GEOPHYSICS, VOL. 86, NO. 4 (JULY-AUGUST 2021); P. R485–R495, 16 FIGS.

10.1190/GEO2020-0383.1

Target-oriented time-lapse waveform inversion using deep

learning-assisted regularization

Yuanyuan Li1, Tariq Alkhalifah1, and Qiang Guo2

ABSTRACT time-lapse virtual data for the target-oriented inversion. Consid-

ering that the injection and production wells are often present in
Detection of the property changes in the reservoir during in- the target zone, we can incorporate the well velocity information
jection and production is important. However, the detection proc- with the time-lapse inversion by using regularization to comple-
ess is very challenging using surface seismic surveys because ment the resolution and illumination at the reservoir. We use a
these property changes often induce subtle changes in the seismic deep neural network to learn the statistical relationship between
signals. The quantitative evaluation of the subsurface property the inverted model and the facies interpreted from well logs. The
obtained by full-waveform inversion allows for better monitoring trained network is used to map the property changes extracted
of these time-lapse changes. However, high-resolution inversion from the wells to the target inversion domain. We then perform
is usually accompanied with a large computational cost. Besides, another time-lapse inversion, in which we fit the predicted data
the resolution of inversion is limited by the bandwidth and aper- difference to the redatumed one from observation, as well as fit
ture of time-lapse seismic data. We have applied a target-oriented the model to the predicted velocity changes. The numerical re-
strategy through seismic redatuming to reduce the computational sults demonstrate that our method is capable of inverting for
cost by focusing our high-resolution delineation on a relatively the time-lapse property changes effectively in the target zone
small zone of interest. The redatuming technique generates by incorporating the learned model information from well logs.

INTRODUCTION required to stabilize the simulation of high-frequency wavefields

(Alford et al., 1974; Routh et al., 2017). Its corresponding time-
Time-lapse seismic data have been widely used to monitor the lapse applications require even more computational cost because
fluid injection and the hydrocarbon extraction in oilfield develop- we often need two FWI runs on the monitor and baseline data. Sec-
ment and evaluate the storage of CO2 in carbon capture and storage ond, due to the limitation that we encounter with surface seismic
(Lumley, 2001; Arts et al., 2003; Chadwick et al., 2010; Urosevic data in terms of signal-to-noise ratio (S/N), frequency band, aper-
et al., 2010). Time-lapse full-waveform inversion (TLFWI) pro- ture, and coverage, seismic data nearly always admit incomplete
vides a quantitative estimation of the subsurface property changes subsurface information with limited resolution (Sirgue and Pratt,
induced by the aforementioned operations by fitting the time-lapse 2004; Alkhalifah, 2016; Li et al., 2018). TLFWI is a severely
seismic data directly (Hicks et al., 2016; Maharramov et al., 2016; ill-posed inverse problem due to the incomplete measured data
Kazei and Alkhalifah, 2018). High-resolution delineation of prop- and the large search space (Tarantola, 1984; Zhang and Huang,
erty changes is a major goal of time-lapse seismic inversion. How- 2013). In this work, we improve the performance of TLFWI by
ever, there are two primary issues that hinder the time-lapse addressing these two issues.
inversion in practice. First, high-resolution full-waveform inversion The target-oriented inversion strategy is capable of reducing the
(FWI) is computationally intensive because fine discretization is computational cost by inverting for a relatively small target zone of

Manuscript received by the Editor 3 June 2020; revised manuscript received 11 January 2021; published ahead of production 9 April 2021; published online
23 June 2021.
1
King Abdullah University of Science and Technology, Physical Science and Engineering Division, 4700 KAUST, Thuwal 23955-6900, Saudi Arabia.
E-mail: yuanyuan.li@kaust.edu.sa (corresponding author); tariq.alkhalifah@kaust.edu.sa.
2
Formerly King Abdullah University of Science and Technology, Physical Science and Engineering Division, 4700 KAUST, Thuwal 23955-6900, Saudi
Arabia; presently Shearwater Geoservices, Tunbridge Wells TN4 8BS, UK. E-mail: qiang.guo@kaust.edu.sa.
© 2021 Society of Exploration Geophysicists. All rights reserved.

R485

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 R486 Li et al.

