Current state and future directions for deep learning based automatic seismic fault interpretation: A systematic review

Full text

Earth-Science Reviews 243 (2023) 104509
Available online 20 July 2023
0012-8252/© 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Current state and future directions for deep learning based automatic
seismic fault interpretation: A systematic review
☆
Yu An
a
,
*
, Haiwen Du
b
,
1
, Siteng Ma
a
, Yingjie Niu
a
,
f
, Dairui Liu
a
, Jing Wang
c
,
1
, Yuhan Du
a
,
Conrad Childs
d
, John Walsh
d
,
e
, Ruihai Dong
a
,
*
a
School of Computer Science, University College Dublin, Beleld, Dublin, Ireland
b
School of Astronautics, Harbin Institute of Technology, Harbin, Heilongjiang, China
c
School of Computer Science and Engineering, North China Institute of Aerospace Engineering, Guangyang, Langfang, Hebei, China
d
Fault Analysis Group, School of Earth Sciences, University College Dublin, Beleld, Ireland
e
iCRAG (Irish Centre for Research in Applied Geosciences), Ireland
f
SFI Centre for Research Training in Machine Learning, Ireland
ARTICLE INFO
Keywords:
Systematic literature review
Deep learning
DL
Seismic fault interpretation
Articial intelligence
Convolutional neural network
ABSTRACT
Automated seismic fault interpretation has been an active area of research. Since 2018, Deep learning (DL) based
seismic fault interpretation methods have emerged and shown promising results. However, to date, these
methods have not been reasonably summarised, making it difcult for those involved to make sense of the
current development process. To close this gap, we systematically reviewed the DL-based fault interpretation
literature published between 2012 and 2022, and searched seven digital libraries. Fault interpretation has been
considered an image-processing task using only convolutional neural networks (CNN)-based DL methods, and
most of them have been trained in a supervised manner. U-Net and its variants designed for the image seg-
mentation task are the most commonly used network structures. A total of 73 seismic datasets were summarised
from the 56 articles included, of which only three eld datasets and four synthetic datasets were publicly
available benchmarks. The study reported benets of using DL, such as its outstanding learning and general-
isation capabilities or predicting faults in a fast, cheap and repeatable manner, which ultimately led to an in-
crease in the acceptability of these methods and the potential to incorporate them into oil and industry
workows. However, we identied 12 challenges that hinder its integration into industrial workows, including
the most discussed lack of sufcient annotated data. We conclude with an in-depth discussion of current research
trends and potential future research directions to promote research on less studied areas and collaboration be-
tween computer scientists and geoscientists.
1. Introduction
Seismic data interpretation is a critical process of obtaining subsur-
face geological information, including fault interpretation which is of
great importance in hydrocarbon exploration and natural hazard
assessment. Faults are prominent geological structures formed in the
upper part of the earth’s crust due to brittle deformation (Fossen, 2010;
An et al., 2021). Seismic data in this context refers to seismic reection
data, which images the subsurface geology by showing reections of
density contrasts between rock layers. Fault interpretation refers to the
process of mapping faults marked by discontinuities within otherwise
continuous seismic reections (An et al., 2021).
Conventionally, fault interpretation has only been possible through
manual interpretation by domain experts, which is labour-intensive,
time-consuming, and expensive (An et al., 2021). However, the
advancement in data acquisition and processing techniques (Reilly et al.,
2023), along with the reduced cost of seismic data collection (Baraniuk
and Steeghs, 2017), has led to an increase in seismic data resolution and
a rise in the number of seismic projects conducted. As a result, the
overall volume of seismic data has experienced exponential growth.
☆
This document is the results of the research project funded by the Science Foundation Ireland (SFI) [SFI/12/RC/2289_P2]; Beijing Dublin International College
Fund; China scholarship council Grant 202106120101.
* Corresponding author.
E-mail addresses: yu.an@insight-centre.org (Y. An), ruihai.dong@insight-centre.org (R. Dong).
1
Work done while visiting University College Dublin.
Contents lists available at ScienceDirect
Earth-Science Reviews
journal homepage: www.elsevier.com/locate/earscirev
https://doi.org/10.1016/j.earscirev.2023.104509
Received 24 October 2022; Received in revised form 27 June 2023; Accepted 15 July 2023

Earth-Science Reviews 243 (2023) 104509
2
Effectively harnessing this vast amount of seismic data within the
limited available time by the conventional approach has become pro-
gressively more challenging (Kozhenkov et al., 2019). Furthermore,
manual interpretation, which is susceptible to biases and inuenced by
researchers’ interests, can introduce considerable uncertainties to the
interpretation results (Randle et al., 2019; Bond, 2015). Hence, over the
last few decades, there has been signicant research into automatic
interpretation methods as an alternative approach.
During the same period, articial intelligence experienced rapid
development. Compared to the earlier manual selection of a single or a
few features to assist in fault interpretation, geologists began utilizing
the ability of machine learning (ML) to simultaneously consider multiple
attributes in the processing of seismic data. Early methods included
traditional machine learning techniques such as Support Vector Ma-
chines (SVM) and Multilayer Perceptrons (MLP) (Di et al., 2017; Guitton
et al., 2017; Di et al., 2019).
Although these early multi-feature methods reduced the need for
manual intervention to some extent, traditional attribute-based or semi-
automatic ML methods still required substantial human intervention.
For example, some noise-sensitive attributes required parameter ad-
justments prior to use (Marfurt et al., 1998), and feature engineering, i.
e., the extraction of suitable features, was necessary for complex datasets
like seismic data. Additionally, the best-performing ML attribute sets
vary across datasets. Attributes and parameters often must be tuned by
trial and error. These algorithms are not guaranteed to perform satis-
factorily on new datasets, even if they include considerable human
intervention (Guitton et al., 2017).
Over the past decade, DL demonstrating remarkable results in
various elds such as science and business. Unlike conventional
methods, DL algorithms eliminate the need for feature engineering by
autonomously learning and extracting relevant features from datasets in
their raw form, leveraging increased computational power and abun-
dant data (LeCun et al., 2015). This automated feature extraction pro-
cess alleviates the limitations associated with manual feature
engineering, including subjective feature selection, time-consuming it-
erations, and potential information loss due to human bias (Bengio et al.,
2013; An et al., 2021).
A quantitative comparison conducted by Wang et al. (2021b) be-
tween a traditional ML approach (i.e., support vector machine SVM) and
CNN method revealed that CNN achieved a high accuracy of 0.98, a 10%
improvement over SVM on the MNIST dataset. Moreover, under the
same hardware conditions, this lightweight CNN even required less
processing time than SVM (23.2 and 27.6 min). The study also demon-
strated a larger accuracy boost with increased image size and categories.
However, it is important to note that different CNN architectures can
result in a wide range of computational costs, and performance may vary
across different datasets.
In terms of fault interpretation task, Di et al. (2018a) conducted a
qualitative comparison between SVM and CNN, concluding that
although CNN extracted features that were challenging to interpret, they
enabled more precise fault labelling and were less susceptible to noise
interference.
However, DL methods also exhibit certain limitations compared to
traditional approaches including ML methods. Firstly, DL methods
eliminate the need for a feature engineering step, as the relevant features
are computed by the model itself. While this enhances automation, it
also reduces interpretability, making it difcult for geologists to assess
the quality of results from DL models (Kozhenkov et al., 2019). Sec-
ondly, DL imposes higher requirements on the quantity and quality of
training data compared to traditional methods, which can result in
signicant training costs.
Overall, given the growing interest and concerns about DL in fault
interpretation tasks, a systematic literature review is needed to
comprehensively summarize the latest research advancements, chal-
lenges, trends, and future opportunities. This review aims to address this
gap in the existing literature.
Our review is structured as follows. Section 2 outlines the principles
of DL and fault interpretation and provides a description of the moti-
vation and relevant surveys of this paper. Section 3 details the meth-
odology of our systematic review. Section 4 summarises data we
extracted from related research articles, including DL models, datasets,
advantages, challenges of using DL in fault interpretation and corre-
sponding candidate solutions. Section 5 discusses current research
trends and potential future research directions and Section 6 provides
concluding remarks.
2. Background
In this section, we introduced the concepts of fault interpretation and
DL, together with a consideration of conventional approaches. This is
followed by a list of relevant literature reviews that motivate this
review.
2.1. Seismic Fault Interpretation
Seismic fault interpretation is an essential subset of the broader
concept of seismic interpretation, focusing specically on geological
faults. The interpretation on faults is not a simple process like the
annotation of objects from natural camera-captured images/videos. It
involves the construction of a 3-dimensional (3D) geological model of
fault geometry and displacement, then performing annotations and
subsequent iterative corrections (Alcalde et al., 2019; Gibbs, 1984;
Badley et al., 1991; Walsh et al., 1996; An et al., 2021). Faults often form
a complex network of surface or narrow zones (i.e. <ca 100 m wide and
often <10 m wide), that usually intersect each other. Arising from
limitations in the display capabilities of annotation software and the
ability of human eye to process 2D images, fault interpretation is usually
performed on different 2D cross-sectional slices of a 3D seismic volume.
It considers whether there are consistent reector discontinuities and
the continuity of displacement in adjacent cross-sections, with dense (i.
e. narrow spacings between) 2D interpretations required to create
smooth and accurate fault surfaces (An et al., 2021). Overall, the
complexity of fault interpretation lies in the incorporation of geological
knowledge and the iterative nature of the annotation process, dis-
tinguishing it from straightforward object annotation in 2D or 3D
images.
Since seismic data is signal-recorded data, the quality of the signals
collected can vary dramatically depending on the sophistication of the
signal acquisition equipment, the subsurface conditions and properties
and the quality of the manual processing. Despite signicant efforts to
enhance seismic data quality, there are inherent constraints in fault
illumination and imaging. Identifying larger faults is relatively simpler
due to their noticeable seismic horizon offsets resulting from larger
throws. However, fault dip signicantly affects the visibility of the fault
plane (Faleide et al., 2021a). As a result, challenging cases of fault
interpretation heavily rely on the expertise and judgment of geological
modellers.
