Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

GEOPHYSICS, VOL. 85, NO. 4 (JULY-AUGUST 2020); P. O47–O58, 19 FIGS., 2 TABLES.

10.1190/GEO2019-0627.1
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Seismic facies classification using supervised convolutional neural


networks and semisupervised generative adversarial networks

Mingliang Liu1, Michael Jervis2, Weichang Li3, and Philippe Nivlet2

ABSTRACT The proposed models can predict 3D facies distribution based


on actual well log data and core analysis, or other prior geologic
Mapping of seismic and lithologic facies from 3D reflection knowledge. Therefore, they provide a more consistent and
seismic data plays a key role in depositional environment analy- meaningful implication to seismic interpretation than commonly
sis and reservoir characterization during hydrocarbon explora- used unsupervised approaches. The two deep neural networks
tion and development. Although a variety of machine-learning have been tested successfully on a realistic synthetic case based
methods have been developed to speed up interpretation and im- on an existing reservoir and a real case study of the F3 seismic
prove prediction accuracy, there still exist significant challenges data from the Dutch sector of the North Sea. The prediction
in 3D multiclass seismic facies classification in practice. Some results show that, with relatively abundant well data, the super-
of these limitations include complex data representation, limited vised CNN-based learning method has a good ability in feature
training data with labels, imbalanced facies class distribution, learning from seismic data and accurately recovering the 3D fa-
and lack of rigorous performance evaluation metrics. To over- cies model, whereas the semisupervised GAN is effective in
come these challenges, we have developed a supervised convo- avoiding overfitting in the case of extremely limited well data.
lutional neural network (CNN) and a semisupervised generative The latter seems, therefore, particularly adapted to exploration
adversarial network (GAN) for 3D seismic facies classification or early field development stages in which labeled data from
in situations with sufficient and limited well data, respectively. wells are still very scarce.

INTRODUCTION process. Examples of approaches include K-means, the support vec-


tor machine (Zhao et al., 2014; Wrona et al., 2018), the random
Seismic facies analysis aims to interpret the depositional environ- forest (Ao et al., 2019), self-organizing maps (Strecker and Uden,
ment and facies distribution from reflection seismic data (Dumay 2002; de Matos et al., 2006; Saraswat and Sen, 2012; Zhao et al.,
and Fournier, 1988), which is a fundamental and important step in 2017), generative topographic mapping (Roy et al., 2014), indepen-
hydrocarbon exploration and development. In conventional work- dent component analysis (Lubo-Robles and Marfurt, 2019), and ar-
flows, facies are interpreted manually via collaboration among geo- tificial neural networks (West et al., 2002; Saggaf et al., 2003) based
physicists, geologists, and petrophysicists. However, with 3D seismic on different seismic attributes. One common issue of those ap-
volumes drastically increasing in size, such manual interpretation be- proaches is that the choice of the particular input seismic attributes
comes extremely time-consuming. Moreover, the results are typically has a significant impact on the classification results. This requires
subjective and heavily rely on the experience of skilled interpreters the interpreter to choose the most optimal seismic attributes and
(Bond et al., 2007). analysis window size in practice. However, the optimal attributes
To overcome these limitations, great efforts have been made in can vary from one data set to the next depending on the local
computer-assisted seismic interpretation, aiming to automate this geologic environment and data quality.

Manuscript received by the Editor 22 September 2019; revised manuscript received 23 March 2020; published ahead of production 2 June 2020; published
online 10 June 2020.
1
Aramco Research Center — Aramco Services Company, Houston, Texas, USA and University of Wyoming, Department of Geology and Geophysics,
Laramie, Wyoming, USA. E-mail: mliu4@uwyo.edu (corresponding author).
2
EXPEC ARC, Saudi Aramco, Dhahran, Saudi Arabia. E-mail: albajervis@yahoo.com; philippe.nivlet@aramco.com.
3
Aramco Research Center — Aramco Services Company, Houston, Texas, USA. E-mail: weichang.li@aramcoservices.com.
© 2020 Society of Exploration Geophysicists. All rights reserved.

O47
O48 Liu et al.

Recently, deep learning has become an emerging subfield of challenges, we use a realistic 3D facies model and use only the in-
machine learning, benefitting from the advancement in computing formation at well locations (lithofacies upscaled from well logs and
resources, especially graphics processing units (GPUs), and access derived from core analysis) as labels to perform facies classification
to large amounts of data. This has led to a series of major break- from seismic reflection data with deep neural networks. Because the
throughs, especially in computer vision and speech recognition. The amount of well data can vary from relatively abundant in older fields
to extremely rare in new prospects, we propose two types of deep-
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

