IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.

60, 2022 4511311

Deep Learning Seismic Inversion Based on Prestack

Waveform Datasets
Jian Zhang , Hui Sun , Gan Zhang, and Xiaoyan Zhao

Abstract— Prediction of elastic parameters (e.g., P-, S-wave as well-logs data, and indirect measurements of the Earth’s
velocity, and density) from observed seismic data is one of the response at the surface referred to as seismic data. Estimating
most common means of reservoir characterization. Recently, deep elastic parameters from observed seismic data and well-logs
learning (DL), as a data-driven approach, has been attracting
increasing interest in seismic inversion. DL is proven to have the data is an inverse problem from the mathematical point of view
potential to learn complex systems and solve inverse problems [5]. For conventional seismic inversion (i.e., model-driven
efficiently. One of the most key components of DL is the training inversion), the physical system, or forward model, i.e., the
dataset, and an effective training dataset is a prerequisite for the relationship between the physical parameters and the observed
success of DL-based methods. In seismic inversion, the training data must be established in advance. The elastic parameters are
dataset needs to be artificially expanded due to the limited
number of actual training data pairs. Traditional approaches of then predicted from seismic and well-logs data by combining
using the exact Zoeppritz equation (EZE) or its approximations inverse theory with known physical system. However, in many
for training dataset construction have limitations, principally, practical cases, the physical system is difficult or impossible
the single interface assumption and the neglect of wave prop- to be built precisely. Moreover, when the forward operator
agation effects. Alternatively, the analytical solution of the 1-D is highly nonlinear, the inverse problem is difficult to solve.
wave equation (i.e., reflectivity method [RM]) can simulate the
full wave, including transmission losses and internal multiples, There are also problems such as limited data bandwidth,
and can be executed in a target-oriented manner. Inspired data noise, and incomplete data coverage, all of which cause
by this, we develop a data-driven elastic parameter prediction various troubles for traditional seismic inversion [6], [7].
method based on waveform formulation. The method uses RM Different from traditional seismic inversion based on model-
to construct training dataset, which both compensates for the driven, deep learning (DL) is a type of data-driven method
inadequate training dataset in data-driven seismic inversion and
improves the accuracy of the inversion results. We implement that can learn a complex nonlinear mapping between inputs
the method in a synthetic model as well as field data. The results and outputs with adjustable parameters (e.g., the weight and
are compared with model-driven methods (EZE and RM) and bias) based on training dataset. DL is a subset of machine
data-driven method based on EZE, and it is shown that the learning, which is essentially a neural network with three or
proposed method outperforms these three methods. more layers [8]. DL method drives many artificial intelligence
Index Terms— Deep learning (DL), reflectivity method (RM), applications and services (e.g., image classification, computer
reservoir characterization, waveform inversion. vision, and natural language processing) that improve automa-
tion, and performing analytical and physical tasks without
I. I NTRODUCTION human intervention [9]–[11]. In recent years, the geophysical
community has also shown great interest in DL, while the DL
E LASTIC parameters (e.g., impedance, P-, S-wave veloc-
ity, and density) are critical for reservoir characterization
because they reveal high-resolution rock properties which help
method has provided several astonishing results in geophysical
applications. For instance, for fault recognition and horizon
indicate lithology, porosity, hydrocarbon, etc. [1]–[4]. Typi- interpretation, DL-based methods are no less accurate than
cally, the available data include direct measurements of elastic manual, while being far more efficient [12], [13]. DL is also
parameters of interest from sparse well locations, referred to widely used in some geophysical problems such as noise
attenuation [14], [15], lithology prediction [16], [17], and
Manuscript received 22 November 2021; revised 24 March 2022, 14 May seismic full waveform inversion (FWI) [18], [19].
2022, and 21 June 2022; accepted 30 July 2022. Date of publication 1 August
2022; date of current version 17 August 2022. This work was supported in Reservoir characterization based on DL is also a research,
part by the pre-research project on Civil Aerospace Technologies from the both for poststack seismic data and for prestack seismic data.
China National Space Administration (CNSA) under Grant D020201, in part Das et al. [20] estimate impedance from normal-incidence
by the Fundamental Research Funds for the Central Universities under Grant
2682022CX030, and in part by the National Natural Science Foundation of seismic data using training data constructed based on poststack
China under Grant 41672295. (Corresponding author: Jian Zhang.) formulation, and a two-layer convolutional neural network
Jian Zhang is with the Faculty of Geosciences and Environmental Engineer- (CNN). Kazei et al. [21] apply CNN to transform rele-
ing, Southwest Jiaotong University, Chengdu 611756, China, and also with
the Ministry of Education (MOE) Key Laboratory of High-speed Railway vant seismic data cubes (i.e., multiple common mid point
Engineering, Southwest Jiaotong University, Chengdu 610036, China (e-mail: (CMP) gathers) to respective velocity logs (i.e., target loca-
[email protected]). tion or central midpoint log). Mustafa et al. [22] perform
Hui Sun and Xiaoyan Zhao are with the Faculty of Geosciences and
Environmental Engineering, Southwest Jiaotong University, Chengdu 611756, acoustic impedance inversion based on temporal convolu-
China (e-mail: [email protected]; [email protected]). tional network. Puzyrev et al. [23] test different networks
Gan Zhang is with Sichuan Water Resources and Hydroelectric Investi- including CNN, recurrent neural network (RNN), and fully
gation & Design Institute Company Ltd., Chengdu 610072, China (e-mail:
[email protected]). connected network (FCN) for velocity estimation from seismic
Digital Object Identifier 10.1109/TGRS.2022.3195858 data and demonstrate the potential of DL seismic inversion.
1558-0644 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
4511311 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Wu et al. [24] use the poststack impedance formulation strategy [41] to implement the three-parameter (i.e., P-, S-wave
to construct training dataset and implement impedance pre- velocity, and density) prediction in real data.
diction in two synthetic data based on fully convolutional
residual network and transfer learning strategy. Zhang et al. II. M ETHOD
[25] construct the training dataset using the exact Zoep-
pritz equation (EZE) in the DL-based prestack inversion, The key factors for the success of DL are the training
while introducing the initial model to obtain stable inversion dataset, network architecture, and parameter optimization.
results. It is difficult and costly to obtain sufficient training datasets
The seismic inversion of DL-based algorithms usually directly in the actual seismic inversion. In this section, we first
requires sufficient training datasets in order to obtain robust introduce two physically driven methods (i.e., EZE and RM)
and reliable results. For the field seismic data, however, the to augment training datasets. Then, the network architecture
labeled data-model pairs (i.e., well-logs data) are limited. used in this article is introduced. Finally, the transfer learning
Many strategies such as data augmentation [20], [24]–[26], strategy about how the proposed method can be better applied
semi-supervised learning [27]–[29], and unsupervised learning from synthetic data to real data is reported.
[30], [31] are proposed to alleviate the requirement of large In general, the relationship between elastic parameters m
amount of training data for DL algorithms. The construction and seismic data d can be expressed as
of training datasets using EZE or its approximations to achieve d = G(m) + ε (1)
data augmentation for improved prediction results is currently
the most common means in data-driven prestack seismic where ε represent the noise component of the observed data.
inversion. However, both approximate and exact formulations G is the forward modeling operator. The goal of seismic
are derived under certain assumptions (e.g., the single interface inversion is to obtain unknown elastic parameters m from the
assumption and the neglect of wave propagation effects) and recorded seismic data d, given the forward modeling operator
have many restrictions [32]–[34]. This requires that the input G. However, G is usually unknown and its exact form is
prestack seismic data has undergone specific processing, e.g., difficult or impossible to obtain. We are able to obtain an
the processed data contains only the primary reflections, approximation (G̃) of G by physically driven methods (e.g.,
otherwise the training data constructed based on EZE or EZE and RM).
its approximations will be unreliable. However, it is almost
impossible to meet such requirements with current processing
A. Exact Zoeppritz Equation
technology.
Wave equation considering full wave-field information is EZE defines the relationship between the elastic parame-
a good choice to construct the training dataset and ensure ters (i.e., V p , Vs , and density) and the reflection coefficient
better similarity between the simulated and observed data. under the single interface assumption, which can avoid the
There are two types of waveform formulas, one is the full error caused by its approximate equation. The expression
waveform formula and the other is the prestack waveform of EZE is [34] (2), as shown at the bottom of the next
formula. The full waveform formula is less commonly used in page, where R and T represent reflection coefficients and
reservoir characterization due to low computational efficiency, transmission coefficients, respectively. · pp represents P-wave
high memory requirements, and the inability to target reser- incident, P-wave reflected (or transmitted), and · ps repre-
voirs [35], [36]. Alternatively, the prestack waveform formula sents P-wave incident, S-wave reflected (or transmitted). θ1
or the analytical solution of the 1-D wave equation (i.e., and θ2 denote the incidence and transmission angles of the
reflectivity method [RM]) is another method to simulate the P-wave, respectively. ϕ1 and ϕ2 denote the incidence and
full wave, including transmission losses and internal multiples, transmission angles of the S-wave, respectively. v P1 , v P2 ,
and can be executed in a target-oriented manner [37]–[40]. v s1 , v s2 , ρ1 , and ρ2 denote the P-wave velocity, S-wave
Thus, we develop a data-driven elastic parameter prediction velocity, and density of the upper and lower layers of an
method based on prestack waveform formulation (i.e., RM). interface.
We use RM to construct training dataset, which both compen-
sates for the inadequate training dataset in data-driven seismic B. Reflectivity Method
inversion and to ensure a good match between the simulated
and actual data. In this way, we improve the accuracy of the RM has the ability to simulate the full wavefield including
prediction results. transmission losses, internal multiples, and other effects, thus
In this article, we use the example of a hybrid (CNN + making the simulation results closer to the real observation.
FCN) network to validate the proposed method for the problem Under the local 1-D or horizontal layer assumption, based
of prestack seismic inversion. We first implement the proposed on the RM, reflection coefficients in the frequency slowness
method in a synthetic model to verify its effectiveness. The domain can be derived using the following equation [35]:
predicted P-, S-wave velocity, and density are compared with  T
ν = , −Rsp , −Rss , R pp , R ps , |R| (3)
the results of model-driven methods (EZE and RM) and the
results of data-driven method based on EZE, and it is shown where  is a coefficient factor. R represents reflection coef-
that the proposed method outperforms these three methods. ficients. · pp and · ps represent P-wave incident, P-wave, and
Then, the proposed method is combined with transfer learning S-wave reflected, respectively. ·sp and ·ss represent S-wave