Given the limitations and complexity of conventional workows,
domain researchers rst invented some discontinuity-sensitive attri-
butes to help illustrate potential faults and fault zones (Marfurt et al.,
1999; Roberts, 2001; Marfurt et al., 1998). Building upon this founda-
tion, further advancements were made in developing data-driven
methods for horizon and fracture identication (Bugge et al., 2019;
Bugge et al., 2018). Later, ML techniques were proposed to highlight
fault zones by exploiting multiple attributes simultaneously (Di et al.,
2017; Guitton et al., 2017).
Today, the advancements in seismic resolution have resulted in the
generation of massive seismic data sets, varying in size from a few
gigabytes to several terabytes (Baraniuk and Steeghs, 2017). Consid-
ering the enormous amount of data that has to be interpreted and the
valuable information that can be derived from previously coarsely
interpreted datasets, we expect that more automated workows will
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
3
enable better and objective fault interpretation in a short time frame. DL
methods, particularly suitable for addressing big data problems, have
been introduced into fault interpretation in recent years.
2.2. Deep learning
DL is a subset of the ML eld, which also belongs to the broader eld
of AI. In the early days of AI, machines were primarily programmed
based on predened rules. Later, the concept of ML was introduced,
focusing on enabling machines to learn patterns from raw data without
the need for hard-coded knowledge. Traditional ML methods such as
linear regression, decision trees, and random forests perform well in
many tasks but have limitations when it comes to handling complex
high-dimensional data such as images, speech, and natural language
(Goodfellow et al., 2016). DL draws inspiration from biological neurons
and utilizes multiple layers of neurons (i.e. computational nodes) to
guide and automatically extract useful high-dimensional features from
large amounts of raw data (LeCun et al., 2015). With the advancement of
computational resources, DL methods can leverage deep computational
graphs for predictions. Despite the complexity of DL methods, which
makes them challenging to interpret, they have achieved breakthrough
performance in various tasks, especially in the realm of unstructured
data like images (Alzubaidi et al., 2021).
The success of DL is often underpinned by the well-designed neural
network architectures (Schmidhuber, 2015). One of the most well-
known network architectures, especially for tasks involving image pro-
cessing, is the convolutional neural network (CNN). Over the past
decade, CNNs have primarily improved the performance of various
computer vision systems, from high-level tasks such as image classi-
cation (Russakovsky et al., 2014) and object detection (Redmon et al.,
2015) to low-level tasks such as semantic image segmentation (Ronne-
berger et al., 2015; Chen et al., 2018) and edge detection (Liu et al.,
2016; Xie and Tu, 2015).
Image classication refers to the type of task that classies an entire
image. One of the best-known examples is the ImageNet challenge,
which classies millions of images into 1,000 categories. Object detec-
tion refers to framing a target object in an image or video clip (Voulo-
dimos et al., 2018). Image segmentation refers to the task of performing
accurate object localisation as well as pixel-level classication. It divides
a digital image into multiple segments (sets of pixels) (Shapiro and
Stockman, 2001). Similarly, instance segmentation is the task of
detecting and delineating each distinct object in an image. Edge detec-
tion focuses on the edges of objects in an image, i.e. pixels with sharply
changing values (Bugge et al., 2019; Xie and Tu, 2015). Image synthesis
is also a popular task aiming to generate super-resolution images or
realistic images (Kumar et al., 2020).
A typical CNN, such as the classical VGG network (Simonyan and
Zisserman, 2014) contains convolutional layers, activation layers (e.g.
ReLU), pooling layers (e.g. maximum pooling), and fully connected
layers. The convolution layer extracts useful features by computing the
convolution (i.e. cross-correlation) between the learned kernel and the
input matrix. The activation layer introduces a non-linear property to
solve the non-linear correlation problem. The pooling layer down-
samples the features, merges them and reduces the complexity of the
computation. The fully-connected layer converts the high dimensional
feature map into a one dimensional feature vector and is mainly used for
classication problems.
The success of DL methods also benets from a well-designed
training/learning methodology. Depending on the number of labels
the learning process provides, they can be broadly classied into three
categories: supervised learning, unsupervised learning and semi-
supervised learning (Alzubaidi et al., 2021). Deep supervised learning
is the most straightforward and typical category of DL. It requires fully
labelled datasets to supervise the updating of randomly initialised
neural network parameters. The update mechanism is an error-based
feedback mechanism that optimises the parameters by computing the
difference between the model predictions and the labels. This category
of DL method is the best performing of the above 3 DL learning cate-
gories due to the prior knowledge learned from the labelled dataset
(Alzubaidi et al., 2021). However, fully labelled datasets are often not
readily available, circumstances which lead to the other two types of DL
approach.
Deep unsupervised learning focuses on the fact that some tasks have
data without corresponding labels. It clusters or associates data by
identifying exciting structures or patterns in the data. Although it may
not directly solve the target task, it can provide some analysis for tasks
without labels, giving an unbiased and robust unrestricted basis for
subsequently complex tasks (Raza and Singh, 2021).
Deep semi-supervised learning or weakly supervised learning falls
between the rst two categories. It is ideal for tasks with only a limited
number of labels and a considerable amount of unlabelled data (Ouali
et al., 2020). In recent years, it has also been extended to other use cases,
such as where only incomplete and inaccurate labels are available
(Nashaat et al., 2021; Zhang et al., 2021) or where coarse-grained labels
are easily available while ne-grained labels are difcult to obtain (Garg
et al., 2022; Zhou, 2018).
Additionally, our researched articles involve some other learning
approaches cannot simply be classied as any of the above. Transfer
learning aims to use prior knowledge learned from a known domain to a
new, different but similar domain (Zhuang et al., 2020). Knowledge
distillation involves training a smaller and simpler student model to
replicate the performance of a larger and more complex teacher model
(Gou et al., 2021). By mimicking the behavior and predictions of the
teacher model, the student model aims to achieve comparable perfor-
mance while being computationally more efcient. Multi-task learning
represents an ML category that deals with multiple tasks simultaneously.
In addition to the advantage of handling multiple tasks simultaneously
with a single model, it assumes that similar tasks can build on each other
and achieve optimal overall performance (Zhang and Yang, 2018).
These learning methods: transfer learning, knowledge distillation, and
multi-task learning are techniques that can be applied within the
frameworks of supervised, unsupervised, or semi-supervised learning to
improve model performance, knowledge transfer, and learning ef-
ciency. Each technique addresses specic challenges and goals within
the broader spectrum of machine learning.
2.3. Related works and motivation
Deploying DL to fault interpretation is a relatively new topic, with
relevant literature appearing since 2017. We found only seven surveys
or case studies broadly mentioning DL-based fault interpretation. We
describe their contributions and limitations in more detail below,
leading to our motivation for this review.
Wang et al. (2018) specically reviewed research on treating seismic
data as images rather than signals and leveraging ML-based image
processing techniques in fault and salt body interpretation. The authors
note the rise of DL, particularly the use of CNNs in seismic structural
interpretation, and mention its potential use in other seismic data-
related tasks. The review was, however, published at a time when
very little work had been performed on DL-based fault interpretation.
Since 2018, there has been a growth in associated research, and there
has not been a subsequent literature review covering these research
articles.
Di et al. (2018b) investigated what factors make CNNs outperform
conventional multilayer perceptrons (MLP) in interpreting seismic faults
and salt bodies. The authors argued that CNNs could automatically
extract helpful high-level abstract features from image patches, whereas
MLPs require expert intervention for feature preparation and can only
compute sample-level features. We consider this article a case study
rather than a conventional literature review.
Kozhenkov et al. (2019) gave a brief history of the development of
ML, reviewed several examples of ML applications in the geoscience
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
4
domain, and suggested that geoscientists should not fear the emerging
ML techniques but rather take advantage of machine predictions to
achieve better interpretation results. The main emphasis of this article
was on showcasing the advantages of employing machine learning (ML)
techniques for geoscience applications, with limited attention given to
the specic topic of DL-based fault interpretation.
Kuang et al. (2021) surveyed the AI applications in petroleum
exploration and discussed development trends and future directions.
With DL and fault interpretation subsets of the broad content of AI and
petroleum exploration, this review focused on the general picture and
mentioned only one article about DL in fault interpretation (Wu et al.,
2019b) (an abstract paper later extended to (Wu et al., 2019c)).
Dimililer et al. (2021) reviewed applications applying ML and the
Internet of Things (IoT) techniques in geophysical engineering, focusing
on earthquake-related geophysical engineering applications with only
one research article covering CNN-based fault interpretation (Zheng
et al., 2019).
Tariq et al. (2021) systematically reviewed the application of data
science and ML in the oil and gas industry. With the brief review of only
one research article on DL-based fault interpretation (Xiong et al.,
2018). This study does not adhere to the recommended guidelines or
best practices of a standard SLR, as it omits a methodology for the article
selection process.
Li et al. (2021) surveyed the development of the seismic coherence
attribute. Seismic coherence is a measure of quantifying the degree of
discontinuity of seismic data, which is often used to support the inter-
pretation on faults. They mentioned AI-based seismic discontinuity
interpretation in their appendix, with six research articles surrounding
CNN and fault interpretation (Huang et al., 2017; Qi et al., 2020; Wu
et al., 2019a, 2020; Xiong et al., 2018; Zhao and Mukhopadhyay, 2018).
Overall, we recognised some defects in the above mentioned related
literature reviews:
•Most related reviews address a broad research area with little focus
on DL-based fault interpretation (Kuang et al., 2021; Li et al., 2021;
Tariq et al., 2021; Dimililer et al., 2021).
