You are on page 1of 4

Introduction

Reliable seismic-to-well tying is a crucial step in seismic interpretation to correlate subsurface geology
to observed seismic data. The traditional tying procedure often starts with a statistical estimation of a
wavelet from the seismic data. This wavelet is then convolved with the reflectivity calculated from the
well logs (sonic log and bulk density log) to generate the first synthetic trace. With these two traces,
the interpreter finds the best matching reflectors by visually tying two time series. Following these steps
does not guarantee the correct tie (Anderson and Newrick, 2008) because the entire process is prone
to pitfalls due to subjectivities in interpretation and procedures.
In this paper we compare two nonlinear approaches to automatically tie well logs to seismic data. Our
procedures substitute the manual stretching and squeezing step by an optimization algorithm, which
is still supervised by the interpreter. The first alternative to perform the automated tying is based on
Dynamic Time Warping (DTW), first introduced by Herrera and Van der Baan (2012) and extended
and completed in this work. The second approach solves the nonlinearity correction using the local
similarity attribute (LSIM) (Fomel, 2007a). Both techniques share the quality control step by monitoring
the relative velocity change produced by the tying.

Theory
The correlation coefficient is the most common similarity metric between two time series at and bt :
c( ) =

n
at bt
t=1
,
n
n
at2 t=1
bt2
t=1

(1)

where the denominator supplies the energy normalization terms and the numerator is the dot product of
two time series. The optimal time lag is generally set at the maximum correlation coefficient. This
metric works well if is constant, i.e, for a constant time shift. However, this measure fails to find the
best matching in nonstationary cases.
The first alternative to cross correlation is Dynamic Time Warping (DTW) (Berndt and Clifford, 1994).
DTW can accommodate stretching and squeezing in time series by linear programming. It uses the
Euclidean distance as the initial metric but allows for the one-to-many alignment (Herrera and Van der
Baan, 2012). In the warping matrix the squared distance in the elements (ith , jth ) is calculated by:

(ai , b j ) = (ai b j )2 .

(2)

The optimal path that minimizes the total warping cost (Berndt and Clifford, 1994) is:
p

DTW (a, b) = min (wk ),


W

(3)

k=1

where each wk corresponds to a point (i, j)k and each grid point (i, j) corresponds to an alignment or
connection between ai and b j . The dynamic programming approach uses the following recurrence to
find the warping path (Berndt and Clifford, 1994):

(i, j) = (ai , b j ) + min[ (i 1, j), (i 1, j 1), (i, j 1)],

(4)

where (ai , b j ) is the distance defined in (2), and the cumulative distance (i, j) is the sum of the distance
between the current elements and the minimum cumulative distance of the three neighboring cells.
This mapping process produces stretched versions of the original signals with length k. Only if the two
signals are fully aligned, i.e., i = j = k. For the final warping process we compute a new argument , by
extracting the indices of the intersection of the two sets {ik } and { jk }. With the new argument we can
now get the warped signal a by selecting the th samples from ai . This signal is the best approximation
76th EAGE Conference & Exhibition 2014
Amsterdam RAI, The Netherlands, 16-19 June 2014

of b j following the optimum warping path (a b j ). We use a global distance constraint r to limit
the maximum allowed amount of stretching and squeezing (Sakoe and Chiba, 1978). Elements of the
warping matrix are restricted by |ik jk | < r, where r is the window width.
The second algorithm involves the local similarity (LSIM) attribute. It uses the fact that the squared
correlation coefficient given in equation 1 can be split as the product of two factors c2 = pq (Fomel and
Jin, 2009):
at bt
at bt
p = t 2 and q = t 2 ,
b
t t
t at
with the first factor being the solution of a least-squares minimization problem min p t (at pbt )2 , and
the second factor, q is the solution that minimizes the difference minq t (bt qat )2 between the seismic
trace and the synthetic trace. This is a two-way minimization problem and we can provide p and q with
local properties, such that pt and qt become the solution of a regularized least square problem (Fomel
and Jin, 2009):
(
)
min pi t (at pt bt )2 + R[pt ]
(5)
(
)
2
minqt t (bt qt at ) + R[qt ]
(6)
where R is a regularization operator designed to enforce a desired behavior such as smoothness, estimated from shaping regularization (Fomel, 2007b).
The application of local similarity to the well-seismic tying problem consists of squeezing and stretching
the synthetic trace with respect to the seismic trace while computing the local similarity attribute. By
picking the strongest similarity trend from the attribute panel we identify the corresponding shift to
correct the synthetic trace (Fomel and Jin, 2009). These time shifts form a vector which is similar to the
warping path used in dynamic time warping.
The procedure is then, to create a warping function w(t) from the local similarity scan by means of a
shortest-path ray-tracer (Fomel, 2007a). At this point the warping function is just a new time scale , like
the one estimated with dynamic time warping, i.e., w(t) = t. Both methods are described in more
detail in Herrera et al. (2014) and a case study using similarity-based registration can be found in Zhang
et al. (2013).
Both methods create a warping function w(t) that maps the synthetic trace onto the seismic traces by the
following transformation a(w(t)) b(t). The two-way travel time is related to velocity by:
t =2