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: DL SEISMIC INVERSION BASED ON PRESTACK WAVEFORM DATASETS 4511311

Fig. 1. Unequal thickness multilayer model (P-, S-wave velocity, and

density). An extreme example to highlight the impact of transmission losses Fig. 2. Angle-gather comparisons between EZE and RM. (a) Synthetic
and internal multiples. angle-gathers using EZE. (b) Synthetic angle-gathers using RM. (c) Difference
between RM and EZE. The sky-blue box highlights internal multiples and the
gray arrows indicate transmission losses.

incident, P-wave, and S-wave reflected, respectively. |R| rep-

resents the determinant of the reflection coefficient. For a
particular layer n, the total response below that layer can be
expressed as
ν n = Qn ν n+1 , Qn = T+ −
n En Tn (4)
where Qn represents the wave propagator matrix. En is the
phase shift matrix. T+ −
n and Tn are the downward and upward
interface energy evolution matrices, respectively. Refer to the
Appendix for the specific representations of the matrices En , Fig. 3. Comparison of 12◦ gather from Fig. 2 red and black arrows.
The sky-blue box highlights internal multiples and the gray arrows indicate
T+ −
n , and Tn . The propagation from the bottom interface to the transmission losses.
top interface is obtained using the following equation:
ν 0 = Q0 Q1 , . . . , Q N −1 ν N (5)
and the difference profiles between them are shown in Fig. 2,
where ν 0 denotes the final response obtained at the top where the angle range is 0–27◦ with a gap of 3◦ .
interface, and ν N = [1, 0, 0, 0, 0, 0]T . The synthetic seismograms of RM and EZE are almost
In general, only PP wave is considered. Finally, the total indistinguishable at the first interface. However, due to trans-
reflectivity coefficient of PP in the frequency-angle domain mission loss, the amplitude of RM at the interface is smaller
can be given by ν 0 than that of EZE from the second interface downward. More-
over, there are obvious internal multiples between the inter-
ν 0 (4)
R pp (ω, θ ) = (6) faces. These internal multiples are difficult to remove, and
ν 0 (1) their presence can mislead the interpretation and inversion.
where ν 0 (1) and ν 0 (4) denote the first and fourth terms of ν 0 , Fig. 3 shows a comparison of the angle gather of 12◦ for
respectively. The specific representation of ν 0 can be found RM and EZE. In Fig. 3, the internal multiples of RM around
in (3). 0.4, 0.7, and 1 s (blue dashed box) and their differences
A multilayer model including P-, S-wave velocity, and den- between RM and EZE in amplitude at the interfaces (gray
sity is used to verify the advantages of RM over EZE, as shown arrows) are obvious. Nowadays, transmission loss and inter-
in Fig. 1. The reflection coefficients R pp are first calculated nal multiples are difficult to be accurately corrected during
based on RM and EZE, and subsequently convolve with a seismic data preprocessing due to the limitation of processing
20 Hz zero-phase Ricker wavelet to obtain the seismograms. technology [40]. Thus, the synthetic seismograms using RM
The angle gathers generated by the RM and EZE methods can be better matched with the actual data. That is, RM-based

