Convolutional neural networks for

automated seismic interpretation

Anders U. Waldeland1, Are Charles Jensen1, Leiv-J. Gelius2, and Anne H. Schistad Solberg1
https://doi.org/10.1190/tle37070529.1.
Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Abstract CNNs as a black-box algorithm can easily result in overfitting or

Deep-learning methods have proved successful recently for make the algorithms not perform as well as expected.
solving problems in image analysis and natural language process- In this paper, we explain CNNs and how the different parts
ing. One of these methods, convolutional neural networks (CNNs), of CNNs work together to solve a given classification task. We
is revolutionizing the field of image analysis and pushing the state give examples with natural images to better explain the general
of the art. CNNs consist of layers of convolutions with trainable concepts. We then explain how CNNs are commonly trained in
filters. The input to the network is the raw image or seismic a robust and efficient manner. We will discuss how this method
amplitudes, removing the need for feature/attribute engineering. can be used in the context of seismic interpretation. As an example,
During the training phase, the filter coefficients are found by a CNN is applied to seismic data for salt body classification.
iterative optimization. The network thereby learns how to compute Finally, we discuss how this method can be applied to other
good attributes to solve the given classification task. However, problems in automated seismic interpretation and what limitations
CNNs require large amounts of training data and must be carefully can be expected when applying CNNs to seismic images.
designed and trained to perform well. We look into the intuition
behind this method and discuss considerations that must be made Convolutional neural networks
in order to make the method reliable. In particular, we discuss To explain how CNNs work, we will consider face detection
how deep learning can be used for automated seismic interpretation. as an example. In a CNN, classification problems are solved by
As an example, we show how a CNN can be used for automatic dividing a problem into multiple smaller problems. We know that
interpretation of salt bodies. a face consists of two eyes, a mouth, a nose, and two ears. The nose,
eyes, ears, and mouth can be decomposed into primitive features
Introduction like edges, curves, circles, or blobs of colors. These objects can be
Interpretation of seismic data is an important part of the detected using simple edge detectors with different orientations or
seismic workflow, but it can be tedious and difficult. Due to large Gabor-like filters (Gabor, 1946). The CNN utilizes this hierarchical
amounts of 3D data in modern seismic, manual interpretation is structure of objects by first detecting low-level features and then
highly time-consuming and subject to human bias. Tools for putting them together to detect higher level features (Figure 1).
automated seismic interpretation can speed up the process and, We recognize that objects other than faces can be decomposed
to some extent, reduce the subjective bias. into the same low-level features. CNNs use the same detectors
Convolutional neural networks (CNNs) have become increas- to find the shared simple features and then build specific detectors
ingly popular for solving problems in image analysis (Goodfellow for the different classes on top of these. We observe this property
et al., 2016). CNNs now hold the record in the ImageNet Large in the Alex-net (Krizhevsky et al., 2012) in which simple features
Scale Visual Recognition Challenge (Russakovsky et al., 2015), are put together to form higher level features (Figure 2).
pushing the error rates for classification from 25.8% (2011) to 3.0% CNN architecture for classification. A CNN mainly consists
(2016). CNNs differ from traditional feedforward neural networks, of two types of layers — convolutional layers and fully connected
which have long been used to analyze seismic attributes (van der layers. For classification networks, it is common to first have some
Baan and Jutten, 2000). CNNs are applied directly to the input convolutional layers followed by some fully connected layers
image, not to a set of attributes/features. During the training (Figure 3). The convolutional layers act as feature extractors, and
phase, the network learns how to compute the best attributes to the fully connected layers act as a classifier. When the network
solve a given task. This makes CNNs fundamentally different from is trained, it is the (filter) weights of the convolutional and fully
many classical image analysis approaches in which feature selection/ connected layers that are optimized. This means we do not have
engineering is an integral part of the methodology. to specify how the object classes should be detected; the network
Despite successful applications within natural and medical will learn how to do this from the training data.
image analysis, only a few publications exist in which CNNs are Convolutional layers. The convolutional layers in the neural
applied to seismic data (e.g., Araya-Polo et al., 2017; Huang et network consist of nodes that perform a convolution between an
al., 2017). However, we expect an increased interest in this topic input image and a filter (Figure 4a). The convolutions act as feature
due to the success of applying CNNs for natural image analysis. extractors since each filter is tuned to extract a specific charac-
CNNs are complex and may appear as a black-box algorithm teristic from the image. The nodes in the first convolutional layer
where the user lacks intuition of what is happening inside the receive the raw image as their input. Each node applies one filter
algorithm. As there are many details that need to be correct, using to the input image, which generates one channel in the

