TYPE Original Research

PUBLISHED 27 February 2023

DOI 10.3389/feart.2023.1104488

Multichannel seismic impedance

OPEN ACCESS inversion based on Attention
EDITED BY
Peng Zhenming,
University of Electronic Science and
U-Net
Technology of China, China

REVIEWED BY Juan Ning 1, Shu Li 1,2*, Zong Wei 1 and Xi Yang 1

Gulan Zhang, 1
Southwest Petroleum University, China School of Communication and Electronic Engineering, Jishou University, Jishou, China, 2School of
Bibo Yue, Biomedical Engineering, Guangzhou Medical University, Guangzhou, China
Southwest Petroleum University, China
Cai Hanpeng,
University of Electronic Science and
Technology of China, China
Recently, seismic inversion has made extensive use of supervised learning
*CORRESPONDENCE
Shu Li,
methods. The traditional deep learning inversion network can utilize the
shuli@jsu.edu.cn temporal correlation in the vertical direction. Still, it does not consider the
SPECIALTY SECTION
spatial correlation in the horizontal direction of seismic data. Each seismic
This article was submitted to trace is inverted independently, which leads to noise and large geological
Environmental Informatics variations in seismic data, thus leading to lateral discontinuity. Given this, the
and Remote Sensing,
a section of the journal
proposed method uses the spatial correlation of the seismic data in the horizontal
Frontiers in Earth Science direction. In the network training stage, several seismic traces centered on the
RECEIVED 21 November 2022 well-side trace and the corresponding logging curve form a set of training sample
ACCEPTED 13 February 2023 pairs for training, to enhance the lateral continuity and anti-noise performance.
PUBLISHED 27 February 2023 Additionally, Attention U-Net is introduced in acoustic impedance inversion.
CITATION Attention U-Net adds attention gate (AG) model to the skip connection
Ning J, Li S, Wei Z and Yang X (2023), between the encoding and decoding layers of the U-Net network, which can
Multichannel seismic impedance
inversion based on Attention U-Net. give different weights to different features, so the model can focus on the features
Front. Earth Sci. 11:1104488. related to the inversion task and avoid the influence of irrelevant data and noise
doi: 10.3389/feart.2023.1104488 during the inversion process. The performance of the proposed method is
COPYRIGHT evaluated using the Marmousi2 model and the SEAM model and compared
© 2023 Ning, Li, Wei and Yang. This is an
with other methods. The experimental results show that the proposed method
open-access article distributed under the
terms of the Creative Commons has the advantages of high accuracy of acoustic impedance value inversion, good
Attribution License (CC BY). The use, transverse continuity of inversion results, and strong anti-noise performance.
distribution or reproduction in other
forums is permitted, provided the original
author(s) and the copyright owner(s) are KEYWORDS
credited and that the original publication
Attention U-Net, acoustic impedance inversion, spatial correlation, deep learning,
in this journal is cited, in accordance with
accepted academic practice. No use,
multichannel inversion
distribution or reproduction is permitted
which does not comply with these terms.
1 Introduction
Seismic inversion can be defined as the process of obtaining subsurface model
parameters, such as formation velocity, density, or impedance, from seismic data by
comprehensively available geological and logging data (Treitel and Lines, 2001). For
conventional seismic inversion methods, i.e., model-driven inversion methods, the
mathematical theory is based on the convolution model or other mathematical and
physical models. The convolution model is essentially a simplification and
approximation of the seismic wave transmission process. The subsurface structure is
usually very complex, and errors will inevitably arise when describing the wave
propagation with the convolution model, which leads to inaccurate inversion results. On
the other hand, in order to get a good inversion result, the model-driven method needs a
better initial model and an accurate wavelet. In practical applications, it is usually challenging
to obtain good initial models and accurate wavelets. In addition, problems such as limited

Frontiers in Earth Science 01 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