⎡ ⎤−1
− sin θ1 − cos ϕ1 sin θ2 − cos ϕ2
⎡ ⎤ ⎢
⎢ cos θ
⎥
⎥ ⎡ ⎤
R pp ⎢ 1 − sin ϕ1 cos θ2 sin ϕ2 ⎥ sin θ1
⎢ R ps ⎥ ⎢ ⎥ ⎢ cos θ1 ⎥
⎢ ⎥=⎢ ⎥ ⎢ ⎥
⎣ T pp ⎦ ⎢ v p1 ρ2 v s2 v p1 ρ2 v s2 v p1 ⎥
2 (2)
⎢ sin 2θ1 cos 2ϕ1 sin 2θ2 − cos 2ϕ ⎥ ⎣ sin 2θ1 ⎦
⎢ v s1 ρ1 v s1 v p2
2
ρ1 v s1
2 2 ⎥
T ps ⎢ ⎥ cos 2ϕ1
⎣ v s1 ρ2 v p2 ρ2 v s2 ⎦
− cos 2ϕ1 sin 2ϕ1 cos 2ϕ2 sin 2ϕ2
v p1 ρ1 v p1 ρ1 v p1

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
4511311 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 4. Workflow of the DL-based prestack seismic inversion. The true elastic parameters including P-wave, S-wave, and density serve as the output targets
(i.e., labels), and both prestack seismic data and the initial model of elastic parameters as inputs. The inputs and corresponding outputs are split into patches
and put into the network for training. The network is then applied to the filed data using transfer learning strategy.

method can provide more reliable training datasets for data- complete trace as input, but divide the data into small patches.
driven methods, which, on the one hand, can alleviate the Since the favorable reservoir thickness is usually thin (50–
requirements of data-driven methods for large amounts of 500 ms), these small patches are set to a size of 200 sample
training data, and on the other hand, can provide accurate points. Then, the input sample consists of a tensor of size
training data as much as possible to improve prediction results. 200 × 6, including the angle-gathers (i.e., small, middle, and
far) and the elastic initial model, and the corresponding output
samples are the actual elastic parameters of size 200 × 3.
C. DL-Based Seismic Inversion The intrinsic relationship between the three parameters (i.e.,
Fig. 4 shows the complete framework of the DL-based P-, S-wave velocity, and density) is obtained by the network
seismic inversion. This framework consists of two steps, one based on the training dataset.
is the testing and application of synthetic data, and the other is The seismic inversion of DL-based algorithms usually
the application of field data. In this article, the outputs of the requires sufficient training datasets in order to obtain robust
network are the true elastic parameters including P-, S-wave and reliable results. For the field seismic data, however, the
velocity, and density, while the inputs contain both the prestack labeled data-model pairs (i.e., well-logs data) are limited.
angle gathers and the initial model of the elastic parameters. Thus, we first use the model data [see Fig. 5(a)–(c)] to con-
We first pretrain the network using synthetic data and verify struct training data to meet the data volume requirements of the
the effectiveness of the pretrained network, and then, fine- DL-based algorithm. The model data consists of 746 time sam-
tune the pretrained network using transfer learning strategy ples and 3201 traces with a time sampling rate of 2 ms. The
and apply the fine-tuned network to real data. corresponding initial model is shown in Fig. 5(d)–(f). In syn-
Considering the computational cost and data cost, in this thetic tests, the initial model is obtained from the true model
example, we do not directly take the whole data cube or a using low-pass filtering techniques, while in practical exam-