1
University of Oslo, Department of Informatics, Oslo, Norway. E-mail: [email protected]; [email protected]; [email protected].
2
University of Oslo, Department of Geosciences, Oslo, Norway. E-mail: [email protected].

July 2018 THE LEADING EDGE 529

Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 1. A visualization of how faces may be represented in a CNN. We observe that (a) edges with different orientations form (b) eyes, noses, and mouths that can be
used to represent (c) faces. This figure originally appeared in Lee et al. (2011) and is printed with permission from the author of the original article.

Figure 2. Filter responses for some of the nodes of the first four layers of the Alex-net (Krizhevsky et al., 2012) visualized using the technique by Zeiler and Fergus (2014).
This figure originally appeared in Zeiler and Fergus (2014) and is printed with permission from the authors of the original article.

Figure 3. A CNN applied for image classification. The input is an image, and the output is a probability vector in which each element contains the estimated probability
of the image containing an object of the given class.

530 THE LEADING EDGE July 2018

 multichannel output image from this
layer. Each channel in the output image
highlights a different feature in the
input image depending on the filter
(Figure 5).
In the second layer, multichannel
Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

filters are applied to the output of the

first layer (Figure 4b). The nodes are
analyzing the simple features detected
in the first layer and combining them
to form higher level features.
We can describe these operations
mathematically. The input image to an
arbitrary layer i is denoted by Xi(x,y,c)
and a filter j in this layer is denoted by Figure 4. (a) A convolution between the input image (green) and a single filter (red). The filter is slid across the
input image to compute the output image. (b) If the previous layer had multiple filters the input image has multiple
Wi,j (x,y,c). Here, x and y are the image channels. The nodes, therefore, apply multichannel filters to the input image. In this case, the layer has three
coordinates, and c is the channel. The filters, which produce three output channels.
input image to layer i has Ci−1 channels,
corresponding to the number of filters in the previous layer. The a weight to each of the input elements, sums, and applies the activa-
output channel j from a convolutional layer is then tion function. The output of the final fully connected layer is a
vector (x) with one element for each class. Each element in this
⎛Ci−1 ⎞ vector then indicates the probability of the input image belonging
X i+1 (x, y, j ) = f ⎜⎜ ∑ X i (x, y, k)∗Wi , j (x, y, k)⎟⎟ , (1) to a given class. Finally, the output is subject to a softmax function
⎝ k=1 ⎠ to scale and normalize the probabilities:

where * is the 2D convolution operator in x,y space and f is a

e xi
nonlinear activation function. We use zero padding on the input yi = , (3)
image so that the output image has the same size. ∑ e xk
k
The activation function introduces nonlinearity to the network.
One of the most common activation functions is the rectified where the summation runs over all elements in x.
linear unit (RELU) (Nair and Hinton, 2010): If we consider the fully connected part of the CNN, we
see that this is the same as a conventional (nonconvolutional)
f (x) = max(0,x). (2) neural network.