•According to Keele et al. (2007), an SLR is a specic type of literature
review that aims to answer specic research questions in an objec-
tive, transparent and reproducible manner. Almost none of these
reviews are SLR (Li et al., 2021; Kuang et al., 2021; Kozhenkov et al.,
2019; Di et al., 2018b; Wang et al., 2018; Dimililer et al., 2021), and
two of them are not typical literature reviews (Di et al., 2018b;
Kozhenkov et al., 2019). None of the above literature reviews con-
tains a clear description of the research questions, search strategy
and article selection steps.
•None cover the recent (i.e. 2018 to present) development of DL-
based fault interpretation models comprehensively.
Considering the research area and limitations of reviews in the cur-
rent literature, we present the rst SLR on DL-based fault interpretation,
covering research articles that were published between 2012 and 2022.
A detailed methodology section is described below.
3. Methodology
This review of DL-based fault interpretation was conducted in
accordance with Kitchenham’s best practice systematic review guide-
lines (Keele et al., 2007). The systematic review followed a structured
process in order to extract and analyse relevant literature in an unbiased
and transparent manner. This structured process includes the following
steps:
Step 1.Identify specic research questions
Step 2.Develop a clear search strategy
Step 3.Devise a clear article selection and quality assessment
standard
Step 4.Data extraction procedures
Step 5.Data analysis
The following subsections detail our literature screening process (steps
1–4).
3.1. Research questions
In this review, we have identied the following research questions as
we aim to explore the current state and future directions of DL methods
in fault interpretation.
•RQ1: What DL models have been developed to achieve or assist
automatic fault interpretation?
•RQ2: What datasets were used in this area?
•RQ3: What are the reported benets of using DL in fault
interpretation?
•RQ4: What are the identied challenges and candidate solutions in
this area?
3.2. Search strategy
We performed an electronic search among seven famous digital
literature databases: ACM Digital Library,
2
IEEE Xplore,
3
ScienceDir-
ect,
4
Scopus,
5
, SpringerLink,
6
Web of Science (WoS),
7
SEG Digital Li-
brary.
8
Based on the research questions we identied, we came up with
the following search terms: deep learning and fault interpretation. We
considered replacement terms: neural network, fault, and seismic image
to ensure complete search results. Plural forms are also considered.
On 5 March 2022, the keyword combinations shown in Table 1 were
used to search seven digital libraries. As digital libraries provide date
and search eld restrictions, we searched for records published from
2012 to the present containing keywords throughout the records except
for Scopus. Scopus is restricted to searching only the title, abstract and
keywords as it is an abstract and citation database.
3.3. Article selection and quality assessment
Following the search strategy, 721 records were returned from the
seven digital libraries, with the number of records returned for each
search shown in Fig. 1. Next, we devised some inclusion and exclusion
criteria to identify the actual relevant research articles that fell within
our review objectives. Only articles that used DL models to implement or
directly assist the seismic fault interpretation workow were included.
As modern DL techniques have only shown groundbreaking success
since 2012, our SLR only needed to examine articles published from
2012 to the present (i.e. 5 March 2022). Besides, we only consider ar-
ticles that are peer-reviewed as they are considered to be of high quality.
Records in ineligible formats, such as extended abstracts and archived
Table 1
Keyword search command.
(“deep learning” OR “neural network”)
AND
((fault AND “seismic image”) OR “fault interpretation”)
2
https://dl.acm.org.
3
https://ieeexplore.ieee.org.
4
https://www.sciencedirect.com.
5
https://www.scopus.com.
6
https://link.springer.com.
7
https://www.webofscience.com.
8
https://library.seg.org.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
5
articles, are not included. Finally, articles for which we do not have
access to the full-text and articles not written in English are not included
in this review.
To streamline the manual screening process and ensure data integ-
rity, a format and language ltering step was implemented. By
leveraging features offered by certain digital libraries, a signicant
number of records that didn’t meet the required criteria were auto-
matically removed. This automated ltering resulted in the exclusion of
three non-English records and 234 records that didn’t conform to the
specied format. In addition, unique Digital Object Identiers (DOIs)
were checked by a customised Python program to remove duplicates.
Leaving 405 articles for full-text retrieval. However, four of 405 records
are not available in full-text from our end, leaving 401 records to be
manually screened.
We conducted two rounds of manual screening of the remaining 401
records at this stage. Round 1 was a rapid screening process conducted
by the rst author alone, which excluded three non-English articles and
34 articles that did not meet the formatting requirements.
The rst author then manually screened the remaining 364 articles
as the rst reviewer. 229 (62.9%) of these 364 articles were indepen-
dently reviewed by co-authors as second reviewers for cross-checking.
The rst and second reviewers achieved a high agreement rate of
93%. For the 16 articles disagreed by reviewers, we involved an inde-
pendent third reviewer and agreement was achieved by the majority
vote. Altogether, 285 articles were excluded, leaving 79 articles for the
quality assessment step (see Fig. 1).
Next, at least two reviewers checked if these articles contained suf-
cient details: Was the DL method well described? Did the article
describe the model architecture? Did the article describe how the DL
model was trained? Did the article describe the dataset used? Articles
that failed to provide satisfactory answers to any of these questions were
categorized as “unclear method” and excluded from this review. An
exception is that if it is a review article, the article was included as
literature support for Section 2. Two research articles focusing on un-
certainties in fault interpretations were also included as literature sup-
port for RQ4.
As a result, a total of 56 research articles met the inclusion criteria for
this systematic review, while 9 additional articles were included solely
as literature support for the relevant sections.
3.4. Data extraction
Furthermore, to ensure precise documentation of the necessary in-
formation, one author cross-validated the data extraction results of
Fig. 1. Flowchart showing our literature search inclusion process.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
6
another author.
4. Results
The distribution of all 56 articles by year is shown in the Fig. 2a.
Although we searched for relevant articles in the last decade, the nal
set was only distributed between 2018 and 2022. It suggests that DL for
fault interpretation is a new research area that emerged in 2018. The
increasing number of articles published each year indicates the growing
interest of the researchers involved. As shown in Fig. 2b, researchers in
the eld prefer to publish their latest research through journals (73.2%)
rather than conference proceedings (26.8%). Of the selected journal
articles, the top 5 journal choices in order were IEEE Transactions on
Geoscience and Remote Sensing (19.5%), Geophysics (19.5%), Inter-
pretation (12.2%), The Leading Edge (9.8%) and Computers & Geo-
sciences (7.3%), as shown in Fig. 2c.
4.1. RQ1: What DL models have been developed to achieve or assist
automatic fault interpretation?
The use of DL to assist fault interpretation is an interdisciplinary task
with widely varying approaches, which have not yet been classied and
summarised in a relevant literature review (see Section 2.3). Consid-
ering an interdisciplinary audience, we summarise the literature from
the perspective of computer science (i.e. DL tasks, specic models, input
dimensions) and geosciences (i.e. whether direct fault interpretation is
performed, geological targets), respectively. The classications of the
surveyed articles are presented in Table 2.
Based on Table 2, it can be observed that all articles consider fault
interpretation as an image-processing task and use CNN-based methods.
The preference for CNN-based methods in the reviewed articles can be
attributed to factors such as the perceived alignment of seismic image
interpretation with image processing principles, the inuence of early
literature, the ease and effectiveness of image processing DL models, and
the visualization benets they offer. The majority (32, 57.1%) used 2D
CNNs directly for 2D seismic images, followed by a minority (21, 37.5%)
that used 3D CNNs for 3D seismic volumes, and very few (3, 5.4%)
offered both 2D and 3D versions (see Fig. 3a).
Regarding image processing tasks (see Section 2.2), we identied
ve different DL image processing tasks in the selected 56 articles, which
are image segmentation (36, 60.0%), image classication (16, 26.7%),
image synthesis (4, 6.7%), instance segmentation (2, 3.3%) and edge
detection (2, 3.3%). As shown in Fig. 3b, the dominance of image
classication in 2018 has been replaced by image segmentation from
2019 onwards, and the difference is getting wider.
The DL tasks show a strong association with the DL model used. For
image segmentation, the vast majority (31, 86.1%) of the network
structures are UNet or its variants, with the other ve model structures
used in only one or two articles. UNet is a classical U-shaped segmen-
tation network with an encoder for feature extraction and a decoder
structure with transposed convolution to upsample (Ronneberger et al.,
2015). For image classication, more than half (9, 56.3%) of the articles
present their custom CNN, followed by VGG variant, LeNet variant and
GAN variant architecture. VGG and LeNet are two classical structures for
image classication networks. GAN stands for generative adversarial
network, a classical generative network consisting of two simulta-
neously trained sub-networks that minimise algorithm competes with
learning the data distribution and generating images (Goodfellow et al.,
2014) for classication. Similarly, half of the four image synthesis ar-
ticles use the GAN-based approach, while the other half use VGG and
UNet variant architecture. All instance segmentation approaches
employ the UNet-based neural network. Overall, as shown in Fig. 3d, the
most popular DL model architecture is the UNet variant (34, 54.0%),
followed by custom CNN (11, 17.5%), VGG variant (5, 7.9%), GAN
variant (4, 6.3%), and remaining six architecture (9, 14.4%).
In addition to DL tasks and DL models, DL methods can also be
classied as supervised, semi-supervised, unsupervised, and others (see
Section 2.2), depending on the training/learning method. Here, we
Fig. 2. General statistics of included articles.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
7
Table 2
Articles included in this review with information about publication year, model input dimension, DL task,
a
DL model,
b
DL category (DL Cat), fault interpretation
category (FI Cat), geo-target
c
and target-model category (TM Cat).
d
Reference Dimension DL Task DL Model DL Cat FI Cat Geo-target TM Cat
Guitton (2018) 3D Classication VGG variant Supervised Semidirect Fault STSN
Di et al. (2018a) 2D Classication Custom CNN Supervised Direct Fault STSN
Lu et al. (2018) 2D Synthesis GAN Supervised Indirect Super-resolution STSN
Xiong et al.
(2018)
2D Classication Custom CNN Supervised Direct Fault STSN
Haroon et al.
(2018)
2D Segmentation UNet variant Supervised Direct Fault STSN
Pochet et al.