data bandwidth, data noise, and incomplete data coverage cause to a certain extent while improving training efficiency. Seismic
various troubles for model-driven inversion methods. inversion also faces the problem of a small number of labels (few
Unlike traditional model-driven seismic inversion, deep logging data) and a very large amount of seismic data. In view of this,
learning is a data-driven approach that can learn complex Cao et al. (2022) proposed an inversion network consisting of a
non-linear mappings between inputs and outputs based on U-Net combined with three fully connected networks and named it
training datasets, and the parameters are adjustable. Deep the UCNN, which was used to predict elastic parameters from pre-
learning is a subset of machine learning that has recently stack seismic data. To further reduce the reliance on labeled data,
made breakthroughs in image classification (Krizhevsky et al., they use Sequential Gaussian Co-Simulation and Elastic Distortion
2017), object detection (Ren et al., 2015), image segmentation algorithms to generate adequate and diverse pre-stack seismic
(Chen et al., 2017), image and video captioning (Vinyals et al., inversion datasets. Similarly, Wang et al. (2020) proposed a
2022), speech recognition (Graves et al., 2013), and machine closed-loop CNN structure with a U-Net network as the main
translation (Cho et al., 2014). The success of deep learning in the body to make CNN less dependent on the amount of labeled
fields of computer vision and natural language processing has data in seismic inversion. The proposed closed-loop CNN can
led to widespread interest among scholars in data-driven simulate both seismic forward and inversion processes from the
intelligent seismic inversion methods. This class of methods training dataset.
does not require an initial model and does not require the Given the excessive and repeated extraction and utilization of
estimation of seismic wavelets. Using the powerful learning similar features for each cascaded CNN structure in U-Net, this
ability of deep neural networks to establish non-linear results in a significant computational effort and network parameter
mapping relationships between seismic data and parameters scale. Oktay et al. (2018) proposed the Attention Gate (AG) model
to be inverted has become a trendy research direction in the and integrated it into U-Net to obtain the Attention U-Net network.
field of seismic inversion. The AG model can implicitly learn to emphasize prominent features
Currently, the application of deep learning methods in the field that are helpful for inversion while suppressing irrelevant regions in
of seismic inversion is expanding, involving acoustic impedance the input data. In addition, AG is easily integrated into standard
inversion, pre-stack elastic and lithological parameter inversion, full CNN architectures such as U-Net, which can reduce the
waveform inversion, and so on. Recently, seismic inversion has computational overhead while improving the sensitivity and
made extensive use of supervised learning methods. Alfarraj and prediction accuracy of the network.
AlRegib (2018) used recurrent neural networks for petrophysical In conclusion, this paper proposes a multichannel acoustic
parameter estimation. Das et al. (2019) and Wu et al. (2020) trained impedance inversion based on Attention U-Net to address the
the convolutional neural networks (CNNs) to invert seismic issues with conventional deep learning inversion networks, such
impedance using synthetic seismic records on the earth model as poor continuity of inversion results and susceptibility to noise due
constrained by petrophysical relationships. The results show that to the trace-by-trace inversion method. The horizontal spatial
the type of sediment phase and source wavelet parameters used in correlation is applied to the inversion network by mapping
the training dataset affect the inversion process of the network. multiple seismic traces to one logging curve. Under the
Mustafa et al. (2019) used the temporal convolution network (TCN) supervision of limited logging data, the inversion network is
to estimate the acoustic impedance. This method not only trained. The training samples consist of several seismic traces
successfully captured the long-term trend but also preserved the centered on the well-side traces and associated well-logging
local patterns while overcoming the gradient disappearance problem curves. The inversion network simultaneously performs the
in the inversion of recurrent neural network (RNN) and the duties of predicting acoustic impedance and forwarding seismic
overfitting problem in convolutional neural networks. Du et al. data. This paper is structured as follows: In Section 2, the theory and
(2019) proposed SeisInv-ResNet for pre-stack seismic inversion to network structure of Attention U-Net are briefly introduced, and
obtain p-wave impedance, s-wave impedance, and other then the architecture of the inversion network consisting of three
petrophysical parameters. Aleardi and Salusti (2021) proposed an modules and their specific internal parameter settings are presented.
elastic pre-stack seismic inversion method based on CNN. In Section 3, the experimental results of the inversion of two typical
Although the above inversion networks based on deep learning seismic models (the Marmousi2 model and the SEAM model) are
can well utilize the temporal correlation in the vertical direction, presented, analyzed, and discussed. The experimental results are
they do not consider the spatial correlation of seismic data in the compared with other deep learning inversion methods, and the
horizontal direction, and each seismic trace is inverted noise immunity of the inversion network is discussed in this paper.
independently. However, in subsurface seismic profiles, adjacent Finally, Section 4 concludes this paper.
traces are highly correlated. The inversion method based on trace by
trace does not exploit the spatial correlation in the horizontal
direction, which may lead to poor horizontal continuity of 2 Methods
inversion results. To improve the continuity, Wu et al. (2021)
proposed a 2D network-based inversion method. 2.1 Inversion framework
Traditional CNN networks take a long time to train and need a
lot of labeled data. To address these drawbacks of classical CNN Geological structures are spatially correlated. The closer the
networks, Ronneberger et al. (2015) proposed the U-Net network in distance, the stronger the correlation, and conversely, the weaker the
their study of biomedical image segmentation problems. Their correlation. The correlation of seismic data is reflected in the
research shows that U-Net can reduce the need for labeled data temporal correlation in the vertical direction of seismic traces