Conceptually, we can think of it as a soft classifier that is applied Training the network
to the extracted feature image. The RELU is determining if the Now that we have investigated how a trained CNN can solve
previous filter detected the given feature (input is higher than classification tasks, the most prominent question is: how is the
zero) and, if detected, returns how strongly the feature is present. network trained? Defining the filters manually is not an option
By applying convolution plus activation, each node becomes a since there might be thousands of filters in a network. Instead,
small feature extractor and classification unit. we train the CNN by iteratively updating the filter weights to
Between some of the convolutional layers, the spatial size of minimize the error on the training set. For simplicity, we use the
the images is sometimes reduced by downsampling or pooling. term weights for all tunable parameters in the network. This
This gradually introduces spatial invariance as it helps the network includes the filter coefficients in the convolutional layers and the
to move information from the spatial arrangement of pixel values weights in the fully connected layers.
(image domain) to features containing information relevant for Cost function. The network is trained by minimizing a dif-
the classification task (feature domain). ferentiable cost function. The cost function is used instead of
Fully connected layers. After the final convolutional layer, we the error rate because the error rate is discontinuous and will be
have hopefully detected high-level features that will help identify harder to minimize numerically. Because the softmax function
the class of the object. The next step is to predict the class of the gives a “soft” label with values between 0 and 1, the output of
image based on these features. For example, if we have detected this layer is used to construct a continuous cost function. A
a mouth and eyes, we can predict that the image contains a human common cost function, especially for multiclass problems, is
face. If we have detected eyes and fur, we can predict that the cross entropy:
image contains an animal. Such hierarchical decision rules are
modeled in the fully connected part of the network.
The output vector from the final convolutional layer is the input E = −∑ y ʹj log( y j ) , (4)
to the first fully connected layer. Each of the subsequent layers ∀j

receives the output vector from the previous layer. Each node applies

July 2018 THE LEADING EDGE 531

 regularization to the network. When
the final network is trained, including
experiments with different network
architectures and hyperparameters, a
third independent test set is used to
obtain a final error estimate. This indi-
Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

cates how well the network will perform

on new previously unseen data.

Recommendations for successful

training
Training neural networks robustly
without overfitting has always been a
major problem. This is also true for
CNNs, and many solutions have been
proposed to improve the training
(Bengio, 2012). Here we have collected
some of the most important methods
and recommendations for training
modern CNNs successfully.
Figure 5. An input image is convolved with orientation filters. This produces new images where edges with different Large amount of training data.
orientations are highlighted. Training large networks requires large
amounts of training data to avoid over-
where yj is the (predicted) softmax output for class j, and y 'j is 1 if fitting. The amount of training data needed is always relative to
the (true) label for the sample belongs to class j and 0 otherwise. the number of weights in the model. If we have a small model,
Stochastic gradient descent and back-propagation. An efficient we can avoid overfitting even with a small training set. However,
method for minimizing the cost function is by using the stochastic to solve a complex classification task, a larger model (and thus
gradient descent (SGD) method with (mini) batches. A batch is more training data) may be needed.
a small random subset of the training set. Batch sizes typically Initialization of weights. The initial values for the weights are
vary between eight and 256 samples. Iteratively, a random batch important. If all filter coefficients are set to the same value (for
of samples is drawn from the full training set and fed through example zero), all filters will be trained using similar gradients
the network. Using the method of back propagation (Rumelhart and become almost identical. In addition, the initialization also
et al., 1986; Simard et al., 2003), the gradients of the cost function affects how fast the network is trained. It is common to initialize
with respect to the weights are computed. The gradients are the weights with random values — i.e., by the approach proposed
efficiently computed using vectorization and averaged over the by Glorot and Bengio (2010).
samples in the batch to approximate the gradients for the full data Choice of activation function. Conventional neural networks
! often use a sigmoid or tanh function as the activation function.
∂E
set ( ). Each weight (wi) is then updated by taking a step in These activation functions may cause the network to converge
∂wi slowly. Modern deep networks often use the RELU activation
the direction of the negative gradient: function (equation 2) or variants of this. By using such activation
functions, the networks are often trained faster.
! Batch norm. When training a deep network, the weights
∂E
wi* = wi − τ . (5) become increasingly dependent on weights in previous layers.
∂wi
When a weight changes in a shallow layer, it may imply a large
The τ parameter is the learning rate and controls how much the change in the input to a deeper layer (covariate shift). This makes
weights are updated at each iteration and is often set to a fixed it hard to train the weights in deeper layers. The batch norm
value. Adaptive schemes for determining τ can lead to quicker algorithm (Ioffe and Szegedy, 2015) has proven to be a powerful
convergence (i.e., Kingma and Ba, 2015). strategy to cope with this challenge.
Training, validation, and test set. Due to the high number of Balancing class labels. When training a network with SGD,
weights, CNNs are prone to “overfitting.” Overfitting occurs when the weight updates will be biased if the different classes are
the network becomes so specialized in classifying the samples in unevenly represented in the training set (Hensman and Masko,
the training set that it does not perform well on new samples. To 2015). To avoid this, we have to either compensate for uneven
detect possible overfitting, we need to withhold some data from distributions when computing the cost function or ensure that
the training set to use for validation. If the error on the training the batches (on average) have an equal number of samples from
data decreases but the error rate on the validation set increases, each class.
it is a sign of overfitting. If this happens, we must decrease the Dropout. Dropout (Srivastava et al., 2014) is another popu-
number of layers or nodes in our network and/or apply some lar regularization technique designed to prevent overfitting

