Journal of Geophysics and Engineering

Journal of Geophysics and Engineering (2023) 20, 1180–1193 https://doi.org/10.1093/jge/gxad077

Advance access publication 23 September 2023

An integrated workflow for pre-stack gather

optimization
Jie Zhou 1 ,2
, Huailai Zhou1 ,2 , ∗ , Junping Liu1 ,2 and Yaoguang Sun 3

1
Key Laboratory of Earth Exploration & Information Techniques of Ministry of Education, Chengdu

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

University of Technology, 1 Erxianqiao Dongsan Road, Chengdu, Sichuan 610059, China
2
College of Geophysics, Chengdu University of Technology, Chengdu, 1 Erxianqiao Dongsan Road, Chengdu,
Sichuan 610059, China
3
College of Geophysics, China University of Petroleum, No. 18, Fuxue Road, Changping District, Beijing,
China 102249
∗
Corresponding author: Huailai Zhou. E-mail: [email protected]

Received 11 April 2023, revised 28 August 2023

Accepted for publication 22 September 2023

Abstract
High-quality seismic data contribute to improving the prediction accuracy of complex reservoirs.
Currently, there are several methods for enhancing pre-stack gather resolution and
signal-to-noise ratio (SNR). Meanwhile, to avoid incorrect seismic interpretation, we must
evaluate whether the processed data maintains its original characteristics and reliability. This
paper establishes an integrated workflow for optimizing common-reflection-point (CRP) gather
that is divided into processing and quality control. Furthermore, we summarize the
comprehensive quality control means and explain their specific significance in each step. Data
processing includes four parts: gather flattening, SNR enhancement, energy compensation, and
resolution improvement. Simultaneously, we use well log data, forward simulation, stack data,
and inversion to guarantee the processed data is optimized and keeps its original characteristics.
Applications in carbonate reservoir CRP gathers demonstrate that this workflow can
comprehensively optimize pre-stack seismic data and improve SNR and resolution. Importantly,
the quality controls guarantee results have improved accuracy in reservoir prediction and
stronger correlation coefficients with well data.
Keywords: pre-stack gather; data optimization; quality control; resolution improvement

1. Introduction weak deep energy, narrow frequency band and low resolution
(Singleton 2009). When solving these problems, it is also
In recent decades, complex unconventional resources have
necessary to evaluate whether the gather retains its original
become the focus of geophysical exploration. Consequently,
characteristics or whether the effective signal is destroyed.
seismic interpretation has placed higher demands on the
Furthermore, we cannot blindly pursue high resolution and
quality of seismic data than ever before. The pre-stack gather
wide bandwidths, which are likely to introduce more noise
contains offset information, which can reflect the amplitude
and reduce the stability of seismic data.
variation with offset. Hence, using pre-stack data for AVO
Many scholars have proposed methods for improving the
inversion or obtaining pre-stack properties can provide a
quality of pre-stack gather that have been maturely developed
guidance for reservoir prediction. However, there are unde-
and widely used. For the CRP gather, the medium- and far-
sirable pre-stack seismic phenomena that need to be dimin-
offset events may not be perfectly flat, which can contaminate
ished or removed, such as unequal far-offset gather, low SNR,

1180 © The Author(s) 2023. Published by Oxford University Press on behalf of the SINOPEC Geophysical Research Institute Co., Ltd. This is an Open Access article distributed under the
terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any
medium, provided the original work is properly cited.
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

seismic interpretation results. At this point, we need a brute- The purposes of the optimization methods mentioned
force method to flatten the events. The general method is to above are to enhance the SNR and improve the resolution
determine a reference channel and obtain the residual move- of seismic data. We can show the visual improvement of the
out between it and the seismic event (Hinkley et al. 2004; Gu- data by comparing the pre-stack and post-stack profiles and
lunay et al. 2007). Also, some scholars improved the accuracy the frequency spectra of the data before and after process-
of moveout correction by researching residual moveout cal- ing. However, these methods lack a thorough, objective, and
culation methods (Gamar-Sadat et al. 2008; Qian et al. 2016; convincing quality control process to guarantee the reliabil-
Shi et al. 2022). ity of the result (Singleton 2009). Therefore, we establish
In general, noise can be divided into regular and irregular an integrated workflow that includes processing and qual-
noises. Derivative methods based on f–k filtering can be used ity control. The processing and quality controls are carried
to remove regular noise with simple spatial characteristics out alternately, which can comprehensively evaluate whether
(Duncan & Beresford 1995; Hosseini et al. 2015; Lowney the processed data are high-quality and reliable for seismic
et al. 2020). Multiple waves can be removed by using the interpretation.
developed radon transform (Fiddy 1985; Beylkin 1987; Xie