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: DL SEISMIC INVERSION BASED ON PRESTACK WAVEFORM DATASETS 4511311

ples, the initial model is obtained using commercial software

based on known well-logs data and picked horizon. We then
randomly select a certain number of traces from Fig. 5(a)–(c)
to construct the training and validation datasets. The seismic
records (angle-gathers of 5◦ , 15◦ , and 25◦ ) corresponding to
the selected traces are calculated based on physical forward
model (e.g., RM). The synthetic seismic data traces have
different signal-to-noise ratios (SNRs) to improve network per-
formance. The SNR is defined as 10 log((||d||2 )/(||d − d||˜ 2 )),
where d and d̃ represent noise-free and noisy data, respec-
tively. Finally, these traces include randomly selected true
models, the initial models corresponding to the selected traces,
and the seismic records corresponding to the selected traces
are used to generate training data pairs. These training datasets
are fed into the neural network, which is iteratively updated
to achieve a minimal error between the network outputs
and the labels, resulting in a pretrained network. During the
training process, 16.7% of the training datasets are used as the
Fig. 5. Overthrust models of P-wave velocity, S-wave velocity, and density.
validation datasets. Using the pretrained network, the elastic (Left) The true model. (Right) The initial model.
parameters can be predicted given the inputs.
a stable and reliable network. The closer the simulated data is
D. Hybrid Network (CNN + FCN) Architecture to the actual data, the more reliable the pretrained network will
In this article, a hybrid neural network framework is used as be. The forward operator (i.e., G̃) constructed by the model-
an example for seismic inversion, which consists of convolu- driven approach is an approximation of the real physical model
tion and fully connected layers, as shown in Fig. 4. It contains (i.e., G) between elastic parameters (i.e., P-wave velocity,
two convolution layers (32, 3 × 1, 1) → (32, 3 × 1, 1), S-wave velocity, and density) and the observed seismic data.
followed by an ReLU activation layer, a Maxpooling layer, Due to its similarity (G̃ ≈ G), synthetic data can be used
and a dropout layer after each convolutional layer. Here, the to pretrain our network and reduce the difficulty of network
Maxpooling layer is optional, depending on the computational training in practical applications. Therefore, if well-logs data
efficiency and accuracy required. The outputs of the second are available, the transfer learning strategy [41], [43] is a
convolution layer is flattened to 1-D vectors before being fed good choice to fine-tune the pretrained network to make it
into the first fully connected layer. The first FCN layer has more suitable for the actual data. Transfer learning is not a
256 elements and is immediately followed by an ReLu acti- distinct type of machine learning algorithm, instead, it is a
vation layer and a dropout layer. The outputs of the last fully technique or method used whilst training models. The process
connected layer corresponds to the elastic parameters to be of transfer learning from synthetic data to real data is shown
predicted. During the training process, we adopt the adaptive in Fig. 4. Transfer learning means taking the relevant parts of
moment estimation (Adam) method [42] as the optimizer to a pretrained neural network model and applying it to a new
minimize the mean square error between the predicted outputs but similar problem (i.e., Step 2: the actual seismic inversion).
and the labels to obtain optimal parameters (i.e., the weight During transfer learning, in this article, all network parameters
and bias values). The batch size and epochs are 128 and 500, that have been pretrained are fixed except for the last fully
respectively. We obtain the relatively optimal parameters by connected layer. In general, transfer learning is also susceptible
trial-and-error method. Although the network parameters we to the type of pretraining samples, the number of training
use produce good inversion results, they may not be optimal. samples and the choice of fine-tuning strategies, more details
Combining DL with an optimization algorithm may be a of which can be found in [44] and [45].
solution to adaptively set the number of hidden layers and III. A PPLICATION
the number of neurons in each layer. Besides, the proposed
We carry out tests on synthetic data from the SEG/EAGE
method can be easily extended to other networks by simply
overthrust model and a real dataset to verify the effective-
replacing the hybrid network with a fully CNN, an RNN,
ness of the proposed method. The numerical experiments are
a fully convolutional residual network, a generative adversarial
performed using Keras with TensorFlow as the backend on a
network, etc.
PC with Windows 10, AMD Ryzen 7 5800X, and NVIDIA
GeForce RTX 3080Ti. The results are compared with model-
E. Transfer Learning Strategy driven methods and data-driven method based on EZE. For the
In practical applications, it is difficult to train a robust and convenience of subsequent comparisons, the inversion method
reliable network directly using these small amounts of datasets using the EZE as the forward operator is denoted as EZE and
due to the limited labeling data (i.e., well-logs data). We use the RM-based method is denoted as RM in the model-driven
the advantages of RM to construct (or augment) the training inversion method, while the EZE-based method is denoted as
dataset in a synthetic model and pretrain the network to obtain DL-EZE and the RM-based method is denoted as DL-RM in

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
4511311 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 6. Effects of the amount of training data on different methods. The NRMSE between the inverted (a) V p , (b) Vs , (c) Rho, and the true values.
(d) Computational time versus the amount of training data.

Fig. 7. Angle gathers from the overthrust model with SNR = 5 of (a) 5◦ , (b) 15◦ , and (c) 25◦ .

the data-driven inversion method. In addition, to quantitatively

evaluate the inversion results, we introduce the normalized root
mean square error (NRMSE) as follows:

1
N
 2
NRMSE = m − m̃
i i (m max − m min ) (7)
N i=1

where m and m̃ represent true (or reference) values and

inverted results, respectively. Mmax and m min represent the
maximum and minimum values in data m, respectively. N
represents the number of samples.

A. Synthetic Data Example

As a benchmark model, the SEG/EAGE overthrust model
[see Fig. 5(a)–(c)] is used as an inversion example. Based
on the training data construction strategy reported in the
theoretical section, we first compare the NRMSE of the two
methods for the whole profile prediction results in the case Fig. 8. Inverted (Top) V p , (Middle) Vs , and (Bottom) Rho obtained
of different training data amounts for DL-EZE and DL-RM. using (Left) EZE and (Right) RM, respectively.
Fig. 6(a)–(c) shows the variation of NRMSE of P-, S-wave
velocity, and density predicted by DL-EZE and DL-RM
with the amount of training data. In this experiment, the data increases the time cost, as shown in Fig. 6(d). Therefore,
angle gathers are noise-free during the prediction process. in the subsequent experiments of synthesis, we set the number
As expected, both DL-EZE and DL-RM gradually improve of traces for constructing the training datasets to 100. The
the prediction performance as the amount of training data training datasets are then used to pretrain the network for
increases. However, the DL-RM method consistently predicts model testing, thus providing the basis network for transfer
better results than the DL-EZE method. It can be seen that the learning in real data.
NRMSE decreases rapidly with the increase in the amount of Furthermore, we perform experiments on the noisy angle
training datasets at the beginning of the gradual increase in the gathers (SNR = 5), as shown in Fig. 7. Three angle gathers
amount of training data (<100 traces), and then the NRMSE of 5◦ , 15◦ , and 25◦ are used to invert elastic parameters
decreases slowly with the further increase in the amount of (i.e., V p , Vs , and rho). We compare the inversion results (see
training datasets. However, increasing the amount of training Figs. 8 and 9) of four methods (i.e., EZE, RM, DL-EZE,

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: DL SEISMIC INVERSION BASED ON PRESTACK WAVEFORM DATASETS 4511311

Fig. 11. Comparison of elastic parameter curves (P-, S-wave velocity, and
density) at CDP 1600.

Fig. 9. Training and validation convergence curves (loss and accuracy)

corresponding to different methods for model data. (a) and (b): DL-EZE,
(c) and (d): DL-RM. The left graphs use a log-10 scale for the Y -axis.

Fig. 12. Real case test: 3-D cube data of the partial stacked angle-gathers.
(a) Small-angle gathers with angle ranges of 0◦ –10◦ , (b) middle-angle gathers
with angle ranges of 10◦ –20◦ , and (c) large-angle gathers with angle ranges of
20◦ –30◦ . The white and red markers show the well locations. The red marked
well is used as a blind well to verify the reliability of the inversion results.