the subsurface (van Manen et al., 2007; Ayeni and Biondi, 2010; mation, can be incorporated into the inversion to complement the
Borisov and Singh, 2015; Yuan et al., 2017; Biondi et al., 2018; resolution and illumination at the target monitoring zone (Asnaa-
da Costa et al., 2019; Guo and Alkhalifah, 2020; Li et al., shari et al., 2013; Singh et al., 2018; Zhang et al., 2018; Aragao
2020b). The target-oriented strategy can be applied to time-lapse and Sava, 2020). The model information in the sparsely sampled
seismic inversion because the monitored property changes usually wells should be projected to the inversion region to provide a prior
occur in the reservoir of interest (Li et al., 2020b, 2020c). Redatum- model for regularization of the inversion.
ing techniques generate a virtual data set for target-oriented inver- Deep learning (DL) (LeCun et al., 2015) can efficiently learn a
sion by projecting sources and receivers to a desired datum level just statistical relationship between the input and output features in a
above the target area (Wapenaar and Fokkema, 2006; Yang et al., data-driven manner (Zhang and Alkhalifah, 2019; Li et al., 2020a;
2012; Guo and Alkhalifah, 2020). In addition, the wavefield distor- Song et al., 2021). Zhang and Alkhalifah (2019) use deep neural
tion from the complex overburden, above the datum level, is also sup- networks (DNNs) to build the proper statistical connection that con-
pressed by waveform redatuming; thus, it has less of an effect on the verts seismic estimates to facies interpreted from well logs, and they
target inversion (Zhao et al., 2020). The redatuming is expected to derive a prior model that incorporates the well information. The
improve time-lapse repeatability by suppressing 4D unrepeatable prior model helps the subsequent elastic FWI converge to a more
noise such as that induced by near-surface diurnal and seasonal accurate inversion result. Facies are defined as groups of seismic
weather cycles as well as small changes in acquisition geometry properties and conformity layers that share a particular relationship
and shot coupling (Bakulin and Calvert, 2006). Li et al. (2020b) in- with geologic and lithologic properties (Kemper and Gunning,
vestigate the impacts of the overburden model and 4D noise on the 2014). To retain the high-resolution well velocity in the prior model,
process of redatuming and target-oriented time-lapse inversion. Li et al. (2020a) extract the statistical distribution information, de-
In TLFWI, the time-lapse property changes are given by the differ- fined in a Gaussian window, from the wells and map the interpreted
facies featured by the distribution information (mean and variance)
ence of the monitor and baseline FWI resulting models. The two FWI
to the inversion region to recover a prior model. The high-resolution
runs for monitor and baseline can be performed in parallel or in a
prior information is then added to the elastic FWI scheme as a regu-
coupled fashion (Denli and Huang, 2009; Zheng et al., 2011; Routh
larization term.
et al., 2012; Asnaashari et al., 2014; Hicks et al., 2016), often depend-
In our proposed method, we adopt waveform redatuming to re-
ing on their assumptions about the repeatability of the acquisition
trieve the target-oriented data for TLFWI focused on the target zone.
survey. Double-difference waveform inversion (DDWI) (Denli and
Thus, we reduce the computational burden of TLFWI by inverting
Huang, 2009) renders the estimated property changes less contami-
for a relatively small zone of interest. We also incorporate well
nated by the imperfect inverted baseline model. However, 4D unre-
velocity information to constrain the TLFWI scheme by using a
peatable noise deteriorates its performance (Yang et al., 2015).
DL-assisted regularization technique to improve the inversion accu-
Prior information needs to be incorporated into the inversion
racy and resolution. The required prior model is predicted from a
through regularization to steer inversion toward geologically real- trained DNN that learns the statistical relationship between the
istic models. Total variation (TV) regularization has been used to time-lapse seismic estimation and the facies interpreted from well
preserve sharp boundaries in the inversion result (Lin and Huang, logs. Numerical tests on the time-lapse Marmousi and SEG Ad-
2015; Esser et al., 2018; Kalita et al., 2019). Zhang and Huang vanced Modeling Program (SEAM) models are used to demonstrate
(2013) incorporate the location of a monitoring region as prior in- the performance of the proposed method.
formation into DDWI to improve the inversion quality. Considering
that injection and production wells are often present in the target
zone, well velocity information, containing fine-scale model infor- THEORY
In this section, we first briefly review the re-
datuming scheme in the time-space domain,
which provides the virtual data set at the prede-
signed datum level for the target-oriented FWI.
Then, we introduce a regularized TLFWI method
that incorporates well velocity information into
inversion using DL.

Waveform redatuming
A typical configuration for waveform reda-
tuming is shown in Figure 1. The sources and
receivers for a surface acquisition system are lo-
cated near the earth’s surface. The redatuming
operation aims to redatum the surface seismic
Figure 1. Diagram demonstration for the redatuming operation. The sources (xs ) and data from the earth’s surface to a desired datum
receivers (xr ) for a typical acquisition survey are located near the earth surface. The level just above the target area by removing the
redatuming is applied to redatum the surface seismic data from the earth surface to effects of the overburden on the wavefields. The
the datum level. The virtual sources (xvs ) and receivers (xvr ) for the redatumed data
set are deployed at the datum level (the red line). The term xvr ¼ xvs þ h (h: the subsur- overburden denotes the zone above the datum
face offset vector). The wavefield distortions caused by the complex overburden are level. It is an inverse problem that predicts the vir-
desired to be removed by the redatuming implementation. tual data set (Green’s function) at the datum level

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 Target-oriented TLFWI using DL R487

using the observed seismic data and a reasonable estimate of the over- propagation above the datum level. The gradient of the objective
burden model. To solve this inverse problem, the corresponding mod- function with respect to the redatumed data (gD ) is expressed as
eling operator needs to be constructed. Based on the convolution-type ∂J X
representation of Green’s functions (Wapenaar and van der Neut, ðxvs þ h; xvs ; tÞ ¼ − huðxvs ; xs ; tÞ; u† ðxvs þ h; xs ; tÞit ;
2010), the upgoing wavefield v at a virtual receiver xvr excited ∂gD s
by the source at xs can be written as a convolution-type representation (7)
(Wapenaar and van der Neut, 2010):
Z where u† is the backward propagating wavefield from the adjoint
vðxvr ; xs ; tÞ ¼ gD ðxvr ; xvs ; tÞ  uðxvs ; xs ; tÞdxvs ; (1) source at the receivers. We can compute the gradient for the virtual
SD data (gD ) by crosscorrelating the forward- and backward-propagat-
ing wavefields at the datum level followed by a summation over the
where  refers to a temporal convolution operator, SD denotes the sources (Guo and Alkhalifah, 2019). Given a prior estimate of the
surface at datum level, xvs represents the virtual source location, u overburden model, the redatuming process is performed iteratively
represents the downgoing wavefield from surface to datum, and gD using a gradient-based optimization algorithm. To some extent, the
indicates the unknown Green’s function at the datum level describ- redatuming implementation is similar to least-squares extended im-
ing scattering from the underlying model. The downgoing wave- aging (Symes, 2008), which can be related to earlier findings by
field (u) and the upgoing wavefield (uD ) from the datum are Vasconcelos et al. (2010). We will apply the redatuming algorithm
obtained by solving the following wave equations: to the time-lapse seismic data to prepare the corresponding virtual
data sets for the following target-oriented TLFWI.
Fuðx; xs ; tÞ ¼ fðx; xs ; tÞ; (2)
Regularized TLFWI using DL
Z
FuD ðx; xs ; tÞ ¼ δðx − xvr Þvðxvr ; xs ; tÞdxvr ; (3) Once the base and monitor data sets of the time-lapse experiment are
SD redatumed to the datum level, just above the target zone, we then per-
form DDWI (Denli and Huang, 2009) to estimate the time-lapse
where F is the acoustic wave modeling operator changes in the monitoring (target) zone from the redatumed time-lapse
F ¼ ðð1∕c2 Þð∂2 ∕∂t2 Þ − ∇2 Þ, c is the velocity, f denotes the source data sets. In the DDWI algorithm, the baseline virtual data set is first
function, and δðx − xvr Þ is a Dirac delta function. Substituting equa- used to invert for the baseline model. Then, we update the monitor
tion 1 into equation 3, the upgoing wavefield (uD ) is written as model starting from the inverted baseline model (mibase ) by minimiz-
Z Z ing the misfits between the simulated time-lapse data difference for the
FuD ðx; xs ; tÞ ¼ δðx − xvr Þ gD ðxvr ; xvs ; tÞ virtual survey and the data difference of the monitor and baseline re-
SD SD datumed data. The double-difference objective function is expressed as
 uðxvs ; xs ; tÞdxvs dxvr : (4) 1
Jd ðmÞ ¼ kuðmÞ − uðmibase Þ − δdvirtual k22 ; (8)
2
The wavefield that is related to the response of the overburden pro-
vides a pilot trace at the virtual source location. Then, the modeling where u is the simulated seismic data for the virtual survey using the
operator can be given as (Guo and Alkhalifah, 2019) acoustic modeling operator, mibase is the estimated baseline model, m
ZZ is the monitor model, and δdvirtual is the data difference for the virtual
FuD ðx;xs ;tÞ ¼ uðxvs −h;xs ;tÞ survey. The model updates (m − mibase ) are considered to represent the
model changes (δm) in the target zone between the monitor and
gD ðxvs −h;xvs ;tÞδðx−xvs þhÞdhdxvs ; (5) baseline.
Prior information in the monitoring area from other geophysical
where h is the subsurface offset vector pointing from the virtual surveys can help constrain or regularize the inverse problem to
source. The simulated upgoing wavefield (uD ) is generated from deliver a more accurate and robust estimate. Here, we assume that
the secondary source, located at the datum level, given by the con- some well logs provide reliable measurements of the property
changes in the monitoring zone, but they have very limited cover-
volution of the downgoing wavefield (u) with the Green’s function
age. To incorporate the prior information from well logs, a model-
(gD ) for the underlying model.
space regularization term is added to the objective function:
The redatumed virtual data (the Green’s function gD ) can be re-
trieved in an optimization problem by minimizing the data misfits
JðmÞ ¼ Jd ðmÞ þ βJm ðmÞ; (9)
(Alkhalifah and Wu, 2016; Li et al., 2019):
1X
where
min J ¼ kðuðmo Þ þ uD ðmo ; gD ÞÞ − dk22 ; (6)
gD 2 s 1
J m ðmÞ ¼ kWm ðm − ðmibase þ δmprior ÞÞk22 ; (10)
2
where d is the observed seismic data, mo denotes the overburden
model, and u and uD represent the simulated wavefield data at the and β is a weighting parameter balancing the contributions from the
receivers obtained by the acoustic wave equation 2 and the model- data and model misfit terms, Wm is a diagonal weighting matrix,
ing equation 5, respectively. It is noteworthy that an estimate of the and δmprior is a prior model for the property changes in the target
overburden model is required to account for the seismic wave zone that incorporates the well velocity information by DL.