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

et al. 2021). For irregular noise that is difficult to remove
in the time domain, many scholars converted seismic data 2. Pre-stack gather processing methods
into the multiscale frequency domain to separate the noise
from the effective signal, especially frequency-specific noise 2.1. Strategies to improve the data quality
(Huang & Wu 2008; Han & van der Baan 2013; Sun et al. First, we analyze the problems in the data to generate a strat-
2020). In addition, multi-channel tuning can eliminate egy to improve the quality of the data. Figure 1 shows CRP
discontinuous random noise or wild values, which is the gathers used in this paper to demonstrate the entire process.
core idea of mean filtering and median filtering (Baddari These data are from Sichuan, China, and the carbonate reser-
et al. 2011; Wu et al. 2018). With the fast development of voir areas are bioherms and beaches. The top and bottom of
hardware and algorithms, artificial intelligence has made an the reservoir are marked by black lines in Fig. 1. As shown in
extraordinary performance in eliminating noise ( Jin et al. Fig. 1, these data have the following problems that need to be
2018; Yang et al. 2020). Real seismic data contains multiple addressed: (i) the in-phase axis is irregular, especially the far-
complex noises. Depending on the features of the noises, offset, (ii) there is a lot of random noise contaminating the
we select the appropriate one or even several methods to events, (iii) the resolution of the reservoir area is low and the
suppress the noises. in-phase axis direction is not clear.
As the seismic wave travels, its energy is absorbed and at- We have established the following integrated workflow of
tenuated, which has a direct impact on the imaging of deep processing and quality control (Fig. 2) to solve the problems.
reservoirs and structures (Futterman 1962; Johnston et al. The data processing part includes gather flattening, SNR en-
1979; Jones 1986; Gist 1994). There are numerous meth- hancement, energy compensation, and improved resolution.
ods for recovering seismic wave energy from the time and fre- The quality control part includes gather comparison, AVO
quency domains, but the improved methods based on inverse analysis, well-seismic calibration, post-stack comparison, and
Q and deconvolution are the most popular and effective ones wavelet extraction. In addition, each processing step has a
(Hale 1981; Hargreaves & Calvert 1991). The key to inverse corresponding quality control work and the focus is different.
Q filtering is to find the Q quality factor. Initially, the quality If the quality control fails, the parameters need to be read-
factor is a constant, it compensates energy without discrimi- justed to repeat the previous step until the quality control
nation for seismic data and noise, which decreases the stabil- meets the requirements. This workflow can apply to most
ity of the data (Bickel & Natarajan 1985). Thus, some schol- data, and processing steps can be added or subtracted de-
ars have expressed strong interest in proposing various im- pending on the specific issue of data.
proved methods to obtain accurate quality factors and guar- The fundamental strategy is to first obtain flat and clean
antee the stability of data (Wang 2002; Wang 2006; Zhang pre-stack gathers, so flattening and denoising are the basic
& Ulrych 2007; Sun 2015). Likewise, with the develop- and crucial steps. At the same time, we carry out quality con-
ment of time-frequency methods, non-stationary deconvolu- trol to ensure that the effective signal and AVO characteris-
tion was developed to improve the resolution of the seismic tics are maintained. Then, making further improvements to
data (Rosa & Ulrych 1991; Mundim et al. 2006; Margrave the data by energy balancing and frequency band expansion,
et al. 2011), such as the time-varying deconvolution, which here we examine the well-seismic correlation and spectrum
can adapt the time-varying characteristics of seismic data to ensure the reliability of the optimized gathers. Finally, we
and achieve better application results than the conventional apply the optimization data to the inversion to compare the
method (Clarke 1968; Van der Baan 2008; Jiang et al. 2020). results of seismic interpretation.

1181
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 1. The raw pre-stack gathers.

Figure 2. The integrated workflow for pre-stack gather optimization.

2.2. Gather flattening et al. 2007; Zhou et al. 2016). The principle is to calculate
According to the workflow in Fig. 2, we first perform the correlation between each trace and the reference trace to
gather flattening. Seismic data contains complete informa- define the optimal value of time shift in a given time window.
tion before denoising so that more event information can be The detailed processes are as follows:
retrieved during the flattening.
In the theoretical AVO model, the gather event is horizon-
tal. Unfortunately, the actual pre-stack gather event is not per- (i) Determine the reference trace R(t) for each gather, usu-
fectly flat due to many factors, such as the normal moveout ally, the reference trace is a full-stack trace or a near-
(NMO) correction algorithm, the inaccurate NMO velocity, offset stack trace.
the wavelet deformation, and others. To flatten the event, a (ii) Determine the time window for calculation, the inter-
residual moveout needs to be corrected. We selected the dy- val needs to be estimated based on the non-flat event,
namic flattening method based on statistics to address the which contains at least one complete seismic wave, the
small fluctuations in events (Hinkley et al. 2004; Gulunay start and end time points are ta and tb , respectively.