532 THE LEADING EDGE July 2018

Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 6. (a) The inline slice used for training. The manual interpretation was reproduced from Rojo et al. (2016). (b) The test slice. No postprocessing was applied to the
predicted salt pixels.

during training. During the training phase, 50% of the nodes The network is trained on one manually labeled inline slice
of each layer are randomly dropped (put to zero) at each itera- (Figure 6a). We select 3D cubes around the pixels in this slice
tion. The weights in each of the nodes then become less depen- (including amplitudes from the neighboring inline slices) and use
dent on the other nodes, which prevents overfitting. It also these cubes as training samples. When the network has been trained,
forces the network to learn redundant features and increases we go through the full 3D volume and select all possible 65 × 65 × 65
the robustness. cubes and apply the network to predict the class at each location.
Augmentation. Given that we have gathered the largest train- We used the network configuration presented in Figure 7.
ing set practically possible, random augmentation can be used to This is a further development of the network we proposed in
simulate a larger training data set. Small random geometrical Waldeland and Solberg (2017). Batch norm was applied before
transforms (rotation, flipping of axes, scaling, etc.) are added to the RELU activations. The network was trained for 2000 itera-
the training images, which artificially generate more training tions, with a batch size of 32 with 16 samples from each class.
samples. The technique prevents the network from overfitting to Having only one training slice does not give enough training
a limited set of training examples and makes it invariant to the data to avoid overfitting. To remedy this, we applied random
types of augmentation that are applied. The amount of augmenta- augmentation to simulate a larger training set. The augmentation
tion should be chosen such that we get samples with as much was random scaling (± 20%), random flipping of the nondepth
variety as possible but should not alter a sample so much that it axes, random rotation (± 180°), and random tilting (± 15°). The
is not representative of its original class. chosen network and training configurations were a result of
some experimentation. 3 The training lasted for approximately
Example: Salt classification 20 minutes on a Nvidia Titan GPU (2013 model).
In this section, a CNN is applied in the context of automated Results and discussion. To assess the quality of the automatic
seismic interpretation for delineation of salt bodies. Interpretation interpretation, we compare it with a manual interpretation of an
of salt bodies is of interest for two reasons. Due to the low perme- inline test section (Figure 6b). The automatic interpretation gener-
ability associated with salt bodies, they may form seals for reser- ally coincides well with the interpreted salt body. At the flanks,
voirs. In addition, they have a relatively high sound velocity, which the manual interpretation is slightly more conservative than the
makes it important to obtain an accurate velocity model in the
vicinity of the salt bodies.
Data set and network configuration. In our example, we use 3
Initially, we investigated how the number of layers and number of
a 3D data set acquired in the Barents Sea at the 7228/7 block, which nodes affected the performance. The results were very similar for network
configurations with everything between four and 10 layers and with
contains one salt wall and four salt stocks (Rojo et al., 2016). We 40–80 nodes in each layer. The network seemed to converge after
define two classes — “salt” and “not salt.” We select small 1000–2000 training iterations and did not improve notably when training
65 × 65 × 65 cubes of seismic amplitudes from the full cube. The for more iterations. (Up to 20,000 iterations were tested.) We also tested
dropout, max, and average pooling, ELU activation function, but the
goal is to have the network predicting the class of the center pixel results did not change much. The network did not perform well, however,
of the small cubes. without augmentation and batch norm.