H0
dz
0

v0 (z)

(7)

where vo (z) is the base velocity as a function of depth z and H0 is the base depth, where base stands for
reference values. The warping path is related to the new velocity v1 (z) by (Fomel and Jin, 2009):
w(t) = 2

H1
dz
0

v1 (z)

t+t
v0 ( )
0

v1 ( )

d ,

(8)

where H1 is the depth in the well logs, vo (t) and v1 (t) are the sonic
velocity as a function of time, and

t is the time shift caused by the stretching or squeezing t = 2 HH01 v1dz(z) . The relative velocity change
can be estimated by simple differentiation of equation 8:
dw v0 (t)

.
dt
v1 (t)

(9)

In a perfect correlation, the estimated velocity ratio should be close to one, i.e., the relative velocity
changes due to the local stretching and squeezing imply little to no velocity variations. These variations
are quantified by the relative stretch measure s(t) = w(t)/t. Deviations of s(t) from one indicate possible
misalignment and therefore velocity changes.
76th EAGE Conference & Exhibition 2014
Amsterdam RAI, The Netherlands, 16-19 June 2014

Example
Both approaches were used to obtain well-seismic ties between observed seismic data and synthetic
traces created from well logs. The test dataset consists of a 3D poststack time-migrated seismic profile
(Hampson-Russell, 1999). Figure 1 shows a seismic section with one well at CDP 39.

Time (s)

The well-seismic tie for this well is


shown in the inset figure. We scanned
the seismic section to find the best
matching location, following White and
Simm (2003), which is at CDP 41.
The generated synthetic trace (red) and
the corresponding seismic trace (blue)
were used to generate the manual tie.
The sampling rate is 2 ms and both signals have the same length, i.e., the seismic trace has been shortened to the well
log length.

Figure 1 Seismic section with study well annotated at CDP 39. The inset shows the synthetic
seismogram with the initial manual correlation
of 0.77 in the time window 8001144 ms.

The estimated warping path by DTW is shown in Figure 2, where we used r = 10 samples to limit the
amount of permitted stretching/squeezing. With a sampling rate of 2 ms, two peaks can then still be
aligned even if they are 20 ms apart. The relative velocity change is then estimated by computing the
derivative of the smoothed warping path in Figure 2. The LSIM method performs a local-similarity scan
shown in Figure 3. The red color indicates high similarity and from these pick values we estimate the
relative stretch measure s(t), represented as a black curve in Figure 3. This curve is already the derivative
of the warping path, thus from equation 9 we can compute the velocity ratio.
Stretch/Squeeze Scan

50
100

Samples

150
200
250
300
350
400
0.9

Figure 2 DTW. Warping path r = 10 connecting synthetic (red) with seismic trace (blue).

0.95

1
Relative Stretch

1.05

1.1

Figure 3 LSIM. Local similarity scan for detecting the warping function. Red indicates
strong similarity. The black curve shows an
automatically detected trend.