Fig. 10. Inverted (Top) V p , (Middle) Vs , and (Bottom) Rho obtained

using (Left) DL-EZE and (Right) DL-RM, respectively. Fig. 13. Initial model of (a) P-wave velocity, (b) S-wave velocity, and
(c) density.

DL-RM) separately. Fig. 8(a)–(c) and (d)–(f) shows the inver-

TABLE I
sion results based on EZE and RM, respectively. The complete
NRMSE B ETWEEN THE T RUE AND THE I NVERTED E LASTIC PARAMETERS
inversion process based on RM and EZE is from [35] and U SING EZE, RM, DL-EZE, AND DL-RM FOR THE F ULL OVERTHRUST
[46]. Both EZE-based and RM-based inversions have a heavy M ODEL
shadow of the initial model in their inversion results. The
inversion results of RM are slightly better than those of EZE.
Although the inversion results shown in Fig. 8 can roughly
reflect the trend of the strata, there are many details that
cannot be revealed. This is due to the fact that the model-based
inversion method relies heavily on the initial model, and when
the accuracy of the initial model is low, its corresponding It can be seen that as the number of iterations increases, the
inversion results are also poor. value of the loss function in both the training and valida-
Fig. 9 shows the convergence curves of the training and val- tion sets gradually decreases, while the prediction accuracy
idation sets for different methods (i.e., DL-EZE and DL-RM). gradually increases and finally reaches a stable state, which

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
4511311 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

Fig. 14. NRMSE of inverted V p (a), Vs , (b), and Rho (c) for EZE, RM, DL-EZE, and DL-RM when using different well as the blind testing well.

TABLE II
NRMSE B ETWEEN THE T RUE AND THE I NVERTED E LASTIC PARAMETERS
U SING EZE, RM, DL-EZE, AND DL-RM FOR THE S YNTHETIC M ODEL
(CDP = 1600)

further validates the effectiveness of the network. Due to the

low accuracy of the input data for the DL-EZE method, its
validation set shows a consistently lower accuracy than DL-
RM. Fig. 10(a)–(c) and (d)–(f) shows the inversion results
based on DL-EZE and DL-RM, respectively. It is clear to
see that the inversion results of the data-driven approaches
(see Fig. 10) outperform those of the model-driven approaches Fig. 15. Training and validation convergence curves (loss and accuracy)
(see Fig. 8). This also validates the potential of the data-driven corresponding to different methods during transfer learning. (a) and (b): DL-
EZE. (c) and (d): DL-RM.
approach in seismic inversion and its advantages over the
model-driven approach. From Fig. 10, it can be seen that both
the DL-RM predictions and the DL-EZE predictions match i.e., the DL-RM inversion results (red lines) are in the best
well with the true values [see Fig. 5(a)–(c)]. However, because agreement with the true values (black lines). In addition, the
the training dataset used in the DL-EZE method ignores effects DL-RM is verified quantitatively to be optimal by comparing
such as transmission loss and internal multiples, some detailed the NRMSE of the results recorded in Table II.
information in the prediction results is obscured, as shown by
the green arrow and the oval box in Figs. 10(a)–(c). In partic-
ular, at the 1.3 s position [black arrows in Fig. 10(a)–(c)], the B. Field Data Example
DL-EZE predictions make it difficult to identify a stratum that Encouraged by the testing of the synthetic data, we apply the
is already real. In comparison, the DL-RM is clearly superior proposed method to real data to further validate the method.
to the DL-EZE in these respects. Furthermore, in order to The size of the real dataset is 201 (time) × 101 (inline) × 301
quantitatively compare the quality of the inversion results of (crossline), as shown in Fig. 12. The study zone is defined
the four methods (i.e., EZE, RM, DL-EZE, and DL-RM), the from 1700 to 2100 ms, with temporal sampling of 2 ms.
NRMSEs between the inversion results of the four methods In Fig. 12, the partially stacked angle gathers, referred to as
and the true values are recorded in Table I. The magnitudes small, middle, and large angle gathers with angle ranges of
of NRMSE are different in P-, S-wave velocity, and density 0◦ –10◦ , 10◦ –20◦ , and 20◦ –30◦ , respectively. The data have
associated with their different ranges of values. In line with been preprocessed for denoising and amplitude calibration.
the conclusions, we obtain earlier, the data-driven methods are Based on the interpreted horizon as well as the well-
preferable to the model-driven methods, that is, the NRMSE logs data, the low-frequency (low-pass filter of 10 Hz) ini-
of the data-driven methods (i.e., DL-EZE and DL-RM) are tial P-, S-wave velocity, and density models are shown in
smaller than the NRMSE of the model-driven methods (i.e., Fig. 13(a)–(c). In this example, there are ten wells available,
EZE and RM). However, the NRMSE of DL-RM is the named “w1”–“w10.” We randomly selected 9 out of the
smallest. 10 wells to build the training dataset to fine-tune the pretrained
To further assess the DL-RM results, the true, initial, EZE, network, while the remaining one well is used as a blind well
RM, DL-EZE, and DL-RM inverted V p , Vs , and rho curves to verify the inversion results.
at CDP 1600 are shown together in Fig. 11. The DL-RM To test the generalization capability of the network, we use
approach seems to have yielded the most reliable results, different wells as the blind well in turn for fine-tuning. Fig. 14

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: DL SEISMIC INVERSION BASED ON PRESTACK WAVEFORM DATASETS 4511311

TABLE III
NRMSE B ETWEEN THE T RUE AND THE I NVERTED E LASTIC PARAMETERS
U SING EZE, RM, DL-EZE, AND DL-RM FOR THE R EAL D ATA (B LIND
W ELL - W 5)

Fig. 20. Comparison of elastic parameter curves (P-, S-wave velocity, and
density) of blind well-w5.

