Journal of Applied Geophysics 136 (2017) 298–314

Contents lists available at ScienceDirect

Journal of Applied Geophysics

journal homepage: www.elsevier.com/locate/jappgeo

Comparison between deterministic and statistical wavelet estimation

methods through predictive deconvolution: Seismic to well tie
example from the North Sea
Isadora A. S. de Macedo a , Carolina B. da Silva a , J.J.S. de Figueiredo a, b, * , Bode Omoboya c
a
UFPA, Faculty of Geophysics, Petrophysics and Rock Physics Laboratory – Prof. Dr. Om Prakash Verma, Belém, PA, Brazil
b
National Institute for Petroleum Geophysics (INCT-GP), Brazil
c
University of Houston, Department of Earth and Atmospheric Sciences, Houston, TX, USA

A R T I C L E I N F O A B S T R A C T

Article history: Wavelet estimation as well as seismic-to-well tie procedures are at the core of every seismic interpretation
Received 8 January 2016 workflow. In this paper we perform a comparative study of wavelet estimation methods for seismic-to-
Received in revised form 1 November 2016 well tie. Two approaches to wavelet estimation are discussed: a deterministic estimation, based on both
Accepted 4 November 2016 seismic and well log data, and a statistical estimation, based on predictive deconvolution and the classi-
Available online 14 November 2016
cal assumptions of the convolutional model, which provides a minimum-phase wavelet. Our algorithms,
for both wavelet estimation methods introduce a semi-automatic approach to determine the optimum
Keywords:
parameters of deterministic wavelet estimation and statistical wavelet estimation and, further, to estimate
Well tie
the optimum seismic wavelets by searching for the highest correlation coefficient between the recorded
Predictive deconvolution
Wavelet estimation trace and the synthetic trace, when the time–depth relationship is accurate. Tests with numerical data show
Well log some qualitative conclusions, which are probably useful for seismic inversion and interpretation of field
data, by comparing deterministic wavelet estimation and statistical wavelet estimation in detail, especially
for field data example. The feasibility of this approach is verified on real seismic and well data from Viking
Graben field, North Sea, Norway. Our results also show the influence of the washout zones on well log data
on the quality of the well to seismic tie.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction
The earth’s subsurface is composed of rock layers with different
lithology and physical properties. From the seismic point of view,
these layers manifest as density and velocity contrasts as seismic
waves propagate through them. The product of density and velocity
generates the seismic impedance. The impedance contrast between
the adjacent rock layers causes the reflections that are recorded
along a surface profile. These signals are what we will refer to as
real seismic data. The recorded seismogram can also be modeled
by the convolution of the earth’s reflectivity, using well log data,
with the seismic wavelet. The result of the modeled seismogram
is the synthetic seismic trace. The process of seismic-to-well tie
* Corresponding author at: UFPA, Faculty of Geophysics, Belém, PA, Brazil.
E-mail addresses: isadora.s.macedo@gmail.com (I. Santana de Macedo),
cbarros.geof@gmail.com (C. Barros da Silva), jadsom@ufpa.br (J. de Figueiredo),
bodeomoboya@outlook.com (B. Omoboya).
involved mostly the comparison between the real seismic trace and
the synthetic trace (Rob Simm, 2014).
Seismic-to-well tie is a useful tool used to relate recorded seismic
waveforms to the lithology, stratigraphy and rock properties of the
subsurface (White et al., 1998). Accurate seismic-to-well tie is perti-
nent to a successful seismic lithological interpretation (White, 2003).
As long as the geology in the vicinity of the well is not unduly com-
plex, the main factors controlling accuracy of seismic-to-well tie are
the quality of the seismic processing as well as the accurate replica-
tion of the earth model from well logs. The link between a primary
reflection signal and the reflectivity constructed from a well log is the
seismic wavelet. In general, most studies of this subject agree that
methods to estimate the seismic wavelet are divided in two cate-
gories: deterministic methods and statistical methods. Deterministic
methods require direct measurements of the source wavefield or the
use of the well log data (Oldenburg et al., 1981; Yilmaz, 2000). Statis-
tical methods estimate the wavelet from the seismic trace itself and
require some assumptions about the characteristics of the wavelet
(Buland and Omre, 2003; Edgar and Der Baan, 2011; Lundsgaard
et al., 2015). The latter, is based on mathematical tools to solve

http://dx.doi.org/10.1016/j.jappgeo.2016.11.003
0009-2541/© 2016 Elsevier B.V. All rights reserved.
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314 299

and a random process reflectivity. We introduce a semi-automatic

Statistical Deterministic approach to determine the optimum parameters in deterministic
Estimation Estimation and statistical wavelet estimation, and further obtain the optimum
wavelets by searching for the best seismic-to-well tie. Tests with
Seismic Traces Seismic numerical data using our semi-automatic algorithm show the esti-
Traces Reflectivity mation of the seismic wavelet with a reasonable degree of accuracy
Time window for both cases. The feasibility of the algorithms is verified on the real
seismic and well data from Viking Graben field, North Sea, Norway.
PREDICTIVE Our results also show the influence of washout zones on the quality
FILTER
DECONVOLUTION of the seismic-to-well tie.
WAVELETS F Inc
α N
2. Seismic-to-well tie
SYNTHETIC TRACE
Seismic-to-well tie is an important part of a seismic interpreter’s
CORRELATION
THE BEST trade once they have the following prerequisites of 1) correctly iden-
tifying horizons to pick and 2) estimating the wavelet for inverting
ESTIMATED
HIGHEST CORRELATION seismic data to impedance (White and Simm, 2013). It is a basic
WAVELET
tool to estimate the connection of subsurface geology and seis-
mic. Borehole measurements such as sonic and density logs are
Fig. 1. A schematic illustration of the algorithm for the estimation of the wavelet that recorded in depth while seismic measurements are in time. To con-
produces the best well to seismic tie.
vert the borehole measurements from depth to time, a time–depth
relationship need to be established. This time–depth relationship is
usually acquired as checksots at the borehole location. In general, the
seismic-to-well tie workflow include the following steps:

(1) Edit the sonic and density logs.

the problem of wavelet estimation. Beyond these, there are other
(2) Generate a reflectivity series.
methods based on intelligent optimization algorithms used to esti-
(3) Apply a time–depth relationship.
mate the seismic wavelet (Yuan et al., 2009) and the seismic phase
(4) Convolve the reflectivity series with a wavelet.
wavelet (Wang et al., 2015), with a fixed filter length. Both works are
(5) Compare the output of the convolution with the real seismic
based on swarm intelligence inversion methods.
data.
In this paper we compare the quality of seismic-to-well tie based
on two different wavelet estimation methods. The first method is
the traditional deterministic method, which selects a segment of the The first step is necessary in order to avoid the introduction of
reflectivity sequence and a segment of the seismic data. The best noise or spikes to the generated reflectivity series. Both the sonic
wavelet estimated is the one that leads to the best match between and density curves may have some spikes and null values that needs
the seismic trace and the synthetic trace. The second method is based to be dealt with. The reflectivity series is generated by changes of
on classical predictive deconvolution assumptions about the convo- impedance I = qVp within the earth. The reflectivity (r(i)) on the
lutional model of the earth, which infers a minimum-phase wavelet depth axis can be calculated from the sonic and bulk density logs,

a) Intervalar Velocity b) Density c) Reflectivity Resampled

0 0 0

100 100
0.5
200 200

300 300 1
Depth (m)