1182
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

(iii) Calculate the cross-correlation coefficient Cx(𝜏) be- corresponding weights wk (i, j) of each sample point in Xk and
tween each trace x(t) and the reference trace R(t), and normalize them, the equation is
obtain the time-shift value 𝜏shift according to the max- 1∕ (1 + Max(Dk , T ))
imum cross-correlation coefficient Cmax x(𝜏shift ). No- wk (i, j) = , (2)
tably, we calculate the cross-correlation coefficient with ∑
N
1∕ (1 + Max(Dk , T ))
absolute values, which avoids the phase inversion of the k=1
event, using where Dk is the absolute value of the difference between the
ta values of the sample points Xk and their means, according
Cx(𝜏) = |R(t)||x(t + 𝜏)|dt. (1) to equation (3). T is the average of all Dk and represents a
∫t b
threshold, according to equation (4). The weighting algo-
rithm adopts the threshold optimization principle, that is, if
(iv) Set the minimum correlation coefficient value and max-
Dk is greater than the threshold T, the weight is determined
imum time-shift value when calculating the time shift
by Dk , and if Dk is less than the threshold T, the weight is
for each trace to eliminate outliers.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

determined by T:
(v) Calculate the entire time-shifted field over the entire
time interval by rolling the time window and apply to Dk = |Xk − Mean(Xk )|, (3)
flatten the gathers. Since this is a brute-force method
for small fluctuations, it is also possible to smooth the
entire time-shifted field to avoid overstretching. ∑
N
|Xk − Mean(Xk )|
k=1
T= . (4)
2.3. Improved weighted mean filtering algorithm N

To avoid noise increases during energy compensation and All sample points in the Xk [x(i, j)] set are weighted, and
high resolution, we need to perform denoising immediately the output result x′ (i, j) is the center point of the filter
after flattening to ensure the input data has high SNR. Ac- window W:
cording to the data quality analysis, we know that this data is
∑
N
contaminated by random noise. Random noise is usually dis- x′ (i, j) = Xk [x(i, j)] ⋅ wk (i, j). (5)
tributed over the entire frequency band, the time-frequency k−1
conversion denoising method is not the best choice for this.
Fortunately, random noise in the time domain shows obvi-
ous discontinuous features, stacking the gathers can remove 2.4. Inverse Q filtering
most random noise. In the same way, we can achieve pre-stack Seismic wave energy is absorbed and attenuated during
random noise attenuation according to the irregularity in the travel, so we perform energy compensation to restore deep
time domain. energy and improve image quality. Moreover, the next high-
Mean filtering is an effective means to suppress random resolution processing step will redistribute the event energy
noise by averaging a local range. However, its disadvantages as it is necessary to ensure that the energy is sufficient. En-
are that edge information will be significantly blurred as the ergy compensation is likely to increase the noise energy at the
window widens, and some abnormal amplitude values can- same time, hence denoising is needed before that.
not be removed. Therefore, we use the improved weighted Many applications have proved that inverse Q filtering is
mean filtering algorithm to suppress random noise (Li et al. an effective method for energy recovery. Stability factor in-
2022). First, we remove the maximum and minimum values verse Q filtering is widely used in amplitude and phase com-
in the local range, and then assign different weights to each pensation (Wang 2006), this method can solve the stability
point for mean filtering. The details of the algorithm are as problems of traditional compensation methods.
follows: According to Wang (2006, 2008), the equation for ampli-
With each sample point as the center x(i, j), select the filter tude compensation and phase correction is
window W[x(i, j)], and remove the maximum and minimum
( )
| 𝜔 |−
1
points in the window. Then record the set of remaining sam- U(T + ΔT, 𝜔) = U(T, 𝜔)Λ(T, 𝜔) exp j| | 𝜋Q 𝜔ΔT
ple points as Xk , and its mean is Mean(Xk ), k is the number | 𝜔c |
of sample points in the filter window. For 3 × 3 windows, the ( )
| 𝜔 |− 𝜔ΔT
1

window is the center point and the surrounding eight num- × exp | | 𝜋Q , (6)
| 𝜔c | 2Q
bers, we remove the maximum and minimum points, so k
assumes the value from 1 to 7; for 5 × 5 in the window, k where U(T, 𝜔) is the wavefield, T is the propagation time,
assumes the value from 1 to 23. After that, we calculate the 𝜔 is the radial frequency, 𝜔c is the radial frequency center,