Frontiers in Earth Science 02 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 1
Structure of inversion network.

and the spatial correlation in the horizontal direction between the feature extraction module is the seismic data of the well-side trace
central trace and the adjacent traces. Based on the spatiotemporal and the nearby 2 k traces centered on it, and the output feature size is
characteristics of the seismic data, the inversion framework in the same as the input size. Attention U-Net is improved by using
Figure 1 is constructed using a supervised learning approach. U-Net as the base framework, as shown in Figure 1, adding AG at the
The inversion framework shown in Figure 1 consists of three jump connection between the encoding-decoding layers of the
main modules: the feature extraction module, the regression U-Net network, so that the originally up-sampled features are
module, and the forward module. connected with the encoded layer AG-processed signal. By
In the training phase, the input of the inversion network is the assigning different weights to different features, the model is
seismic data of the well-side trace and the 2 k nearby seismic data better able to pay attention to the features relevant to the
centered on it. The feature extraction module extracts the inversion task, which improves the sensitivity and prediction
temporal and spatial features of the seismic data of the well- accuracy of the model.
side trace and the 2 k nearby seismic traces by Attention U-Net. Attention U-Net is divided into an encoding part and a decoding
The regression module is used to map the data from the feature part, as shown in Figure 1. The encoding part of the Attention U-Net
domain (spatiotemporal feature series) to the target domain framework used in this paper contains four downsampling layers.
(predicted acoustic impedance), while the forward module is The downsampling layer includes two consecutive convolutional
used to map the data from the feature domain to the target blocks and a 2 × 1 max-pooling layer, and each convolutional block
domain (forward 2 k + 1 traces seismic data). Referring to the consists of a 3 × 3 two-dimensional convolutional layer (Conv2d), a
structure of the multi-task learning of Mustafa et al. (2021), the batch normalization layer (BN) (Ioffe and Szegedy, 2022), and a
inverse network learns two tasks simultaneously: the predicted rectified linear unit (ReLU) (Nair and Hinton, 2010) activation
acoustic impedance data and forward seismic data. By sharing function. Batch normalization is used to accelerate the convergence
representations between the two tasks, especially if they are of the network, and ReLU is used to enhance the non-linear
related to each other, we bias the network to learn more approximation capability of the model. The decoding part
generalizable features. corresponds to the encoding part, and the decoding part also
contains four upsampling layers. Each upsampling layer consists
of a 4 × 3 deconvolution layer, an AG model, and two convolution
2.2 Network model blocks.
The input of AG is the feature in the encoding part and the
2.2.1 Feature extraction module feature after deconvolution in the decoding part. The specific
The Attention U-Net is used as a feature extraction module to structure of AG is shown in Figure 2. The features extracted
extract spatial and temporal features of seismic data. The input of the from the decoding part after deconvolution are used as the

Frontiers in Earth Science 03 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 2
Attention gating model.

FIGURE 3
Block diagram of the regression module.

gating signal g, and the features from the matching layer’s coding normalization layer, and the ReLU activation function. Group
portion are used as x. The 1 × 1 convolution is done for g and x, and normalization groups the outputs of the convolutional layers and
the two results A and B are added element by element to highlight normalizes each group using the learned mean and standard
the features. Then, the non-linear ability of the added result is deviation, which have been shown to reduce covariate bias in the
increased by the ReLU activation function to obtain C, and the learned features and speed up learning (Wu and He, 2012).
channel of C is reduced to 1 channel by a convolution operation. D is As shown in Figure 3, the input of the regression module is the
processed using a sigmoid activation function such that its value falls output of the feature extraction module, and the output is the predicted
within the range of (0, 1), and the result is an attention weight that is acoustic impedance. Calculate the mean square error between the actual
the same size as the input feature and has one channel. Finally, the acoustic impedance and the output of the regression module. In other
attention weight is multiplied by x. words, the mean square error between the predicted and the actual
acoustic impedance data is calculated to update the learnable
2.2.2 Regression module parameters in the feature extraction module and the regression
The regression module maps the output of the feature extraction module. The following Eq. 1 illustrates this:
module from the feature domain to the target domain. The
^ i,t 
l1
MSE
mi,t , m (1)
regression module’s structure, as shown in Figure 3, consists of
two convolutional blocks and a 2D convolutional layer. Each ^ i,t is the predicted
where mi,t is the actual acoustic impedance, m
convolutional block consists of a 2D convolutional layer, a group acoustic impedance, and MSE is as in Eq. 4.

Frontiers in Earth Science 04 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 4
Block diagram of the forward module.