July 2018 THE LEADING EDGE 533

Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 7. The full 3D cube is partitioned into small amplitude cubes of 65 × 65 × 65 samples, which are input into the salt classification network. The network then
predicts the class of the center pixel.
automatic interpretation using the
CNN. It should be noted that it is not
trivial to determine the boundaries at
the salt flanks and that manual inter-
pretations might vary from interpreter
to interpreter. The resulting classifica-
tion of the full 3D data set shows that
the network has successfully delineated
the salt wall and the four salt stocks
(Figure 8). Although the automatic
interpretation is somewhat more con-
servative for the four salt stocks (visible
in the time slices), the automatic inter-
pretation is very close to the manual 3D
interpretation. In our example, we only
trained on one inline containing the salt
wall. It is likely that by including train-
ing slices from the four salt stocks, the
classification would improve.
Learned attributes. We can con-
struct sections with the learned attri-
butes to gain insight into how the clas-
sification task is solved. To compute the
attribute sections, we select the small
3D cube for a given center pixel and run
it through the network. The output from
a given node can then be collected as
the attribute value at this location.
Figure 5 shows five of the learned attri-
butes for the test slice. The first attribute
(Figure 9a) is sensitive to horizontally
layered geology. This is a good indicator
of regions outside the salt but does not
give a good separation close to the
Figure 8. (a–c) The predicted salt body is marked with red. (d) The predicted full 3D salt body is visualized with
color, indicating the time. No postprocessing was applied to the predicted salt pixels. Panels (e) and (f) show the
boundary. In these regions, we often
manual interpretation done in Rojo et al. (2016) and are printed with permission from the authors of the original have dipping salt flanks. The next attri-
article. Note that the manual 3D interpretation was conducted only for the salt wall and one of the four salt stocks. butes (Figures 9b and 9c) are sensitive

534 THE LEADING EDGE July 2018

Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

Figure 9. Some of the learned attributes from the last convolutional layer computed for the test inline slice.

July 2018 THE LEADING EDGE 535

 to the dipping reflectors and complement the first attribute. The phase, which means CNNs can be used without the need for
attributes in Figures 9d and 9e are sensitive to different parts of predefined attributes. This means that the network may detect
the salt body. In the final fully connected layers, these attributes which characteristics of the classes are useful for classification.
are combined to give the final classification. The CNNs are commonly trained by iterative optimization using
the SGD algorithm. This requires a sufficient amount of training
Implications of moving from natural to seismic images data and the use of the correct regularization techniques. We have
Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