1183
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

ΔT is the travel time increment, and Q is the frequency- Deconvolution using L1 constraints is an inverse problem
independent quality factor (Yang et al. 2016): to find exactly the location and amplitude of the reflection co-
efficient from seismic data and certain wavelets. Deconvolu-
𝜋Tfc fr2
Q= . (7) tion using L1 constraints adds a sparse prior constraint to the
fr2 − fc2 traditional least squares deconvolution to solve the objective
functions. The objective function is expressed as (Gholami
Proposed originally by Wang (2006), the stabilization ampli-
2016):
tude compensation factor Λ is expressed as
arg min‖S − Wr‖22 + 𝜆‖r‖1 , (12)
𝛽(T, 𝜔) + 𝜎 2 r
Λ(T, 𝜔) = 2 , (8)
𝛽 (T, 𝜔) + 𝜎 2 where 𝜆 is a regularization parameter used to control the
1
− 𝜋Q fitting error and solution dispersion; it is crucial to test it
where 𝛽(T, 𝜔) = exp(−| 𝜔𝜔 | 𝜔ΔT
2Q
), 𝜎2 = repeatedly to select the right value.
c
exp[−(0.23Glim + 1.63)], Glim is the gain limit, and the

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

general value is 0–70 dB.
3. Application
In this section, we apply the integrated workflow to the pre-
2.5. Deconvolution using L1 constraints
stack data, and the raw data are shown in Fig. 1. The data
High-resolution data facilitates the identification of thin set comes from Sichuan, China, and it has some problems
reservoirs and the characterization of complex structures. we mentioned before, such as uneven far-offset event, noise
Improving the resolution is the “icing on the cake,” we contamination, and low resolution. After completing the pro-
must verify the reliability of the processed data and avoid cessing for these issues, we check the results quality from var-
blindly pursuing high resolution. Deconvolution using L1 ious aspects to guarantee that the quality of processed data is
constraints is widely used to improve the resolution of improved and reliable.
seismic data.
In general, the convolution model can be expressed as fol- 3.1. Processing result
lows:
For the pre-stack gather, we flatten the events, enhance SNR,
s(t) = w(t) ∗ r(t) + n(t), (9) compensate energy, and increase resolution by using the pre-
which represents seismic data s(t) as a convolution rela- viously described algorithms. Figure 3 shows the selected raw
tionships ∗ between the reflection coefficients r(t) and the gather and the results of the individual processes, the pro-
wavelet w(t) = {w1 , w2 , … , wk }, k is the wavelet length, and cessing is done in a progressive way as described. Figure 4
then adds noise n(t) to simulate the real seismic data. The shows another selected gather and the results. Overall, the
convolutional model can be understood as a high-frequency final result has less noise, its events are flatter, and the reso-
pulse signal r(t) that propagates underground and becomes lution is much higher than the raw data. Figures 5 and 6 are
low-frequency seismic data s(t), where the passing strata the enlarged processed results of the reservoir location, the
acts as a filter w(t) = {w1 , w2 , … , wk }, and the process is optimized data is clear, energy-focused, and contains more
disturbed by the noise n(t). details. However, this is only a demonstration of visual im-
The matrix-vector form of the seismic convolution model provement. We still need to evaluate the result to determine
can be written as follows: whether it maintains the original characteristics and whether
the quality has significantly improved in different aspects.
S = WR + N, (10)
where S is the seismic data, R is the reflection coefficient, N 3.2. AVO analysis
is the random noise, and the wavelet matrix W is: AVO analysis is an important part of quality control. In seis-
⎡w 1 mic interpretation, such as intercept and gradient properties
0 ⋯ 0⎤
⎢ ⎥ and pre-stack inversion, AVO characteristics indicate signif-
⎢⋮ w1 ⋱ ⋮⎥ icant oil and gas information. Therefore, we must guarantee
⎢w ⋮ ⋱ 0⎥ that the processed data has the same AVO characteristics as
W=⎢ ⎥.
k
(11) the original data to reduce reservoir prediction errors.
⎢ 0 wk ⋱ w1 ⎥
⎢ ⎥ We selected two gathers across Well A and Well B sepa-
⎢⋮ ⋱ wk ⋮⎥ rately to analyze the AVO curves before and after optimiza-
⎢0 ⋯ wk ⎥⎦
⎣ 0 tion. Figure 7 shows the comparison results of the gather
through Well A. Figure 7 parts a and b are the raw data and

1184
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 3. A CRP gather in each processing step. (a) Raw data, (b) gather flattening, (c) SNR enhancement, (d) energy compensation, and (e) improve
resolution.