(2019)
2D Classication VGG variant Supervised Direct Fault STSN
Wu et al. (2019a) 3D Segmentation UNet variant Supervised Direct Fault STSN
Lapteva et al.
(2019)
2D Segmentation UNet variant Supervised Direct Fault STSN
Chang et al.
(2019)
2D Segmentation UNet variant (ResUNet) Supervised Direct Fault STSN
Zheng et al.
(2019)
3D Classication Custom CNN Supervised Direct Fault, inversion STSN
Yuan et al.
(2019)
2D Synthesis VGG variant Supervised Indirect Super-resolution STSN
Di et al. (2019) 2D Classication Custom CNN Supervised Direct Fault STSN
Wu et al. (2019d) 2D, 3D Classication Custom CNN Supervised Direct Fault, strikes, dips MTSN-
MCC
Egorov (2019) 3D Segmentation Vnet variant Supervised Direct Fault STSN
Wu et al. (2019c) 3D Segmentation UNet variant (ResUNet) Supervised Direct Fault, Noise reduction,
seismic normal vectors
MTSN-
MTL
Cunha et al.
(2020)
2D Classication Custom CNN Transfer learning Direct Fault STSN
Alfarhan et al.
(2020b)
2D Segmentation UNet variant Supervised Direct Fault, Salt MTSN-
MCC
Wang and Ma
(2020)
3D Classication VGG variant Supervised Direct Fault STSN
El Zini et al.
(2020)
2D Segmentation Autoencoder variant Transfer learning Direct Bright spot, Fault, Facies STSN
Liu et al. (2020b) 2D Segmentation Custom CNN Supervised Direct Fault STSN
Wu et al. (2020) 3D Segmentation UNet variant Supervised Direct Fault, Horizon, RGT
volumes
STSN
Liu et al. (2020a) 3D Segmentation UNet variant (ResUNet) Supervised Direct Fault STSN
Di et al. (2020) 3D Segmentation UNet variant Supervised Direct Fault, Stratigraphy STMN
Alfarhan et al.
(2020a)
2D Segmentation UNet variant (ResUNet) Transfer learning Direct Salt, Fault MTSN-
MCC
Aribido et al.
(2020)
2D Classication GAN variant Unsupervised Semidirect Chaotic-Horizon, Faults,
Horizon, Salt
MTSN-
MCC
Mosser et al.
(2020)
3D Segmentation UNet variant (BNN) Supervised Direct Fault STSN
Manral (2020) 2D Segmentation UNet variant Supervised Direct Fault STSN
Jiang and
Norlund
(2020)
3D Synthesis GAN Supervised Indirect Super-resolution STSN
Zhou et al.
(2020)
3D Instance
Segmentation
UNet variant Supervised Direct Fault STSN
An et al. (2020) 2D Segmentation, Edge
detection
UNet variant, Deeplab, RCF
variant
Supervised Direct Fault STSN
da Silva et al.
(2021)
2D Classication LeNet variant Supervised Semidirect Chaotic-Horizon, Faults,
Horizon, Salt
MTSN-
MCC
Shi et al. (2021) 2D Instance
Segmentation
UNet variant Supervised Direct Salt, Fault STSN
Wu et al. (2021) 3D Segmentation Custom CNN Supervised Direct Fault STSN
Di et al. (2021) 2D Segmentation UNet variant Supervised Direct Fault, Facies, Stratigraphy STSN
Feng et al. (2021) 2D Segmentation UNet variant (BNN) Supervised Direct Fault STSN
Yan et al. (2021) 2D Segmentation UNet variant Transfer learning Direct Fault STSN
Wang et al.
(2021a)
3D Segmentation UNet variant (ResUNet) Supervised +
knowledge distillation
Direct Fault STSN
Wrona et al.
(2021)
2D Classication,
Segmentation
Custom CNN, UNet Supervised Direct Fault, Salt, Horizon STSN
An et al. (2021) 2D Segmentation, Edge
detection
UNet variant, Deeplab, HED
variant, RCF variant
Supervised Direct Fault STSN
Bi et al. (2021) 3D Segmentation UNet variant Supervised Semidirect RGT volumes, Horizon,
Fault
STSN
Ao et al. (2021) 3D Segmentation U-Net, SegNet variant Supervised Semidirect curvature STSN
Ahmad and Tsuji
(2021)
3D Segmentation UNet variant Supervised Direct slump, Faults STSN
Aribido et al.
(2021)
2D Classication GAN variant Unsupervised Semidirect Chaotic-Horizon, Faults,
Horizon, Salt
MTSN-
MCC
Ma and Li (2021) 3D Segmentation UNet variant Supervised Direct Fault STSN
(continued on next page)
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
8
identied four DL learning categories, namely supervised learning (46,
83.9%), unsupervised learning (2, 3.5%), transfer learning (7, 12.5%)
and knowledge distillation (1, 1.8%). As shown in Fig. 3e, at the
beginning of this area, only supervised learning methods were used, but
from 2020 onwards, other learning methods gradually emerged.
Nevertheless, supervised learning methods are still dominant
throughout these years.
From a geoscientic perspective, these articles can be classied into
three categories based on whether they directly provide fault probability
maps: direct, semi-direct and indirect. As shown in Fig. 3f, the majority
of articles (45, 80.4%) presented DL models that directly provided fault
probability maps, followed by a minority (7, 12.5%) that highlighted
fault-rich zones or fault-related attributes, and a few (4, 7.1%) articles
that focused on improving the quality of seismic data and indirectly
assisting fault interpretation. Of the direct approaches, nearly three-
quarters (33, 73.3%) focused solely on faults, while others (12,
26.7%) provided approaches to faults and other targets.
Furthermore, based on the number of geo-targets directly predicted
by each DL-based approach and the number of neural networks used for
one target, we propose a new classication, namely the target-model
category. They are Single Target Single Network (STSN), Multiple
Target Single Network for Multi-Class Classication (MTSN-MCC),
Multiple Target Single Network for Multi-Task Learning (MTSN-MTL),
and Single Target Multiple Network (STMN). As illustrated in Fig. 3g,
the vast majority (46, 82%) of articles were in the STSN category, fol-
lowed by a minority in the MTSN-MCC, STMN and MTSN-MTL cate-
gories. There is an increasing trend to use complex DL models, such as
ensemble models for fault interpretation (STMN) or single neural net-
works to predict multiple geo-targets (MTSN).
In summary, 2D methods are preferred over 3D methods. Most
selected articles consider fault interpretation an image segmentation
task and use UNet or its variants as the network structure. As far as
learning methods are concerned, the vast majority of these methods are
trained in a supervised manner, and other learning methods are just
emerging. A signicant proportion of these selected articles (45, 80.4%)
can produce fault probability maps directly. Finally, although the most
common approach is to use a single neural network to predict faults,
there are signs of using complex DL models to predict multiple geo-
targets or employing ensemble neural networks.
4.2. RQ2: What datasets were used in this area?
We extracted 73 seismic datasets used in the nal 56 articles and
summarised them into eld datasets and synthetic datasets, as shown in
Tables 3 and 4. A signicant proportion (47, 64.4%) of them are eld
datasets, while others are synthetic datasets. Among all 73 datasets, only
15 (20.5%) datasets published their seismic data, including one article
that published its code for generating the synthetic data. Although most
(50, 68.5%) of these datasets have labels, only 7 (9.6%) datasets made
their label open-accessible (OA). Labelled datasets can be used as
training sets in supervised learning models and can also be used to verify
the accuracy of the models. Because fault was selected as a keyword in
the search command, the most common label is fault labels. DL tasks
were also included in the table to show each seismic dataset’s utility and
the label’s granularity. Image segmentation appears to be the dataset’s
most popular DL task (42, 57.5%). Most seismic datasets are 3D (53,
72.6%), while others are in 2D form. These 73 datasets are used 123
times, about 1.68 times per article. As shown in Fig. 4b, more than a
third (21, 37.5%) of the 56 articles use two datasets, followed by 18
(32.1%) articles that use one dataset. Wu et al. (2019a) uses seven
datasets in their article, the most datasets included in any study. About
two-fth (29, 39.7%) of all datasets did not provide their origin (11,
15.1%) or generating code (18, 24.7%).
For eld seismic datasets, they (count =47) are used 85 times in 52
articles. A minority of these eld datasets (11, 23.4%) are provided as
OA seismic data, which are used 47 (50.6%) times, whereas there is a
fourfold relative decrease in the use of non-OA eld seismic datasets (36
datasets are used only 38 times). More than half (24, 51.1%) of the eld
datasets were labelled, but only three datasets made their labels OA.
These three label-OA datasets were used on average 3.7 times, which
was more popular than the other 21 label non-OA datasets that were
used on average 1.2 times. The vast majority (37, 78.7%) of the eld
datasets are 3D in shape, while the rest are 2D in shape. Although 3D
datasets are the majority, it does not mean that most researchers use 3D
seismic volumes as input for their model since they can also use 2D
patches from 3D datasets. Besides our main target of this survey, i.e.,
fault (23, 48.9%), one article takes “fault, horizon, chaotic and salt” (1,
2.1%) as the target of the dataset, while others are not applicable (23,
48.9%). For the DL task, most eld datasets are used for image seg-
mentation (20, 42.6%) tasks, only three are used for image
Table 2 (continued )
Reference Dimension DL Task DL Model DL Cat FI Cat Geo-target TM Cat
Zhou et al.
(2021)
2D Classication Custom CNN +DANN Transfer learning Direct Fault STSN
Gao et al.
(2022b)
2D, 3D Segmentation UNet variant Supervised Direct Fault STSN
Wei et al. (2022) 2D Classication LeNet variant Transfer learning Direct Fault STSN
Michie et al.
(2022)
2D Segmentation UNet variant Supervised Direct Fault STSN
Smith et al.