Time (s)

400 400
1.5
500 500

600 600 2

700 700
2.5
800 800

900 900 3
0 2000 4000 1500 2000 2500 −0.1 0 0.1
[m/s] [kg/m3]

Fig. 2. a) P-wave log related to a synthetic model. b) Density log related to a synthetic model with 6 layers. c) Calculated reflectivity resampled to fit the time axis.
300 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

according to the equation or in matrix terms

qi+1 vi+1 − qi vi
r(i) = , (1) ⎛ ⎞
qi+1 vi+1 + qi vi r0 r1 r2 · · · rn−1 ⎛ ⎞ ⎛ ⎞
⎜ r1 ⎟ a0 g0
⎜ r 0 r1 · · · rn−2 ⎟ ⎜ a1 ⎟ ⎜ ⎟
⎜ r ⎟⎜ ⎟ ⎜ g1 ⎟
⎜ 2 r1 r0 · · · rn−3 ⎟⎜ a ⎟ ⎜ ⎟
⎜ . ⎟⎜ 2 ⎟=⎜ g2 ⎟, (4)
where i represents a sample on the depth axis. Since well logs ⎜ . .. .. .. .. ⎟⎜ . ⎟ ⎜ ⎟
⎜ . . . . . ⎟⎜ . ⎟ ⎜ .. ⎟
are more densely sampled than seismic traces, the next step is to ⎜ ⎟⎝ . ⎠ ⎝ . ⎠
⎝ rn−1 rn−2 rn−3 ··· r0 ⎠
resample the reflectivity, so it can compare adequately with seismic an−1 gn−1
traces. The newly sampled reflectivity in time domain (r(t)) can be
convolved with a seismic wavelet (w(t)) to create a synthetic seis-
mogram (s(t)), according to the equation of the convolutional model
given by
where ri represents the autocorrelation lags of the input wavelet,
ai is the Wiener filter coefficients, n is the length of the filter
coefficient, gi is the crosscorrelation lag of the desired output with
s(t) = w(t) ∗ r(t). (2)
the input wavelet. The autocorrelation matrix of the input wavelet
is a symmetric Toeplitz matrix that can be solved by Levinson
recursion. The demonstration of Eq. (4) is depicted in Appendix A.
This synthetic trace will now be further compared to the real The Wiener filter is optimum because the least squares error
seismic data. If there is a good set of logs, a well approximated between the actual and desired output is minimum. Moreover, this
wavelet and a good time–depth relationship, results of seismic-to- filter can be used to solving a large group of problems with desired
well tie should be good. The synthetic will be a good match to outputs such as, zero-lag spike (spiking deconvolution), spike at
the seismic, provided they have a similar frequency content. The arbitrary lag, zero-phase wavelet and time-advanced form of the
deterministic wavelet estimation method discussed in this paper input series. To ensure numerical stability when computing the filter
was through a filtering process in which the filter coefficient are operators, an artificial level of white noise is added to the data before
estimated and then convolved with the reflectivity to produces a the deconvolution, process called pre-whitening.
trace as similar to the real seismic trace as possible in terms of the
correlation coefficient. The statistical wavelet estimation method in
turn, was through predictive deconvolution. This deconvolution is
based on the classical assumptions about the convolutional model
3.1. Predictive deconvolution
whose theory is discussed in Yilmaz (2000). It is a special form of the
Optimum Wiener Filters, described below.
When the desired output on the normal equations is a time-
advanced form of the input series, a predictive process is suggested.
If a input series x(t) is given and the goal is to predict its value at
3. Optimum Wiener filters some future time x(t + a), where a is the prediction lag, a predictive
deconvolution process is suggested. Wiener showed that the filter
Determining the Wiener filter coefficients requires a solution to used to estimate x(t + a) is a special form of the general normal
these equations: equations, and it is given by


 ra+t = xt xt−(a+t) , (5)
ri−t at = gi , (3)
t

a) Causal wavelet b) Reflectivity c) Seismic Trace 1 d) Seismic Trace 2

−1 0 1 −0.1 0 0.1 −0.2 0 0.2 −0.2 0 0.2
0 0 0 0

0.5 0.5 0.5 0.5

1 1 1 1
Time (s)

Time (s)

Time (s)

Time (s)

1.5 1.5 1.5 1.5

2 2 2 2

2.5 2.5 2.5 2.5

3 3 3 3

Fig. 3. a) Causal wavelet used to produce the seismic trace. b) Reflectivity resampled to the time axis. c) Seismic trace produced by the convolution of the causal wavelet with the
reflectivity. d) Seismic trace produced by the convolution of the causal wavelet with the reflectivity with noise.
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314 301

or in matrix terms a non-minimum trace delay can be represented by the convolution

of an all-pass filter (which has a flat magnitude spectrum) and a
minimum-delay wavelet. Both the trace and the wavelet have the
same magnitude spectrum (the same color). It is well established that
⎛ ⎞
r0 r1 r2 · · · rn−1 ⎛ ⎞ ⎛ ⎞ white refers to anything that has a flat magnitude spectrum, whereas
⎜ r1 ⎟ a0 ra
⎜ r0 r1 · · · rn−2 ⎟ ⎜ a1 ⎟ ⎜ ra+1 ⎟ colored refers to anything that has a curved or non-white magnitude
⎜ r r1 r0 · · · rn−3 ⎟⎜ ⎟ ⎜ ⎟ spectrum. Thus, the trace model can be defined as follows:
⎜ 2 ⎟⎜ a ⎟ ⎜ ⎟
⎜ . .. .. .. ⎟⎜ 2 ⎟ = ⎜ ra+2 ⎟. (6)
⎜ . .. ⎟⎜ . ⎟ ⎜ .. ⎟
⎜ . . . . . ⎟⎜ . ⎟ ⎜ ⎟
⎜ ⎟⎝ . ⎠ ⎝ . ⎠ (1) Seismic trace = mixed-delay wavelet * reflectivity.
⎝ rn−1 rn−2 rn−3 ··· r0 ⎠
an−1 ra+n−1 (2) A mixed-delay wavelet has both a colored and a white
component: mixed-delay wavelet = minimum-delay
wavelet * all-pass wavelet.
That is the case for a n-long prediction filter and a a-long predic- (3) It is possible to rewrite the seismic trace as seismic trace =
tion lag. The prediction filter requires only the autocorrelation of the minimum-delay wavelet * (all-pass filter * reflectivity).
input series. The demonstration of Eq. (6) is depicted in Appendix A.
Jones et al. (2008) explained how the predictive deconvolution The convolution of the all-pass filter with the reflectivity makes
works. The canonical representation of a seismic trace states that up the white components of the seismic trace. The process of

(a)
Wavelets
1
Amplitude (a.u.)