Figure 4. Another CRP gather in each processing step. (a) Raw data, (b) gather flattening, (c) SNR enhancement, (d) energy compensation and
(e) improve resolution.

optimized data, and the red and blue lines indicate the loca- 3.3. Well-seismic calibration
tions to be extracted for AVO analysis, they are at the same
If we seek high-resolution seismic data blindly, the processed
time. Figure 7c shows the comparison of the AVO curves be-
seismic data may have false axes and wrong structures. It is
fore and after processing; the dots are exact values at different
necessary to guarantee the reliability of data quality. Well data
offsets and the line is their regression curves. Figure 8 shows
can reflect actual underground characteristics; well-seismic
the comparison results of gather through Well B, and the con-
calibration is a common method to verify the reliability of
tents of Fig. 8 are the same as in Fig. 7. In Figs 7c and 8c,
data by examining the correlation between seismic data and
the blue curves are the AVO curves of raw data, these blue
well data.
dots are scattered, and their regularity is not evident. The
We used the well log data to build an AVO model and com-
red curves are the AVO curves of optimized data: these red
pare the cross correlation between it and raw and processed
dots are concentrated close to the regression curve and have
data separately. Figure 9 shows the raw data, the optimized
a distinct rhythm. We use comparison to demonstrate that
data, and the AVO model of the gather through Well A.
the trend of the AVO curve for the optimized data is consis-
Figure 10 parts a and b show the cross correlation between
tent with the trend of the original AVO curve, and its ampli-
the AVO model and the data. The maximum correlation co-
tude increases, and features become more pronounced than
efficient of the raw data is 0.723, and the maximum correla-
before.
tion coefficient of the optimized data is 0.805.

1185
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Figure 5. A processing result of reservoir locations. (a) Raw data and (b) optimized data.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 6. Another processing result of reservoir locations. (a) Raw data and (b) optimized data.

Figure 7. (a) Raw and (b) optimized gather through Well A, and (c) the comparison of their AVO curves.

Figure 11 shows the raw data, the optimized data, and the 3.4. Post-stack data
AVO model of the gather across Well B. Figure 12 parts a
The previous quality controls were performed based on the
and b show the cross correlation between the AVO model
pre-stack gathers. Comparing the post-stack data is a widely
and the data. The maximum correlation coefficient of the raw
used method for checking the data quality, which can intu-
data is 0.682. The maximum correlation coefficient of the op-
itively show stratigraphic formations. We stacked the gath-
timized data is 0.746. The two well-crossing examples show
ers to form post-stack data and evaluated its profile and slice
that the processed data are more correlative with the well data
quality.
and more reliable than the raw data.
1186
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 8. (a) Raw and (b) optimized gather through Well B, and (c) the comparison of their AVO curves.

Figure 9. A selected (a) raw gather, (b) optimized gather, and (c) AVO model gather based on the well curve.

Figure 10. (a) The cross correlation of the raw gather. The maximum correlation coefficient is 0.723. (b) The cross correlation of the optimized gather.
The maximum correlation coefficient is 0.805.

To demonstrate the advantages of the optimized data, we lationship of the composite wave is distinctive, and the event
stacked the full-angle gathers of raw and processed data sep- amplitude is compensated. In short, the quality of the stacked
arately (Figs 13 and 14). Obviously, vertically, the number data is significantly improved after the pre-stack processing.
of events increases and the resolution of optimized data is Figure 15 shows the layer slices along the top of the reser-
greatly improved; laterally, the event is smoother and its con- voir. The results demonstrate that many structure details can-
tinuity is improved. Furthermore, the background random not be recognized in the raw data due to weak energy or noise
noise of the optimized data is significantly reduced, the re- contamination. However, the optimized slice contains richer

1187
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 11. Another selected (a) raw gather, (b) optimized gather, and (c) AVO model gather based on the well curve.

