You are on page 1of 8

APPLIED GEOPHYSICS, Vol.7, No.4 (December 2010), P.

384 - 391, 12 Figures


DOI: 10.1007/s11770-010-0264-y

Wavelet estimation by matching well-log, VSP, and surface-


seismic data
Ma Hong-Da1, Roy E. White2, and Hu Tian-Yue1♦

Abstract: In this paper, we present a method of wavelet estimation by matching well-log,


VSP, and surface-seismic data. It’s based on a statistical model in which both input and
output are contaminated with additive random noise. A coherency matching technique is
used to estimate the wavelet. Measurements of goodness-of-fit and accuracy provide tools
for quality control. A practical example suggests that our method is robust and stable. The
matching and estimation of the wavelet is reliable within the seismic bandwidth. This method
needs no assumption on the wavelet amplitude and phase and the main advantage of the
method is its ability to determine phase.
Keywords: coherency matching technique, VSP data, wavelet estimation, matching

Introduction about the noise. The method we present is deterministic


only in the sense that it uses input and output signals.
The purposes of matching well-log, VSP, and surface-
Methods of wavelet estimation can be divided seismic data are: correctly identifying horizons to pick
into two categories (Tygel and Bleistein, 2000): (1) and estimating the wavelet for inversion. Activities
Deterministic methods require direct measurements such as “stretch and squeeze” to fit a seismic trace are
of the source wavefield or input sequence (well-log unscientific and not good practice. We need a method
synthetic seismogram), such as Bayesian wavelet that makes no assumptions about the wavelet and
estimation (Buland and Omre, 2003) and linear inversion provides quantitative measures of accuracy to get a good
(Cooke and Schneider, 1983); and (2) Statistical tie (White and Simm, 2003).
methods estimate the wavelet from the trace itself and On the basis of a previous study (White, 1980;
requires a statistical assumption about the data, as in Walden and White, 1998), we present a method of
SRME wavelet estimation (Matson, 2000), time-varying wavelet estimation by matching well-log, VSP, and
wavelet estimation by kurtosis maximization (Van der surface-seismic data. We start from a statistical model
Bann, 2008), and the “Euclid” algorithm (Riestch, 1997a in which both input and output are contaminated with
and b). In practice a purely deterministic method would additive random noise. Then, a coherency matching
be seriously undermined by noise and the so-called technique is used to estimate the wavelet. This method
deterministic methods require statistical assumptions needs no assumption on the wavelet amplitude and

Manuscript received by the Editor June 9, 2010; revised manuscript received October 27, 2010.
This research is supported by the Natural Science Foundation of China (Grant Nos. 40974066 and 40821062) and by the
National Basic Research Program of China (Grant No. 2007CB209602).
1. School of Earth and Space Sciences, Peking University, Beijing 100871, China.
2. School of Earth Sciences, Birkbeck, University of London, London, WC1, UK.
♦Corresponding author: Hu Tianyue, tianyue@pku.edu.cn.

384
Ma et al.

phase. The main advantage is its ability to determine time. The power spectra of xt, yt, and zt can be obtained
phase. Measurement of goodness-of-fit and accuracy is as three equations. Also, the cross spectra between xt
introduced. The method is demonstrated on a practical and yt, yt and zt, and zt and xt can be measured as another
example from the Tarim oilfield in western China. three equations. Therefore, we can have six equations as
the constraints for the six unknowns rt, ht, gt, ut, vt and wt.
As a result, we can get a solution without any additional
Theory information.

Figure 1 is a diagrammatic representation of the Goodness-of-fit and accuracy


model. We assume that the error and noise sequences in
this model are random and stationary. xt is the computed
synthetic reflectivity: The goodness-of-fit measure and accuracy are very
important for assessing the reliability of the matching
xt = rt + u t , (1) and wavelet estimation.
PEP, the proportion of energy in the surface-seismic
where rt is the actual reflectivity and ut is random noise.
data segment predicted by the VSP data or well-log
The recorded surface-seismic trace yt can be written as:
synthetic, is a measure of goodness-of-fit:
y t = ht rt + vt , (2)
¦ ( y  sˆ )
t t
2

where ht is the surface-seismic wavelet and vt is random PEP = 1 t


, (4)
noise. Similarly, the recorded VSP trace zt is given by: ¦y t
2
t

z t = g t rt + wt , (3) where ŝt is the predicted surface-seismic data and yt is


the recorded surface-seismic data.
where gt is the VSP wavelet and wt is random noise.
NMSE, the normalized mean square error, is a measure
Noise ut of accuracy:

+ Well-log synthetic xt
Noise vt ¦ (sˆ  s )
t t
2

NMSE = t
, (5)
Seismic wavelet ht + Seismic trace yt ¦s t
2
t

Reflectivity rt
VSP wavelet gt + VSP trace zt
where st is the true surface-seismic reflection signal. We
can’t know the true surface-seismic reflection signal but
Noise wt estimation theory tells us the expected NMSE is (Walden
and White, 1984):
Fig. 1 Model for matching and wavelet estimation.
1 1  PEP
The method for wavelet estimation is based on a E{NMSE} | , (6)
bT PEP
coherency matching technique (White, 1980; Walden
and White, 1998). The tool appropriate to the analysis of where T is the length of the matched data segment and
this model is the spectral theory of stationary stochastic b is the analysis bandwidth (White, 1984). In the time
processes. The original method only uses the synthetic domain, b is proportional to the reciprocal of the length
reflectivity xt and surface-seismic data yt to estimate the of the wavelet.
wavelet. The power spectra of xt and yt and the cross The standard error of phase response S.E{Phase} is
spectra between them can be measured which gives a crude measure of the phase error within the seismic
three equations but there are four unknown sequences: bandwidth. It can be found from:
rt, ht, ut and vt. In order to solve these equations, extra
information is required to obtain a solution. S .E.{Phase} = NMSE / 2 . (7)
In our case, the synthetic reflectivity x t, surface-
seismic data yt, and VSP data zt are available at the same However, the amplitude and phase spectra of the

385
Wavelet estimation

wavelet with error bars are more informative. Well-log calibration and construction of
synthetic reflection series
A practical example White et al. (1998) showed that editing and
conditioning the log can significantly improve matching.
Timing errors in the well-log synthetic are more
The field data for matching is from well TZ72 in detrimental than amplitude errors (White, 1997). After
the Tarim oil field of West China. The VSP data is a editing and conditioning the log from well TZ72, we
corridor stack section and the surface-seismic data is calibrate the timing of the sonic log into agreement with
a time migrated post-stack section through well TZ72. the VSP. As illustrated in Figure 2, a calibration curve is
The procedure for matching comprises the following fitted to the difference between integrated sonic log time
steps: and VSP time.

4000 4000 4000


1.95 us/ft

8000 8000 8000


-4.91 us/ft
Depth (ft)

Depth (ft)

Depth (ft)

12000 12000 12000

0.85 us/ft

16000 16000 16000


2.62 us/ft

150 100 50 -20 0 20 150 100 50


Uncorrected DT (us/ft) TL-TC (ms) Corrected DT (us/ft)
(a) (b) (c)
Fig. 2 Well-log calibration of well TZ72. (a) Uncalibrated sonic log. (b) Integrated log time minus VSP time
fitted using knee points. (c) Calibrated sonic log 1 ft = 0.3048 m.

0.5
Reflectivities

-0.5
500 1000 1500 2000 2500 3000 3500
Two-way time (ms)
Fig. 3 Synthetic reflectivities of well TZ72.

The calibrated sonic and density logs are used to Determine the best match location
construct the synthetic. The reflection series created For various reasons (White and Simm, 2003), the
from well TZ72 is shown in Figure 3. best match often does not occur at the well location.
386
Ma et al.

A volume of surface-seismic data around a well is the red circle (inline 371, crossline 810) is the well
scanned for the best match location. Figure 4 shows location. Both plots contours show that the best match
PEP contours over a 2200-3200 ms data segment with location (blue circle in the figure) is about 50 m NE of
a cube of data centered on the well. The position of the well.

10 20 30 40 50 60 70 80 90 100

Predictability (percent) Predictability (percent)


500 500
Inlines (cross-line direction) (m)

Inlines (cross-line direction) (m)


250 250

0 0

-250 -250

-500 -500
-500 -250 0 250 500 -500 -250 0 250 500
Crosslines (in-line direction) (m) Crosslines (in-line direction) (m)
(a) (b)
Fig. 4 Contour plots of goodness-of-fit over a 2200-3200 ms data segment. (a) Scanning PEP by matching VSP and surface-
seismic data. (b) Scanning PEP by matching well-log synthetic and surface-seismic data.

Surface-seismic wavelet estimation 2600


By matching well-log, VSP, and seismic data for well
2700
TZ72, we obtain surface-seismic wavelets and VSP
wavelets at different depths. Figure 5 shows the tie over
Time (ms)

2800
the interval 2584-3084 ms at the best match location.
83% of the surface-seismic trace energy is predicted by 2900
well-log synthetic. The normalized mean square energy
is 0.029 and the standard error in phase is 7.0°. The ties 3000
between the two data types and the estimated surface-
seismic wavelet are reliable. 1 2 3 4 5 6 7

Figure 6 shows part of the seismic section through Fig. 5 Tie over the interval 2584-3084 ms at the best match
well TZ72. Three synthetic seismic traces predicted by location. PEP = 0.83, NMSE = 0.029, and S.E. (phase) =7.0°. 1.
the well-log synthetic are spliced into the data at the Impedance, 2. well-log synthetic, 3. surface-seismic wavelet, 5.
location of best match over 2200-3200 ms. filtered synthetic, 4 and 6 are seismic trace, 7. residual.