performance on different wells. However, the data-driven

approaches outperform the model-driven approaches and the
RM-based approaches outperform the EZE-based approaches,
Fig. 16. Inverted results obtained using EZE: (a) V p , (b) Vs , and (c) density. which is consistent with the conclusions obtained from the
synthetic tests. In the case of w5 as a blind testing well (i.e.,
the red marker in Fig. 12), the fine-tuned network is applied
to the entire 3-D cube. Fig. 15 shows the convergence curves
of the training and validation sets (loss and accuracy) cor-
responding to different methods (i.e., DL-EZE and DL-RM)
during transfer learning. DL-RM is superior to DL-EZE in that
the loss of DL-RM converges to a smaller value first and the
accuracy reaches a larger value faster.
Figs. 16–19 show the inverted results of EZE, RM, DL-EZE,
and DL-RM, respectively. All the inversion results shown in
Fig. 17. Inverted results obtained using RM: (a) V p , (b) Vs , and (c) density. Figs. 16–19 can reveal the change of the strata, but those (see
Figs. 17 and 19) based on the RM method reflect more details.
To further compare the differences between the inversion
results of the four methods in detail, we show a comparison
of the blind well results in Fig. 20. From Fig. 20, we see
that the inverted V p , Vs , and rho of DL-EZE and DL-RM are
better match the well-logs. Table III lists the NRMSE between
the blind testing well and the inverted results of above four
methods. It can be found that the NRMSE of DL-RM is the
smallest. In conclusion, a series of numerical tests, qualitative,
Fig. 18. Inverted results obtained using DL-EZE: (a) V p , (b) Vs , and
and quantitative, show that the proposed method has the best
(c) density. performance. Although there are some differences between
this and the real data-driven method, the proposed method is
certainly a good choice in seismic inversion under the existing
technology.

IV. C ONCLUSION
Data-driven methods show great potential in seismic inver-
sion, which is then made difficult to apply in field data by
the limited number of labeled training data pairs. Augmenting
training data based on known physical models is a common
Fig. 19. Inverted results obtained using DL-RM: (a) V p , (b) Vs , and strategy. We propose to pretrain the deep neural network
(c) density. by providing training datasets for the data-driven method
based on the prestack waveform formulation, thus alleviating
shows the NRMSE of the four methods between inversion the requirement for a large amount of training data for the
results and true values when using different well as the blind data-driven method, while improving the stability and predic-
testing well. The result shows that the performance of the tion accuracy of the network. It has been shown through a
DL-RM is optimal regardless of which well is used as the blind single-trace model testing that RM can simulate more accurate
testing well. EZE, RM, DL-EZE, and DL-RM have different wavefield (e.g., transmission losses and internal multiples)

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
4511311 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 60, 2022

⎡         ⎤
− p2 + q p qs /μ −2 pq p /μ − p2 − q p qs /μ p2 − q p qs /μ −2 pqs /μ − p2 + q p qs /μ
⎢ ⎥
⎢ 1i qs /v s2 0 −1i qs /v s2 −1i qs /v s2 0 −1i qs /v s2 ⎥
⎢ ⎥
⎢         ⎥
⎢ ⎥
⎢ −1i p + 2q p qs −4i p2q p −1i p − 2q p qs 1i p − 2q p qs −2i qs −1i p + 2q p qs ⎥
T+ = ⎢ ⎥ (9)
n
⎢         ⎥
⎢ −1i p + 2q p qs −2i q p −1i p − 2q p qs 1i p − 2q p qs −4i p2 qs −1i p + 2q p qs ⎥
⎢ ⎥
⎢ ⎥
⎣ ⎦
 −1i q p /v s2  −1i q p /v s  −1i q p /v s2  /v
2 2
 0  0 1i
 2 q p s 
−μ 2
+ 4 p 2 q p qs −4μ pq p −μ 2 − 4 p 2 q p qs μ 2
− 4 p 2 q p qs −4μ pqs −μ + 4 p q p qs
2

than EZE, which helps to construct more realistic training ACKNOWLEDGMENT

datasets. We then demonstrate the proposed method using
The authors would like to thank the Editor John Wright
synthetic models and real data. In all experiments, qualitative
and Simon Yueh, and two anonymous reviewers for a thorough
and quantitative comparisons of the inversion results of the
reading and many constructive comments on this article which
four methods (i.e., EZE, RM, DL-EZE, and DL-RM) show
significantly improves this article.
that: 1) the results of the data-driven method outperform the
results of the model-driven method and 2) the results of
the RM-based method outperform the results of the EZE-
R EFERENCES
based method. Overall, the DL-RM is optimal. This article
further validates the advantages of the data-driven approach [1] R. B. Latimer, R. Davidson, and P. van Riel, “An interpreter’s guide
over the model-driven approach. Furthermore, it is shown to understanding and working with seismic-derived acoustic impedance
data,” Lead. Edge, vol. 19, no. 3, pp. 242–256, Mar. 2000.
that the more realistic systems or training data pairs are [2] M. Bosch, T. Mukerji, and E. F. Gonzalez, “Seismic inversion for
utilized, the better prediction results can be obtained for reservoir properties combining statistical rock physics and geostatistics:
both model-driven and data-driven methods. Since the method A review,” Geophysics, vol. 75, no. 5, pp. 75A165–75A176, 2010.
[3] Q. Guo, J. Ba, and C. Luo, “Prestack seismic inversion with data-driven
we propose in this article is a fully data-driven approach MRF-based regularization,” IEEE Trans. Geosci. Remote Sens., vol. 59,
where the training dataset is requested in advance. If physical no. 8, pp. 7122–7136, Aug. 2021.
models can be added directly to deep neural networks, which [4] C. Luo, J. Ba, J. M. Carcione, G. Huang, and Q. Guo, “Joint PP and PS
pre-stack seismic inversion for stratified models based on the propagator
helps alleviate the need of purely data-driven methods to matrix forward engine,” Surv. Geophys., vol. 41, no. 5, pp. 987–1028,
the amount of training data and even enable unsupervised Sep. 2020.
learning. [5] A. Tarantola, Inverse Problem Theory and Methods for Model Parameter
Estimation. SIAM, 2005.
A PPENDIX [6] J. Zhang, J. Li, X. Chen, and Y. Li, “Geological structure-guided hybrid
MCMC and Bayesian linearized inversion methodology,” J. Petroleum
A. Representation of the Matrices En , T+ −
n , Tn
Sci. Eng., vol. 199, Apr. 2021, Art. no. 108296.
[7] L. P. de Figueiredo, D. Grana, M. Roisenberg, and B. B. Rodrigues,
The phase shift matrix En is written as “Gaussian mixture Markov chain Monte Carlo method for linear seismic
 inversion,” Geophysics, vol. 84, no. 3, pp. R463–R476, May 2019.
En = diag e−iωh (1/v p +1/vs ) 1 e−iωh (1/v p −1/vs ) [8] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
 no. 7553, pp. 436–444, 2015.
× eiωh (1/v p −1/vs ) 1 eiωh (1/v p +1/vs ) (8) [9] L. Perez and J. Wang, “The effectiveness of data augmentation in image
classification using deep learning,” 2017, arXiv:1712.04621.
where h represents the thickness of the n th layer of media. [10] A. Voulodimos, N. Doulamis, A. Doulamis, and E. Protopapadakis,
“Deep learning for computer vision: A brief review,” Comput. Intell.
v p and v s represent P-wave velocity and S-wave velocity, Neurosci., vol. 2018, pp. 1–13, Feb. 2018.
respectively. [11] D. W. Otter, J. R. Medina, and J. K. Kalita, “A survey of the usages
The downward energy distribution matrix T+ n is expressed
of deep learning for natural language processing,” IEEE Trans. Neural
Netw. Learn. Syst., vol. 32, no. 2, pp. 604–624, Feb. 2021.
as (9), shown at the top of the page, where ρ is the density, [12] X. Wu, Y. Shi, S. Fomel, L. Liang, Q. Zhang, and A. Z. Yusifov,
and “FaultNet3D: Predicting fault probabilities, strikes, and dips with a
single convolutional neural network,” IEEE Trans. Geosci. Remote Sens.,
= 2 p2 − 1/v s2 , μ = ρv s2 vol. 57, no. 11, pp. 9138–9155, Nov. 2019.
 