Figure 12. (Left) The cross correlation of the raw gather. The maximum correlation coefficient is 0.682. (Right) The cross correlation of the optimized
gather. The maximum correlation coefficient is 0.746.

structural information, has a more obvious amplitude energy Figure 16a shows that the shapes of the raw wavelets are in-
contrast, and has better spatial continuity than the raw data, consistent, especially large-angle wavelet is shifted at the zero
which helps to highlight the structural boundary. In Fig. 15, point. The dominant frequency of raw wavelets is ∼20 Hz;
the zones indicated by yellow circles in the optimized slice the phase spectrum gradually becomes chaotic and changes
have a higher resolution to identify the structure than the raw drastically after 20 Hz, indicating that this component after
slice. 20 Hz is unstable. However, compared to the partial-angle
stack data, the full-angle stack data are much more reliable.
In particular, the phase spectrum stays smooth up to 50 Hz,
3.5. Wavelet extraction which proves that the full-angle stack can eliminate some of
the noise and improve the data stability.
Seismic wavelets are the basic units in seismic data, they can To better compare the improvements, we show the am-
reflect the frequency and phase information of seismic data. plitude, frequency, and phase spectra of optimized data in
Moreover, seismic wavelets play a critical role in seismic pro- Fig. 17. From Figure 17a and b, the waveforms of all wavelets
cessing and interpretation, such as seismic deconvolution, are similar; the main frequency of optimized wavelets is
seismic inversion, and others (Singleton 2009). around 40 Hz, which is significantly wider and contains more
Here, we extract the statistical wavelets from the three frequency information than before. What is more, the phase
partial-angle stacks (0–15°, 10–25°, and 20–35°) and one spectra remain stable before 60 Hz (Fig. 17c), which proves
full-angle stack for the raw and optimized data after the that this part of the information is reliable. The results il-
well-seismic tie. Meanwhile, we obtain the frequency and lustrate that the optimized seismic data has less noise and
phase spectra of these wavelets, as shown in Figs 16 and 17. good stability on near-, middle-, and far-angle stacks. The
1188
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 13. Raw full-stack profile.

Figure 14. Optimized full-stack profile.

optimized data are expected to help achieve more accurate inversion result of the raw data shows that the formation is
seismic interpretation than before. blurry, and much information is too small to detect. How-
ever, the inversion result of the optimized data has higher res-
olution, the structure is more laterally continuous, and the
3.6. Inversion analysis
boundary of the bioherm is clearer than the raw data result. In
The reservoir in this application is mainly derived from bio- addition, the inversion results of the optimized data are more
herms and beaches; it is thin and its boundary is difficult consistent with the well data than the raw data, which also
to delineate, so successful reservoir prediction has a high proves the reliability of the inversion results.
requirement for data quality (Singleton 2009; Zhang et al. We extract the P-velocity inversion layer slices of the raw
2015). Previously, we used a variety of methods to guaran- and optimized data (Fig. 19a and b). The inverted slice of
tee the quality of the optimized data. Then, we carry out the optimized data contains detailed information and has a clear
pre-stack inversion to examine the reservoir identification boundary between high and low energy values, while the in-
capability of the raw and optimized data. verted slice of raw data only provides an approximate con-
Figure 18 parts a and b show the P-velocity inversion pro- struction boundary. Furthermore, the inversion result of the
files of raw and optimized data through Well A and Well B. raw data shows low velocities in the non-reservoir area (the
The red curves in Fig. 18 indicate high-cut P-velocity and the northeast corner of the work area), which may mislead the
black lines are the top and bottom of the reservoir area. The seismic interpreters. The inversion results of optimized data

1189
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 15. Layer slice of (a) raw and (b) optimized full-stack data.

Figure 16. (a) Raw statistical wavelets, (b) the frequency spectrum, and (c) the phase spectrum for the three partial-stack and full-stack.

have fewer low values at non-reservoir locations, demonstrat- choose different denoising methods as well as single or mul-
ing that improving the quality of pre-stack data can reduce in- tiple methods to improve the SNR depending on the types
correct seismic interpretations and enhance the accuracy of of noise. There are a lot of parameters that need to be repeat-
reservoir predictions. edly adjusted and compared to select the optimal one during
processing. In general, the quality of the optimized data is ex-
amined by comparing the post-stack profiles and their spec-
4. Discussion
tra, but this is one-sided. Especially for the pre-stack gathers,
From the processing flowchart, we comprehensively improve we must evaluate AVO characteristics, consistency with well
the quality of seismic data in four aspects. In practice, to de- data, and other factors. To show the whole process, this paper
termine the processing flow appropriately, we need to ana- introduced processing and quality control in turn, and these
lyze the characteristics and problems of the original data. The two parts should appear alternately to improve data quality
methods for each step can be replaced. For example, we can and ensure result reliability.

1190
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 17. (a) Optimized statistical wavelets, (b) the frequency spectrum, and (c) the phase spectrum for the three partial-stack and full-stack.

Figure 18. P-velocity inversion profile of (a) raw data and (b) optimized data.

1191
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

Figure 19. P-velocity inversion layer slice of (a) raw data and (b) optimized data.