Original Wavelet
0.5 Estimated Wavelet

−0.5

−1
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)
Real trace and synthetic trace for the model
1
Real
Amplitude (a.u.)

0.5 Synthetic

−0.5

−1
0 0.5 1 1.5 2 2.5 3
Time (s)

(b)
Wavelets
1
Original Wavelet
Amplitude (a.u.)

0.5 Estimated Wavelet

−0.5

−1
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)
Real trace and synthetic trace for the model
1
Real Synthetic
Amplitude (a.u.)

0.5

−0.5

−1
0 0.5 1 1.5 2 2.5 3
Time (s)

Fig. 4. a) Estimated deterministic wavelet for the modeled seismic trace with noise. The correlation coefficient between the synthetic trace and the real trace is 0.981. b) Estimated
deterministic wavelet for the modeled seismic trace without noise. The correlation coefficient between the synthetic trace and the real trace is 1.
302 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

predictive deconvolution can be separated into two components:
“minimum-delay” and “white”. The prediction error filter (the error
series of the predictive deconvolution) yields the white components
of the seismic trace, which is the convolution of the all-pass filter
and reflectivity. The prediction filter, which is the real output of
the predictive deconvolution, yields the predictable components of
the seismic trace, that is the “minimum-delay” component that
constitute the estimated wavelet.
4. Semi-automatic algorithm for wavelet estimation
Our semi-automatic algorithm to perform seismic-to-well tie
searches for the best wavelet that produces the highest correlation
coefficient between the synthetic trace and the real seismic trace.
Fig. 1 illustrates the procedure for the statistical and deterministic
estimation of the wavelet.
The first step needed to estimate the wavelet through the predic-
tive deconvolution (left side of Fig. 1) is to select a range of traces
in the vicinity of the well location. Then, the predictive deconvolu-
tion is performed in a time window within the selected traces. This
deconvolution is dependent on the prediction lag (a) and the oper-
ator length (N). The choice of the optimum values of prediction lag
and operator length will produce the best result for the predictive
deconvolution. For that reason, on the selected traces, the predictive
deconvolution is applied using a range of prediction lag and operator
length. Each combination a-N generates a wavelet for all the range
of CMPs that will be convolved with the reflectivity calculated from
the wells. The correlation coefficient between the real seismic trace
and all the synthetic traces generated with all the different wavelets

(a)
Wavelets
1
Original Wavelet
Amplitude (a.u.)

0.5 Estimated Wavelet

−0.5

−1
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)
Real trace and synthetic trace for the model
1
real synthetic
Amplitude (a.u.)

0.5

−0.5

−1
0 0.5 1 1.5 2 2.5 3
Time (s)

(b)
Wavelets
1
Original Wavelet
Amplitude (a.u.)

0.5 Estimated Wavelet

−0.5

−1
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)
Real trace and synthetic trace for the model
1
real synthetic
Amplitude (a.u.)

0.5

−0.5

−1
0 0.5 1 1.5 2 2.5 3
Time (s)

Fig. 5. a) The estimated wavelet by the predictive deconvolution for the modeled seismic trace with noise. The correlation coefficient between the synthetic and real traces is
0.911. b) The estimated wavelet by the predictive deconvolution for the modeled seismic trace without noise. The correlation coefficient between the synthetic and real traces is
0.956.
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314
are estimated. The algorithm identifies the highest correlation
coefficient and the trace (or corresponding CMP), the prediction
lag and the operator length that yields it and generates the best
statistical wavelet among all.
The other statistical alternative was to calculate the average
wavelet. In this case, the procedure mentioned before is done with
three or more different time windows of the seismic trace. When
the algorithm gives the best wavelet for each segment, an average
wavelet is calculated. This average wavelet is the one that will be
convolved with the reflectivity to generate the final synthetic seismic
trace.
The right side of Fig. 1 shows the well tie flow for the deter-
ministic wavelet estimation. The inputs of the algorithm are the
real seismic traces (from stacked or migrated sections) in the
vicinity of the well location and the reflectivity from the wells in
a corresponding scale. In this case, a filter is built such that the
convolution of the filter coefficients with the reflectivity, generates
a synthetic trace that is as close to the real seismic trace as possi-
ble. This algorithm is dependent on the length of the filter (F) and
an increment factor (Inc). This increment in time corresponds to the
best shift between the reflectivity series and the seismic trace. The
algorithm finds the best shift (Inc) that produces the best match
between the real trace and the synthetic trace. Each combination of
filter length and increment generates a wavelet for all the range of
traces. Each wavelet will be convolved with the reflectivity calcu-
lated from the wells and then the correlation between the synthetic
and the real trace is performed. The algorithm identifies the highest
correlation coefficient and the trace, the filter length and the incre-
ment that generated it and produce the best deterministic wavelet
among all.
The best wavelets calculated through the algorithms are the ones
used to generate the synthetic traces for the seismic-to-well tie.
The algorithm is not entirely automatic in the estimation of the
best wavelet because of the statistical estimation stage, it is nec-
essary to carefully select a segment of the seismic trace in which
the predictive deconvolution will be performed. It is recommended
Fig. 6. Time migrated and interpreted seismic section of the Viking Graben field with the main geological structures.
Source: From Monroe et al. (2015)
303
that this segment represents a reflection signal from an interface of
interest, so that the predictive deconvolution can separate the
reflectivity component from the wavelet component. As for
the deterministic estimation, the process is semi-automatic because
the seismic trace and the reflectivity are in corresponding domains
(time domain) and are not created by the algorithm, the domain
conversion is created by the time–depth relationship. The advan-
tages of this method is that several wavelets are generated and
have the correlation coefficient tested and only the ones that
produce the best match are selected. The implementation is not
computationally expensive, even when using a large range of
traces.
5. Application on synthetic data
The methodologies of seismic wavelet estimation presented
early, were tested first in a synthetic model consisting of six layers,
which logs are depicted in panels a and b in Fig. 2.
The reflectivity was calculated according to Eq. (1) and then
resampled to the time domain. We used a function to resample the
reflectivity in depth to a new fixed rate, according to the length of the
seismic trace in time. To avoid additional spikes due to the resam-
pling process on the reflectivity in the time domain, we applied a
low pass filter. The total length of signal is 3 s and the sample rate is
0.004 s. Fig. 2c shows the corresponding reflectivity of the model
depicted in Fig. 2a and b. We choose a causal wavelet with a peak
frequency of 20 Hz to compute the synthetic trace, as our method
deals with the classical assumptions of the convolutional model. Two
seismic traces were modeled, one with white noise and one without
white noise, according to Fig. 3.
5.1. Application of the deterministic extraction of the wavelet
Concerning the deterministic wavelet extraction, the inputs to
the algorithm are the reflectivity series, the seismic trace, the length
of the wavelet (which is the length of the filter) and the increment.
304 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