Figure 4a shows the relative stretch (dw/dt) for both methods. The LSIM method (red) shows few
variations, while the DTW result (blue) shows reliable stability in the bottom half of the display. These
variations are reflected in the velocity changes shown in Figure 4b. It is known that DTW will find
the best match between two events and improve the final correlation, but at the cost of more velocity
variations (Herrera and Van der Baan, 2014).
Finally we compare both methods with the manual tie shown in Figure 5a. The manual tie produces
a correlation c = 0.77 in the window 8001144 ms, the bold vertical lines indicate the correlation
window. The DTW output with a warping window r = 10 is shown in b); with this approach and using
the full trace length the correlation is 0.85. The LSIM output using the entire trace length as shown in
c) achieves a correlation of 0.68. In all displays the blue signal is the seismic trace and the red signal is
76th EAGE Conference & Exhibition 2014
Amsterdam RAI, The Netherlands, 16-19 June 2014

the synthetic trace, and black arrows indicate correlative reflections.


a) Relative Stretch
0.5

Vertical depth from surface (m)

400

1.5
DTW
LSIM

a)- Manual method.


Correlation coefficient = 0.77

b) Vp (m/s)
2000

4000

6000
Vp
DTW
LSIM

b)- DTW.
Correlation coefficient = 0.85

600

800

c)- LSIM.
Correlation coefficient = 0.68

1000

1200

1400

300

500

1600

Figure 4 Quality control. a)- Relative stretch


curves for LSIM (red) and DTW with r = 10
(blue). b)- Local velocity change for LSIM
(red) and DTW (blue) together with the original sonic log (light gray).

800
Time (ms)

1144

Figure 5 Comparison of the automated approaches and manual well-seismic ties. a)


Manual tie, c= 0.77 in the window 800
1144 ms. b)- DTW with a warping window
r = 10, c=0.85. c)- LSIM using the entire trace
length c=0.68.

Conclusions
The LSIM method produces a more correct correlation by allowing fewer changes in the velocity pattern. The constrained DTW method performed in good agreement with the manual method inside the
correlation window but introduces more suspect correlations in low-quality areas. The matching outside
this window represents the best possible, but not necessarily most correct tie. Fortunately, the relative
velocity variations help identify when these changes are acceptable or not.
The two methods vary both in the matching criteria and extraction of the warping paths. DTW has
one-to-many (point by point) sample matching which is possibly less realistic. DTW finds the locally
optimal path from point to point but LSIM method is allowed to jump across sequences (its locally
non-optimal). This causes more local stretching/squeezing in DTW. Both methods aim to guide the
interpreter by supporting standard methodologies of well-seismic tying, but careful quality control is a
must to avoid unrealistic changes. We therefore advise against unsupervised implementations as these
methods can greatly aid in an interpretation but they are not a replacement for best practices in well ties.

References
Anderson, P. and Newrick, R. [2008] Strange but True Stories of Synthetic Seismograms. CSEG Recorder, 12,
5156.
Berndt, D.J. and Clifford, J. [1994] Using dynamic time warping to find patterns in time series. KDD Workshop,
359370.
Fomel, S. [2007a] Local seismic attributes. Geophysics, 72(3), A29A33.
Fomel, S. [2007b] Shaping regularization in geophysical-estimation problems. Geophysics, 72, R29R36.
Fomel, S. and Jin, L. [2009] Time-lapse image registration using the local similarity attribute. Geophysics, 74,
A7A11.
Hampson-Russell [1999] Theory of the Strata program. Tech. Rep. May 1999, CGGVeritas Hampson-Russell.
Herrera, R.H., Fomel, S. and Van der Baan, M. [2014] Automatic approaches for seismic to well tying. Interpretation, accepted. Special Section: Well ties to seismic data.
Herrera, R.H. and Van der Baan, M. [2012] Automated seismic-to-well ties? 74th EAGE Conference & Exhibition.
Herrera, R.H. and Van der Baan, M. [2014] A semi-automatic method to tie well logs to seismic data. Geophysics,
accepted.
Sakoe, H. and Chiba, S. [1978] Dynamic programming algorithm optimization for spoken word recognition. IEEE Transactions on Acoustics, Speech, and Signal Processing, 26, 4349, ISSN 0096-3518, doi:
10.1109/TASSP.1978.1163055.
White, R.E. and Simm, R. [2003] Tutorial: Good practice in well ties. First Break, 21, 7583.
Zhang, R., Song, X., Fomel, S., Sen, M. and Srinivasan, S. [2013] Time-lapse seismic data registration and
inversion for sequestration study at cranfield. Geophysics, 78(6), B329B338.
76th EAGE Conference & Exhibition 2014
Amsterdam RAI, The Netherlands, 16-19 June 2014

You might also like