5. Conclusion Duncan, G. & Beresford, G., 1995. Some analyses of 2-D median F-k filters,
Geophysics, 60, 1157–1168.
Accurate reservoir characterization depends on high-quality Fiddy, M.A., 1985. The radon transform and some of its applications,
seismic data, and directly optimizing pre-stack gather is an Optica Acta: International Journal of Optics, 32, 3–4.
effective way to improve seismic data quality. Meanwhile, Futterman, W.I., 1962. Dispersive body waves, Journal of Geophysical
quality control is required to guarantee resolution enhance- Research, 67, 5279–5291.
Gamar-Sadat, F., Gulunay, N., Hoeber, H., Dyce, M., McKenzie, C. & Whit-
ment and avoid seismic data distortion that might result in
combe, D., 2008. Robust residual gather flattening, in 70th EAGE (Eu-
incorrect interpretation. The conventional quality control ropean Association of Geoscientists & Engineers) Conference and Exhibition
method is to compare the post-stack profile and its spectrum Incorporating SPE EUROPEC 2008, Rome, Italy.
width, but this method cannot adequately quantify the data Gholami, A., 2016. Projected Gabor deconvolution, Geophysics, 81,
quality and its reliability. In this paper, we established a novel V151–V157.
integrated workflow, which contains the complete process- Gist, G.A., 1994. Fluid effects on velocity and attenuation in
sandstones, The Journal of the Acoustical Society of America, 96,
ing chain and corresponding quality control at each step. We 1158–1173.
applied it to the field data and demonstrated that the pro- Gulunay, N., Magesan, M. & Roende, H.H., 2007. Gather flattening, The
cessed data has obviously higher SNR and resolution and is Leading Edge, 26, 1538–1543.
more conducive to obtaining accurate inversion results than Hale, D., 1981. Q and adaptive prediction error filters, Stanford Exploration
raw data. The quality control certified that the processed Project Report, 28, 209–231.
Han, J. & van der Baan, M., 2013. Empirical mode decomposition for seis-
data not only keeps the original AVO characteristics but also
mic time-frequency analysis, Geophysics, 78, O9–O19.
enhances the correlation with the original well logs. As a re- Hargreaves, N.D. & Calvert, A.J., 1991. Inverse Q filtering by fourier trans-
sult, we believe that this workflow can generate high-quality form, Geophysics, 56, 519–527.
and reliable pre-stack data for broad seismic interpretation Hinkley, D., Bear, G.W. & Dawson, C., 2004. Prestack gather flattening for
and reservoir prediction. AVO, in SEG (Society of Exploration Geophysicists) International Exposi-
tion and 74th Annual Meeting 2004, Denver, Colorado.
Conflict of interest statement. We declare that we do not have Hosseini, S.A., Javaherian, A., Hassani, H., Torabi, S. & Sadri, M., 2015.
any commercial or associative interest that represents a con- Shearlet transform in aliased ground roll attenuation and its compari-
flict of interest in connection with the work submitted. son with f-k filtering and curvelet transform, Journal of Geophysics and
Engineering, 12, 351–364.
References Huang, N.E. & Wu, Z., 2008. A review on Hilbert-Huang transform:
method and its applications to geophysical studies, Reviews of Geo-
Baddari, K., Ferahtia, J., Aïfa, T. & Djarfour, N., 2011. Seismic noise atten- physics, 46.
uation by means of an anisotropic non-linear diffusion filter, Computers Jiang, Y., Cao, S., Chen, S., Wang, H., Dai, H., Li, X. & Zheng, D., 2020.
& Geosciences, 37, 456–463. Time-varying wavelet estimation and deconvolution for nonstation-
Beylkin, G., 1987. Discrete radon transform, IEEE Transactions on Acous- ary data based on a FWE function, Journal of Applied Geophysics, 183,
tics, Speech, and Signal Processing, 35, 162–172. 104198.
Bickel, S.H. & Natarajan, R .R ., 1985. Plane-wave Q deconvolution, Jin, Y., Wu, X., Chen, J., Han, Z. & Hu, W., 2018. Seismic data denoising
Geophysics, 50, 1426–1439. by deep-residual networks, in SEG (Society of Exploration Geophysicists)
Clarke, G.K.C., 1968. Time-varying deconvolution filters, Geophysics, 33, International Exposition and 88th Annual Meeting 2018, Anaheim, Cali-
936–944. fornia.