It builds a vector of filter coefficients that when convolved with the
reflectivity yields a synthetic trace that produces the best match
with the real seismic trace, given the filter length and the incre-
ment. Therefore, the filter is the wavelet estimated. We selected a
range of filter length and increment to perform the extraction of the
wavelet.
The similarity between the synthetic trace and the real trace show
that this approach was very effective for estimating the wavelet for
the synthetic layer model. For the modeled seismic trace with noise,
the correlation coefficient between the synthetic and the real trace,
was 0.981 (see Fig. 4a). For the seismic trace without noise, the
wavelet extracted was exactly the same as the causal wavelet used to
generate the real trace. The correlation coefficient was 1 (see Fig. 4b).
In Fig. 4b the algorithm returns several pairs of filter length (F) and
the time increment (INC). We can use the shorter filter length pro-
vided by the algorithm to estimate the best wavelet. The estimated
wavelets depicted in panels a and b in Fig. 4 are similar to that esti-
mated by Yuan et al. (2009) using the optimization method based on
ant colony optimization (ACO) algorithm.

(a)
P−Wave Velocity − Well A RHOB − Well A Reflectivity − Well A

(a)
1500 1500 1500 CMP x Correlation Coeficient
0.7

Correlation Coeficient
Depth [m]

2000 2000 2000 0.65

0.6
2500 2500 2500
0.55

3000 3000 3000

0.5
2000 4000 6000 1000 2000 3000 −0.1 0 0.1
[m/s] 3 0.45
[Kg/m ]
804 806 808 810 812
CMP
(b)
P−Wave velocity − Well B RHOB − Well B Reflectivity − Well B (b)
Prediction lag X Correlation coeficient for a fixed CMP and operator length N
2800 2800 2800 0.8
Correlation Coeficient

3000 3000 3000

0.6
Depth [m]

3200 3200 3200

0.4
3400 3400 3400

0.2
3600 3600 3600

3800 3800 3800 0

0 5000 0 2 4 −0.2 0 0.2 0 10 20 30 40 50 60 70
[m/s] [g/cm3] Prediction lag

(c) (c)
Time−Depth relationship for Well A and Well B Operator length X Correlation coeficient for a fixed CMP and prediction lag
0.8
Time (ms)
1000 1500 2000 2500 3000 3500 4000
Correlation Coeficient

1000 0.6
Well A
Well B
1500
0.4
Depth (m)

2000

2500 0.2

3000
0
0 50 100 150
3500 Operator length N

Fig. 7. Logs used to construct the synthetic seismogram at well A (a) and well B (b). Fig. 8. a) Correlation coefficient×CMPs for the well A using the predictive decon-
From left to right: sonic log, density log and reflectivity log. The red lines are the output volution to estimate the wavelet. The best match location is on CMP 809 with a
from the despiking process. c) Time picks of the VSP direct arrivals that constitute the correlation coefficient of 0.658. b) Prediction lag×correlation coefficient for the CMP
time–depth relationship. 809. c) Operator length×correlation coefficient for the CMP 809.
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314
5.2. Application of the predictive deconvolution
The predictive deconvolution aims to separate the “white” com-
ponents from the “minimum-delay” components of the trace which
contains the wavelet. We selected a segment of the trace that is
a good representation of the reflector of interest, for this case, we
choose the segment between 0.28 s and 0.41 s. That segment is
deconvolved in order to separate the white components from the
wavelet, by choosing a proper prediction lag and the operator length.
We selected a range of operator length and a range of prediction lag
to perform the deconvolution and then estimated the wavelet. The
algorithm calculates the prediction lag and the operator length that
produces the best match between the synthetic trace and the real
trace. Using those parameters, we can generate the best wavelet to
convolve with the reflectivity. Panels a and b in Fig. 5 show the com-
parison between the traces. As it can be seen, they are similar even
in amplitude magnitude.
The deterministic method produces a more accurate wavelet than
the statistical method by the predictive deconvolution. This happens
because the deterministic method uses both well log and seismic
data. The statistical method makes assumptions about the charac-
teristic of the wavelet (minimum-phase) and the reflectivity of the
earth (random process), therefore generating a statistical model that
is purely based on mathematics. The input data for the deterministic
case is the entire trace and for the statistical case is a segment of the
seismic trace, which ranged from 0.28 s to 0.41 s. This difference of
accuracy between the two methods is also expected when applying
the algorithms to real data set.
6. Results on Viking Graben dataset
The real data used in this study comes from the northern North
Sea basin, Viking Graben. It is a north–south trending linear trough
straddling the boundary between the Norwegian and UK sectors of
the northern North Sea (Madiba and McMechan, 2003).
The North Viking Graben formed as a result of late Permian to
Triassic rifting; extensional episodes and accompanying sedimenta-
tion continued through the Jurassic into the Early Cretaceous. The
extensional episode in the beginning of Jurassic causes the Pangea
to break into two continents, Gondwana and Laurasia. During this
period the sea level rose. The burial of algae and bacteria below the
mud of the seafloor resulted in the formation of the oil and gas in the
1
Amplitude (a.u.)
−1
0 0.05
1
Amplitude (a.u.)
real synthetic
−1
1.4
Fig. 9. The wavelet estimated through the predictive deconvolution and the synthetic and real migrated seismic traces on CMP 809. The correlation coefficient is 0.658.
305
North Sea. The Jurassic transgressive system has periods of regres-
sion that provided the coarse clastic input that forms the reservoir
intervals separated by deep water shales (Madiba and McMechan,
2003). The Jurassic was a period of active faulting; hydrocarbon traps
are usually fault-bounded structures, but some are associated with
stratigraphic truncation at the unconformity at the base of the Creta-
ceous (Keys, 1998). This unconformity — referred as BCU — is located
above the Jurassic synrift sediments that are also overlain by Cre-
taceous and Tertiary basin fill. The Paleocene reservoir in the North
Viking Graben is characterized by deep water clastics. The Paleocene
interval is undisturbed by the rift tectonism and dips gently into
the basin (Keys, 1998). The sediments were deposited in a slope
environment and so contains turbidites (Madiba and McMechan,
2003).
6.1. Seismic and well logs data
The seismic line consists of 1001 shot records, each shot recorded
on 120 channels for 6 s. The sample rate is 4 ms. An air gun was
used as the seismic source. The seismic data has 2142 CMP loca-
tions with 1501 samples each. The well log information is from two
wells designed well A and well B located on seismic lines CMP 808
and CMP 1572, respectively. According to Madiba and McMechan
(2003), the major base Cretaceous unconformity (BCU) is located at
approximately 1.9 s in well A and at 2.4 s in well B. Fig. 6 shows
the interpreted seismic section migrated in time by Monroe et al.
(2015).
Due to unrecorded points and noises in the density and sonic logs,
it was necessary to edit them for better accuracy of results. Because
of that, a despike process was applied on both well logs (see panels a
and b in Fig. 7). As for well B density log information was insufficient,
it was completed using the Gardner’s relation Gardner (1974) that
relates compressional velocity with the bulk density of the lithology
where the wave travels. It is given by,
q = 0.31Vp0.25 , (7)
where Vp is in m/s and q in g/cm3 .
Three-component zero-offset vertical seismic profile were
recorded in wells A and B (Keys, 1998). In this case, the Vertical
Seismic Profiles (VSPs) are useful in the seismic-to-well tie process
because they provide a link between wells and seismic at the correct
Estimated wavelet
0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time (s)
Real seismic trace and synthetic trace
1.6 1.8 2 2.2 2.4 2.6 2.8
Time (s)
306 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