main advantage of deep learning over traditional machine-learning


methods, such as the support vector machine, random forest, or any learning frameworks: (1) a conventional 3D CNN (LeCun et al.,
form of shallow artificial neural network, is its powerful ability for 1989) which relies on a substantial quantity of labeled data and
learning features and hierarchical representations from large sample (2) a state-of-the-art semisupervised GAN (Salimans et al., 2016)
data sets in high-dimensional space and for handling arbitrary which requires only limited well data to be incorporated in the net-
nonlinear complexity. It allows automatic extraction of the salient work training process. Both approaches use convolutional layers
feature representation that is most sensitive for specific tasks of that take seismic subsections or subcubes as input instead of the
interest, instead of hand-crafted attributes. For this reason, deep conventional trace-by-trace classification. As a result, the predicted
learning has been applied to numerous geophysical applications. facies models are more consistent with the well and the seismic re-
For instance, Li (2018) uses a stacked autoencoder and the convolu- flection data than the conventional methods.
tional neural network (CNN) for automated detection and classifi-
cation of multiclass geologic structure elements. The CNNs also
METHODOLOGY
have been used to detect faults from seismic data (Huang et al.,
2017; Di et al., 2018a; Guitton, 2018; Xiong et al., 2018; Wu et al., Supervised CNN
2019a, 2019b), as well as applied to delineate salt bodies (Walde-
land et al., 2018; Di et al., 2018b; Shi et al., 2019). Qian et al. (2018) The CNN is a specialized class of deep neural networks that uses
apply a deep convolutional autoencoder in unsupervised seismic convolutional layers to filter input data with a grid-like topology, such
facies classification; Dramsch and Lüthje (2018) and Zhao (2018) as natural images and time series data, for useful information extrac-
tion. Because AlexNet won the ImageNet Large Scale Visual Rec-
investigate and compare different CNNs in supervised seismic
ognition Challenge with remarkable accuracy in 2012 (Krizhevsky
facies classification. Araya-Polo et al. (2018) apply the CNN to
et al., 2012), CNNs have become dominant and have achieved
estimate velocity models. Alfarraj and AlRegib (2019) introduce
state-of-the-art performance in various computer vision and speech
recurrent neural networks to estimate petrophysical properties from
recognition tasks. The basic idea behind CNNs is inspired by the
poststack seismic data.
structure of the human visual system to automatically and adaptively
Most machine- and deep-learning methods used in geophysics
learn spatial hierarchies of features from low- to high-level patterns:
applications can be categorized into unsupervised clustering and
the first convolutional layers aim to extract low-level meaningful
supervised classification. In the unsupervised setting, there is no
features (i.e., edges, corners, textures, and lines) from which the sub-
prior labeled data, which means that only the input variables are
sequent convolution layers are able to compose increasingly abstract
given, but with no corresponding output variables. The goal is to
and complex features (Zeiler and Fergus, 2014). It is the mechanism
model the underlying structure or distribution in the data space.
of hierarchical feature learning that enables CNNs to capture the most
However, the main drawbacks are that it is usually difficult to de- salient and sensitive feature representation for a specific task, such as
termine the number of clusters and, more critically, because clusters image classification and segmentation.
are deemed to not have any well-defined geologic meaning, the re- A CNN is different from a regular neural network, such as the
sults of unsupervised methods would require geologic interpretation multilayer perceptron, in that it uses multiple convolutional layers
or calibration. As an alternative, supervised methods consider la- instead of fully connected layers to filter input data. This has sig-
beled data directly from well data or prior partial seismic strati- nificant advantages in terms of computing efficiency and feature
graphic interpretation. Results obtained by such an approach representation. Neuron units in a fully connected layer typically in-
then are interpretable automatically, whether it is in terms of rock teract with all activations in the previous layer. It would become
type, lithology, fluid type or stratigraphic environment. However, computationally prohibitive and memory demanding for high-di-
supervised methods require sufficient labeled training samples mensional data. For example, to handle images at the size of 100 ×
(in number and/or in diversity). In new exploration areas where only 100 neuron units, the number of trainable parameters in a hidden
limited well data are available to generate labels, supervised learn- layer with 100 neuron units can be as high as 1,000,000. Instead, a
ing methods may perform poorly in predicting properties away from convolutional layer takes advantage of the fact that one pixel is
well locations. To overcome these types of challenges, semisuper- more related to its neighbors than faraway locations by connecting
vised learning is highly desirable because it makes use of limited each neuron to only a local region of the input volume. To do that,
labeled samples and the large number of unlabeled samples. To neuron units in the convolutional layer are organized as a group of
date, semisupervised learning methods have been ignored to a large filters and a parameter sharing scheme is used to reduce the number
extent in geophysics (Alaudah et al., 2018). of parameters and make the computation more efficient. Parameter
Although deep learning has become increasingly popular in geo- sharing in CNNs basically means that the weight parameters of the
physics, significant challenges remain for 3D facies classification, filters are constant when performed on different parts of the input
including data representation associated network complexity, image. This also makes the extracted feature maps general and
limited labeled data for training, imbalanced facies class distribu- meaningful, which is helpful in generalizing the network to other
tion, and lack of rigorous performance evaluation in realistic set- tasks via transfer learning.
tings. In addition, facies class distribution in the training set is Because seismic volumes can be regarded as a particular type of
not necessarily consistent with realistic settings. To address these image, in this study, we attempt to investigate the feasibility and
Seismic facies analysis with CNN and GAN O49