1192
Journal of Geophysics and Engineering (2023) 20, 1180–1193 Zhou et al.

Johnston, D.H., Toksöz, M.N. & Timur, A., 1979. Attenuation of seismic Sun, S.Z., Sun, X., Wang, Y. & Xie, H., 2015. Q estimation using
waves in dry and saturated rocks: II. Mechanisms, Geophysics, 44, 691– modified S transform based on pre-stack gathers and its applica-
711. tions on carbonate reservoir, Journal of Geophysics and Engineering, 12,
Jones, T.D., 1986. Pore fluids and frequency-dependent wave propagation 725–733.
in rocks, Geophysics, 51, 1939–1953. Van der Baan, M., 2008. Time-varying wavelet estimation and deconvolu-
Li, W., Chen, J., Zhang, Y. & Shi, X., 2022. Research of medical image de- tion by kurtosis maximization, Geophysics, 73, V11–V18.
noising based on improved weighted mean filtering, Journal Of Liaoning Wang, Y., 2002. A stable and efficient approach of inverse Q filtering,
University (Natural Sciences Edition), 49, 30–35. (in Chinese) Geophysics, 67, 657–663.
Lowney, B., Lokmer, I., O’Brien, G.S., Amy, L., Bean, C.J. & Igoe, M., 2020. Wang, Y., 2006. Inverse Q-filter for seismic resolution enhancement,
Enhancing interpretability with diffraction imaging using plane-wave Geophysics, 71, V51–V60.
destruction aided by frequency-wavenumber f-k filtering, Interpretation, Wang, Y., 2008. Seismic Inverse Q Filtering, Blackwell, Oxford.
8, T541–T554. Wu, S., Wang, Y., Di, Z. & Chang, X., 2018. Random noise attenuation by
Margrave, G.F., Lamoureux, M.P. & Henley, D.C., 2011. Gabor deconvolu- 3D multi-directional vector median filter, Journal of Applied Geophysics,
tion: estimating reflectivity by nonstationary deconvolution of seismic 159, 277–284.
data, Geophysics, 76, W15–W30. Xie, J., Wang, X., Wang, X., Wu, D., Zeng, H. & Jin, B., 2021. Multiple-
Mundim, E.C., Schots, H.A., de Araújo, J.M. & Tavares, D.M., 2006. suppression method using the 𝜆-f domain high-resolution parabolic

Downloaded from https://academic.oup.com/jge/article/20/6/1180/7281720 by guest on 18 March 2025

WTdecon, a colored deconvolution implemented by wavelet transform, radon transform with curvature magnification, Applied Geophysics, 18,
The Leading Edge, 25, 398–401. 1–10.
Qian, F., Chen, L., Zhang, F. & Hu, G., 2016. Prestack gather flattening us- Yang, D., Qin, C., Wang, R., He, M., Wang, S. & Li, J., 2016. Estimation of
ing segmental dynamic time warp, in SEG (Society of Exploration Geo- Q value based on the energy spectrum centroid frequency shift method,
physicists) International Exposition and 86th Annual Meeting 2016, Dallas, Oil Geophysical Prospecting, 51, 863–867.
Texas. Yang, L., Chen, W., Liu, W., Zha, B. & Zhu, L., 2020. Random noise attenua-
Rosa, A.L.R. & Ulrych, T.J., 1991. Processing via spectral modeling, tion based on residual convolutional neural network in seismic datasets,
Geophysics, 56, 1244–1251. IEEE Access, 8, 30271–30286.
Shi, Z., Zhang, Z., Zhou, H. & Wang, Y., 2022. Inversion-based pre-stack Zhang, B., Chang, D., Lin, T. & Marfurt, K.J., 2015. Improving the qual-
gather flattening by exploiting temporal sparsity, Digital Signal Process- ity of prestack inversion by prestack data conditioning, Interpretation, 3,
ing, 132, 103783. T5–T12.
Singleton, S., 2009. The effects of seismic data conditioning on prestack Zhang, C. & Ulrych, T.J., 2007. Seismic absorption compensation: a least
simultaneous impedance inversion, The Leading Edge, 28, 772–781. squares inverse scheme, Geophysics, 72, R109–R114.
Sun, M., Li, Z., Li, Z., Li, Q., Liu, Y. & Wang, J., 2020. A noise attenua- Zhou, P., Zhang, Y., Liu, Z. & Liu, C., 2016. Seismic gather optimization,
tion method for weak seismic signals based on compressed sensing and Oil Geophysical Prospecting, 51, 232–237 +203–204.
CEEMD, IEEE Access, 8, 71951–71964.

© The Author(s) 2023. Published by Oxford University Press on behalf of the SINOPEC Geophysical Research Institute Co., Ltd. This is an Open Access article distributed under the 1193
terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any
medium, provided the original work is properly cited.