear systems, Technical Report,
DTIC Document, 1965.
[15]. D.R. Velis, T.J. Ulrych, Simulated
annealing wavelet estimation via fourth-
order cumulant matching, Geophysics,
61(6), 1996, 1939-1948.
[16]. J. Zheng, S. ping Peng, M. chu Liu, Z.
Liang, A novel seismic wavelet estimation
method, Journal of Applied Geophysics,
90(0), 2013, 92-92.
[17]. B. Yi, G. Lee, H.J. Kim, H.T. Jou, D. Yoo,
B. Ryu, K. Lee, Comparison of wavelet
estimation methods, Geosciences Journal,
17(1), 2013, 55-63.
[18]. Z. Wang, B. Zhang, J. Gao, The residual
phase estimation of a seismic wavelet using

a rényi divergence-based criterion, Journal
of Applied Geophysics, 106(0), 2014, 96-
105.
[19]. S.A.M. Oliveira, W.M. Lupinacci, L1 norm
inversion method for deconvolution in
attenuating media, Geophysical
Prospecting, 61(4), 2013, 771-777.
[20]. S. Levy, D. Oldenburg, Automatic phase
correction of common-midpoint stacked
data, Geophysics, 52(1), 1987, 51-59.
[21]. O. Yilmaz, Seismic data analysis:
processing, inversion and interpretation of
seismic data (Society of Exploration
Geophysicists, 2nd edition, 2001).
[22]. E.A. Robinson, S. Treitel, Geophysical
signal analysis (Society of Exploration
Geophysicists, 2000).
[23]. L.R. Lines, T.J. Ulrych, The old and the
new in seismic deconvolution and wavelet
estimation, Geophysical Prospecting, 25(3),
1977, 512-540.
[24]. N. Wiener, Extrapolation, interpolation,
and smoothing of stationary time series:
with engineering applications (The MIT
Press, 1966).
Fig.21 -Marmousi estimated with Bayesian inversion for full seismogram and 1000 iterations per trace.

Fig.22 -Marmousi estimated with Bayesian inversion. Trace 276.
Fig.23 - Marmousi estimated with Bayesian inversion and noise 𝜎 = 10−3
.
Fig.24 - Marmousi estimated with Bayesian inversion using Hilbert transform method for wavelet estimation
and noise (𝜎 = 10−3
).

Fig.25 - Marmousi estimated with Bayesian inversion using Smooth spectra method for wavelet estimation and
noise (𝜎 = 10−3
).
Fig.26 - Marmousi estimated with Bayesian inversion using Hilbert transform method for wavelet estimation
without noise.
Fig.27 -Marmousi estimated with Bayesian inversion using Smooth spectra method for wavelet estimation and
without noise.

Evaluation of the Sensitivity of Seismic Inversion Algorithms to Different Statistically Estimated Wavelets

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to Evaluation of the Sensitivity of Seismic Inversion Algorithms to Different Statistically Estimated Wavelets (20)

Recently uploaded (20)

Evaluation of the Sensitivity of Seismic Inversion Algorithms to Different Statistically Estimated Wavelets