2.2.3 Forward module where α and β are weighting factors that control the effects of
The forward module maps the output of the feature extraction acoustic impedance losses and seismic losses, respectively.
module from the feature domain to the target domain. As shown in
Figure 4, the input of the forward module is the output of the feature
extraction module, and the output is the predicted well-side trace 2.4 Evaluation of inversion results
and 2 k nearby seismic data. The structure of the forward module
consists of two convolutional blocks plus a 2D convolutional layer. The inversion results are evaluated quantitatively by calculating
Each convolutional block consists of a 2D convolutional layer, a the mean square error (MSE) and the coefficient of determination
group normalization layer, and a ReLU activation function to (R2) of the actual and predicted acoustic impedance.
achieve reconstruction. Mean Squared Error (MSE): MSE is the average of the squared
Calculate the mean square error between the feature extraction sum of the errors of the corresponding points of the predicted data
module’s input and the forward module’s output. To put it another and the real data, and the smaller the value indicates that the
way, the mean square error between the 2 k + 1 seismic data in the predicted data fits better with the original data, which is defined as:
well-side trace and nearby traces and the predicted 2 k + 1 seismic
1 N  2
data is calculated in order to update the learnable parameters in the MSE
y − y^i 2 (4)
N i
1 i
feature extraction module and the forward module. The following
Eq. 2 illustrates this:
where yi , y^i denote the actual acoustic impedance and predicted
l2
MSE
xi,t , x^i,t  (2) impedance, respectively, and N is the number of data.
Determination Coefficient (R2): R2 is a measure of the goodness
where xi,t is the seismic data of the well-side and nearby 2 k traces, of fit between variables that takes into account the mean square error
and x^i,t is the predicted 2 k + 1 seismic data. between predicted and actual data. Its range of values is [0, 1], and
the larger the value, the better the fit between the variables, the more
the independent variable explains the dependent variable, and the
2.3 Loss function more the independent variable contributes to the overall variation. It
is defined as:
The loss of the entire inversion network is the mean square error
N 2
between the predicted acoustic impedance data and the actual
2
i
1
yi − y^i 
acoustic impedance (l1 ), and the mean square error between the R
y, y^
1 − N 2 (5)
i
1 yi − μy 
2 k seismic data in and around the well-side traces and the predicted
2 k + 1 seismic data (l2 ), with the total loss shown in Eq. 3: where yi , y^i , and μy represent the actual acoustic impedance,
l
αl1 + βl2 (3) predicted acoustic impedance, and the average of the actual

Frontiers in Earth Science 05 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 5
The Marmousi2 model. (A) Seismic data profile; (B) real acoustic impedance profile.

acoustic impedance, respectively. When R2 is closer to 1, the stronger validate inversion and imaging algorithms. The researchers added
the correlation between the predicted and actual acoustic more complex structures representing hydrocarbon regions to the
impedance is. model and increased the number of strata, resulting in a new model,
the Marmousi2 model, which has a width of 17 km and a depth of
3.5 km. The model is accompanied by synthetic seismic data, which
3 Experiments are obtained by convolutional forward simulations of the model’s
reflection coefficients using seismic wavelets.
The Marmousi2 and the SEAM models are widely used to The acoustic impedance model was obtained by multiplying the
validate the performance of deep learning inversion methods. density and p-velocity models of the Marmousi2 data. The seismic
This subsection will use these two models to validate the data and acoustic impedance profiles are shown in Figures 5A, B,
performance of the inversion network architecture proposed in with 2,721 traces and 688 sampling points per trace in the seismic
this paper for acoustic impedance inversion. profile and 2,721 traces and 688 sampling points per trace in the
acoustic impedance profile. The colors in Figure 5A represent
seismic amplitude values, and the colors in Figure 5B represent
3.1 Marmousi2 model inversion experiments the acoustic impedance values. Twenty traces of acoustic impedance
are uniformly extracted as pseudo-well data, and for each pseudo-
The Marmousi2 model is an extension of the original Marmousi logging curve, 2 k + 1 seismic traces centered on the well-side trace
model for amplitude variation with offset (AVO) analysis (Martin and with k as the radius will be obtained. This paper sets k to 3, and
et al., 2002). The original Marmousi model has been widely used to each pseudo-logging curve corresponds to 7 seismic traces with a

Frontiers in Earth Science 06 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 6
Acoustic impedance inversion profiles and residual profiles on the Marmousi2 model. (A) Inversion result of CNN method and its residual (B). (C)
Inversion result of TCN method and its residual (D). (E) Inversion result of U-Net method and its residual (F). (G) Inversion result of the method in this paper
and its residual (H).