One difference between natural images and seismic data is demonstrated the use of CNNs in the context of seismic images
that the objects we typically want to detect in seismic images are by using them to delineate salt bodies. One manually labeled slice
less complex than objects we want to detect in natural images. was used to train the network, and the network was used to
This means the network architecture needed to solve seismic successfully delineate the full 3D salt body. This was confirmed
interpretation problems can have fewer layers than the architectures by a comparison with a manual interpretation.
used for natural images, which is beneficial because smaller
networks are more easily trained with less data. Acknowledgments
Much of the success of CNNs has come due to efficient This work is funded by the Norwegian Research Council,
implementation of convolutions on powerful GPUs, which have Grant 234019. The CNN used in this paper was implemented
made it possible to train networks with more iterations within a using the TensorFlow framework by Google. Example code is
feasible time frame. For 3D seismic data, this poses a challenge available at https://github.com/waldeland/CNN-for-ASI.
due to the limited amount of memory on GPUs. For 3D data,
this quickly limits the sizes of networks, samples, and batches Corresponding author: [email protected]
that can be used.
In the context of seismic, we need to be aware that the underly- References
ing geology might be very different depending on the region where Araya-Polo, M., T. Dahlke, C. Frogner, C. Zhang, T. Poggio, and
the data were acquired. Also, the data quality is dependent on D. Hohl, 2017, Automated fault detection without seismic process-
the processing workflow, sampling interval, and eventual errors ing: The Leading Edge, 36, no. 3, 208–214, https://doi.org/10.1190/
in the velocity model. If we want to train a network that should tle36030208.1.
Bengio, Y., 2012, Practical recommendations for gradient-based
generalize to new data sets, we need to have training samples
training of deep architectures, in G. Montavon, G. B. Orr, and
from data sets covering a wide range of these differences. However,
K.-R. Müller, eds., Neural networks: Tricks of the trade, second
when a new seismic data set is processed and interpreted, a large edition: Springer Heidelberg, 437–478.
amount of manual work and quality control is invested in each Gabor, D., 1946, Theory of communication. Part 1: The analysis of
data set. Therefore, the cost of labeling some parts (e.g., some information: Journal of the Institution of Electrical Engineers
slices) of a data set is relatively small compared to the entire — Part III: Radio and Communication Engineering, 93, no. 26,
workflow. This makes it possible to train the network on a few 429–457, https://doi.org/10.1049/ji-3-2.1946.0074.
slices from the data set and use it to interpret the remaining. Ganin, Y., and V. Lempitsky, 2015, Unsupervised domain adaptation
by backpropagation: Proceedings of the International Conference
Suggested reading on Machine learning, 1180–1189.
The network architecture we used is rather simple and is based Glorot, X., and Y. Bengio, 2010, Understanding the difficulty of
training deep feedforward neural networks: Proceedings of the
on texture classification. State-of-the-art models for classification
13th International Conference on Artificial Intelligence and Sta-
(Szegedy et al., 2015; He et al., 2016) should be considered for
tistics, 249–256.
complex classification tasks and segmentation networks Goodfellow, I., Y. Bengio, and A. Courville, 2016, Deep learning:
(Ronneberger et al., 2015) for non-texture-based problems. In MIT Press.
cases with little training data, reuse of existing networks (Razavian He, K., X. Zhang, S. Ren, and J. Sun, 2016, Deep residual learning
et al., 2014) or domain adaption (i.e., Ganin and Lempitsky, 2015) for image recognition: IEEE Conference on Computer Vision
could improve results, but it is not trivial to succeed with such and Pattern Recognition, 770–778, https://doi.org/10.1109/
methods. Visualization of trained networks (i.e., Zeiler and Fergus, CVPR.2016.90.
2014; Mordvintsev et al., 2015) is important in order to understand Hensman, P., and D. Masko, 2015, The impact of imbalanced training
how a network solves a given classification task. It should be noted data for convolutional neural networks: PhD thesis, KTH Royal
that the field of deep learning is actively being developed and that Institute of Technology.
Huang, L., X. Dong, and T. E. Clee, 2017, A scalable deep learning
best practices for network configurations and choosing hyperpa-
platform for identifying geologic features from seismic attributes:
rameters and training strategies change rapidly.
The Leading Edge, 36, no. 3, 249–256, https://doi.org/10.1190/
tle36030249.1.
Conclusions Ioffe, S., and C. Szegedy, 2015, Batch normalization: Accelerating
The use of CNNs on seismic data is promising and may lead deep network training by reducing internal covariate shift: Proceed-
to increased accuracy for automated seismic interpretation, as it ings of 32nd International Conference on Machine Learning.
has done for tasks in natural and medical image analysis. CNNs Kingma, D., and J. Ba, 2015, Adam: A method for stochastic opti-
consider the spatial aspect of images and exhibit the hierarchal mization: Proceedings of the 3rd International Conference on
structure for detecting complex objects through layers of convo- Learning Representations, 1–15.
lutional nodes. The filter weights are optimized during the training