[13] Z. Bi, X. Wu, Z. Geng, and H. Li, “Deep relative geologic time: A deep
q p = 1/v 2p − p2 , qs = 1/v s2 − p2 , p = sin θ /v p . (10) learning method for simultaneously interpreting 3-D seismic horizons
and faults,” J. Geophys. Res., Solid Earth, vol. 126, no. 9, Sep. 2021,
The upward energy distribution matrix T− n can be obtained Art. no. e2021JB021882.
by rearranging the elements ti j of the matrix T+
n , as follows:
[14] L. Yang, W. Chen, H. Wang, and Y. Chen, “Deep learning seismic
⎡ ⎤ random noise attenuation via improved residual convolutional neural
t61 t51 t31 t31 t21 t11 network,” IEEE Trans. Geosci. Remote Sens., vol. 59, pp. 7968–7981,
⎢ −t65 0 −t45 −t35 0 −t15 ⎥ 2021.
⎢ ⎥
⎢ −t13 ⎥
[15] O. M. Saad and Y. Chen, “A fully unsupervised and highly general-
T− = ⎢ −t63 −t51 −t33 −t33 t12 ⎥ (11) ized deep learning approach for random noise suppression,” Geophys.
n ⎢ t63 −t51 t33 t33 t21 t13 ⎥
⎢ ⎥ Prospecting, vol. 69, no. 4, pp. 709–726, May 2021.
⎣ −t62 0 −t42 −t32 0 −t12 ⎦ [16] J. Zhang, J. Li, X. Chen, Y. Li, and W. Tang, “A spatially coupled
data-driven approach for lithology/fluid prediction,” IEEE Trans. Geosci.
t61 −t51 t31 t31 −t21 t11 Remote Sens., vol. 59, no. 7, pp. 5526–5534, Jul. 2021.