VSP-to-seismic wavelet estimation data types and the estimated wavelet are reliable.
From the estimated surface–seismic wavelet and Figure 8 shows the VSP-to-seismic wavelets estimated
VSP wavelet, we can get the VSP-to-seismic wavelet over different time windows. Figure 9 shows their
as a filter to convert the VSP to surface seismic data for energy and phase response in the frequency domain.
comparison. Figure 7 shows the tie over the interval The responses within the seismic bandwidth are reliable.
2200-2700 ms at the best match location. 76% of the The surface-seismic wavelet is attenuated faster than the
surface-seismic trace energy is predicted by VSP data. VSP wavelet with depth. High frequency components
The normalized mean square energy is 0.047. The and wavelet amplitudes are both rapidly attenuated from
standard error in phase is 8.8°. The tie between the two 2200 to 3300 ms.

387
Wavelet estimation
CHAN 372
790 800 810 820 830 835

2250

2450
Time (ms)

2650

2850

3050

3150

Fig. 6 Part of the seismic section through well TZ72. Three synthetic seismic traces predicted by the well-log
synthetic are spliced into the data at the location of best match over 2200-3200 ms.

2200
(a) 2200-2700 ms
2300 4

(b) 2400-2900 ms
2400
Time (ms)

2500 (c) 2600-3100 ms


2

2600 (d) 2800-3300 ms


1

1 2 3 4 5 6 5 -1 -0.5 0 0.5 1 1.5


Time (ms)
Fig. 7 Tie over the interval 2200-2700 ms at the Fig. 8 VSP-to-seismic wavelets in the time domain.
best match location: PEP = 0.76, NMSE = 0.047, (a) 2200-2700 ms, PEP = 0.76, NMSE = 0.047;
S.E. (phase) = 8.8°. 1. VSP data, 2. VSP-to-seismic (b) 2400-2900 ms, PEP=0.66, NMSE=0.074; (c)
wavelet, 4. filtered VSP data, 3 and 5 are seismic 2600-3100 ms, PEP=0.65, NMSE=0.078; (d)
trace, 6. Residual. 2800-3300 ms, PEP=0.60, NMSE=0.099.

Power transfer function of wavelet (2200-2700 ms) Power response of wavelet (2200-2700 ms)
Phase (radians)

2
Power transfer

10-5
0

-2
0 40 80 0 40 80
Frequency (Hz) Frequency (Hz)
(a)
Power transfer function of wavelet (2400-2900 ms) Power response of wavelet (2400-2900 ms)
Power transfer

2
Phase (radians)

10-6 0

-2

0 40 80 0 40 80
Frequency (Hz) Frequency (Hz)
(b)

388
Ma et al.

Power transfer function of wavelet (2600-3100 ms) Power response of wavelet (2600-3100 ms)

Phase (radians)
Power transfer
2
0
10-6
-2
0 40 80 0 40 80
Frequency (Hz) Frequency (Hz)
(c)
Power transfer function of wavelet (2800-3300 ms) Power response of wavelet (2800-3300 ms)
Power transfer

Phase (radians)
2

10-6 0

-2

0 40 80 0 40 80
Frequency (Hz) Frequency (Hz)
(d)
Fig. 9 Energy and phase responses of the VSP-to-seismic wavelets. Red lines are the response in the frequency
domain; Blue dash lines are error bars. (a) 2200-2700 ms, PEP = 0.76, NMSE = 0.047. (b) 2400-2900 ms, PEP = 0.66,
NMSE = 0.074. (c) 2600-3100 ms, PEP = 0.65, NMSE = 0.078. (d) 2800-3300 ms, PEP = 0.60, NMSE = 0.099.

CHAN 372 372


790 805 820 0 800 815 830 0

2650

2750
Time (ms)

2850

2950

3050

(a) (b)
Fig. 10 Part of the seismic section through well TZ72. Three VSP corridor stack traces are spliced into the data at the
best match location over 2600-3100 ms. (a) Before deconvolution. (b) After deconvolution.