Frontiers in Earth Science 07 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 7
Acoustic impedance inversion results of trace no. 570. (A) Inversion result of the CNN method. (B) Inversion result of the TCN method. (C) Inversion
result of the 1D U-Net method. (D) Inversion result of the method in this paper.

depth of 688 sampling points. The inverse network is trained using method based on CNN is shown in Figure 6A, the inversion result of
seismic data and pseudo-well data, the training epoch is set to 700, the method based on 1D TCN is shown in Figure 6C, the inversion
and the batch size is 20 for each iteration. In each training iteration, result of the method based on 1D U-Net is shown in Figure 6E, and
the weight coefficients in the loss function α and β are set to 1. The the inversion result of the method proposed in this paper is shown in
total training loss of the previously described inverse network is Figure 6G. Figures 6B, D, F, H correspond to the residual difference
calculated and back-propagated through the network. between each network’s inverse acoustic impedance and the real
The Marmousi2 model has a complex stratigraphic structure acoustic impedance.
and contains many different subsurface layered media models. The As shown in Figure 6, the inversion results shown in Figures 6E, G
mean square error function is chosen as the loss function to measure have a higher similarity to the real model than the inversion results in
the mean square error of the predicted and real acoustic impedance. Figures 6A, C. Moreover, Figure 6G has stronger horizontal continuity
ADAM is chosen as the optimizer, and ADAM adaptively sets the and weaker visible jitter in both horizontal and vertical directions for
learning rate during training, with the initial learning rate set to the inverse acoustic impedance profile than Figure 6E, the water layer
0.001. A weight decay of 0.0001 is chosen to limit the L2 norm of the at the top of the figure also clearly shows a relatively better inversion of
weights from becoming too large, reducing the risk of overfitting the Figure 6G. The partition interface and fault location in different strata
network. The network’s training is implemented in the PyTorch are the main locations where the inversion results show errors,
framework, and GPUs are applied to accelerate the computation. according to the residual profiles. In comparison to other figures
Finally, the trained inverse network is used for acoustic impedance in Figure 6, the inversion method in this paper can also invert the
inversion. convolution structure in the model well, and the inversion results are
In order to prove the effectiveness of this paper’s method, the more continuous and closer to the actual acoustic impedance, as well
inversion results of this paper’s inversion method are compared with as more accurate in predicting the location of the faults. In most
the inversion results of the commonly used deep learning inversion locations, the error is lower than that of other inversion methods. This
methods, including the inversion method based on CNN (Das et al., is due to the effective use of the inversion network proposed in this
2019), the inversion method based on 1D TCN (Mustafa et al., paper for the spatial correlation of seismic data’s horizontal direction.
2019), and the inversion method based on 1D U-Net. This 1D U-net In order to compare the details of the inversion results of
model is constructed into the same network structure as the U-net different methods from the microscopic level, the representative
proposed in this paper, but it lacks an attention mechanism. These Trace No. 570 (corresponding to the position around x = 3,565 m)
inverse networks are set up with the same training conditions, and Trace No. 1400 (corresponding to the position around x =
training data, and hyperparameters. The inversion result of the 8,747 m) are selected for inversion.

Frontiers in Earth Science 08 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 8
Acoustic impedance inversion results of trace no. 1400. (A) Inversion result of the CNN method. (B) Inversion result of the TCN method. (C) Inversion
result of the 1D U-Net method. (D) Inversion result of the method in this paper.

TABLE 1 MSE, R2 between inversion results and actual acoustic impedance. inversion results of Figures 8C, D are in better agreement with the
actual curves, but between sampling points 0 and 100, the inversion
Methods MSE R2
of Figure 8D is better, while the curve change of the inversion result
CNN 0.0897 0.9090 of Figure 8C is too drastic. This further validates the performance of
the inversion network proposed in this paper.
TCN 0.0540 0.9452
In order to objectively and quantitatively evaluate the reliability
U-Net 0.0343 0.9653 of the inversion results of the four methods, the coefficients R2 and
Attention U-Net 0.0199 0.9800 MSE are used as evaluation criteria. Table 1 shows the MSE and R2
between the acoustic impedance inversion results of different
methods in Figure 6 and the actual acoustic impedance.
Table 1 shows that this paper employs multichannel inversion,
At these two locations, the acoustic impedance values obtained by and the method of acoustic impedance inversion by Attention
four inversion methods were compared. Figures 7A–D shows the U-Net using spatial correlation performs best in terms of MSE
inversion results of Trace No. 570 using the conventional CNN and R2, demonstrating the method’s efficacy.
inversion method, the 1D TCN inversion method, the 1D U-Net Gaussian noise of 4%, 8%, and 12% was added to the seismic
inversion method, and the method proposed in this paper, with the data to test the adaptability of the method proposed in this paper to
red and black lines representing the true impedance and acoustic noise. Table 2 shows the quantitative evaluation of the inversion
impedance inversion results, respectively. Similar to the inversion results obtained from the different inversion networks in Figure 6
results of the four networks mentioned above for all seismic traces, under different noise conditions. As shown in Table 2, the
the inversion result of the method in this paper is relatively better. The performance of each method’s inversion results decreases as
inversion result in Figure 7A has a large inversion error at a large depth, noise increases relative to a noiseless environment, but the
the inversion result in Figure 7B is very different from the true value, and performance index of the method proposed in this paper
the inversion result in Figure 7C changes too drastically, whereas the decreases the least. For example, when the noise of the seismic
difference between the inversion result and the actual acoustic impedance data increases from 4% to 12%, the R2 coefficients of the inversion
in Figure 7D is very small, with the two curves almost overlapping. results of CNN, TCN, U-Net, and the proposed method decreased
Figures 8A–D corresponds to the inversion results of the above by 7.43%, 2.85%, 4.83%, and 1.38%, respectively. Observing the
four methods for Trace No. 1400 seismic trace, respectively, and the changes in MSE data leads to a similar conclusion. It can be seen that
conclusions are consistent with Figure 7. The inversion results of the proposed method in this paper has better noise immunity
Figures 8A, B in the figure deviate more from the true values. The performance compared with other methods.