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 R488 Li et al.

The prior model from DNN statistical connection between the features of the inverted model
and the facies classes. We then apply the trained network to the target
We derive the prior model by mapping the inverted time-lapse inversion zone to predict the facies distribution. Considering that the
model to the interpreted facies from wells based on their statistical
velocity changes for each of the facies have been defined from the
relationship predicted by a DNN (Li et al., 2020a). The architecture
well logs, the prior model for the velocity changes can be extracted
of the DNN is shown in Figure 2. The neural network consists of
from the facies distribution. Finally, we implement a regularized
input and output layers, with fully connected hidden layers in be-
DDWI by incorporating the prior model into the inversion workflow.
tween. The data flow from the input layer, through the hidden
layers, and to the output layer in a feedforward fashion. For a vector
The algorithm
of inputs a0 , a general forward-propagation equation is written as
al ¼ gl ðWl al−1 þ bl Þ, where Wl and bl refer to the weighting ma- We implement the proposed target-oriented TLFWI algorithm
trix and the bias vector for the lth layer, respectively. The activation following the workflow shown in Figure 3:
function gl is used to induce nonlinearity in the formulation. Here,
we adopt rectified linear unit (ReLU) functions and a softmax ac- 1) Using a reasonable overburden model, usually estimated from
tivation function for the hidden layers and the output layer, respec- FWI, we redatum the baseline and monitor data sets from the
tively (Nair and Hinton, 2010). The softmax function admits a earth’s surface to the new datum to provide the target-oriented
probability distribution that categorizes the input classification with time-lapse data sets.
regard to the output facies, as denoted by ðp1; · · · ; pi; · · · ; pnf Þ in the 2) We then invert for the baseline model in the target zone (mibase )
output layer in Figure 2. by matching the redatumed baseline data sets.
The inverted baseline model (mibase ) and the recovered velocity 3) We use the baseline inversion result (mibase ) as a starting model
changes (δm) from seismic data act as input features to discriminate to invert for the monitor model (m) using the data difference
the class of facies. Considering that the property changes often oc- between monitor and baseline plus the synthetic data from
cur in a limited depth range, the depth of the training samples is also mibase . The model updates (δm ¼ m − mibase ) are considered
used as an input feature to improve the prediction accuracy of the as the estimated property changes.
DNN-based facies classifier. We can identify at least two facies, 4) We then prepare the training data set for optimizing the DNN
representing the injection and noninjection areas, from the wells parameters: label the selected training samples from the inverted
to set labels for the training data set. In some complicated cases, model using the interpreted facies from the wells. Using the la-
the injection area needs to be described by more than one facies beled data, we train the network.
for a detailed interpretation. 5) We apply the trained DNN to the target inversion zone and build
To improve the performance of the trained DNN, the synthetic the prior model for property changes.
minority over-sampling technique is applied to enlarge the training 6) We apply an FWI similar to that described in step 3, but based
data sets and reduce the imbalance of different facies (Chawla et al., on the objective function defined in equation 9, which includes
2002). We also use a random dropout of 30% to avoid overfitting a regularization term given by the prior model.
(Srivastava et al., 2014). For the training process, we initialize the 7) The resulting model changes represent the output of the approach.
weights and biases using a normal distribution and update them using
the Adam optimizer, with the loss function given by a sparse cross
entropy (Kingma and Ba, 2014). We set the batch size and the total
number of epochs to 64 and 2000, respectively, for training the net-
work. Once the training process is finished, the DNN constructs a

Figure 3. The workflow for the proposed inversion method. The

input and output for the workflow are shown in the blue boxes.
Figure 2. The architecture of the DNN. The DNN consists of input The DDWI part is described by the red boxes. The DL part for
and output layers, with fully connected hidden layers in between. building the prior model is shown by the black boxes; the dashed
There are three hidden layers (denoted as L1 ; L2 ; and L3 ) and 64 line indicates that it is activated according to the inversion require-
neurons (denoted as N 1; · · · ; N i; · · · ; N 64 ) in each layer. Three input fea- ment, and for this case, the two dashed lines pointing to “DNN-
tures for the input layer include the inverted baseline model (mibase ), based facies classification” are activated after the first implementa-
the inverted property changes (δm), and the depth. The output layer tion of DDWI, and the dashed line pointing to “output the model
delivers the probabilities ðp1; · · · ; pi; · · · ; pnf Þ of being categorized into changes” is activated after including regularization into another im-
each facies given the input. plementation of DDWI.