performance of CNNs in seismic facies interpretation. From the overfitting. Overfitting is an undesirable behavior in which the
available drilled wells, seismic samples at well locations can be as- trained model achieves a perfect fit on the training data but does
signed with reliable facies labels derived from well logs and core not generalize well on new test data. However, it is often expensive
analysis. With sufficient samples with labels, seismic facies inter- or impractical to obtain a large number of labeled samples for train-
pretation can be formulated as a supervised classification problem. ing in seismic interpretation. In new prospects, drilled wells are
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Generally, CNNs use a set of convolutional layers to automatically rather limited. Additionally, due to class imbalance, some of the
extract the hierarchical feature maps of the input images. Each con- underrepresented facies could involve very few numbers of sam-
volutional layer consists of a stack of learnable filters that detect ples. Fortunately, 3D seismic volumes, however, provide millions
specific features present in the input data. Each convolution output to billions of unlabeled samples. It is known that the use of unla-
is passed through a typically nonlinear activation function to intro- beled data together with a small number of labeled samples can pro-
duce nonlinearity into the model. To reduce the dimension of ex- duce considerable improvement in learning accuracy by using
tracted feature maps and speed up the computation, a pooling layer, semisupervised learning. In this work, we propose a semisupervised
such as max pooling, is periodically inserted between successive method based on the generative adversarial network (GAN) (Sali-
convolutional layers. Pooling is a subsampling procedure that repla- mans et al., 2016) to deal with the challenging case of few drilled
ces the values over a certain area with their summary statistics, wells and insufficient labeled samples.
which is also helpful in making the feature representation approx- Semisupervised learning is a class of techniques that makes use
imately invariant to small translations and scale changes (Goodfel- of labeled and unlabeled data for training, typically a small amount
low et al., 2016). Then, the feature maps extracted from the of labeled data with a large amount of unlabeled data. Figure 1 il-
intermediate convolutional layer are fed into the subsequent fully lustrates how unlabeled data can help in classification. The seismic
connected layers to predict the probability of each class. input data or attributes are represented in a multivariate space, but
Training a neural network is an optimization problem, which is for the sake of simplicity, only two dimensions are shown. If we
equivalent to minimizing a predefined loss function. The loss func- only consider the labeled data shown in Figure 1a, the optimal de-
tion measures how good a model is and guides how to update the cision boundary could be a straight line (or a hyperplane) and then
model parameters during training, so it is critical to select proper the sample shown in gray would be classified as channel facies.
loss functions for different tasks. Specifically, seismic facies clas- However, if in addition to the labeled samples we also consider
sification is a typical multiclass classification problem in which the all of the unlabeled samples (the gray circles shown in Figure 1b),
multiclass cross-entropy loss (equation 1) commonly is used: it is obvious that there exists some structure to the underlying dis-
tribution of those unlabeled data, which would help us update the
1X N X K
decision boundary as shown in Figure 1c. Then, the gray sample
L¼− Iðk; yi Þ log Pðyi ¼ kjxi Þ; (1) shown in Figure 1a would be included in the nonchannel class.
N i¼1 k¼1
The semisupervised GAN is a variant of the original GAN, which
where K is the number of classes; N is the number of training sam- is specialized to deal with learning problems with very few labeled
ples; xi and yi represent the input (seismic patch or subcube) and data. The GANs are a class of generative models devised by Good-
label (facies type) of training sample i, respectively; Iðk; yi Þ is the fellow et al. (2016), and they have since become widely known for
indicator function that is equal to 1 if and only if sample i belongs to their application versatility and their often remarkable ability in gen-
class k; and Pðyi ¼ kjxi Þ is the output probability of the neural net- erating realistic data. Unlike traditional neural networks, GANs
work that sample i belongs to class k calculated from the output comprise two competing players each represented by a neural net-
layer through the softmax activation function. work: the generator and the discriminator. The generator tries to
create samples that come from a targeted, yet unknown, probability
distribution (for instance, data points from Gaussian distribution
Semisupervised generative adversarial networks
and channelized reservoir realizations) using random noise in a low-
Supervised CNNs typically have millions of trainable parameters dimensional space as the input, whereas the discriminator acts
to fit high-dimensional and nonlinear models. As a result, a tremen- like a judge to distinguish those samples produced by the generator
dous amount of labeled training samples is required to avoid from the real samples. In other words, the training objective of the

Figure 1. Illustration of semisupervised learning: (a) a possible decision boundary (dashed red line) using only one channel sample (white
circle) and one nonchannel sample (black circle), (b) the data distribution of the labeled and unlabeled samples (gray circles), and (c) the
updated decision boundary (dashed red curve) using the labeled and unlabeled samples.
O50 Liu et al.

generative network is to increase the error rate of the discriminative discarded after training. To make it applicable for semisupervised
network by producing samples that are as similar as possible to the learning, Salimans et al. (2016) transform the discriminator into a
real data, whereas it is exactly the opposite for the discriminative multiclass classifier that has dual roles in the label prediction as well
network that aims to minimize the same error rate. When trained in as in the discrimination of samples generated by the generator from
an alternative manner, the two networks would guide each other to the real data. To be more specific, the discriminator takes an input
sample and classifies it into K þ 1 classes, where K is the number of
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

become more skilled in their respective jobs. Eventually, the gen-


erator determines the underlying probability distribution of the data classes of the classification task: the first K dimensions represent
and starts generating samples with realistic appearances. Figure 2 the classes to be classified, and the last dimension indicates whether
depicts the general architecture of a GAN. the input sample is real or fake. For example, in the case with four
The generator and the discriminator can be expressed by two dif- seismic facies for classification, we need to convert the discrimina-
ferentiable functions Gðz; θg Þ and Dðx; θd Þ, respectively. The term tor into a five-class classifier: The first four classes are for the indi-
Gðz; θg Þ maps the input noise variables z into the desired data space vidual class probabilities of facies, and the last class is an indicator
x; conversely, Dðx; θd Þ outputs the probability that the data come of whether the input sample is generated from the generator or from
from the real sample data set. Here, θg and θd represent the model real seismic data.
parameters that define each neural network, respectively. Training The loss function of semisupervised GAN is defined in such a
of GANs is equivalent to playing a minimax game with value func- way that the discriminator can help the generator learn to produce
tion VðD; GÞ: realistic images and, by distinguishing between real and fake sam-
ples, the discriminator uses the fake data generated from the gen-
min max VðD; GÞ ¼ Ex∼pdata ðxÞ ½log DðxÞ erator along with the labeled and unlabeled real data to help classify
G D
the data set. Therefore, the overall loss function in the discriminator
þ Ez∼pz ðzÞ ½logð1 − DðGðzÞÞÞ: (2) has three terms associated with (1) real samples with labels, (2) real
samples without labels, and (3) fake samples from the generator, as
Because GANs are formulated as a game between two competing described in the following equation:
networks, it is important to keep them in balance during training. If
either the generator or the discriminator is much better than the LD ¼ Lsupervised þ Lunsupervised ; (3)
other, the training would not converge. For this reason, GANs
are generally more difficult to train than a regular deep neural net- where
work. Radford et al. (2015) introduce deep convolutional generative
adversarial networks to help bridge the gap between the success of Lsupervised ¼ −Ex;y∼pdata ðx;yÞ ½log pmodel ðyjx; y < K þ 1Þ; (4)
CNNs for supervised learning and the challenge in unsupervised
learning, and they introduce a combination of techniques as the
key to stabilize the training of GANs and to deal with poor weight
initialization problems, including (1) using CNNs to build the gen- Lunsupervised ¼ −fEx∼pdata ðxÞ ½logð1 − pmodel ðy ¼ K þ 1jxÞÞ
erator and the discriminator, (2) batch normalization (Ioffe and
Szegedy, 2015), and (3) convolutions with stride instead of pooling þ Ex∼G ½log pmodel ðy ¼ K þ 1jxÞg: (5)
layers.
The original design of GANs is for data generation rather than The loss function of the discriminator includes the unsupervised
classification, for example, for image and signal synthesis, in which loss that is the same as the loss function defined in equation 2 and an
the discriminator is used only to help training the generator and then additional supervised loss term that measures the predicted error
rate on real samples with labels. In the training
stage, the discriminator will predict the probabil-
ities of the K þ 1 classes for all of the input sam-
ples including the labeled and unlabeled samples
from the training set and samples generated from
the generator. When a real sample from the train-
ing set is fed regardless of being labeled or un-
labeled, it will be assigned a high probability of
real facies classes and a low probability of fake
classes, whereas it is the opposite for samples
generated from the generator. In the prediction
stage, we only keep the predicted probabilities
of the first K classes and drop the last dimension.
The combination of these losses associated with
different sources of data will enable the classifier
to learn from a broader perspective. That, in turn,
allows the model to perform inference much
Figure 2. Schematic architecture of the original GAN consisting of the generator, which
tries to create samples that come from a targeted yet unknown probability distribution, more precisely than only using very few labeled
and the discriminator to distinguish those samples produced by the generator from the examples. Although we use the same weights for
real samples. the supervised and unsupervised loss in this
Seismic facies analysis with CNN and GAN O51