Frontiers in Earth Science 09 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

TABLE 2 MSE, R2 between inversion results and actual acoustic impedance under different noise conditions.

Indicator、SNR MSE R2

methods 4% 8% 12% 4% 8% 12%

CNN 0.1146 0.1404 0.1787 0.8836 0.8571 0.8180

TCN 0.0622 0.0742 0.0886 0.9375 0.9253 0.9108

U-Net 0.0454 0.0729 0.0911 0.9540 0.9261 0.9079

Attention U-Net 0.0244 0.0285 0.0375 0.9751 0.9711 0.9616

FIGURE 9
SEAM model. (A) Seismic data profile. (B) Real acoustic impedance profile.

3.2 SEAM model inversion experiments with very drastic lateral variations in density and longitudinal wave
velocity, which is challenging for the inversion algorithm. The SEAM
To further verify the feasibility of the method, this paper conducts model is constructed based on basic rock properties, such as the volume
experiments with the SEAM model. The SEAM model is open source of shale and sand. It follows the changing trend of shale porosity
and also widely used for the validation of deep learning inversion characteristics in the Gulf of Mexico, which is a better simulation of the
methods (Mustafa et al., 2021). The SEAM model is a 3D seismic survey actual geological conditions. The density of the SEAM model and the

Frontiers in Earth Science 10 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 10
Acoustic impedance inversion profiles and residual profiles on the SEAM model. (A) Inversion result of CNN method and its residual (B). (C) Inversion
result of U-Net method and its residual (D). (E) Inversion result of the method in this paper and its residual (F).

longitudinal wave velocity model are multiplied to obtain the clearer. Some thin stratigraphic variations can be clearly observed
real acoustic impedance model. The seismic data and the real in the upper left part of Figure 10E diagram between 5,000 and
acoustic impedance profiles are shown in Figures 9A, B, respectively, 9,000 m depth. For example, at 2,500 m depth in the real model,
with 501 traces and 688 sampling points per trace in the seismic profile there is a thin arc-shaped stratigraphy that can be seen more
and 501 traces and 688 sampling points per trace in the acoustic clearly in Figure 10E, whereas it is difficult to see in Figures 10A,
impedance profile. 12 traces of acoustic impedance are uniformly C, and Figure 10A does not outline the central uplifted area in the
extracted from the acoustic impedance model as pseudo-well data, real model better. Although the method in this paper has some
and k is also set to 3, so that each pseudo-logging curve corresponds to errors in the inversion of the SEAM model, the overall effect is
7 seismic traces with a depth of 688 sampling points. The training epoch better than the other two methods.
is set to 400, and the batch size is 12 for each iteration. The network is Trace No. 179 (corresponding to the vicinity of x = 12,500 m) was
then trained in the same way as the Marmousi2 model, and the trained selected for the inversion experiment, and the acoustic impedance
network is used to perform acoustic impedance inversion on all seismic inversion results of the three inversion methods are shown in Figure 11.
traces. Figures 11A–C shows the inversion results of Trace No. 179 using the
The inversion results are shown in Figure 10. Figures 10A, C, conventional CNN inversion method, the 1D U-Net inversion method,
E correspond to the results of the inversion based on the and the method proposed in this paper, with the red and black lines
conventional CNN inversion method, the 1D U-Net inversion representing the true impedance and acoustic impedance inversion
method, and the inversion of the proposed method in this paper, results, respectively. The proposed method has better inversion results
respectively. Figures 10B, D, F correspond to the residuals compared with other methods. From Figure 11C, we can see that the
between the acoustic impedance and the real acoustic inversion result obtained by the proposed method almost completely
impedance inverted by each method, respectively. As can be overlaps with the true impedance, while the inversion result of the CNN
seen from the figure, compared with Figures 10C, E has a deviates from the true value, and the result obtained by the 1D U-Net
better effect in displaying the stratigraphic interface in the left inversion method also has large deviations, with a large deviation at a
half of the depth range of 10,000 m to 14,000 m, and the strata are small depth.