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 Target-oriented TLFWI using DL R489

NUMERICAL EXAMPLES survey (Figure 6b). We first conduct a DDWI on the surface time-
lapse seismic data, and only the target zone is updated for a fair
We apply the proposed inversion algorithm to the Marmousi and comparison with the following inversion results. Figure 7a shows
SEAM model examples. In the SEAM model example, the time- the DDWI result using the surface seismic data. Figure 7b shows the
lapse data set is provided by the SEAM project. In these two exam- target-oriented time-lapse velocity changes resolved from DDWI
ples, we use an eighth-order in space finite-difference scheme to using the redatumed data. We can see that the target-oriented
simulate the seismic wave propagation. The absorbing boundary time-lapse inversion approach captures the velocity changes at
condition is used on all sides of the model. We ignore the time-lapse the correct position, and it shows better inversion accuracy than
changes in the overburden model and use the same overburden the time-lapse inversion approach using the surface seismic data.
model to redatum the baseline and monitor data. We refer to Li et al. However, the target-oriented DDWI result still includes some arti-
(2020b) for a detailed study of the impact of the overburden model
and 4D noise on the process of redatuming and target-oriented time-
lapse inversion. a)

Marmousi synthetic example

Figure 4a shows the baseline Marmousi model. We introduce
velocity changes into the baseline model to construct the monitor
model, and the position of these changes is highlighted by the red
arrow in Figure 4a. We set the datum at depth 2.18 km (the dashed
red line in Figure 4a), just above the target monitoring zone. We
only need to focus the inversion on a small area, indicated by
the box in Figure 4a. Figure 4b shows the time-lapse velocity b)
changes in the target region. A Ricker wavelet with peak frequency
of 15 Hz is used as the source wavelet. We deploy 92 shots spread
evenly at a depth of 50 m to excite the seismic wave energy, and we
use 461 receivers located at a depth of 0.2 km to record the seismic
data for the baseline and monitor models.
Starting from the initial model (Figure 5a), we invert for the over-
burden model above the datum using two frequency bands 3–12 and
3–30 Hz, sequentially. Figure 5b shows the final inverted overbur-
den model. Given the overburden model, we redatum the recorded
baseline and monitor seismic data to the new datum to provide the Figure 5. (a) The initial model and (b) the inverted overburden
virtual data sets for the following target-oriented inversion. The vir- model using standard FWI for the Marmousi example.
tual shot gathers for the baseline at the datum level are shown in
Figure 6a. Subtracting the baseline virtual data from the monitor
virtual data, we obtain the time-lapse data difference for the virtual
a)

a)
Time(s)

b)

b) Distance (km)
5 5.5 6 6.5 7 7.5 8
2.2
Time(s)

0.1
2.4
Depth (km)

0.05
km/s

2.6 0

–0.05
2.8
–0.1

Figure 4. (a) The true baseline model and (b) the true time-lapse Figure 6. (a) The virtual shot gathers for baseline at the datum level
velocity changes in the target zone for the Marmousi example. and (b) the time-lapse data difference (δdvirtual ) for the virtual survey
The dashed white line shows the well location. for the Marmousi example.

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 R490 Li et al.

facts and the magnitude of the changes is underestimated. To pre- noninjection areas (zero velocity changes), are identified from
serve the sharp boundaries and suppress the artifacts in the DDWI the well and are used as the labels for the training data set. Once
result (Figure 7b), we also apply a TV regularization (Lin and the training is finished, the DNN learns a statistical relationship be-
Huang, 2015) to the target-oriented DDWI, and the inversion result tween the inverted velocity and the facies. Then, we apply the
is shown in Figure 7c. We can see that the inversion result is im- trained neural network to the target inversion zone to predict the
proved, but it still fails to describe the true velocity changes cor- facies distribution and we derive the prior model for velocity
rectly especially below the depth of 2.85 km. changes (Figure 7d). Finally, we conduct a regularized DDWI using
A vertical profile at 6.5 km (pointed out by the dashed white line the predicted prior model with the objective function shown in
in Figure 4b) represents the well-log information, which provides equation 9. To suppress the artifacts in the noninjection area (no
complementary high-resolution information to the target zone. Two property changes), we penalize the relevant model misfits from
facies, representing the injection (nonzero velocity changes) and the prior model more by increasing the weighting factor in the ma-
trix Wm corresponding to the noninjection area. Specifically, Wm is
Distance (km)
a) 5 5.5 6 6.5 7 7.5 8 not an identity matrix; the weighting factors for the noninjection and
2.2 injection areas are set to 100 and 1, respectively. We consider a 20%
0.1
2.4
misfit contribution from the model-space regularization term by ad-
Depth (km)

0.05
justing the balancing factor (β). Figure 7e shows the estimated

km/s
2.6 0
velocity changes from the regularized DDWI using the redatumed
–0.05 data set. We can see that the image artifacts are alleviated and the
2.8
–0.1 inversion result is cleaner than that from the target-oriented DDWI
with the TV regularization (Figure 7c). The shape of the recovered
b) Distance (km) changes is similar to that of the prior model thanks to the constraints
5 5.5 6 6.5 7 7.5 8
2.2 from the prior model. For a more detailed comparison, we plot the
0.1 vertical profiles of the inversion results at location 6.8 km (Figure 8).
2.4
0.05 The notch shown in the red line mainly comes from the model up-
Depth (km)

dates from the time-lapse seismic data. These figures show that the
km/s

2.6 0

–0.05 inversion result using the proposed DL-based regularization scheme

2.8
–0.1 interprets the true velocity changes with higher accuracy and res-
olution compared with the other inversion results.
c) Distance (km) Figure 9 shows the predicted time-lapse data difference
5 5.5 6 6.5 7 7.5 8 (uðmibase þ δmÞ − uðmibase Þ) and the residuals (uðmibase þ δmÞ−
2.2
0.1
uðmibase Þ − δdvirtual ) using the estimated velocity changes (δm)
2.4 in Figure 7b and 7e. The predicted data difference matches the
Depth (km)