domain. The time–depth relationship in this study was generated by
time picks of the VSP direct arrivals. Fig. 7c shows the time–depth
relationship for wells A and B.
It is possible to measure the quality of the tie through the correla-
tion coefficient. According to White (2003), the best match location
(higher correlation coefficient of the synthetic seismogram with the
real seismic trace from a CMP) often does not occur at the well loca-
tion. If the seismic data was time migrated, since velocity typically
increases with depth, time migration would commonly move to the
best match location, up-dip from the well. The best match location
is found by scanning a number of traces around the well for the
correlation coefficient (Rob Simm, 2014).
On the well to seismic tie from the work of White (2003)
the best match location of the well tie was about 100 m distant
from the well, with a correlation coefficient between the seismic
trace and the synthetic trace equal to 0.790. The well to seismic
tie in this work was performed on both well logs A and B with
a time migrated seismic section. The processing of the seismic
section used in this work is described in Gonzales et al. (2016). The
reflectivity series used to create the synthetic seismogram was the
same for both deterministic and statistical method to estimate the
wavelet.
6.2. Migrated seismic section — well A
The first analysis conducted in this study was to compare the
quality of the seismic-to-well tie made with statistical method
versus deterministic method in the vicinity of well A. We followed
the recommendation of White (2003) and selected traces on the
vicinity of the posted well location (CMP 808) to get the best match.
The results are shown in Figs. 8 and 9. The correlation coefficient
using the statistical estimation of the wavelet was 0.658. In order to
have a frequency content on the synthetic trace similar to the real
seismic trace, we compute an average wavelet. It was selected from

(a)
Estimated wavelets from each segment and the average wavelet
0.3
Seg1
Seg2
0.2 Seg3
Seg4
0.1 Average
Amplitude (a.u.)

−0.1

−0.2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Time (s)

(b)
Estimated average wavelet
0.2
Amplitude (a.u.)

−0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (s)
Real seismic trace and synthetic trace
1
Amplitude (a.u.)

real synthetic
0.5

−0.5
1.4 1.6 1.8 2 2.2 2.4 2.6 2.8
Time (s)

Fig. 10. a) The wavelets estimated through the predictive deconvolution from different segments of the seismic trace. b) The average statistical wavelet estimated and the
synthetic and real traces for the CMP 809. The correlation coefficient is 0.674.
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314 307

four different segments of the seismic trace to perform the predictive
deconvolution and consequently four wavelets were estimated. The
time range of the first segment ranged from 1.38 s to 1.6 s, the second
from 1.66 s to 1.88 s, the third from 1.88 s to 2.10 s and the last one
from 2.30 s to 2.56 s. It is worth to mention, that the size of the
wavelet is related to the operator length of predictive deconvolution.
In our case, the choice of this operator over the selected time window
was performed by the algorithm.
These wavelets as well as the average wavelet are shown in
Fig. 10. This approach provided better results on the quality of the tie.
The correlation coefficient increased from 0.658 (using the predic-
tive deconvolution of one segment alone — see Fig. 9) to 0.674 when
using the average statistical wavelet as shown in Fig. 10. The results
of the seismic-to-well tie using the deterministic method to estimate
the seismic wavelet is shown in Fig. 11. The highest correlation
coefficient was 0.712 also on CMP 809.
As our semi-automatic method determines the CMP, as well as
the operator length and the prediction lag by searching for the
highest correlation coefficient to further estimate the wavelet, we
depicted the results of the search in Fig. 8. The fixed values of
operator length and prediction lag in panels a and b in Fig. 8 are the
ones that produced the highest correlation. Those values for the well
A were N = 14 and a = 13.
6.3. Migrated seismic section — well B
Applying the statistical predictive deconvolution method to esti-
mate the wavelet along with informations of the well B yields to a
best match location on CMP 1573. The correlation coefficient of the
tie was 0.741, as shown in Figs. 12 and 13. We also estimated the
average statistical wavelet for the well B. The correlation coefficient

(a)
CMP x Correlation Coeficient
0.71

0.7

0.69
Correlation Coeficient

0.68

0.67

0.66

0.65

0.64

0.63
804 805 806 807 808 809 810 811 812
CMP

(b)
Estimated wavelet
1
Amplitude (a.u.)

−1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (s)
Real seismic trace and synthetic trace − CMP 809
1
Amplitude (a.u.)

real synthetic
−1
1.4 1.6 1.8 2 2.2 2.4 2.6 2.8
Time (s)

Fig. 11. a) Correlation coefficient×CMPs for the well A through the deterministic method to estimate the wavelet. The best match location is also on CMP 809 with a correlation
of 0.712. b) The deterministic estimated wavelet and the comparison of the synthetic and real traces on CMP 809.
308 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

(a) The best match location for the well B using the deterministic
CMP x Correlation Coeficient estimative of the wavelet was also on CMP 1573 (see Fig. 15). It was
0.75 obtained the best result on the quality of the well to seismic tie, a
correlation of 0.811 between the real migrated seismic trace and the
synthetic trace, using the deterministic wavelet estimation. Fig. 16
Correlation Coeficient

0.7
shows the synthetic seismic traces located on the migrated seismic
section, as shown in Fig. 6.
0.65