(2022)
2D Segmentation UNet variant Supervised Direct Fault STSN
Dou et al. (2022) 3D Segmentation UNet variant Supervised Direct Fault STSN
Li et al. (2022b) 3D Segmentation UNet variant Supervised Direct Fault STMN
Li et al. (2022a) 2D Synthesis UNet variant Supervised Indirect Super-resolution, Noise
reduction
STSN
Wang et al.
(2022)
2D Classication,
Segmentation
Custom CNN, UNet variant Supervised Direct Fault STMN
Gao et al.
(2022a)
2D, 3D Segmentation UNet variant (ResUNet) Supervised Direct Fault STSN
Ao et al. (2022) 3D Segmentation UNet variant Transfer learning Semidirect dip STSN
Hu et al. (2022) 2D Segmentation VGG variant Supervised Direct Fault STSN
a
DL task, classication of the DL-based image processing tasks used for fault interpretation in each article.
b
DL model, model architecture category.
c
Geo-target, geological targets covered by each article.
d
Target-model category, classify each article based on the number of geo-targets directly predicted by each DL-based approach and the number of neural networks
used for one target. STSN: Single Target Single Network, MTSN-MCC: Multiple Target Single Network for Multi-Class Classication, MTSN-MTL: Multiple Target Single
Network for Multi-Task Learning, STMN: Single Target Multiple Network.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
9
Fig. 3. Figures for RQ1.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
10
Table 3
Field dataset used in the selected articles with information about name,
a
seismic OA,
b
label,
c
label OA,
d
origin/geo-location, dimension, geo-target,
e
DL task and
articles that used them.
Name Seismic
OA
Label Label
OA
Origin/
Geo-
location
Dimen-
sion
Geo-target DL task Used by
F3 Y
f
N - EU-North
Sea
3D - - Ao et al. (2021, 2022), Aribido et al. (2020, 2021); Cunha
et al. (2020); Feng et al. (2021); Gao et al. (2022b); Ma and
Li (2021); Pochet et al. (2019); Wei et al. (2022); Wu et al.
(2019a, 2020)
F3 with label Y Y N EU-North
Sea
2D Fault Segmentation Alfarhan et al. (2020a); Haroon et al. (2018); Hu et al.
(2022); Lapteva et al. (2019); Wang et al. (2021a);
Alfarhan et al. (2020b)
LANDMASS-1 Y Y Y
g
EU-North
Sea
2D Fault,
Horizon,
Chaotic, Salt
Classication Aribido et al. (2020), Alfarhan et al. (2020b), Aribido et al.
(2021), da Silva et al. (2021), Alfarhan et al. (2020a)
Great South Basin
(GSB)
Y Y Y
h
New
Zealand
2D Fault Segmentation An et al. (2021), Di et al. (2018a), Cunha et al. (2020), Di
et al. (2019)
Thebe Y Y Y
i
Australia 3D Fault Segmentation An et al. (2021), An et al. (2020)
Groningen PSDM
survey
Y
j
Y N Com/
Institue
3D Fault Segmentation Smith et al. (2022)
Opunake 3-D
survey
Y
k
N - New
Zealand
3D - - Gao et al. (2022a,b); Jiang and Norlund (2020); Wu et al.
(2019a,d, 2020)
Canning Basin Y
l
N - Australia 3D - - Mosser et al. (2020)
Utah FORGE Y
m
N - US 3D - - Gao et al. (2022a)
Kerry-3-D Y
n
N - New
Zealand
3D - - Ao et al. (2022); Dou et al. (2022); Gao et al. (2022a,b); Li
et al. (2022b); Wu et al. (2019a, 2020); Yan et al. (2021)
BroadSeis’s 2D
Field
Y
o
N - Com/
Institue
2D - - Li et al. (2022a)
Lu’s Field N Y N Mexico 3D Fault Super-
resolution
Lu et al. (2018)
GN1101 survey N Y N EU-North
Sea
3D Fault Segmentation Michie et al. (2022)
Xiong’s Field N Y N Com/
Institue
3D Fault Segmentation Xiong et al. (2018)
Wang’s 3D Field N Y N China 3D Fault Classication Wang et al. (2022)
Dou’s 3D Field N Y N China 3D Fault Segmentation Dou et al. (2022)
Opunake 3-D
survey Di’s label
N Y N New
Zealand
3D Fault Segmentation Di et al. (2020)
Exmouth basin N Y N Australia 3D Fault Segmentation Manral (2020)
NH0301 N Y N EU-North
Sea
3D Fault Segmentation Di et al. (2021)
Xiong’s Field # 1 N Y N Com/
Institue
3D Fault Segmentation Xiong et al. (2018)
GX N Y N China 3D Fault Segmentation Zhou et al. (2021)
Yuan’s Field N Y N China 2D Fault Segmentation Yuan et al. (2019)
Wrona’s Field N Y N EU-North
Sea
2D Fault Classication Wrona et al. (2021)
Unknown Field 3D
- with label
N Y N - 3D Fault Segmentation Wang and Ma (2020), Xiong et al. (2018), Guitton (2018),
Lu et al. (2018)
Unknown Field 2D
- with label
N Y N - 2D Fault Segmentation Liu et al. (2020b), da Silva et al. (2021), Chang et al. (2019)
Beatrice N N - UK 3D - - An et al. (2021)
Campos Basin N N - Brazil 3D - - Wu et al. (2019a,d)
Zheng’s eld N N - Mexico 3D - - Zheng et al. (2019)
Wu’s Field # 1 N N - Brazil 3D - - Wu et al. (2019c)
Yan’s Field N N - China 3D - - Yan et al. (2021)
Bozhong Sag N N - China 3D - - Liu et al. (2020a)
Costa Rica margin N N - Costa Rica 3D - - Wu et al. (2019a)
Indian 3D 2000
MSS
N N - Australia 3D - - Jiang and Norlund (2020)
Wu’s Field N N - China 3D - - Wu et al. (2021)
Soda Lake N N - US 3D - - Bi et al. (2021); Gao et al. (2022b)
Clyde Petroleum
Plc.
N N - Com/
Institue
3D - - Wu et al. (2019a)
BZ eld data N N - China 3D - - Li et al. (2022b)
Kumano forearc
basin
N N - Japan 3D - - Ahmad and Tsuji (2021)
Unknown Field 3D
- without label
N N - - 3D - - Bi et al. (2021), Wu et al. (2020), Wu et al. (2019c)
Unknown Field 2D
- without label
N N - - 2D - - Wu et al. (2019d)
a
For datasets with no name provided, we name them by the rst author’s name. If the same author proposes more than one dataset, # is added to distinguish.
Datasets that do not provide origin or geo-location are merged for better illustration.
b
Seismic OA, if seismic data is open accessed. Y for yes and N for No. Links for OA datasets are also provided.
c
Label, if annotations are provided along the seismic data.
d
Label OA, if the annotations are open accessed.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
11
e
Geo-target, what geological targets are annotated for the seismic dataset.
f
https://github.com/xinwucwp/osv,https://terranubis.com/datainfo/F3-Demo-2020.
g
https://github.com/olivesgatech/LANDMASS.
h
https://github.com/augustoicaro/SFD-CNN-TL/tree/master/data.
i
https://doi.org/10.1016/j.dib.2021.107219.
j
https://geo.public.data.uu.nl/vault-nam-geological-model/Publication[1605778324].
k
https://wiki.seg.org/wiki/Opunake-3D.
l
https://researchdata.edu.au/canning-basin-gravity-gravity-point/1358825?source =suggested_datasets.
m
https://gdr.openei.org/submissions/1015.
n
https://wiki.seg.org/wiki/Kerry-3D.
o
One picture from literature Soubaras et al. (2012).
Table 4
Synthetic dataset used in the selected articles with information about name, seismic OA, label, label OA, origin,
a
dimension, geo-target, DL task and articles that used
them.
Name Seismic
OA
Label Label
OA
Origin Dimen-
sion
Geo-target DL task Used by
FaultSeg3D Y Y Y
b
FaultSeg3D 3D Fault Segmentation Ahmad and Tsuji (2021); Dou et al. (2022); Egorov (2019);
El Zini et al. (2020); Feng et al. (2021); Gao et al. (2022a);
Li et al. (2022b); Liu et al. (2020a); Ma and Li (2021);
Wang et al. (2022); Wu et al. (2019a,d, 2021)
Bi’s 3D synthetic Y Y Y
c
FaultSeg3D 3D Fault, RGT Segmentation Bi et al. (2021)
Wu’s 2D SR Y Y Y
d
FaultSeg3D 2D Super-
resolution
Synthetis Li et al. (2022a)
Pochet’s 2D
synthetic
Code Y Code
e
Hale’s method 2D Fault Classication Pochet et al. (2019)
Synthetic 3D using
Faultseg 3D
method
N Y N FaultSeg3D 3D Fault Segmentation Ao et al. (2021, 2022); Gao et al. (2022b); Mosser et al.
(2020); Shi et al. (2021); Wu et al. (2019c, 2020); Zhou
et al. (2020)
Wu’s 2D RGT N Y N FaultSeg3D 3D RGT Segmentation Wu et al. (2020)
GX synthetic N Y N The feature of
GX
2D Fault Segmentation Zhou et al. (2021)
Cunha’s synthetic N Y N Hale’s method
& IPF code
2D Fault Segmentation Cunha et al. (2020)
Wang’s Synthetic N Y N FaultSeg &
Hale’s method
3D Fault Segmentation Wang et al. (2021a)
Synthetic 3D - with
label - single
target
N Y N Unknown 3D Fault Segmentation Yan et al. (2021), Xiong et al. (2018), Jiang and Norlund
(2020), Ahmad and Tsuji (2021)
Synthetic 2D - with
label - multi
target
N Y N Unknown 2D Fault, Dip,
Strike
Classication Wu et al. (2019d); Zheng et al. (2019)
Fault,
Horizon
Segmentation Haroon et al. (2018)
Synthetic 2D - with
label - single
target
N Y N Unknown 2D Fault Segmentation Chang et al. (2019); Gao et al. (2022b); Wei et al. (2022)
a
Origin, method used for generating the synthetic dataset
b
https://github.com/xinwucwp/faultSeg.
c
https://doi.org/10.5281/zenodo.4536561.
d
https://github.com/JintaoLee-Roger/SeismicSuperResolution.
e
https://github.com/dhale/ipf.