Frontiers in Earth Science 11 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

FIGURE 11
Acoustic impedance inversion results of trace no. 179. (A) Inversion result of the CNN method. (B) Inversion result of the 1D U-net method. (C)
Inversion result of the method in this paper.

TABLE 3 MSE, R2 between inversion results and actual acoustic impedance. network. Different from the conventional supervised learning
inversion method, this inversion method applies the spatial
Methods MSE R2
correlation in the horizontal direction to the inversion
CNN 0.2659 0.5436 network, and trains the network with 2 k + 1 seismic traces
U-Net 0.1549 0.7991 centered on the well-side trace and the corresponding logging
curve to enhance the lateral continuity. In addition, the Attention
Attention U-Net 0.1182 0.8250
U-Net network is used as a feature extraction module in the
inversion network, and the attention gating model is added to the
traditional U-Net-based inversion network. The AG is used to
To quantitatively evaluate the performance of the method implicitly learn to suppress irrelevant regions in the input data
proposed in this paper, the MSE and the R2 between the acoustic while emphasizing salient features useful for inversion results,
impedance inversion results and the true acoustic impedance are and it can be easily integrated into the standard CNN architecture
calculated and presented in Table 3. The data are the MSE and R 2 to reduce computational overhead while improving the model’s
between the acoustic impedance and the true acoustic sensitivity and prediction accuracy. The method’s performance is
impedance obtained by the inversion of different inversion evaluated using the Marmousi2 and SEAM models, and it is also
methods in Figure 10. The data in the table show that the compared to several other commonly used deep learning
inversion result of the proposed method performs best in inversion methods. The results show that the inversion results
terms of MSE and R2, which verifies the effectiveness of the of the method proposed in this paper are more consistent with
method. the actual acoustic impedance values, and the anti-noise
performance is the best. In the SEAM model, where the lateral
velocity and density vary drastically, the proposed method can
4 Conclusion better obtain the stratigraphic structure and details in the true
model. These are attributed to the combined application of the
This paper proposes a multichannel seismic acoustic attention gating model and methods such as multichannel
impedance inversion method based on the Attention U-Net simultaneous inversion.

Frontiers in Earth Science 12 frontiersin.org

Ning et al. 10.3389/feart.2023.1104488

Data availability statement the China Postdoctoral Science Foundation (No. 2021M700682)
and the Scientific Research Fund of Hunan Provincial Education
Publicly available datasets were analyzed in this study. This data Department (No. 21B0507).
can be found here: https://github.com/amustafa9/Geophysics-2021-
Joint-learning-for-spatial-context-based-inversion/blob/master/
data.zip. Conflict of interest
The authors declare that the research was conducted in the
Author contributions absence of any commercial or financial relationships that could be
construed as a potential conflict of interest.
JN responses for the experiments of the work and drafting papers.
SL responses for the concept and design of the work and revisions to the
paper. ZW and XY responses for important revisions to the papers. Publisher’s note
All claims expressed in this article are solely those of the authors
Funding and do not necessarily represent those of their affiliated
organizations, or those of the publisher, the editors and the
This research was funded by the National Natural Science reviewers. Any product that may be evaluated in this article, or
Foundation of China (Nos. 42164006 and 62161012), the Hunan claim that may be made by its manufacturer, is not guaranteed or
Provincial Natural Science Foundation of China (No. 2022JJ30474), endorsed by the publisher.