0.05
redatumed one in Figure 6b to some extent, but there are still some
km/s

2.6 0
residuals. These residuals are reasonable because the redatumed
–0.05
2.8 data difference includes some unpredictable artifacts generated
–0.1
from the linear optimization of the redatuming inverse problem us-
ing an imperfect overburden model. Besides, the residuals corre-
d) 5 5.5 6
Distance (km)
6.5 7 7.5 8 sponding to the regularized inversion result are not reduced
2.2 because the regularization renders a steady incorporation of the
0.1
2.4
prior model into the inversion to compete with the contribution
0.05
Depth (km)

of redatumed time-lapse data.

km/s

2.6 0

–0.05
2.8 0.2
TLFWI
–0.1
0.15 TO−TLFWI
TO−TLFWI+TV
Vel (km/s)

0.1
e) Distance (km) DL assisted TO−TLFWI
5 5.5 6 6.5 7 7.5 8 True velocity changes
0.05
2.2
0.1 0
2.4
0.05 −0.05
Depth (km)

km/s

2.6 0 −0.1
2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3
–0.05
Depth (km)
2.8
–0.1

Figure 8. The vertical profiles of the velocity changes at x = 6.8 km

for the Marmousi example. The green line, the DDWI result using
Figure 7. The Marmousi example: (a) the DDWI result using the the surface seismic data shown in Figure 7a; the blue line, the target-
surface seismic data (only the target zone is updated for a fair com- oriented DDWI result shown in Figure 7b; the magenta line, the
parison), (b) the target-oriented DDWI result using the redatumed target-oriented DDWI result using a TV regularization shown in
data, (c) the target-oriented DDWI result using a TV regularization, Figure 7c; the red line, the final inversion result using DL-assisted
(d) the predicted prior model using DL, and (e) the final inverted regularization shown in Figure 7e; and the dashed black line, the
velocity changes regularized by the prior model in (d). true velocity changes.

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 Target-oriented TLFWI using DL R491

Robustness of the DL-assisted prior model building DL-assisted prior model building in the presence of random noise.
The noisy shot gathers for the baseline with S/Ns of 5.0 and 0.5 dB
We add Gaussian white noise with various energy levels to are shown in Figure 10a and 10b. The S/N is defined as
the baseline and monitor data to test the robustness of the

a) b) Figure 9. The Marmousi example: (a) the

predicted time-lapse data difference
(uðmibase þ δmÞ − uðmibase Þ) and (b) the residuals
(uðmibase þ δmÞ − uðmibase Þ − δdvirtual ) for the virtual sur-
vey using the inverted velocity changes (δm) in Figure 7b.
(c) The predicted time-lapse data difference and (d) the
residuals corresponding to the inverted velocity changes
in Figure 7e (these figures are plotted in the same scale).
c) d)

a) b)

Distance (km) Distance (km)

c) 5 5.5 6 6.5 7 7.5 8
d) 5 5.5 6 6.5 7 7.5 8
2.2 2.2
4.5 4.5

2.4 4 2.4 4
Depth (km)

Depth (km)
km/s

km/s
3.5 3.5
2.6 2.6
3 3
2.8 2.5
2.8 2.5

2 2

Distance (km) Distance (km)

e) 5 5.5 6 6.5 7 7.5 8
f) 5 5.5 6 6.5 7 7.5 8
2.2 2.2
0.1 0.1
2.4 2.4
0.05 0.05
Depth (km)

Depth (km)
km/s

km/s

2.6 0 2.6 0

–0.05 –0.05
2.8 2.8
–0.1 –0.1

Distance (km) Distance (km)

g) 5 5.5 6 6.5 7 7.5 8
h) 5 5.5 6 6.5 7 7.5 8
2.2 2.2
0.1 0.1
2.4 2.4
0.05
Depth (km)

0.05
Depth (km)

km/s

0
km/s

0 2.6
2.6
–0.05
–0.05 2.8
2.8 –0.1
–0.1

Figure 10. The robustness test of the DL-assisted prior model building using the Marmousi example: (a and b) the noisy shot gathers for the
baseline with an S/N of 5.0 and 0.5 dB, respectively; (c and d) the inverted baseline models, respectively; (e and f) the recovered time-lapse
changes, respectively; and (g and h) the predicted prior model in the target zone corresponding to the noisy data shown in (a and b).

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 R492 Li et al.