0.6
7. Discussion

0.55
In this study we compared the quality of seismic-to-well tie made
with two different methods to wavelet estimation. Besides that,
0.5 we observed that washout zones can lead to a lower correlation
1568 1570 1572 1574 1576
CMP coefficient between the synthetic and real seismic traces. It is
expected that the average wavelet from the predictive deconvo-
(b) lution produces better statistical results because it was computed
Prediction lag X Correlation coeficient for a fixed CMP and operator length N from wavelets estimated from different segments of the seismic
0.8 trace, therefore, its frequency content tends to be similar to the real
seismic trace. That was evident in well A. The correlation coefficient
Correlation Coeficient

increased from 0.658 to 0.674. As for well B, however, the average

0.6
estimated wavelet produced a correlation coefficient slightly lower
(correlation of 0.734) than the predictive deconvolution using one
0.4 segment alone (correlation of 0.741).
The deterministic wavelet estimation tends to produce better
results for seismic-to-well tie, once it is based on the well log and
0.2 seismic data. However, the need to match well logs and seismic
data domains, makes it more sensitive to errors on the time–depth
0
relationship. For both well logs, the deterministic estimation of the
0 10 20 30 40 50 60 70 wavelet led to the best correlation coefficients as can be seen in
Prediction lag Figs. 11 and 15.
Frequency attenuation caused by the intrinsic attenuation in
(c) rocks causes loss of high frequencies in the propagating waveform
Operator length X Correlation coeficient for a fixed CMP and prediction lag
with increasing travel-time. This gives rise to a non-stationary
0.75
behavior in the shape of the wavelets associated with reflec-
tion events at different times. The wavelet estimated from both
Correlation Coeficient

0.7 methods are considered to be stationary, therefore, it may be a

0.65
0.6
0.55
0 50 100
Operator length N
Fig. 12. a) Correlation coefficient×CMPs for the well B using the predictive deconvo-
lution method to estimate the wavelet. The best match location is on CMP 1573, with
a correlation coefficient of 0.741. b) Prediction lag×correlation coefficient for the CMP
1573. c) Operator length×correlation coefficient for the CMP 1573.
in this case was slightly lower than the predictive deconvolution on
one segment alone, as shown in Fig. 14.
For the average wavelet, the correlation coefficient of the well
tie was 0.734. We selected three segments of the seismic trace to
deconvolve and then estimate the average statistical wavelet. The
first time range was from 2.57 s to 2.79 s, the second was from 2.86 s
to 3.04 s and the last one from 2.87 s to 3.08 s. We also depicted the
results of the search for the optimum parameters of the statistical
estimation of the wavelet for well B in Fig. 12. The best opera-
tor length and prediction lag for this case were N = 2 and a =
37.
factor that contribute to decrease the correlation coefficient val-
ues of the results shown. Moreover, the real seismic data used
shows characteristics of faults and fractures, since it is from a host-
graben structure, which contributes even more to the frequency
attenuation.
The correlation coefficient between the seismic trace and the syn-
thetic for the well A is lower than for the well B. A possible cause
for that difference is because of acquisition problems on the well A
150 (Keys, 1998). To solve that problem, a VSP from a nearby well was
substituted for the well A vertical seismic profile. This may affect
the time–depth relationship for this well, which causes the esti-
mated wavelet, especially for the deterministic method, to not be as
accurate as it could be.
According to Keys (1998) and Madiba and McMechan (2003)
along the well A, an unconformity occurs at the base of the
Cretaceous, at 1.97 s, separating rocks from the Tertiary and from
the Jurassic. This event is visible on the synthetic trace for well A.
Fig. 17 shows a possible lithological interpretation for the wells A and
B based on the information provided by Gleinnie and Parsely (1986),
Keys (1998), and Madiba and McMechan (2003) and the gamma ray
log on the time axis. The heavily faulted upper Jurassic formations
(Kyrkjebo et al., 2004) begin just below the BCU unconformity in
well A at approximately 1.97 s and in well B at approximately
2.55 s. Therefore, due to the structural characteristics of this
formation, we do not expect a high quality seismic-to-well tie.
Well B does not contain Tertiary formations once it begins below
the BCU.
Other important point are the results of the correlation coefficient
of the seismic-to-well tie on different segments of the trace. Fig. 18
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314
1
Amplitude (a.u.)
0
−1
0 0.05
1
Amplitude (a.u.)
−1
2.5 2.6
Fig. 13. The wavelet estimated through the predictive deconvolution and the synthetic and real seismic traces on CMP 1573. The correlation coefficient is 0.741.
shows that the quality of the well tie for both wells is low on
the initial parts of the trace. To understand the possible reason
for that difference of quality of the well tie in different segments
of the trace, we analyzed together with the gamma ray and the
caliper logs. Gamma ray logs are indicators of shale content on
the formation and caliper logs describe the diameter of the bore-
hole along its depth during drilling. The measurements that are
recorded can be an important indicator of cave-ins or shale swelling
in the borehole, which can affect the results of the sonic and density
logs. To perform this analysis, we chose the best ties for well A and
well B.
Fig. 18a shows that for well A the segment with the lower cor-
relation coefficient c1 = 0.409 corresponds to the zone where
the caliper log is unstable, which directly affects the sonic and
density measurements and, consequently, directly affects the qual-
ity of the well to seismic tie. Keys (1998) mentioned problems
during the acquisition of data along the well A, which might be
related with this anomalous zone on the caliper log. This first
segment is related with the lithology of the formation above the
Cretaceous unconformity at 1.97 s. The Paleocene and Cretaceous
rocks above the unconformity are deep water clastic sediments with
the presence of turbidites. On the second c2 = 0.805 and third
segment c3 = 0.839, it is possible to note that when the caliper log
gets more stable, the correlation increases although the shale con-
tent also increases due to the Jurassic rocks associated with deep
water shales.
Fig. 18b shows the relations between the gamma ray log, the
caliper log and the well to seismic tie for the well B. As the Paleocene
and Cretaceous rocks are deposited in a slope in the basin, the Creta-
ceous unconformity that on well A is at approximately 1.97 s, on well
B it is approximately at 2.4 s, so the beginning of the synthetic trace
for the well B is just below the unconformity. Therefore, it covers
the Jurassic rocks. For this well, although the caliper log is relatively
stable along the entire trace, the first segment of the trace shows a
lower correlation coefficient. This might be due to the faults associ-
ated with the Cretaceous unconformity that constitute the oil and gas
traps and the deep water shales associated with it, as mentioned by
Madiba and McMechan (2003) and as can be seen by the gamma ray
log. As the shale content tends to decrease and the caliper log gets
309
Estimated wavelet
0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time (s)
Real seismic trace and synthetic trace
real synthetic
2.7 2.8 2.9 3 3.1 3.2
Time (s)
stable, the quality of the well tie increases and the excellent result
of a correlation of c2 = 0.988 for the second part of the trace was
obtained.
8. Conclusions
The application of our semi automatic algorithm to estimate
the wavelet on the synthetic data shows that both methods were
satisfactory, with values of correlation coefficient higher than 0.90.
In the Viking Graben data set, the predictive deconvolution wavelet
estimation method produced a good seismic-to-well tie for both
wells. Both had correlations above 0.60. Moreover, the average statis-
tical wavelet showed slightly improved results on the quality of the
tie. The deterministic estimation of the wavelet produced the best
results. The highest correlation was of 0.712 on well A and 0.811 on
well B.
The influence of the shale content and the stability of the diam-
eter of the borehole during acquisition was analyzed. Anomalies
on the caliper logs can directly affect the quality of the seismic-
to-well tie and consequently, the estimated wavelet. It was also
possible to relate the lithology of the formations with the results
of the tie. Zones with faults and increasing shale content also
showed the possibility of artifacts affecting the quality of the well
tie.
Because of the acquisition problems on the well A that might have
compromised the time–depth relationship, it would be interesting to
try more accurate statistical methods to estimate the wavelet for this
well, such as the homomorphic deconvolution, that does not imply a
minimum phase wavelet and a white reflectivity.
Acknowledgments
The authors would like to thank the Exxon Mobil for providing
the Viking Graben dataset. This work was kindly supported by the
Brazilian agencies CAPES and CNPq (459063/2014-6). The authors
would like to thank the reviewers for many helpful comments and
suggestions.
310 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