Fig. 4. Figures for RQ2.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
12
classication, one for super-resolution, and others are not applicable.
Datasets for image segmentation provide pixel-level labels, while data-
sets for image classication provide image-level labels. Datasets without
labels are marked as not applicable on target and DL tasks because they
can only be used in the test process. Datasets that have the origin or geo-
location information (36, 76.6%) are used 74 (87.1%) times, with their
origin/geo-location information provided in Fig. 4a. Eleven geo-
locations showed in these datasets with China (8, 17.0%) and EU-
North Sea (6, 12.8%) being the primary geo-location of the eld data-
sets, with 5 (10.6%) of them described as being from companies or in-
stitutions. Some datasets (11, 23.4%) have no origin or geo-location
information, potentially because of privacy or security policies.
Besides, we found that 26 synthetic datasets are used 38 times in 34
articles. Only four (15.4%) datasets were OA, but they were responsible
for almost half (16, 42.1%) of the total synthetic dataset usage. The most
frequently used (13, 34.2%) dataset is Wu’s dataset (Wu et al., 2019a)
called FaultSeg. This dataset simulates the eld seismic signal pattern,
provides a detailed explanation and makes its dataset with code OA.
FaultSeg and similar datasets generated using FaultSeg’s method are,
therefore, used for more than half (25, 65.8%) of total synthetic dataset
usage. The second most popular (used by 3 times, 7.9%) dataset was
based on Hale’s method (Hale, 2013). It might be because Hale pub-
lished their code to generate the dataset. The last two publicly available
datasets were a 2D seismic super-resolution (SR) dataset presented by Li
et al. (2022a) and a 3D seismic fault and RGT dataset presented by Bi
et al. (2021). Both of them are used for just one time. The average use
time of each OA synthetic dataset is 4.0, much larger than 1.0 for the
non-OA synthetic datasets. In contrast to eld datasets, all synthetic
datasets have labels. Almost all of these datasets are used for image
segmentation (22, 84.6%), but only a small number are used for image
classication (3, 11.5%) and image synthesis (1, 3.8%). Besides the
target shown in our keyword, i.e., fault (20, 76.9%), other target in-
cludes super-resolution (1, 3.8%), RGT (1, 3.8%) and “fault and others”
(4, 15.4%). A large proportion (16, 61.5%) of these synthetic datasets
are 3D datasets, while others are 2D datasets.
To summarise RQ2, we found only three eld datasets and four
synthetic datasets open-sourced both seismic data and labels. OA ap-
pears to be the primary attribute affecting datasets’ usage frequency.
Seismic datasets with the label OA are used 3.9 times more on average,
than non-OA datasets, perhaps because OA datasets facilitate bench-
marking. Among the 56 articles, a few use only one synthetic dataset to
train, validate and test their model, which is (3, 5.4%). 3D datasets are
more popular than 2D ones because most eld datasets are 3D in their
original form. Wu’s 3D dataset FaultSeg (Wu et al., 2019a) or datasets
generated using FaultSeg’s code are the most popular synthetic datasets.
4.3. RQ3: What are the reported benets of using DL in fault
interpretation?
Numerous studies in the surveyed articles have highlighted the
benets of using DL techniques for fault interpretation, which can be
broadly categorized into two main aspects: the superior learning and
summarization capability, and the excellent usability of DL models
compared to conventional methods including traditional ML methods.
These advantages primarily stem from the inherent design advantages of
DL methods rather than their specic benets in fault tasks (LeCun et al.,
2015; Li et al., 2022b; Cui et al., 2022; Kumar and Mandal, 2018).
Regarding the powerful summarization capability, DL algorithms are
designed to learn and extract appropriate features for solving target
tasks through large-scale data and a considerable number of parameters,
without the need for extensive domain knowledge or manual feature
engineering as required by traditional methods when dealing with
complex seismic data (Li et al., 2022b; LeCun et al., 2015). Since it is a
fully data-driven approach, the automatic feature extraction process
avoids human subjectivity. When high-quality annotations from multi-
ple experienced seismic interpreters are included in the training set, DL
methods theoretically have the potential to generate high-quality and
objective interpretations, thereby mitigating the uncertainties associ-
ated with manual interpretation (Di et al., 2020; Xiong et al., 2018; Lu
et al., 2018). Additionally, DL methods can automatically extract a large
number of features, eliminating the risk of losing information in the
manual feature extraction process and reducing the chance of artifacts or
misinterpretation associated with fault prediction (Cui et al., 2022;
Kumar and Mandal, 2018). Moreover, DL methods have the capability of
continuous learning and self-optimization with model performance
continuously evolving with more data from multiple interpreters
(Haroon et al., 2018).
Unlike conventional fault interpretation methods, which typically
operate at the 1D sample level or rely on human vision that is particu-
larly effective at nding 2D features, DL-based methods can simulta-
neously consider a range of 2D or 3D samples around a particular
sample, extracting high-dimensional and complex abstract features. This
characteristic reduces sensitivity to noise or artifacts (Yuan et al., 2019;
Guitton, 2018; Di et al., 2019). Consequently, DL methods can produce
smoother and more continuous fault interpretation results compared to
traditional approaches (Liu et al., 2020a; Wrona et al., 2021; Cui et al.,
2022; Ahmad and Tsuji, 2021; Mosser et al., 2020), particularly for noisy
seismic data (El Zini et al., 2020). Lu et al. (2018) argue that this feature
of DL methods is particularly suitable for solving earth science problems
that are not easily formulated by mathematical models or deterministic
algorithms. Furthermore, DL methods can accomplish tasks that tradi-
tional ML methods cannot, as demonstrated by the ability to predict
multiple geological targets simultaneously (Wu et al., 2019c,d) and
reconstruct complex seismic features (Lu et al., 2018).
In addition to the superior learning and summarization capabilities,
several other usability-related benets have been reported. DL methods
offer advantages in terms of fast, relatively inexpensive, repeatable, and
parallel processing of seismic data using powerful computational re-
sources, reducing the need for extensive human–computer interaction (i.
e. manual interpretation on professional software) compared to tradi-
tional methods (Chang et al., 2019; Ao et al., 2021). The neural network
can directly read raw seismic data and produce fault probability maps,
saving signicant pre-processing and post-processing time (Hu et al.,
2022; Wu et al., 2019d). Since trial and error on feature engineering is
eliminated (An et al., 2021; Manral, 2020; Feng et al., 2021; Zheng et al.,
2019), DL-based models can be applied to other seismic datasets without
the need for feature adaptation to different geological and geophysical
conditions (Kumar et al., 2019; Lapteva et al., 2019).
The DL approach can also assist interpreters in improving fault
interpretation workow. Firstly, it can serve as a useful starting point for
geologists to understand and interpret fault zones without constructing
geological models, allowing them to focus more on fault-rich areas
(Michie et al., 2022). DL methods can generate relatively accurate
probabilistic fault maps with reduced noise, making false positives
extraordinarily valuable (Xiong et al., 2018), and even highlighting
hidden irregularities that manual interpreters may have overlooked
(Michie et al., 2022). The probabilistic output nature of DL methods
enables geologists to make further assessments regarding uncertainty
and geological risk (Egorov, 2019). With all these associated benets, DL
models can serve as a valuable additional validation step in standard
fault interpretation workows (Smith et al., 2022).
4.4. RQ4: What are the identied challenges and candidate solutions in
this area?
A total of 12 challenges (C1-12) are reported and broadly classied
into four categories: data related, DL related, evaluation related and
practical use related. Each of them is described in detail, with candidate
solutions suggested by the included articles.
4.4.1. data-related C1–4
The most frequently discussed challenge (C1) is the lack of sufcient
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
13
annotated datasets for training DL models (Hu et al., 2022; Li et al.,
2022a; Pochet et al., 2019; Wang et al., 2022; Ma and Li, 2021; Dou
et al., 2022; An et al., 2021; Wei et al., 2022; Michie et al., 2022; Bi et al.,
2021; Yan et al., 2021; Wu et al., 2020; Lu et al., 2018; da Silva et al.,
2021; Alfarhan et al., 2020a; El Zini et al., 2020; Manral, 2020; Smith
et al., 2022). DL models usually require a large amount of data to train
their millions of parameters (Wang et al., 2020). However, labelling
seismic datasets is still largely manual and requires trained seismic ex-
perts’ interpretation. It is labour-intensive, time-consuming and
expensive. Since seismic data are generally 3D volumes and faults in
seismic data are continuous surfaces, the fault labels on adjacent 2D
slices of seismic data are almost identical. Intensive annotation on each
2D slice is even more expensive and may not signicantly improve the
accuracy of the DL model (Dou et al., 2022; Bi et al., 2021). Worse,
expert annotations are often considered intellectual property by oil
companies and are not made public. Thus, although seismic data is now
available in megabytes to terabytes and keeps growing, domain re-
searchers often only have access to a small labelled dataset (Alfarhan
et al., 2020b; Feng et al., 2021; Di et al., 2020). Usually, a few labelled
2D slices or even have no access to labelled seismic dataset (Yuan et al.,
2019).
Several studies in the literature have addressed challenge C1 in a
variety of ways including: (i) open-sourcing a large eld seismic dataset
(An et al., 2021), (ii) generating realistic synthetic seismic datasets (Bi
et al., 2021; Liu et al., 2020a; Ma and Li, 2021; Wu et al., 2019a, 2020,
2021), (iii) applying advanced ML methods (like semi-supervised
learning (Dou et al., 2022), (iv) unsupervised learning (Aribido et al.,
2020; Aribido et al., 2021), (v) transfer learning (Alfarhan et al., 2020a;
El Zini et al., 2020; Ao et al., 2022), (vi) knowledge distillation (Wang
et al., 2022) and (vii) model compression.