S∕NdB ¼ 10 log10 ðA2signal ∕A2noise Þ, where A denotes the root-mean- describes the kinematics of the medium accurately. Considering that
square (rms) amplitude. We first conduct the target-oriented DDWI the target of interest is the reservoir between 3.5 and 4.2 km depth,
scheme following steps 1–3 in the workflow. Figure 10c and 10d we set the datum at a depth of 3 km, which splits the whole model
shows the inverted baseline models in the target zone using the into the overburden and the underlying model. Using the overbur-
noisy data with an S/N of 5.0 and 0.5 dB, respectively, and den from the starting model, we implement the redatuming process
Figure 10e and 10f shows the corresponding estimated velocity to retrieve the virtual data sets at the datum from the provided base-
changes. As shown in Li et al. (2020b), the inverted baseline model line and monitor data sets. Figure 14a shows the redatumed virtual
(mibase ) is less influenced by the random noise, whereas the esti- shot gathers for baseline. The time-lapse difference between the re-
mated time-lapse changes (δm) deteriorate as more noise is in- datumed monitor and baseline data sets is shown in Figure 14b. We
volved. The terms mibase and δm and the depth of the training then apply the DDWI algorithm to the redatumed data. Figure 15a
samples are used as the input features of the DNN. Once the training and 15b shows the inverted baseline velocity and time-lapse veloc-
is finished, the prior model can be predicted from the trained DNN. ity changes in the target zone, respectively.
The prior model (Figure 10g) corresponding to the S/N of 5.0 dB is Two pseudowells at 7.5 and 8.125 km (pointed out by the dashed
predicted successfully, whereas the prior model (Figures 10h) cor- white line in Figure 12) are used to boost the model information,
responding to the S/N of 0.5 dB is somewhat influenced by the seri- and three facies can be interpreted from the wells representing two
ously contaminated δm. Overall, these test results prove that the injection levels and the none injection area. After training, the DNN
DL-assisted prior model building shows good robustness even with is applied to the target inversion region to generate the prior model
noisy data because three input features, rather than only δm, are
combined to discriminate the facies classes.
a)
The SEAM model
A 3D SEAM time-lapse model was constructed to simulate the
changes in an oilfield reservoir. The model includes a water layer of
267.5 m depth on the top and four fault blocks across the reservoir
with a thickness of approximately 500 m. The provided seismic data
set was simulated by the SEAM project assuming an isotropic elas-
tic medium and recorded by ocean-bottom nodes (OBNs). We ex-
tract a 2D line along the middle of the SEAM model from the OBN
data set to further verify the feasibility of the proposed method. Its b)
acquisition survey geometry includes 60 OBNs with 175 m interval
at 267.5 m depth and 501 shots with 25 m spacing at 10 m depth. A
shot gather representing the pressure field for the baseline model is
shown in Figure 11a. The corresponding time-lapse data difference
is shown in Figure 11b. For simplification, as is usually done with
marine data, we ignore the elastic effects in the seismic data and use
an acoustic assumption.
Figure 12a and 12b shows the true baseline P-wave velocity and
density model, respectively. Figure 12c and 12d shows the velocity
Distance (km)
and density changes in the target zone, caused by the hydrocarbon c) 3 4 5 6 7 8 9 10 11
3
production. The time-lapse changes include changes with sharp edges 0.5
3.2
and smooth changes. In this case, we only invert for the velocity
Depth (km)

3.4
changes in the target zone. After production, the property changes and VP (km/s)
3.6
geometric shifts lead to the time-lapse seismic response (Figure 11b). 0

3.8
The smoothed version of the average of the baseline and monitor
4
models is used as the starting model (Figure 13), which potentially
4.2 –0.5

a) b) d) 3 4 5 6
Distance (km)
7 8 9 10 11
3

3.2
0.05
Depth (km)

3.4
rho (g/cm3 )

3.6 0

3.8

4 –0.05

4.2

Figure 12. The baseline SEAM model: (a) P-wave velocity and
Figure 11. (a) A shot gather of the pressure component for the base- (b) density. The true time-lapse (c) velocity and (d) density changes
line model and (b) the corresponding time-lapse data difference for in the target zone for the SEAM example (the dashed white lines
the SEAM example. denote the pseudowells at 7.5 and 8.125 km).

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 Target-oriented TLFWI using DL R493

(Figure 15c). Considering that the smooth changes are not included using the prior model provides a well-focused inversion result (Fig-
successfully in the prior model, we use an identity weighting matrix ure 15d), which facilitates interpreting time-lapse variations in the
Wm to admit smooth updates coming from the data misfits. We set reservoir. Note that the inverted model somewhat identified the two
the balancing factor (β) to render a 40% misfit contribution from the injection levels separated by a very thin noninjection layer at 3.4 km
model-space regularization term. The regularized DDWI method depth. The vertical profiles at location 7.0 km (Figure 16), which

a) a)

b)
b)

c)

Figure 13. The starting model for the SEAM example: a smoothed
version of the average of the baseline and monitor models (a) veloc-
ity and (b) density.

d)
a)

Figure 15. The SEAM example: (a) the inverted baseline model
in the target zone, (b) the standard target-oriented time-lapse in-
version result, (c) the predicted prior model using DL, and (d) the
final inverted velocity changes regularized by the prior model in (c).

b)
TO−TLFWI
DL assisted TO−TLFWI
0.5
True velocity changes
Vel (km/s)

−0.5

3 3.2 3.4 3.6 3.8 4 4.2

Depth (km)

Figure 16. The vertical profiles of the inverted and true velocity
changes at x = 7.0 km for the SEAM example. The blue line,
Figure 14. The virtual shot gathers for baseline at the datum level the standard target-oriented DDWI result; the red line, the final in-
and (b) the time-lapse data difference for the virtual survey for the version result using DL-assisted regularization; and the dashed
SEAM example. black line, the true velocity changes.

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 R494 Li et al.