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: DL SEISMIC INVERSION BASED ON PRESTACK WAVEFORM DATASETS 4511311

[17] J. Zhang, J. Li, X. Chen, and Y. Li, “Seismic lithology/fluid prediction [41] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans.
via a hybrid ISD-CNN,” IEEE Geosci. Remote Sens. Lett., vol. 18, no. 1, Knowl. Data Eng., vol. 22, no. 10, pp. 1345–1359, Jan. 2009.
pp. 13–17, Jan. 2021. [42] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
[18] W. Zhang and J. Gao, “Deep-learning full-waveform inversion using 2014, arXiv:1412.6980.
seismic migration images,” IEEE Trans. Geosci. Remote Sens., vol. 60, [43] J. El Zini, Y. Rizk, and M. Awad, “A deep transfer learning framework
pp. 1–18, 2022. for seismic data analysis: A case study on bright spot detection,” IEEE
[19] B. Liu, S. Yang, Y. Ren, X. Xu, P. Jiang, and Y. Chen, “Deep- Trans. Geosci. Remote Sens., vol. 58, no. 5, pp. 3202–3212, May 2020.
learning seismic full-waveform inversion for realistic structural models,” [44] A. Mustafa, M. Alfarraj, and G. AlRegib, “Joint learning for spatial
Geophysics, vol. 86, no. 1, pp. R31–R44, Jan. 2021. context-based seismic inversion of multiple data sets for improved gen-
[20] V. Das, A. Pollack, U. Wollner, and T. Mukerji, “Convolutional neural eralizability and robustness,” Geophysics, vol. 86, no. 4, pp. O37–O48,
network for seismic impedance inversion,” Geophysics, vol. 84, no. 6, Jul. 2021.
pp. R869–R880, Nov. 2019. [45] A. Mustafa and G. AlRegib, “A comparative study of transfer learning
[21] V. Kazei et al., “Mapping full seismic waveforms to vertical velocity methodologies and causality for seismic inversion with temporal con-
profiles by deep learning velocity model building by deep learning,” volutional networks,” in Proc. 1st Int. Meeting Appl. Geosci. Energy
Geophysics, vol. 86, no. 5, pp. R711–R721, 2021. Expanded Abstr., Sep. 2021, pp. 1630–1634.
[22] A. Mustafa, M. Alfarraj, and G. AlRegib, “Estimation of acoustic [46] G. Huang, X. Chen, C. Luo, and X. Li, “Application of optimal transport
impedance from seismic data using temporal convolutional network,” to exact Zoeppritz equation AVA inversion,” IEEE Geosci. Remote Sens.
SEG Tech. Program Expanded Abstr., vol. 2019, pp. 2554–2558, Lett., vol. 15, no. 9, pp. 1337–1341, Sep. 2018.
Sep. 2019.
[23] V. Puzyrev, A. Egorov, A. Pirogova, C. Elders, and C. Otto, “Seismic
inversion with deep neural networks: A feasibility analysis,” in Proc.
81st EAGE Conf. Exhibit., vol. 1, 2019, pp. 1–5.
[24] B. Wu, D. Meng, L. Wang, N. Liu, and Y. Wang, “Seismic impedance Jian Zhang received the B.Sc. degree in geophysics
inversion using fully convolutional residual network and transfer learn- from Northeast Petroleum University, Daqing,
ing,” IEEE Geosci. Remote Sens. Lett., vol. 17, no. 12, pp. 2140–2144, China, in 2016, and the Ph.D. degree in geophysics
Dec. 2020. from the China University of Petroleum-Beijing,
[25] J. Zhang, J. Li, X. Chen, Y. Li, G. Huang, and Y. Chen, “Robust Beijing, China, in 2021.
deep learning seismic inversion with a priori initial model constraint,” He is currently a Lecturer with the Faculty of Geo-
Geophys. J. Int., vol. 225, no. 3, pp. 2001–2019, Mar. 2021. sciences and Environmental Engineering, Southwest
[26] L. Wang, D. Meng, and B. Wu, “Seismic inversion via closed-loop Jiaotong University, Chengdu, China. His research
fully convolutional residual network and transfer learning,” Geophysics, interests include seismic data inversion, deep learn-
vol. 86, no. 5, pp. 1–54, 2021. ing (DL), reservoir characterization and evaluation,
and nonlinear parameter estimation.
[27] M. Alfarraj and G. AlRegib, “Semisupervised sequence model-
ing for elastic impedance inversion,” Interpretation, vol. 7, no. 3,
pp. SE237–SE249, Aug. 2019.
[28] B. Wu, D. Meng, and H. Zhao, “Semi-supervised learning for seismic
impedance inversion using generative adversarial networks,” Remote Hui Sun received the B.S. and Ph.D. degrees in
Sens., vol. 13, no. 5, p. 909, 2021. geophysics from Jilin University, Changchun, China,
[29] L. Song, X. Yin, Z. Zong, and M. Jiang, “Semi-supervised learning in 2011 and 2017, respectively.
seismic inversion based on spatio-temporal sequence residual mod- From 2018 to 2019, he was a Post-Doctoral
eling neural network,” J. Petroleum Sci. Eng., vol. 208, Jan. 2022, Researcher with the University of California, Santa
Art. no. 109549. Cruz (UCSC), Santa Cruz, CA, USA. He is cur-
[30] S. Chopra et al., “Seismic characterization of a triassic-jurassic deep rently a Lecturer with the Faculty of Geosciences
geothermal sandstone reservoir, onshore Denmark, using unsupervised and Environmental Engineering, Southwest Jiaotong
machine learning techniques,” Interpretation, vol. 9, no. 4, pp. 1–48, University, Chengdu, China. His research interests
2021. include seismic signal processing, seismic imaging
[31] J. S. Dramsch, A. N. Christensen, C. MacBeth, and M. Luthje, “Deep methods, and forward modeling.
unsupervised 4-D seismic 3-D time-shift estimation with convolutional
neural networks,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–16,
2022.
[32] A. Keiiti and P. G. Richards, Quantitative Seismology: Theory and
Methods, vol. 859. San Francisco, CA, USA: Freeman, 1980. Gan Zhang received the B.Sc. degree in geo-
[33] J. L. Fatti, G. C. Smith, P. J. Vail, P. J. Strauss, and P. R. Levitt, physics from the China University of Petroleum
“Detection of gas in sandstone reservoirs using AVO analysis: A 3-D (East China), Qingdao, China, in 2016, and the
seismic case history using the Geostack technique,” Geophysics, vol. 59, M.Sc. degree in geophysics from the China Univer-
pp. 1362–1376, Sep. 1994. sity of Petroleum-Beijing, Beijing, China, in 2019.
[34] L. Zhi, S. Chen, and X.-Y. Li, “Amplitude variation with angle inver- He is currently a Researcher with Sichuan
sion using the exact Zoeppritz equations—Theory and methodology,” Water Resources and Hydroelectric Investigation &
Geophysics, vol. 81, no. 2, pp. N1–N15, Mar. 2016. Design Institute Company Ltd., Chengdu, China. His
[35] H. Liu, J. Li, X. Chen, B. Hou, and L. Chen, “Amplitude variation with research interests include geophysical data process-
offset inversion using the reflectivity method,” Geophysics, vol. 81, no. 4, ing, full waveform inversion (FWI), and nonlinear
pp. R185–R195, Jul. 2016. parameter estimation.
[36] S. A. M. Oliveira, I. L. S. Braga, M. B. Lacerda, G. F. Ouverney,
and A. W. P. de Franco, “Extending the useful angle range for elastic
inversion through the amplitude-versus-angle full-waveform inversion
method,” Geophysics, vol. 83, no. 3, pp. R213–R226, May 2018. Xiaoyan Zhao received the B.Sc. degree in geolog-
[37] B. L. N. Kennett and N. J. Kerry, “Seismic waves in a stratified half ical engineering and the Ph.D. degree in geotechni-
space,” Geophys. J. Int., vol. 57, no. 3, pp. 557–583, 1979. cal engineering from Southwest Jiaotong University,
[38] R. A. Phinney, R. I. Odom, and G. J. Fryer, “Rapid generation of syn- Chengdu, China, in 2000 and 2005, respectively.
thetic seismograms in layered media by vectorization of the algorithm,” He is currently a Professor with the Faculty of
Hawaii Univ. Honolulu Inst. Geophys., Tech. Rep., 1987. Geosciences and Environmental Engineering, South-
[39] Y. Ma, L. Loures, and G. F. Margrave, “Seismic modeling with the west Jiaotong University. His research interests
reflectivity method,” CREWES Res. Rep., vol. 15, pp. 1–7, 2004. include signal processing, slope instability mecha-
[40] Y. Li, J. Li, X. Chen, J. Zhang, and X. Bo, “Prestack waveform nism, reinforcement, and geologic characterization.
inversion based on analytical solution of the viscoelastic wave equation,”
Geophysics, vol. 86, no. 1, pp. R45–R61, Jan. 2021.

Authorized licensed use limited to: National Geophysical Research Institute. Downloaded on January 25,2024 at 05:58:45 UTC from IEEE Xplore. Restrictions apply.