study, we can tune the hyperparameters to control the trade-off be- Synthetic data with CNN
tween labeled and unlabeled data.
In addition, to avoid the discriminator overconfidence in which The first example is a synthetic case of 3D lithofacies prediction
the network learns very limited features for classification and leads using supervised CNNs. The data set is generated from a realistic
to bad performance in prediction, it is desirable to add regularization synthetic model that is based on an existing on-shore field. The
3D facies model has been built by geostatistics using lithofacies
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

to penalize the discriminator when the predicted probability for any


real sample goes beyond a threshold (e.g., 0.9). This is called one- from actual wells and the structural framework that has been derived
sided label smoothing (Salimans et al., 2016). from an existing reservoir model covering an area of approximately
For the generator, the loss function is supposed to use feature 620 km2 . The reservoir model consists of five units, in which the
matching that minimizes the statistical difference between the fea- three units of most interest are placed in the layers with constant
tures of the real and fake samples from the generator instead of the elastic properties (referred to as BG1). As shown in Figure 3,
difference in sample data space (e.g., between pixels in case of im- the middle unit consists of channel levee systems simulated by
age inputs). This is useful to stabilize training of the networks and the unconstrained object-based simulation technique in a constant
make the classifier have good generalization capability. Typically, background (flood plains), whereas facies in the upper and lower
we measure the L2 distance between the mean feature vectors: units are simulated by the unconstrained stationary sequential indi-
cator simulation technique. The detailed simulation workflow is
 2 given by Pyrcz and Deutsch (2014). The fractions and P-impedance
 
LG ¼ Ex∼pdata ðxÞ ½fðxÞ − Ez∼pz ðzÞ ½fðGðzÞÞ

; (6) values of facies in each unit are listed in Table 1. Because the P
2 impedance of the levee facies is identical to one facies (BG5) in
the top unit, we use the same label for them in the classification.
where fð·Þ is the feature vector extracted in an immediate layer by
the discriminator.
Table 1. Fractions and P-impedance values of facies in the top,
middle, and bottom units.
APPLICATIONS
Facies P impedance
To demonstrate the applicability and performance of the pro- Geologic unit Facies name fraction (g/cm3 * m/s)
posed deep neural networks in 3D seismic facies classification, we
present three examples in this section, including (1) supervised Upper BG3 0.19 10,000
CNNs on a realistic synthetic model with sufficient well data, BG4 0.16 8200
(2) semisupervised GAN on the synthetic model with limited well BG5 0.65 9000
data, and (3) semisupervised GAN on the F3 seismic survey from
Middle Flood plain 0.85 8500
the Dutch sector of the North Sea.
Channel 0.09 9100
Levee 0.06 9000
Lower BG2 0.04 9300
BG3 0.11 10,000
BG4 0.81 8200

Figure 3. Synthetic example shows (a) the lithofacies model and


(b) the corresponding 3D seismic volume. Figure 4. Histogram of facies classes in the training set of CNN.
O52 Liu et al.

In total, there are seven target lithofacies to be classified. Then, the To mimic the scenario of a mature field, we randomly select 256
corresponding 3D poststack seismic data are generated by 1D con- (16 × 16) traces from the synthetic geologic model as
volution with a 25 Hz Ricker wavelet of the zero-offset reflectivity pseudowells and split them into two subsets: 200 wells as a train-
derived from the P-wave velocity and density information. ing set and 56 wells as a validation set. Next, we generate labeled
samples along such pseudowells by taking
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