(a)
Estimated wavelets from each segment and the average wavelet
0.5
Seg1
Seg2
Seg3
Average
Amplitude (a.u.)

−0.5
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)

(b)
Estimated average wavelet
Amplitude (a.u.)

0.2

−0.2

0 0.05 0.1 0.15 0.2 0.25 0.3

Time (s)
Real seismic trace and synthetic trace
1
Amplitude (a.u.)

real synthetic

−1
2.5 2.6 2.7 2.8 2.9 3 3.1 3.2
Time (s)

Fig. 14. a) The wavelets estimated through the predictive deconvolution from different segments of the migrated seismic trace. b) The average wavelet estimated and the
synthetic and real seismic traces for the CMP 1573. The correlation coefficient is 0.734.

Appendix A. Optimum Wiener filters Substituting Eq. (A.2) into Eq. (A.1), it gives

The following discussion of optimum Wiener filters is based on

2
 
Yilmaz (2000). The Wiener filtering is optimum at the least squares L= dt − ft xt−t , (A.3)
sense. In other words, it involves the designing of a filter f(t) so that t t
the least squares error between the actual and the desired output is
minimum. The error L is defined as and expanding the square term, we have


2
    
L= (dt − yt )2 , (A.1) L= d2t −2 dt ft xt−t + ft xt−t . (A.4)
t t t
t t t

where dt is the desired output of the filtering process and yt is the As the goal is to find the filter coefficients ( f0 , f1 , f2 , . . . , fn−1 ) so
actual output. The actual output is the convolution of the filter ft with that the error L is minimum, the variation of L with respect to fi is set
the input xt : to zero:

∂L
yt = ft ∗ xt = ft xt−t . (A.2) = 0, i = 0, 1, 2, 3, · · · , (n − 1), (A.5)
∂ fi
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314 311

(a)
CMP x Correlation Coeficient
0.9

0.85

Correlation Coeficient
0.8

0.75

0.7

0.65
1568 1569 1570 1571 1572 1573 1574 1575 1576
CMP

(b)
Estimated wavelet
1
Amplitude (a.u.)

−1
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (s)
Real seismic trace and synthetic trace
1
real
Amplitude (a.u.)

synthetic
0

−1
2.5 2.6 2.7 2.8 2.9 3 3.1 3.2
Time (s)

Fig. 15. a) Correlation coefficient×CMPs for the well B using the deterministic method to estimate the seismic wavelet. The best match location is on CMP 1573. b) The wavelet
estimated through the deterministic method and the synthetic and real seismic traces from the migrated seismic section on CMP 1573. The correlation coefficient of the well tie
is 0.811.

∂L   
= −2 dt xt−i + 2 ft xt−t xt−i = 0, (A.6) and substituting those relations on Eq. (A.7), it gives
∂ fi t t t

   ri−t ft = gi . (A.10)
xt−t xt−i ft = dt xt−i . (A.7)
t
t t t

By writing Putting Eq. (A.10) into matrix form, it leads to

 ⎛ ⎞
xt−t xt−i = ri−t , (A.8) r0 r1 r2 · · · rn−1 ⎛ ⎞⎛ ⎞
⎜ r1 ⎟ f0 g0
t ⎜ r0 r1 · · · rn−2 ⎟⎜ ⎟ ⎜ g1 ⎟
⎜ r ⎟⎜ f1 ⎟ ⎜ ⎟
⎜ 2 r1 r0 · · · rn−3 ⎟⎜ ⎟ ⎜ ⎟
⎜ . ⎟⎜ f2 ⎟ = ⎜ g2 ⎟. (A.11)
⎜ . .. .. .. ⎟⎜ ⎟ ⎟
and
⎜ . . .
..
. . ⎟⎜ .. ⎟ ⎜ ⎜ . ⎟

⎜
⎝ rn−1 rn−2 rn−3
⎟⎝
⎠ . ⎠ ⎝ .. ⎠
··· r0
dt xt−i = gi , (A.9) fn−1 gn−1
t
312 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

(a)
Synthetic trace on the seismic section − CMP 809
1

1.5

2
Time (s)

2.5

3.5
800 802 804 806 808 810 812 814 816 818 820
CMP

(b)
Synthetic trace on the seismic section − CMP 1573
2

2.5
Time (s)

3.5
1562 1564 1566 1568 1570 1572 1574 1576 1578 1580 1582 1584
CMP

Fig. 16. a) Synthetic seismic trace from well A located on the migrated seismic section on CMP 809. b) Synthetic seismic trace from well B located on the migrated seismic section
on CMP 1573.

Eq. (A.11) configures the normal Wiener equations, where ri are where
the autocorrelation lags of the input and gi are the lags of the cross-
correlation between the desired output and the input. The Wiener
filter applies to a large class of problems in which any desired out- 
put can be considered. The deconvolution problems can be classified rt = xt xt−t . (A.13)
t
according to the desired output dt considered.