was not used in the DNN training, also demonstrate that the inver- cables: A synthetic study: Geophysical Journal International, 201, 1215–
1234, doi: 10.1093/gji/ggv048.
sion quality is enhanced using the proposed inversion algorithm. Chadwick, A., G. Williams, N. Delepine, V. Clochard, K. Labat, S. Sturton, M.
L. Buddensiek, M. Dillen, M. Nickel, A. L. Lima, and R. Arts, 2010, Quan-
titative analysis of time-lapse seismic monitoring data at the Sleipner CO2
CONCLUSION storage operation: The Leading Edge, 29, 170–177, doi: 10.1190/1.3304820.
Chawla, N. V., K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, 2002,
We develop a regularized TLFWI scheme focused on the desired SMOTE: Synthetic minority over-sampling technique: Journal of Artifi-
cial Intelligence Research, 16, 321–357, doi: 10.1613/jair.953.
target zone. The required target-oriented data are extracted from sur- da Costa, C. A., J. C. Costa, W. E. Medeiros, D. J. Verschuur, and A. K.
face seismic data by waveform redatuming. In this case, the impact Soni, 2019, Target-level waveform inversion: A prospective application
of the overburden complexity on the target inversion quality is mi- of the convolution-type representation for the acoustic wavefield: Geo-
physical Prospecting, 67, 69–84, doi: 10.1111/1365-2478.12703.
tigated and the computational cost of the target area inversion is Denli, H., and L. Huang, 2009, Double-difference elastic waveform tomog-
reduced. Considering that well logs provide detailed information raphy in the time domain: 89th Annual International Meeting, SEG, Ex-
for model changes with limited coverage, we incorporate the well panded Abstracts, 2302–2306, doi: 10.1190/1.3255320.
Esser, E., L. Guasch, T. van Leeuwen, A. Y. Aravkin, and F. J. Herrmann,
velocity information into the inversion by using an objective func- 2018, Total variation regularization strategies in full-waveform inversion:
tion that includes a model-space regularization term. The required SIAM Journal on Imaging Sciences, 11, 376–406, doi: 10.1137/
17M111328X.
prior model is predicted from a trained DNN, which identifies the Guo, Q., and T. Alkhalifah, 2019, Datum-based waveform inversion using a
statistical connection between the initial inverted model and facies subsurface-scattering imaging condition: Geophysics, 84, no. 4, S251–
interpreted from well logs. The numerical examples show that the S266, doi: 10.1190/geo2018-0615.1.
Guo, Q., and T. Alkhalifah, 2020, Target-oriented waveform redatuming and
proposed method improves the inversion resolution and accuracy high-resolution inversion: Role of the overburden: Geophysics, 85, no. 6,
with better consistency with the well logs. R525–R536, doi: 10.1190/geo2019-0640.1.
Hicks, E., H. Hoeber, M. Houbiers, S. P. Lescoffit, A. Ratcliffe, and V. Vinje,
2016, Time-lapse full-waveform inversion as a reservoir-monitoring tool
ACKNOWLEDGMENTS — A North Sea case study: The Leading Edge, 35, 850–858, doi: 10
.1190/tle35100850.1.
Kalita, M., V. Kazei, Y. Choi, and T. Alkhalifah, 2019, Regularized full-
We thank KAUST for its support. We also appreciate the seismic waveform inversion with automated salt flooding: Geophysics, 84,
wave analysis group members for the fruitful discussions and help- no. 4, R569–R582, doi: 10.1190/geo2018-0146.1.
ful suggestions. The Shaheen Supercomputing Laboratory in Kazei, V., and T. Alkhalifah, 2018, Time-lapse waveform inversion regular-
ized by spectral constraints and Sobolev space norm: 88th Annual
KAUST provides the computational support. International Meeting, SEG, Expanded Abstracts, 5487–5491, doi: 10
.1190/segam2018-w12-04.1.
Kemper, M., and J. Gunning, 2014, Joint impedance and facies inversion —
DATA AND MATERIALS AVAILABILITY Seismic inversion redefined: First Break, 32, 89–95, doi: 10.3997/1365-
2397.32.9.77968.
Data associated with this research are available and can be Kingma, D. P., and J. Ba, 2014, Adam: A method for stochastic optimiza-
tion: arXiv preprint arXiv:1412.6980.
obtained by contacting the corresponding author. LeCun, Y., Y. Bengio, and G. Hinton, 2015, Deep learning: Nature, 521,
436–444, doi: 10.1038/nature14539.
Li, Y., T. Alkhalifah, and Q. Guo, 2020c, Target-oriented time-lapse wave-
REFERENCES form inversion using a deep learning-assisted regularization: 90th Annual
International Meeting, SEG, Expanded Abstracts, 3744–3748, doi: 10
Alford, R. M., K. R. Kelly, and D. M. Boore, 1974, Accuracy of finite-differ- .1190/segam2020-3425737.1.
ence modeling of the acoustic wave equation: Geophysics, 39, 834–842, Li, Y., T. Alkhalifah, and Z. Zhang, 2020a, High-resolution regularized elas-
doi: 10.1190/1.1440470. tic full waveform inversion assisted by deep learning: 82nd Annual
Alkhalifah, T., 2016, Full-model wavenumber inversion: An emphasis on International Conference and Exhibition, EAGE, Extended Abstracts,
the appropriate wavenumber continuation: Geophysics, 81, no. 3, doi: 10.3997/2214-4609.202010281.
R89–R98, doi: 10.1190/geo2015-0537.1. Li, Y., Y. Choi, T. Alkhalifah, Z. Li, and K. Zhang, 2018, Full-waveform
Alkhalifah, T., and Z. Wu, 2016, The natural combination of full and image- inversion using a nonlinearly smoothed wavefield: Geophysics, 83,
based waveform inversion: Geophysical Prospecting, 64, 19–30, doi: 10 no. 2, R117–R127, doi: 10.1190/geo2017-0312.1.
.1111/1365-2478.12264. Li, Y., Q. Guo, T. Alkhalifah, and V. Kazei, 2020b, Target-oriented time-
Aragao, O., and P. Sava, 2020, Elastic full waveform inversion with prob- lapse waveform inversion using redatumed data: Feasibility and robust-
abilistic petrophysical clustering: Geophysical Prospecting, 68, 1341– ness: 90th Annual International Meeting, SEG, Expanded Abstracts,
1355, doi: 10.1111/1365-2478.12910. 3783–3787, doi: 10.1190/segam2020-3420515.1.
Arts, R., O. Eiken, A. Chadwick, P. Zweigel, L. Van der Meer, and B. Zinsz- Li, Y., Q. Guo, Z. Li, and T. Alkhalifah, 2019, Elastic reflection waveform
ner, 2003, Monitoring of CO2 injected at Sleipner using time lapse seismic inversion with variable density: Geophysics, 84, no. 4, R553–R567, doi:
data: 6th International Conference, Greenhouse Gas Control Technolo- 10.1190/geo2017-0722.1.
gies, 347–352. Lin, Y., and L. Huang, 2015, Quantifying subsurface geophysical properties
Asnaashari, A., R. Brossier, S. Garambois, F. Audebert, P. Thore, and J. changes using double-difference seismic-waveform inversion with a
Virieux, 2013, Regularized seismic full waveform inversion with prior modified total-variation regularization scheme: Geophysical Supplements
model information: Geophysics, 78, no. 2, R25–R36, doi: 10.1190/ to the Monthly Notices of the Royal Astronomical Society, 203, 2125–
geo2012-0104.1. 2149, doi: 10.1093/gji/ggv429.
Asnaashari, A., R. Brossier, S. Garambois, F. Audebert, P. Thore, and J. Lumley, D. E., 2001, Time-lapse seismic reservoir monitoring: Geophysics,
Virieux, 2014, Time-lapse seismic imaging using regularized full-wave- 66, 50–53, doi: 10.1190/1.1444921.
form inversion with a prior model: Which strategy?: Geophysical Pro- Maharramov, M., B. L. Biondi, and M. A. Meadows, 2016, Time-lapse in-
specting, 63, 78–98, doi: 10.1111/1365-2478.12176. verse theory with applications: Geophysics, 81, no. 6, R485–R501, doi:
Ayeni, G., and B. Biondi, 2010, Target-oriented joint least-squares migra- 10.1190/geo2016-0131.1.
tion/inversion of time-lapse seismic data sets: Geophysics, 75, no. 3, R61– Nair, V., and G. E. Hinton, 2010, Rectified linear units improve restricted
R73, doi: 10.1190/1.3427635. Boltzmann machines: Proceedings of the 27th International Conference
Bakulin, A., and R. Calvert, 2006, The virtual source method: Theory and on Machine Learning, 807–814.
case study: Geophysics, 71, no. 4, SI139–SI150, doi: 10.1190/1.2216190. Routh, P., R. Neelamani, R. Lu, S. Lazaratos, H. Braaksma, S. Hughes, R.
Biondi, E., B. Biondi, and G. Barnier, 2018, Target-oriented elastic full- Saltzer, J. Stewart, K. Naidu, H. Averill, and V. Gottumukkula, 2017, Im-
waveform inversion through extended-migration redatuming: 88th An- pact of high-resolution FWI in the Western Black Sea: Revealing over-
nual International Meeting, SEG, Expanded Abstracts, 1228–1232, burden and reservoir complexity: The Leading Edge, 36, 60–66, doi: 10
doi: 10.1190/segam2018-2998407.1. .1190/tle36010060.1.
Borisov, D., and S. C. Singh, 2015, Three-dimensional elastic full waveform Routh, P., G. Palacharla, I. Chikichev, and S. Lazaratos, 2012, Full wavefield
inversion in a marine environment using multicomponent ocean-bottom inversion of time-lapse data for improved imaging and reservoir charac-