Although a large open-source eld seismic dataset is presented by An
et al. (2021), it does not fully address all data-related issues because of
other challenges (C2 and C3). One issue is that the seismic data is often
of poor quality (C2), which strongly limits the fault prediction accuracy.
For example, due to the low signal-to-noise ratio of the model training
data, some DL models predict faults with an insufcient resolution,
resulting in thick or blurry zones with noisy and discontinuous ‘faults’
(Gao et al., 2022a,b; Jiang and Norlund, 2020; Li et al., 2022a; Wang
et al., 2021a). Even seismic data that appear to be of good quality could
have signal disturbance zones (An et al., 2021), with a degradation of
quality due to complex and diverse geological conditions and the
different capabilities of equipment for seismic acquisition (Dou et al.,
2022). This challenge can be addressed by applying more advanced DL
model architecture (Gao et al., 2022a,b) or implementing seismic
denoising algorithms (Lu et al., 2018).
There are also uncertainties in seismic fault interpretation, arising
from data errors, interpretation errors, geological uncertainties, spatial
uncertainties and model uncertainties (C3) (Feng et al., 2021). One of
the well-recognised uncertainties is the uncertainty in fault annotation.
As faults are often manually annotated by experts, annotations may be
incorrect (Wu et al., 2020; Liu et al., 2020a), omitted due to time con-
straints or limited by the specic research focus of a study (An et al.,
2021), and/or the subjective judgements informed by personal experi-
ence (Ahmad and Tsuji, 2021; Mosser et al., 2020; Chang et al., 2019;
Smith et al., 2022). Besides, uncertainties in fault annotation are nega-
tively correlated with the quality of seismic data (An et al., 2021; Ao
et al., 2022; Faleide et al., 2021b). Expert annotations are not, therefore,
the equivalent to ground-truthing but are, to some extent, subjective. DL
models trained on uncertain annotations will inevitably lead to uncer-
tainty in the predicted results (Wang et al., 2022), compromising the
upper limit of accuracy of DL models (Ao et al., 2021; Ao et al., 2022;
Manral, 2020; Lu et al., 2018). Although Guitton (2018) recommends
using the algorithm by Hale (2013) to generate fault annotations in
order to avoid manual annotation uncertainty, Dou et al. (2022) argue
that the accuracy of Hale’s method limits the DL methods trained using
automatic fault annotations. Whatever the circumstances, it is essential
to consider the inherent uncertainties in model interpretation results
(Michie et al., 2022; Smith et al., 2022). As a potential candidate solu-
tion to challenge C3, An et al. (2021) ltered out image patches with
fault labels less than a threshold when preparing the training dataset
because they assumed that all labelled faults were correct and some
faults were missed.
Due to these three challenges (C1-3), the literature suggests using
synthetic seismic datasets as an alternative to training DL models.
However, this leads to another challenge (C4): synthetic seismic datasets
can not fully replace all eld seismic datasets (Wei et al., 2022; Yan
et al., 2021), and models trained using synthetic datasets did not
generalise well to other eld datasets (Dou et al., 2022; Yan et al., 2021;
Wu et al., 2019c; Zheng et al., 2019; Ao et al., 2021). Ao et al. (2021)
argue that this might be because synthetic datasets are simulated with
idealised functions, whereas some suggest this is due to the poor
generalisation of the data-driven DL-based method (Ahmad and Tsuji,
2021). DL-based methods tend to overt synthetic datasets, giving un-
satisfactory results when the eld seismic data differ signicantly from
the trained synthetic data. Cunha et al. (2020) argue that a long tuning
step is required due to CNN’s enormous parameters. Several approaches
have been proposed to address this challenge (C4), including invoking
principles such as transfer learning (Wei et al., 2022; Yan et al., 2021;
Yuan et al., 2019; Cunha et al., 2020; Ao et al., 2021) and progressive
learning (Wu et al., 2019c), by applying model structures with greater
generalisation capabilities (Ma and Li, 2021), implementing lightweight
models (da Silva et al., 2021), and designing appropriate pre-processing
(Wu et al., 2019c; Xiong et al., 2018; Zheng et al., 2019).
4.4.2. DL-related C5-7
One of the candidate solutions for data-related challenges C1, C2,
and C4 is to use advanced DL approaches. However, DL models struggle
with high computational costs (C5). DL models typically use numerous
parameters to learn useful features from the input data. Compared to
conventional algorithms, DL algorithms require longer training time (Hu
et al., 2022; Pochet et al., 2019; Zhou et al., 2021), considerable
computational resources (da Silva et al., 2021; Di et al., 2019), storage
resources and especially GPU resources (An et al., 2020). Besides, some
literature extracts 2D images or 3D cubes around a particular seismic
sample and classies whether a fault has passed through the central
point (Wei et al., 2022; Wu et al., 2019d; Zheng et al., 2019; Di et al.,
2018a). For fault interpretation on vast 3D seismic data, this process is
computationally inefcient, has high memory costs, and dramatically
increases the post-processing effort (Hu et al., 2022; Wrona et al., 2021;
Wu et al., 2019a). In response to C5, Hu et al. (2022), Wrona et al.
(2021) suggests using the image segmentation type of the DL models
instead of the image classication models and da Silva et al. (2021)
suggest using a lightweight CNN to reduce computational cost.
Another drawback of DL methods that is often criticised is the lack of
interpretability (C6) (Liu et al., 2020b; Wrona et al., 2021). DL models
are black-box models, and because experts do not understand how the
labels are predicted, they may not fully trust them. Liu et al. (2020b)
suggest adding a Class Activation Map (CAM) to guide the DL model to
focus on pixels around fault pixels labelled by experts.
Unlike regular image process tasks, fault interpretation is an
extremely unbalanced binary classication problem from a DL
perspective (An et al., 2021; El Zini et al., 2020; Gao et al., 2022b;
Haroon et al., 2018; Wei et al., 2022; Wu et al., 2019c, 2021). Unbal-
anced classication (C7) is a challenge for DL methods, as DL is usually
based on the assumption of evenly distributed classes (Wang et al.,
2016). Without any modications, the model performs poorly for un-
balanced classication. Several countermeasures have been proposed in
the literature, such as resampling the training dataset or using specic
loss functions (e.g. weighted binary cross-entropy loss (An et al., 2021;
Gao et al., 2022b), focal loss (Wei et al., 2022)).
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
14
4.4.3. Evaluation-related C8-9
Due to the uncertainties introduced in challenge C2, fault interpre-
tation is still mainly based on expert evaluation. Much of the literature
points out the lack of appropriate evaluation criteria for DL-based fault
interpretation algorithms (C8). Currently, no proper evaluation criteria
can correctly dene model performance in terms of accuracy, uncer-
tainty, and generalisability. It is challenging to select an optimal algo-
rithm without such an evaluation standard (Li et al., 2022b). Faults can
be seen as edges of the background or thin objects, and it causes data
imbalance that makes general computer vision evaluation methods such
as accuracy inappropriate (see challenge C7). Moreover, a slight
displacement of fault annotation may result in a considerable accuracy
difference (see challenge C2).
To address the uncertainty evaluation issue, Guillon et al. (2020)
proposes an uncertainty-aware metric, which calculates the intersection
over union (IoU) of prediction and ground-truth with a tolerance dis-
tance. Similarly, An et al. (2021) adopted an edge detection evaluation
method that calculates precision, recall, accuracy, and F1 score with a
tolerance distance. Di et al. (2018a) pointed out that some researchers
have tested their models on seismic data partially involved in the
training process and not on independent test data, circumstances
dictating that their evaluations can not correctly justify their model
performance.
Some literature has also reported the lack of open access (C9), in
research within the eld often not publishing their data, code, or
workows (see RQ2) (da Silva et al., 2021), making it extremely difcult
for other researchers to replicate, benchmark and make improvements
(Wrona et al., 2021; An et al., 2021).
4.4.4. Practical use-related C10-12
In addition to the three types of challenges mentioned above, some
articles have identied several challenges preventing the use of pro-
posed DL-based automated fault interpretation models in the industry.
Wrona et al. (2021) points out that DL-based methods are not geologist
friendly and are almost exclusively known to DL experts (C10). Lapteva
et al. (2019), Di et al. (2021) also argue that current DL-based fault
interpretation methods directly analyse input seismic data and omit
known geological constraints (C11). Introducing this knowledge
appropriately into DL models still needs to be investigated. More
importantly, current DL-based fault interpretation methods can, at best,
predict fault probability maps. However, fault probabilities are not
comparable to fault interpretations (C12) (Shi et al., 2021; Smith et al.,
2022; Wu et al., 2019d; Zhou et al., 2020). Polylines and triangle sur-
faces must be digitised from probabilistic maps and stored in a specic
le format for incorporation into current industry workows (Smith
et al., 2022).
As an attempt to address challenges C11 and C12, Wu et al. (2019c,d)
incorporate domain knowledge from the geology eld to predict fault
probabilities and two other fault-related targets with a single CNN. For
example, Wu et al. (2019d) simultaneously predict fault probability,
strikes and dips, while Wu et al. (2019c) predict fault probability, clean
seismic data, and seismic normal vectors. By contrast, Zhou et al. (2020),
Shi et al. (2021) identify and differentiate each fault from the seismic
data using a DL model based on instance segmentation to simulate in-
dustry fault interpretation workows.
5. Discussion
In this section, we provide an in-depth discussion of each research
question. For the four challenges groups summarised in RQ4, we start
our discussion with the least researched group.
5.1. RQ1
All selected 56 articles use CNN-based image processing methods.
CNN is the predominant image-processing neural network in DL.
However, recent breakthroughs, such as visual transformers, have
shown excellent performance in several image-processing tasks. Besides,
seismic data is also time series data. It may be worthwhile to investigate
methods based on something other than CNNs like RNNs, LSTMs, and
Transformers that specialise in processing sequential data.