the seismic samples around a specific point as


input and the associated facies label as output.
Figure 4 shows the facies class distribution.
As shown, the training set consists of approxi-
mately 15,000 labeled samples. Such a large
number of diverse labeled samples would allow
us to perform classification using supervised
CNNs and potentially avoid overfitting. To bet-
ter discover and exploit the spatial structures in
the seismic data, we adopt a 3D CNN model in
this case, which takes seismic subcubes as the
inputs and performs the convolution operations
in 3D.
As illustrated in Figure 5, the network archi-
Figure 5. Architecture of the proposed CNN models: The yellow blocks represent the
convolutional layers, the orange blocks represent the max-pooling layers, and the purple tecture is adapted from the VGGNet (Simonyan
blocks represent the fully connected layers. and Zisserman, 2014). We use three convolu-
tional blocks to extract feature maps from seis-
mic samples, which then are followed by three
fully connected layers and eventually a softmax
layer to predict the class probability. Each con-
volutional layer or fully connected layer is fol-
lowed by the ReLU activation function, which
enables the CNNs to handle nonlinearity. In the
network, the input size is 32 × 32 × 32, and the
filter size of all of the 3D convolutional layers is
(3, 3, 3). Because seismic facies are generally
more continuous horizontally than vertically, the
CNN model is sensitive to the pooling size along
the vertical axis. In this study, the pooling size
of the three max-pooling layers is set to (2, 2, 4),
(2, 2, 4), and (2, 2, 2), respectively.
Figure 6. Loss and accuracy on (a) the training set and (b) the validation set.
Figure 6 shows the training history using the
Adam optimizer (Kingma and Ba, 2014) with a
batch size of 256 and a learning rate of 0.001. Based on the training
history, we choose the ideal point to stop training at approximately
750 epochs to minimize the possibility of overfitting. The prediction
accuracy on the validation set is 0.82, and the F1 score and average
precision score are 0.81 and 0.89, respectively, which are more use-
ful to evaluate the model performance in the case of severely im-
balanced data sets. Figure 7 shows the normalized confusion matrix
from which we can determine the prediction accuracy and the
misclassification of each facies class. We can observe that the
3D CNN works well on most facies, but some channels are difficult
to detect. There are two distinct problems causing the misclassifi-
cations: (1) elastic parameters of levee facies are close to channels
and (2) channels are thin in general close to the resolution limits of
the seismic data.
Following training, we perform prediction on the entire seismic
volume using the trained model. Figure 8 shows several example
stratigraphic slices of the predicted facies along with the ground
truth. We can observe that all facies are predicted with high accu-
racy and most channels are discovered with a convincing accuracy
despite the lower prediction score of channels from the confusion
Figure 7. Confusion matrix of the prediction on the validation set. matrix.
Seismic facies analysis with CNN and GAN O53

One powerful aspect of deep learning is the


ability to map the input data into a learned feature
space where samples belonging to different
classes can be better separated. This effect can
be observed in Figure 9 which shows the plotting
of the T-distributed stochastic neighbor embed-
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

ding (t-SNE; van der Maaten and Hinton, 2008)


visualization of the original seismic data and the
features extracted from the last convolutional
layer in the 3D CNN. As shown, it is evident that
the samples of different lithofacies, initially over-
lapping in the seismic data space, become sepa-
rable groups in the feature space generated in the
trained CNN model.

Synthetic data with GAN


The supervised CNN models perform
adequately in the case with abundant well data.
Figure 8. Stratigraphic slices extracted from the upper (a and d), middle (b and e), and
But the performance will deteriorate quickly and bottom (c and f) units: (a-c) the true model and (d-f) results from 3D CNN.
eventually fail in cases with few drilled wells
where the number of training data samples drops,
and the model becomes overfitted. This greatly
limits their application in practice, especially in
new exploration areas. In this example, we use
the same synthetic model as used in example
1 to demonstrate the feasibility of the semisuper-
vised GAN in seismic facies classification with
very few labeled data. In contrast to example
1, we only select 16 (4 × 4) vertical traces from
the model as pseudowells along which the input
samples are generated by stacking three orthogo-
nal seismic patches with size of 32 × 32 along
the inline, crossline, and time axis similar to a
multichannel image with size of 32 × 32 × 3.
Here, we use all pseudowells to generate samples
and then split the obtained samples into training
set (80%) and validation set (20%) instead of
splitting by wells. The sample distribution is Figure 9. A t-SNE visualization of (a) the original seismic data and (b) the extracted
features by the 3D CNN.
shown in Figure 10. In this case, only approxi-
mately 1000 samples with labels are available
and, more critically, some of the underrepresented facies such as
channels have very few samples (e.g., channel facies = 28 samples)
emphasizing the class imbalance.
Figure 11 shows the network architecture of the semisupervised
GAN used in this study. The generator and the discriminator are
built as 2D CNNs. The generator takes a batch of Gaussian noises
with dimension of 128 as input to synthesize data with the same size
as the training samples (32 × 32 × 3) by one fully connected layer
and three convolutional layers. Unlike the pooling layer in CNNs
for classification, each convolutional layer in the generator is fol-
lowed by an upsampling layer to increase the dimension of feature
maps. To adapt the seismic classification task with seven facies
classes, the discriminator is designed as an eight-class classifier,
which takes the real sample from the training set and the fake sam-
ples from the generator as input. It consists of seven convolutional
layers and two fully connected layers. First, we train the generator
and the discriminator networks on the labeled samples extracted
along pseudowells and unlabeled samples evenly extracted along
the inline, crossline, and time direction from the seismic data. In Figure 10. Histogram of facies classes in the training set of GAN.
O54 Liu et al.

the inference phase, the generator is discarded, and the discrimina- For a fair comparison, we also train a supervised 2D CNN on the
tor is used for the final prediction. Figure 12 shows the accuracy same training set. Given that both networks have difficulty in dis-
curves of the discriminator during training and validation. We criminating channels from levees, potentially due to the limited res-
choose the ideal point to stop training at approximately 20 epochs olution of the seismic data, the similarity of facies in terms of elastic
to minimize the possibility of overfitting. properties, and the insufficient number of total samples (labeled or
unlabeled) of both facies to adequately learn their
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

features, we merge the levees to channel facies in


postprocessing so as to better interpret the results.
The normalized confusion matrices for the semi-
supervised GAN and the CNN model evaluated
on the entire volume are provided in Figure 13.
We can observe that the CNN tends to misclassify
channels and levees as background and floodplain
facies, which are both dominant facies in this case,
an indication of overfitting. The false-negative
channel/levee predictions are from overrepre-
sented facies classes. In comparison, the semisu-
pervised GAN has significantly more true
positives (diagonal values). Figure 14 compares
the predictions from the networks versus the true
Figure 11. Architecture of the proposed semisupervised GAN in which the discrimi- facies. As shown in the 3D views, the semisuper-
nator is extended to an eight-class classifier. vised GAN recovers most facies even with very
limited well data, whereas the CNN performs very
poorly. A close-up view of the central unit containing the channel
facies, as highlighted by the arrows in Figure 15, shows that most
channels are predicted with sufficient accuracy by the semisupervised
GAN and are not well estimated by the supervised CNN.