A.1. Predictive deconvolution

For the a + t lag as in Eq. (A.12), Eq. (A.13) becomes
For the predictive deconvolution, the desired output is a time
advanced form of the input signal. If the input signal is x(t), the
desired output is dt = x(t+a). To design a filter that predicts x(t+a) 
ra+t = xt xt−(a+t) = gt . (A.14)
from x(t), it is necessary to compute the crosscorrelation between
t
the desired output x(t + a) with the input signal x(t). According to
Eq. (A.9), the crosscorrelation for this case will be

   Substituting this result on the general normal Wiener equations

gt = dt xt−t = xt+a xt−t = xt xt−(a+t) , (A.12)
(Eq. (A.11)), the normal equations that must be solved to find the
t t t
 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314 313

(a)

Paleocene

Upper Jurassic

Middle Jurassic

(b)

Upper Jurassic

Middle Jurassic

Lower Jurassic

Fig. 17. a) A possible lithological interpretation for the well A. Formations from the Tertiary and Jurassic are present. b) A possible lithological interpretation for the well B. Since
it begins below the BCU, it contains only Jurassic formations.
314 I. Santana de Macedo et al. / Journal of Applied Geophysics 136 (2017) 298–314

(a) Rob Simm, M.B., 2014. Seismic Amplitude: An Interpreter’s Handbook. Cambridge
Gamma Ray − Well A Caliper − Well A Real and Synthetic Traces − CMP 809 University Press.
c = 0.712 Wang, S.X., Yuan, S.Y., Ma, M., Zhang, R., Luo, C.M., 2015. Wavelet phase estimation
1.4 1.4 1.4 using ant colony optimization algorithm. J. Appl. Geophys. 122, 159–166.
1.6 1.6 1.6 White, R., 2003. Tying well-log synthetic seismograms to seismic data: the key
c1 = 0.409 factors. SEG Technical Program Expanded Abstracts. Society of Exploration
1.8 1.8 1.8
Geophysicists. pp. 2449–2452.
Time (s)

Time (s)

Time (s)
2 2 2
c2 = 0.805 White, R., Simm, R., Xu, S., 1998. Well tie, fluid substitution and AVO modelling: a
2.2 2.2 2.2 North Sea example. Geophys. Prospect. 46 (3), 323–346.
c3 = 0.849 White, R.E., Simm, R., 2013. Tutorial: good practice in well ties. First Break 21 (10).
2.4 2.4 2.4
Yilmaz, O., 2000. Seismic Data Analysis: Processing, Inversion, and Interpretation of
2.6 2.6 2.6
Seismic Data. SEG Books.
2.8 2.8 2.8
real synthetic
Yuan, S.Y., Wang, S.X., Tian, N., 2009. Swarm intelligence optimization and its
50 100 150 200 200 400 600 −1 0 1
API mm Amplitude (a.u) application in geophysical data inversion. Appl. Geophys. 6 (2), 166–174.
(b)
Gamma Ray − Well B Caliper − Well B Real and Synthetic Traces − CMP 1573
2.5 2.5 2.5
c = 0.811

2.6 2.6 2.6

c1 = 0.623
2.7 2.7 2.7
Time (s)

Time (s)

Time (s)

2.8 2.8 2.8

c2 = 0.988
2.9 2.9 2.9

3 3 3

3.1 3.1 3.1

real synthetic
3.2 3.2 3.2
50 100 150 200 250 100 200 300 400 −1 0 1
[API] mm Amplitude (a.u.)

Fig. 18. a) Well tie for the well A with the migrated seismic section and the
deterministic estimated wavelet and the gamma ray and caliper logs. b) Well
tie for the well B with the migrated seismic section and the deterministic
estimated wavelet and the gamma ray and caliper logs.

prediction filter ( f0 , f1 , f2 , f3 . . .) is set. In this case, note that to

find the prediction filter, it is only needed the input signal.
⎛ ⎞
r0 r1 r2 · · · rn−1 ⎛ ⎞ ⎛ ⎞
⎜ ⎟ f0 ra
⎜ r1 r0 r1 · · · rn−2 ⎟⎜ ⎟ ⎜ ra+1 ⎟
⎜ ⎟⎜ f1 ⎟ ⎜ ⎟
⎜ r2 r1 r0 · · · rn−3 ⎟⎜ ⎟ ⎜ ⎟
⎜ . ⎟⎜ f2 ⎟ = ⎜ ra+2 ⎟. (A.15)
⎜ . .. .. . . .. ⎟⎜
⎜ . . . . . ⎟⎜ .. ⎟
⎟
⎜
⎜ . ⎟
⎟
⎜
⎝ rn−1 rn−2 rn−3 · · ·
⎟⎝
⎠ . ⎠ ⎝ .. ⎠
r0
fn−1 ra+n−1

References

Buland, A., Omre, H., 2003. Bayesian wavelet estimation from seismic and well data.
Geophys. 68 (6), 2000–2009.
Edgar, J.A., Der Baan, M.V., 2011. How reliable is statistical wavelet estimation?
Geophys. 76, 59–68.
Gardner, G.H.F., 1974. Formation velocity and density — the diagnostic basics for
stratigraphic trap. Geophys. 39, 770.
Gleinnie, K.W., Parsely, J.A., 1986. Introduction to the Petroleum Geology of the North
Sea. second, Blackwell Science Inc, Oxford Oxfordshire; Boston.
Gonzales, J.A.C., Figueiredo, J.J.S., Coimbra, T.A., Schleicher, J., Novais, A., 2016. Migra-
tion velocity analysis using residual diffraction moveout: a real-data example. J.
Geophys. Eng. 13, 622–633.
Jones, I.F., Bloor, R.I., Biondi, B., Etgen, J.T., 2008. Digital Imaging and Deconvolution:
The ABCs of Seismic Exploration and Processing Society of Exploration. OK, Tulsa.
Keys, R.G., 1998. Comparison of Seismic Inversion Methods on a Single Real Data Set,
Open File Publications, Society of Exploration, OK, Tulsa.
Kyrkjebo, R., Gabrielsen, R.H., Faleide, J.I., 2004. Unconformities related to the
Jurassic–Cretaceous synrift–post-rift transition of the northern North Sea. J. Geol.
Soc. 161 (1), 1–17.
Lundsgaard, A.K., Klemm, H., Cherret, A.J., 2015. Joint Bayesian wavelet and well-path
estimation in the impedance domain. Geophys. 80 (2), M15–M31.
Madiba, G.B., McMechan, G.A., 2003. Processing, inversion, and interpretation of a
2D seismic data set from the North Viking graben, North Sea. Geophys. 68 (3),
837–848.
Monroe, J., de Figueiredo, J., Santos, L., 2015. Seismic attributes analysis for the
identification and characterization of hydrocarbon reservoirs — revisiting the
Viking Graben dataset. 14th International Congress of the Brazilian Geophysical
Society & EXPOGEF, Rio de Janeiro, Brazil, 3–6 August 2015. SEG Global Meeting
Abstracts. Brazilian Geophysical Society. pp. 893–897.
Oldenburg, D., Levy, S., Whittall, K., Nov. 1981. Wavelet estimation and deconvolution.
Geophys. 46 (11), 1528–1542.