20
Power speotrum magnitude (dB)

Before deconv Next we design a filter using the VSP-to-seismic


After deconv
wavelet estimated in Figure 8. The filter is applied to the
-20
seismic line through well TZ72 (Inline372, Crosslines
790-830). Figure 10 shows part of the seismic section
-60
over the interval 2600-3100 ms. The tie between the
VSP corridor stack and seismic section is better after
-100
deconvolution. Figure 11 compares the power spectra
12.5 37.5 62.5 87.5 112.5 of the seismic data before and after deconvolution. The
Frequency (Hz)
high frequency components of filtered seismic section
Fig. 11 Power spectra of seismic data (Inline372, Crosslines
are enhanced.
790-830). Blue and red lines are power spectra before
deconvolution and after deconvolution, respectively. Figure 12 shows part of the seismic section through

389
Wavelet estimation

well TZ72. Three synthetic seismic traces predicted location over 2200-3200 ms. The tie between VSP data
by VSP data are spliced into the data at the best match and surface-seismic data is reliable.

CHAN 372
790 800 810 820 830 835

2250

2450
Time (ms)

2650

2850

3050

3150

Fig. 12 Part of the seismic section through well TZ72. Three synthetic seismic traces predicted by the VSP data
are spliced into the surface-seismic data at the best match location over 2200-3200 ms.

Conclusions We assume that the error and noise sequences in this


model are random and stationary. In cases where there
is strong coherent noise (like multiples) existing in the
In this paper, we discussed wavelet estimation by well-log, VSP, and surface-seismic data, this method
matching well-log, VSP, and surface-seismic data. can’t get good results. In this situation, we need to re-
Starting from a statistical model in which both input and process all the data to eliminate the coherent noise to
output are contaminated with additive random noise, a improve the results.
coherency matching technique is used to estimate the
wavelet. The main advantage of the method is its ability
to determine phase. References
We introduce PEP (the proportion of energy
predicted), NMSE (normalized mean square error),
Buland, A., and Omre, H., 2003, Bayesian wavelet
and S.E{Phase} (Standard error of phase response) to
estimation from seismic and well data: Geophysics,
measure the goodness-of-fit of and matching accuracy. 68(6), 2000 – 2009.
These are tools for assessing the reliability of the tie. Cooke, D. A., and Schneider, W. A., 1983, Generalized
We demonstrated the method on a practical example. linear inversion of reflection seismic data: Geophysics,
It’s robust and stable in the presence of noise in the field 48, 665 – 676.
data. Without any assumption on their amplitude and Matson, K., 2000, An overview of wavelet estimation using
phase, we still can estimate both surface-seismic and free-surface multiple removal: The Leading Edge, 19, 50
VSP wavelets. – 55.
Goodness-of-fit measurement and accuracy shows Rietsch, E., 1997a, Euclid and the art of wavelet estimation,
that matching well-log synthetic, VSP, and surface- Part I: Basic algorithm fro noise-free data: Geophysics,
seismic data of well TZ72 is reliable within the seismic 62(6), 1931 – 1938.
bandwidth. With the wavelet estimated, we can enhance Rietsch, E.,1997b, Euclid and the art of wavelet estimation,
the resolution and the high frequency components of Part II: Robust algorithm and field-data examples:
seismic sections. Geophysics, 62(6), 1939 – 1946.

390
Ma et al.

Simm, R., and White, R. E., 2002, Phase, polarity and the procedures and examples, 67th Ann. Internat. Mtg., Soc.
interpreter’s wavelet: First Break, 20(5), 277 – 281. Expl. Geophys., Expanded Abstracts. 816 – 819.
Tygel, M., and Bleistein, N., 2000, An Introduction to this White, R. E., and Simm, R., 2003, Tutorial: Good practice
Special Section (wavelet estimation): The Leading Edge, in well ties: First Break, 21, 75 – 83.
19, 37. White, R. E., Simm, R., and Xu, S., 1998, Well tie, fluid
Van der Baan, M., 2008, Time-varying wavelet estimation substitution and AVO modeling - a North Sea example:
and deconvolution by kurtosis maximization: Geophysical Prospecting, 46, 323 – 346.
Geophysics, 73(2), 11 – 18.
Walden, A. T., and White, R. E., 1984, On errors of fit and
accuracy in matching synthetic seismograms and seismic
traces: Geophysics Prospecting, 32(5), 871 – 891.
Walden, A. T. and White, R. E., 1998, Seismic wavelet Ma Hong-Da, graduated from School of Earth and
estimation: a frequency domain solution to a geophysical Space Sciences, Peking
noisy input-output problem: IEEE Transactions on University with a PhD in
Geoscience and Remote Sensing, 36, 287 – 297. 2010 and a BS in 2005. His
White, R. E., 1980, Partial coherence matching of research interests are mainly
synthetic seismograms with seismic traces: Geophysical multiple attenuation and
Prospecting, 28, 333 – 358.
wavelet estimation.
White, R. E., 1984, Signal and noise estimation from
seismic reflection data using spectral coherence methods:
Proc. IEEE, 72, 1340 – 1356.
White, R. E., 1997, The accuracy of well ties: practical

391

You might also like