Field data
Next, we apply the semisupervised GAN model to the F3 block
seismic data made publicly available by the Dutch government
through TNO and dGB Earth Sciences. Inline 339 has been inter-
preted manually with label data freely accessed from https://github.-
com/bolgebrygg/MalenoV. Unlike the definition of facies types in
the synthetic example based on lithofacies information from well
log and core analysis, in this case, nine different seismic facies are
defined according to their seismic reflection characteristics: eight
facies with distinct reflection patterns and one facies used to re-
present samples not belonging to the eight target facies (see Table 2).
Because the number and patterns of the seismic facies (i.e., width
and torsion of channels) vary with the seismic data, it is challenging
Figure 12. Training and validation accuracy. to directly transfer the trained network on the synthetic data to the
real field data. For this reason, in this case, we
reuse the network architecture with minor
changes but retrain it from scratch.
The neural network used in this study is sim-
ilar to the model in example 2 except (1) we
adopt 3D convolutional and deconvolutional
layers to build the generator and the discrimina-
tor and (2) we use a larger sample size of
64 × 64 × 64. We also alter the dimension of out-
put of the discriminator to 10 accordingly. To
generate the training set, we extract 20,000 sam-
ples from inline 339 that have been interpreted
previously (Figure 16) and from which we ran-
domly select 50 samples per facies class as la-
beled data with the remaining as unlabeled
Figure 13. Confusion matrix of the prediction on the validation set: (a) semisupervised data. For the latter, the label information is
GAN and (b) supervised CNN. masked only during training, and it is put back
Seismic facies analysis with CNN and GAN O55

and used during validation for model evaluation. After training of DISCUSSION
the networks, the generator is discarded and the discriminator is
used to perform facies prediction on the entire seismic volume. Fig- Although the proposed supervised CNN and semisupervised
ure 17a and 17b shows the classification results on inline 339 and GAN for facies classification work well on the synthetic and real
xline 775. It is evident that the predicted target facies distribution is
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

very consistent with the seismic reflection characteristics. Table 2. Seismic facies defined in the F3 seismic data.
To evaluate the performance of the semisupervised GAN, we com-
pare its results to the supervised 3D CNN by using 10 and 50 labeled
samples per facies class, respectively. Figure 18 shows the t-SNE Facies number Seismic reflection characteristic
visualization of seismic samples in the validation set before being
transformed by deep neural networks, and Figure 19 shows the t- 1 Low coherency
SNE visualization of feature maps extracted from the last convolution 2 High-amplitude continuous
layers in 3D CNN and the discriminator of the GAN. We can see that 3 Low-amplitude dips
the validation samples become separable in the transformed feature 4 High-amplitude dips
space of CNN and GAN. Compared to the supervised CNN, the 5 Chaotic
semisupervised GAN can further separate the facies groups with
6 Low amplitude
the help of unlabeled data. The prediction accuracies of the semisu-
pervised GAN on the validation set with 10 and 50 labeled samples 7 High amplitude
per facies class are 85.8% and 97.5%, respectively, compared with 8 Salt
accuracies of 70.2% and 80.4% for the supervised CNN. 9 Everything else

Figure 14. The 3D view of the true and predicted


lithofacies model: (a) true model, (b) predicted
model from the supervised CNN, and (c) predicted
model from the semisupervised GAN.

Figure 15. Stratigraphic slices (a) 30 ms,


(b) 60 ms, (c) 90 ms, and (d) 120 ms above the
bottom of the channel reservoir; columns from
left to right: the true model, predicted model
from the semisupervised GAN, and predicted
model from the supervised CNN, respectively.
O56 Liu et al.

seismic data sets, running inference on the entire 3D volume is very is to adopt lightweight networks such as MobileNet (Howard et al.,
time consuming, which would severely restrict the applications 2017), MobileNetV2 (Sandler et al., 2018), ShuffleNet (Zhang et al.,
when GPUs are not available. To solve this limitation, some at- 2018), and ShuffleNetV2 (Ma et al., 2018). In the future, we will com-
tempts have been made to convert facies classification from seismic prehensively investigate the performance of the lightweight networks
to an image segmentation problem using deep neural networks such and compare them with the standard CNNs in terms of accuracy and
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

as UNet (Ronneberger et al., 2015; Di et al., 2019). Instead of point- computing time.
wise classification, those segmentation networks take a 2D seismic As discussed in the synthetic case, generally it is challenging to
section or 3D subvolume as input and are able to predict the facies discriminate the lithofacies that have similar P impedances, such as
at each pixel simultaneously. However, that usually requires com- the channel and levee, from poststack seismic data. To further im-
pletely interpreted seismic sections or subvolumes to train the segmen- prove the prediction accuracy, it would be helpful to train the net-
tation networks, which makes them not applicable for identification of works on prestack seismic data that contain more information about
rock and fluid types in which the label information is only available at rock and fluid types. Additionally, in real applications, the seismic
well locations. An alternative solution to speed up the computing time data are contaminated by noise, the lithofacies labels are derived

Figure 16. Seismic data and prior seismic facies


interpretation of inline 339.

Figure 17. Predicted seismic facies include (a) in-


line 339 and (b) xline 775.
Seismic facies analysis with CNN and GAN O57

from well logs, and core analysis usually includes uncertainty. More
effort is necessary to address these challenges.

CONCLUSION
We have proposed two types of deep neural network frameworks
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

for 3D seismic facies classification calibrated from well log data.