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user
 Target-oriented TLFWI using DL R495

terization: 82nd Annual International Meeting, SEG, Expanded Abstracts, Yang, D., M. Meadows, P. Inderwiesen, J. Landa, A. Malcolm, and M. Feh-
doi: 10.1190/segam2012-1043.1. ler, 2015, Double-difference waveform inversion: Feasibility and robust-
Singh, S., I. Tsvankin, and E. Z. Naeini, 2018, Bayesian framework for elas- ness study with pressure data: Geophysics, 80, no. 6, M129–M141, doi:
tic full-waveform inversion with facies information: The Leading Edge, 10.1190/geo2014-0489.1.
37, 924–931, doi: 10.1190/tle37120924.1. Yang, D., Y. Zheng, M. Fehler, and A. Malcolm, 2012, Target-oriented time-
Sirgue, L., and R. G. Pratt, 2004, Efficient waveform inversion and imaging: lapse waveform inversion using virtual survey: 82nd Annual International
A strategy for selecting temporal frequencies: Geophysics, 69, 231–248, Meeting, SEG, Expanded Abstracts, doi: 10.1190/segam2012-1308.1.
doi: 10.1190/1.1649391. Yuan, S., N. Fuji, S. Singh, and D. Borisov, 2017, Localized time-lapse elas-
Song, C., T. Alkhalifah, and U. B. Waheed, 2021, Solving the frequency-do- tic waveform inversion using wavefield injection and extrapolation: 2-D
main acoustic VTI wave equation using physics-informed neural networks: parametric studies: Geophysical Journal International, 209, 1699–1717,
Geophysical Journal International, 225, 846–859, doi: 10.1093/gji/ggab010. doi: 10.1093/gji/ggx118.
Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdi- Zhang, Z., and T. Alkhalifah, 2019, Regularized elastic full waveform in-
nov, 2014, Dropout: A simple way to prevent neural networks from over- version using deep learning: Geophysics, 84, no. 5, R741–R751, doi:
fitting: The Journal of Machine Learning Research, 15, 1929–1958. 10.1190/geo2018-0685.1.
Symes, W. W., 2008, Migration velocity analysis and waveform inversion: Geo- Zhang, Z., and L. Huang, 2013, Double-difference elastic-waveform inver-
physical Prospecting, 56, 765–790, doi: 10.1111/j.1365-2478.2008.00698.x. sion with prior information for time-lapse monitoring: Geophysics, 78,
Tarantola, A., 1984, Inversion of seismic reflection data in the acoustic no. 6, R259–R273, doi: 10.1190/geo2012-0527.1.
approximation: Geophysics, 49, 1259–1266, doi: 10.1190/1.1441754. Zhang, Z. D., T. Alkhalifah, E. Z. Naeini, and B. Sun, 2018, Multiparameter
Urosevic, M., R. Pevzner, A. Kepic, P. Wisman, V. Shulakova, and S. elastic full waveform inversion with facies-based constraints: Geophysical
Sharma, 2010, Time-lapse seismic monitoring of CO2 injection into a Journal International, 213, 2112–2127, doi: 10.1093/gji/ggy113.
depleted gas reservoir — Naylor Field, Australia: The Leading Edge, Zhao, Y., F. Niu, H. Liu, X. Jia, J. Yang, and S. Huo, 2020, Source-receiver
29, 164–169, doi: 10.1190/1.3304819. interferometric redatuming using sparse buried receivers to address com-
van Manen, D.-J., J. O. A. Robertsson, and A. Curtis, 2007, Exact wave field plex near-surface environments: A case study of seismic imaging quality
simulation for finite-volume scattering problems: The Journal of the Acous- and time-lapse repeatability: Journal of Geophysical Research, Solid
tical Society of America, 122, EL115–EL121, doi: 10.1121/1.2771371. Earth, 125, e2020JB019496, doi: 10.1029/2020JB019496.
Vasconcelos, I., P. Sava, and H. Douma, 2010, Nonlinear extended images Zheng, Y., P. Barton, and S. Singh, 2011, Strategies for elastic full waveform
via image-domain interferometry: Geophysics, 75, no. 6, SA105–SA115, inversion of time-lapse ocean bottom cable (OBC) seismic data: 81st An-
doi: 10.1190/1.3494083. nual International Meeting, SEG, Expanded Abstracts, 4195–4200, doi:
Wapenaar, K., and J. Fokkema, 2006, Green’s function representations for 10.1190/1.3628083.
seismic interferometry: Geophysics, 71, no. 4, SI33–SI46, doi: 10.1190/1
.2213955.
Wapenaar, K., and J. van der Neut, 2010, A representation for Green’s func-
tion retrieval by multidimensional deconvolution: Journal Acoustical
Society of America, 128, EL366–EL371, doi: 10.1121/1.3509797. Biographies and photographs of the authors are not available.

Downloaded from http://pubs.geoscienceworld.org/geophysics/article-pdf/86/4/R485/5383318/geo-2020-0383.1.pdf

by CNOOC user