536 THE LEADING EDGE July 2018

 Krizhevsky, A., I. Sutskever, and G. E. Hinton, 2012, ImageNet Rumelhart, D. E., G. E. Hinton, and R. J. Williams, 1986, Learning
classification with deep convolutional neural networks: Proceed- representations by back-propagating errors: Nature, 323, 533–536,
ings of the 25th International Conference on Neural Information https://doi.org/10.1038/323533a0.
Processing Systems, 1097–1105. Russakovsky, O., J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z.
Lee, H., R. Grosse, R. Ranganath, and A. Y. Ng, 2011, Unsupervised Huang, et al., 2015, ImageNet Large Scale Visual Recognition
learning of hierarchical representations with convolutional deep Challenge: International Journal of Computer Vision, 115, no.
belief networks: Communications of the ACM, 54, no. 10, 95–103, 3, 211–252, https://doi.org/10.1007/s11263-015-0816-y.
Downloaded 12/13/18 to 37.224.6.12. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/

https://doi.org/10.1145/2001269.2001295. Simard, P. Y., D. Steinkraus, and J. C. Platt, 2003, Best practices

Mordvintsev, A., M. Tyka, and C. Olah, 2015, Inceptionism: Going for convolutional neural networks applied to visual document
deeper into neural networks: Google research blog, https:// analysis: 7th International Conference on Document Analysis and
research.googleblog.com/2015/06/inceptionism-going- Recognition, 958–963.
deeper-into-neural.html, accessed 4 May 2018. Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever, and R.
Nair, V., and G. E. Hinton, 2010, Rectified linear units improve Salakhutdinov, 2014, Dropout: A simple way to prevent neural
restricted Boltzmann machines: Proceedings of the 27th Interna- networks from overfitting: Journal of Machine Learning Research,
tional Conference on Machine Learning, 807–814. 15, 1929–1958.
Razavian, A. S., H. Azizpour, J. Sullivan, and S. Carlsson, 2014, Szegedy, C., W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D.
CNN features off-the-shelf: An astounding baseline for recogni- Erhan, V. Vanhoucke, and A. Rabinovich, 2015, Going deeper
tion: IEEE Conference on Computer Vision and Pattern Recogni- with convolutions: IEEE Conference on Computer Vision and
tion Workshops, https://doi.org/10.1109/CVPRW.2014.131. Pattern Recognition, https://doi.org/10.1109/CVPR.2015.7298594.
Rojo, L. A., A. Escalona, and L. Schulte, 2016, The use of seismic van der Baan, M., and C. Jutten, 2000, Neural networks in geophysical
attributes to enhance imaging of salt structures in the Barents applications: Geophysics, 65, no. 4, 1032–1047, https://doi.
Sea: First Break, 34, no. 11, 49–57, https://doi.org/10.3997/ org/10.1190/1.1444797.
1365-2397.2016014. Waldeland, A. U., and A. H. S. Solberg, 2017, Salt classification using
Ronneberger, O., P. Fischer, and T. Brox, 2015, U-Net: Convolu- deep learning: 79th Conference and Exhibition, EAGE, Extended
tional networks for biomedical image segmentation: Medical Abstracts, https://doi.org/10.3997/2214-4609.201700918.
Image Computing and Computer-Assisted Intervention — Zeiler, M. D., and R. Fergus, 2014, Visualizing and understanding
MICCAI 2015, 234–241, https://doi.org/10.1007/978-3-319- convolutional networks: European Conference on Computer Vision,
24574-4_28. 818–833, https://doi.org/10.1007/978-3-319-10590-1_53.

July 2018 THE LEADING EDGE 537