Fully supervised CNNs are recommended for mature fields with
relatively abundant well data, and semisupervised GAN work well
for new prospects with limited wells. The CNNs are powerful and
efficient for feature representation using seismic data and, thus, pro-
vide more accurate seismic facies characterization. Semisupervised
GAN can be extremely helpful to avoid model overfitting by best
utilizing labeled and unlabeled samples. Both types of models have
been tested successfully on a realistic synthetic data set and a field
seismic volume, indicating significant potential of these deep-learn-
ing methods in conventional quantitative seismic interpretation
workflows. They can provide a robust method for combining well
data and 3D seismic data to automate and improve seismic facies
interpretation.

Figure 18. A t-SNE visualization of the original seismic data. ACKNOWLEDGMENTS


The authors would like to thank dGB Earth Sciences for provid-
ing the F3 North Sea seismic data to the public and ConocoPhillips
for sharing the MalenoV project for public use.

DATA AND MATERIALS


AVAILABILITY
Data associated with this research are confi-
dential and cannot be released.

REFERENCES
Alaudah, Y., S. Gao, and G. AlRegib, 2018, Learning
to label seismic structures with deconvolution net-
works and weak labels: 88th Annual International
Meeting, SEG, Expanded Abstracts, 2121–2125,
doi: 10.1190/segam2018-2997865.1.
Alfarraj, M., and G. AlRegib, 2019, Semisupervised
sequence modeling for elastic impedance inversion:
Interpretation, 7, SE237–SE249, doi: 10.1190/INT-
2018-0250.1.
Ao, Y., H. Li, L. Zhu, S. Ali, and Z. Yang, 2019, Iden-
tifying channel sand-body from multiple seismic
attributes with an improved random forest algorithm:
Journal of Petroleum Science and Engineering, 173,
781–792, doi: 10.1016/j.petrol.2018.10.048.
Araya-Polo, M., J. Jennings, A. Adler, and T. Dahlke,
2018, Deep-learning tomography: The Leading
Edge, 37, 58–66, doi: 10.1190/tle37010058.1.
Bond, C. E., A. D. Gibbs, Z. K. Shipton, and S. Jones,
2007, What do you think this is? “Conceptual uncer-
tainty” in geoscience interpretation: GSA Today, 17,
4–10, doi: 10.1130/GSAT01711A.1.
de Matos, M. C., P. L. Osorio, and P. R. Johann, 2006,
Unsupervised seismic facies analysis using wavelet
transform and self-organizing maps: Geophysics, 72,
no. 1, P9–P21, doi: 10.1190/1.2392789.
Di, H., D. Gao, and G. AlRegib, 2019, Developing a seis-
mic texture analysis neural network for machine-aided
seismic pattern recognition and classification: Geo-
physical Journal International, 218, 1262–1275, doi:
10.1093/gji/ggz226.
Di, H., Z. Wang, and G. AlRegib, 2018a, Seismic fault
Figure 19. A t-SNE visualization of the extracted features from (a) 3D CNN with 10 detection from post-stack amplitude by convolu-
labeled samples per class, (b) semisupervised 3D GAN with 10 labeled samples per tional neural networks: 80th Annual International
class, (c) 3D CNN with 50 labeled samples per class, and (d) semisupervised 3D Conference and Exhibition, EAGE, Extended Ab-
GAN with 50 labeled samples per class. stracts, 1–5, doi: 10.3997/2214-4609.201800733.
O58 Liu et al.

Di, H., Z. Wang, and G. AlRegib, 2018b, Deep convolutional neural net- Salimans, T., I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X.
works for seismic salt-body delineation: AAPG Annual Convention and Chen, 2016, Improved techniques for training GANs: 30th Annual
Exhibition. Conference on Neural Information Processing Systems, 2234–2242.
Dramsch, J. S., and M. Lüthje, 2018, Deep-learning seismic facies on state- Sandler, M., A. Howard, M. Zhu, A. Zhmoginov, and L. C. Chen, 2018,
of-the-art CNN architectures: 89th Annual International Meeting, SEG, Mobilenetv2: Inverted residuals and linear bottlenecks: Proceeding of
Expanded Abstracts, 2036–2040, doi: 10.1190/segam2018-2996783.1. the 31st IEEE Conference on Computer Vision and Pattern Recognition,
Dumay, J., and F. Fournier, 1988, Multivariate statistical analyses applied to 4510–4520.
seismic facies recognition: Geophysics, 53, 1151–1159, doi: 10.1190/1 Saraswat, P., and M. K. Sen, 2012, Artificial immune-based self-organizing
Downloaded 06/23/20 to 137.111.162.20. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