References
Aleardi, M., and Salusti, A. (2021). Elastic prestack seismic inversion through discrete Mustafa, A., Alfarraj, M., and AlRegib, G. (2019). “Estimation of acoustic impedance
cosine transform reparameterization and convolutional neural networks. Geophysics 86 from seismic data using temporal convolutional network,” in SEG technical program
(1), R129–R146. doi:10.1190/geo2020-0313.1 expanded abstracts 2019 (United States: Society of Exploration Geophysicists). .
Alfarraj, M., and AlRegib, G. (2018). “Petrophysical property estimation from seismic Mustafa, A., Alfarraj, M., and AlRegib, G. (2021). Joint learning for spatial context-
data using recurrent neural networks,” in SEG technical program expanded abstracts based seismic inversion of multiple data sets for improved generalizability and
2018 (United States: Society of Exploration Geophysicists). . robustness. Geophysics 86 (4), O37–O48. doi:10.1190/geo2020-0432.1
Cao, D., Su, Y., and Cui, R. (2022). Multi-parameter pre-stack seismic inversion based Nair, V., and Hinton, G. E. (2010). “Rectified linear units improve restricted
on deep learning with sparse reflection coefficient constraints. J. Petroleum Sci. Eng. 209, Boltzmann machines,” in Proceedings of the 27th International Conference on
109836. doi:10.1016/j.petrol.2021.109836 Machine Learning, Haifa, Israel, June 21-24, 2010.
Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and Yuille, A. L. (2017). Oktay, O., Schlemper, J., Folgoc, L. L., Lee, M., Heinrich, M., Misawa, K., et al. (2018).
Deeplab: Semantic image segmentation with deep convolutional nets, atrous “Attention u-net: Learning where to look for the pancreas,”. arXiv preprint arXiv:
convolution, and fully connected crfs. IEEE Trans. pattern analysis Mach. Intell. 40 1804.03999.
(4), 834–848. doi:10.1109/tpami.2017.2699184
Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster r-cnn: Towards real-time object
Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., detection with region proposal networks. Adv. neural Inf. Process. Syst. 28, 1. doi:10.
et al. (2014). “Learning phrase representations using RNN encoder-decoder for 1109/TPAMI.2016.2577031
statistical machine translation,”. arXiv preprint arXiv:1406.1078.
Ronneberger, O., Fischer, P., and Brox, T. (2015). “U-net: Convolutional networks for
Das, V., Pollack, A., Wollner, U., and Mukerji, T. (2019). Convolutional neural biomedical image segmentation,” in International Conference on Medical image
network for seismic impedance inversion. Geophysics 84 (6), R869–R880. doi:10.1190/ computing and computer-assisted intervention, Germany, October 8-12, 2023.
geo2018-0838.1
Treitel, S., and Lines, L. (2001). Past, present, and future of geophysical inversion—a
Du, J., Liu, J., Zhang, G., Han, L., and Li, N. (2019). “Pre-stack seismic inversion using new millennium analysis. Geophysics 66 (1), 21–24. doi:10.1190/1.1444898
SeisInv-ResNet,” in SEG technical program expanded abstracts 2019 (United States:
Vinyals, O., Toshev, A., Bengio, S., and Erhan, D. (2022). “"Show and tell: A neural
Society of Exploration Geophysicists).
image caption generator,” in Proceedings of the IEEE conference on computer vision
Graves, A., Mohamed, A.-r., and Hinton, G. (2013). “Speech recognition with deep and pattern recognition, San Juan, PR, USA, 17-19 June 1997.
recurrent neural networks,” in IEEE international conference on acoustics, speech and
Wang, Y., Ge, Q., Lu, W., and Yan, X. (2020). Well-logging constrained seismic
signal processing, Singapore, 22-27 May 2022.
inversion based on closed-loop convolutional neural network. IEEE Trans. Geoscience
Ioffe, S., and Szegedy, C. (2022). “Batch normalization: Accelerating deep network Remote Sens. 58 (8), 5564–5574. doi:10.1109/tgrs.2020.2967344
training by reducing internal covariate shift,” in International conference on machine
Wu, B., Meng, D., Wang, L., Liu, N., and Wang, Y. (2020). Seismic impedance
learning: PMLR, USA, 25-27 July 2022, 448–456.
inversion using fully convolutional residual network and transfer learning. IEEE
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2017). Imagenet classification with Geoscience Remote Sens. Lett. 17 (12), 2140–2144. doi:10.1109/lgrs.2019.2963106
deep convolutional neural networks. Commun. ACM 60 (6), 84–90. doi:10.1145/
Wu, X., Yan, S., Bi, Z., Zhang, S., and Si, H. (2021). Deep learning for
3065386
multidimensional seismi impedance inversion. Geophysics 86 (5), R735–R745.
Martin, G. S., Larsen, S., and Marfurt, K. (2002). Marmousi-2: An updated model for doi:10.1190/geo2020-0564.1
the investigation of AVO in structurally complex areas. Salt Lake City: SEG Annual
Wu, Y., and He, K. (2012). “Group normalization,” in Proceedings of the European
Meeting: OnePetro.
conference on computer vision (ECCV)), Florence Italy, October 7 - 13, 3–19.,

Frontiers in Earth Science 13 frontiersin.org