.1442554. maps for seismic facies analysis: Geophysics, 77, no. 4, O45–O53, doi: 10
Goodfellow, I., Y. Bengio, and A. Courville, 2016, Deep learning: MIT Press. .1190/geo2011-0203.1.
Guitton, A., 2018, 3D convolutional neural networks for fault interpretation: Shi, Y., X. Wu, and S. Fomel, 2019, SaltSeg: Automatic 3D salt segmenta-
80th Annual International Conference and Exhibition, EAGE, Extended tion using a deep convolutional neural network: Interpretation, 7, SE113–
Abstracts, 1–5. SE122, doi: 10.1190/INT-2018-0235.1.
Howard, A. G., M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, Simonyan, K., and A. Zisserman, 2014, Very deep convolutional networks
M. Andreetto, and H. Adam, 2017, Mobilenets: Efficient convolutional for large-scale image recognition: arXiv preprint, arXiv:1409.1556.
neural networks for mobile vision applications: arXiv preprint, ar- Strecker, U., and R. Uden, 2002, Data mining of 3D poststack seismic attrib-
Xiv:1704.04861. ute volumes using Kohonen self-organizing maps: The Leading Edge, 21,
Huang, L., X. Dong, and T. E. Clee, 2017, A scalable deep learning platform 1032–1037, doi: 10.1190/1.1518442.
for identifying geologic features from seismic attributes: The Leading van der Maaten, L. V. D., and G. Hinton, 2008, Visualizing data using t-
Edge, 36, 249–256, doi: 10.1190/tle36030249.1. SNE: Journal of Machine Learning Research, 9, 2579–2605.
Ioffe, S., and C. Szegedy, 2015, Batch normalization: Accelerating deep Waldeland, A. U., A. C. Jensen, L. J. Gelius, and A. H. S. Solberg, 2018,
network training by reducing internal covariate shift: arXiv preprint, ar- Convolutional neural networks for automated seismic interpretation: The
Xiv:1502.03167. Leading Edge, 37, 529–537, doi: 10.1190/tle37070529.1.
Kingma, D. P., and J. Ba, 2014, Adam: A method for stochastic optimiza- West, B. P., S. R. May, J. E. Eastwood, and C. Rossen, 2002, Interactive
tion: arXiv preprint, arXiv:1412.6980. seismic facies classification using textural attributes and neural networks:
Krizhevsky, A., I. Sutskever, and G. E. Hinton, 2012, Imagenet classification The Leading Edge, 21, 1042–1049, doi: 10.1190/1.1518444.
with deep convolutional neural networks: 26th Annual Conference on Wrona, T., I. Pan, R. L. Gawthorpe, and H. Fossen, 2018, Seismic facies
Neural Information Processing Systems, 1097–1105. analysis using machine learning: Geophysics, 83, no. 5, O83–O95,
LeCun, Y., B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, doi: 10.1190/geo2017-0595.1.
and L. D. Jackel, 1989, Backpropagation applied to handwritten zip code re- Wu, X., L. Liang, Y. Shi, and S. Fomel, 2019a, FaultSeg3D: Using synthetic
cognition: Neural Computation, 1, 541–551, doi: 10.1162/neco.1989.1.4.541. data sets to train an end-to-end convolutional neural network for 3D seis-
Li, W., 2018, Classifying geological structure elements from seismic images mic fault segmentation: Geophysics, 84, no. 3, IM35–IM45, doi: 10.1190/
using deep learning: 88th Annual International Meeting, SEG, Expanded geo2018-0646.1.
Abstracts, 4643–4648, doi: 10.1190/segam2018-2998036.1. Wu, X., Y. Shi, S. Fomel, L. Liang, Q. Zhang, and A. Yusifov, 2019b, Fault-
Lubo-Robles, D., and K. J. Marfurt, 2019, Independent component analysis Net3D: Predicting fault probabilities, strikes, and dips with a single con-
for reservoir geomorphology and unsupervised seismic facies classifica- volutional neural network: IEEE Transactions on Geoscience and Remote
tion in the Taranaki Basin, New Zealand: Interpretation, 7, SE19–SE42, Sensing, 57, 9138–9155, doi: 10.1109/TGRS.2019.2925003.
doi: 10.1190/INT-2018-0109.1. Xiong, W., X. Ji, Y. Ma, Y. Wang, N. M. AlBinHassan, M. N. Ali, and Y.
Ma, N., X. Zhang, H. T. Zheng, and J. Sun, 2018, ShuffleNet V2: Practical Luo, 2018, Seismic fault detection with convolutional neural network:
guidelines for efficient CNN architecture design: 17th European Conference Geophysics, 83, no. 5, O97–O103, doi: 10.1190/geo2017-0666.1.
on Computer Vision, 116–131. Zeiler, M. D., and R. Fergus, 2014, Visualizing and understanding convolu-
Pyrcz, M. J., and C. V. Deutsch, 2014, Geostatistical reservoir modeling: tional networks: 13rd European Conference on Computer Vision, 818–
Oxford University Press. 833.
Qian, F., M. Yin, X. Y. Liu, Y. J. Wang, C. Lu, and G. M. Hu, 2018, Un- Zhang, X., X. Zhou, M. Lin, and J. Sun, 2018, Shufflenet: An extremely
supervised seismic facies analysis via deep convolutional autoencoders: efficient convolutional neural network for mobile devices: Proceeding
Geophysics, 83, no. 3, A39–A43, doi: 10.1190/geo2017-0524.1. of the 31st IEEE Conference on Computer Vision and Pattern Recogni-
Radford, A., L. Metz, and S. Chintala, 2015, Unsupervised representation tion, 6848–6856.
learning with deep convolutional generative adversarial networks: arXiv Zhao, T., 2018, Seismic facies classification using different deep convolu-
preprint, arXiv:1511.06434. tional neural networks: 88th Annual International Meeting, SEG, Ex-
Ronneberger, O., P. Fischer, and T. Brox, 2015, U-net: Convolutional net- panded Abstracts, 2046–2050, doi: 10.1190/segam2018-2997085.1.
works for biomedical image segmentation: International Conference on Zhao, T., V. Jayaram, K. J. Marfurt, and H. Zhou, 2014, Lithofacies clas-
Medical Image Computing and Computer Assisted Intervention, 234–241. sification in Barnett Shale using proximal support vector machines: 84th
Roy, A., A. S. Romero-Peláez, T. J. Kwiatkowski, and K. J. Marfurt, 2014, Annual International Meeting, SEG, Expanded Abstracts, 1491–1495,
Generative topographic mapping for seismic facies estimation of a carbon- doi: 10.1190/segam2014-1210.1.
ate wash, Veracruz Basin, southern Mexico: Interpretation, 2, SA31–SA47, Zhao, T., F. Li, and K. J. Marfurt, 2017, Constraining self-organizing map
doi: 10.1190/INT-2013-0077.1. facies analysis with stratigraphy: An approach to increase the credibility
Saggaf, M. M., M. N. Toksöz, and M. I. Marhoon, 2003, Seismic in automatic seismic facies classification: Interpretation, 5, T163–T171,
facies classification and identification by competitive neural networks: doi: 10.1190/INT-2016-0132.1.
Geophysics, 68, 1984–1999, doi: 10.1190/1.1635052.

You might also like