The current approach is overly concentrated in some categories, such
as taking fault interpretation as an image segmentation task, using UNet
and its variants. These distributions may be because Wu et al. (2019a)
released their synthetic dataset and UNet-based baseline code in 2019.
However, the UNet model was rst published in 2015, while the most
popular ResUNet variant was published in 2017. Since then, various
advanced model architectures have been developed, which may provide
better results in fault interpretation tasks.
In addition, we’ve noticed that the utilization of 3D methods in fault
interpretation has been relatively limited compared to their 2D coun-
terparts. However, recent studies (e.g. Dou et al., 2022; Bugge et al.,
2018) have highlighted the signicant advantages of employing 3D
seismic data for fault extraction, resulting in more robust and continuous
fault interpretation. These studies emphasize that 3D seismic data allows
for the capture of essential 3D spatial morphological features of faults,
facilitating a more comprehensive understanding of fault structures.
Additionally, considering the inuence of fault dip on fault visibility,
relying solely on a 2D approach may lead to the exclusion of faults that
are not observable from a single perspective (Faleide et al., 2021a).
However, it is important to note that training 3D deep learning models
has its limitations, as it introduces computational and labeling re-
quirements. To address these challenges, intermediate approaches
known as 2.5D methods have been investigated (Dou et al., 2022; Lin
et al., 2022). These methods involve training a 3D convolutional neural
network (CNN) using only a few labelled seismic slices, striking a bal-
ance between capturing the benets of 3D information and managing
computational costs. Given these considerations, future research efforts
should focus on developing techniques that leverage the advantages of
3D seismic data while taking into account computational efciency.
Considering the data-related challenges presented in RQ4, other
learning methods such as transfer learning, semi-supervised learning,
unsupervised learning and self-supervised learning should be one of the
next research focus. In recent years, pretrained models have gained
attention and shown improvements across various domains, including
fault interpretation (Du et al., 2022a). Several studies, categorized as
transfer learning in Table2, all leverage pretrained and ne-tune tech-
niques. Moreover, An and Dong (2023) investigated the impact of prior
knowledge in pretrained models on DL-based fault interpreters. Their
ndings highlighted the common use of synthetic seismic datasets by
domain researchers and emphasized the importance of considering an-
notations in the target dataset when selecting the source dataset.
Nevertheless, there is a lack of widely available general-purpose
pretrained models in the eld of fault interpretation and seismic anal-
ysis, similar to ImageNet in computer vision or the GPT series in natural
language processing. One of the reasons is the limited access to diverse
seismic datasets necessary for training such models. Notably, Meta has
recently made a notable stride in image segmentation, claiming the
ability to accurately “cut out” any object in an image with a single click
(Kirillov et al., 2023). However, when applied to fault interpretation
tasks, its performance is hindered by the substantial disparities between
the training dataset and seismic images. Nevertheless, this highlights the
ongoing trend towards the development of versatile and generalized
models.
5.2. RQ2
The performance of DL models can be signicantly degraded if un-
suitable datasets are used in the training process. Unlike natural images,
obtaining eld seismic data is costly in both time and economics.
Considering the labelling uncertainty (see challenge C3), generating
synthetic seismic datasets with pre-dened labels is a good choice.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
15
However, we believe it is inappropriate if only one synthetic dataset is
used in both training and testing processes (3 in 56 articles) because it
can lead to overtting problems and will not reect the actual perfor-
mance of DL models (An et al., 2021). Although using synthetic datasets
is a potential means of avoiding the fault labelling issue, the quality of
the generated seismic data can also signicantly affect the performance
of DL models on actual eld datasets.
Regardless of the researcher’s chosen seismic dataset type, models in
this eld are designed to address issues relating to actual eld seismic
data. In these circumstances, testing a model’s performance on a eld
dataset allows for direct evaluation of the model. This fact also explains
why eld datasets are more often used than synthetic datasets. We have
summarised three publicly available eld datasets, each with their own
advantages and disadvantages. LANDMASS provides relatively large and
accurate image-level labels that can be used to highlight fault-rich re-
gions. The GSB dataset provides only a small number of 2D cross-section
annotations, which is not ideal for training DL models. Thebe dataset
provides considerable pixel-level expert annotations but is limited by its
expert annotation preferences and the labelling of only large faults.
Despite these drawbacks, the existing OA datasets allow researchers to
compare the approach with other models and thus demonstrate its su-
periority or otherwise, circumstances which explain why label-OA
datasets are preferred over other datasets.
Setting a baseline for eld datasets is difcult due to the uncertainty
challenge (see Challenge C3), as we do not have access to ground truth
labels, and using the authors’ labels to evaluate models is also not
entirely convincing. A promising solution to this problem is to
encourage more authors to share their labels and seismic data (see
Challenge C9), which would allow researchers to assess the quality of
the dataset and thus reach a consensus on a fair baseline.
5.3. Controversy in advantages and challenges (RQ3-4)
We notice that two benets are criticised in the challenges section.
Some studies have praised DL methods because they provide a basis for
fault labelling in a fast, cheap and repeatable manner using powerful
computational resources, but others are critical of DL’s high computa-
tional cost. Whilst the labour costs of using DL approaches are low
compared to conventional approaches; their use can be very challenging
when computational resources are limited.
Another controversy is on interpretability over labour cost. Some
favour DL methods over conventional ones because they do not require
feature engineering, and can take seismic data directly as input to the
production of fault probability maps. However, others criticise DL
methods because they are currently almost exclusively used by DL ex-
perts and are not geoscientist friendly.
Both concerns are valid and they are strongly related to the design
difference between DL methods and traditional ML methods. DL is
designed to automatically extract suitable features directly from com-
plex high-dimensional datasets using large amounts of data and pa-
rameters. On the other hand, traditional ML approaches, particularly
when dealing with complex datasets such as images, require feature
engineering. While DL algorithms reduce the need for human inter-
vention, they sacrice algorithm interpretability. The use of complex
models also increases the barrier to entry for non-experts in the eld of
DL. Overall, the choice between DL and ML methods depends on factors
such as data availability, interpretation objectives, computational re-
sources, and performance. Thus, make a strong case for closer collabo-
ration and the involvement of more interdisciplinary talent in this area
of research in the future (Kuang et al., 2021). For instance, improved
communication can facilitate the design of DL models that take into
account factors such as the desired accuracy of fault interpretation and
existing hardware limitations.
5.4. Challenges and future directions (RQ4)
5.4.1. Practical usage-related
Of the 12 challenges we have summarised for RQ4, the three least
discussed are those related to practical use (C10-12). Current develop-
ment is still at an academic level and cannot be directly applied to in-
dustrial use. There are still multiple hurdles that must be overcome
before this could happen, such as digitising fault probabilities into a
format acceptable for industrial workows (i.e. fault sticks in ASCII
format) and rationalising the incorporation of geological constraints
into DL methods. Future researchers should focus more on developing
geologist-friendly user interfaces to integrate new DL methods into the
standard workows. In that sense, studies exploring potential user in-
terfaces aligned to current industry workows and close collaborations
of algorithm/software developers with industry partners could be
fruitful areas for future research.
In addition, the current DL-based models for fault interpretation
primarily approach the task from an image-processing perspective, with
limited emphasis on incorporating geological constraints. Only a few
studies, such as those conducted by Wu et al. (2019c,d), have made
attempts to simultaneously predict faults and other related geological
targets. By training DL models on multiple geological targets, it becomes
possible to incorporate certain domain knowledge and enhance the fault
interpretation capabilities. Moreover, there have been numerous efforts
to introduce various domain constraints into DL models. For a more
comprehensive understanding of these approaches and potential ideas, I
recommend referring to the review paper by Borghesi et al. (2020).
5.4.2. Evaluation-related
The lack of appropriate evaluation criteria and open access (C9-10)
constraints are two major challenges. Although some considerations of
evaluation metrics have been published in the literature, there is still a
need to establish appropriate evaluation criteria to allow for a fair
comparison. According to An et al. (2021), pixel-wise accuracy is un-
suitable for fault interpretation tasks and would give an overly opti-
mistic evaluation of some algorithms. Future research should take into
account the current evaluation challenges. Attention could be paid to
some computer vision metrics sensitive to imbalanced classication.
Moreover, independent test datasets should be used when testing new
algorithms, as DL models tend to overt the training datasets, resulting
in overly optimistic evaluation scores. The generalisation of DL methods
is a topic worth investigating, and the issue is highly relevant to model
practical utility. An open-source community should be established to
facilitate benchmark replication and improvement and we strongly
recommend that researchers in the eld publish their code and data
(Kuang et al., 2021).
5.4.3. DL-related
Although DL challenges (C5-7) and possible solutions are well dis-
cussed in computer science, the selected articles adequately address the
imbalance classication problem, with some consideration of the po-
tential reduction in the computational cost of DL models, but little effort
in explaining DL methods. Kozhenkov et al. (2019) already reported that
geologists are sceptical about DL approaches as they are black box
models for which they can not perform quality control. Thus, there are
great opportunities for research for interdisciplinary studies to address
the interpretability and explainability of DL-based fault interpretation
methods.
Explainable articial intelligence (XAI), i.e. providing human-
friendly explanations for AI decisions, has attracted increasing atten-
tion in the eld of AI (Gunning et al., 2019). There are three types of post
hoc XAI methods, which provide explanations for models by approxi-
mation of the model behaviours, to explain the least explainable DL
models; these are visual, textual and example-based explanations
(Kenny et al., 2021). Visual interpretations, typically saliency maps, are
able to highlight the pixel regions that contribute most to DL decisions to
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
16
give users insights into why a certain prediction is made. Textual
interpretation describes a given image, offering textual explanation
from the image caption to the le level. Example-based interpretation
provides the images in the training data most similar to the current input
image as model decision explanations (van der Velden et al., 2022).
These popular XAI methods are worthy of investigation. Future re-
searchers should seek to understand what kind of explanations benet
geologists to guide the use of XAI methods. XAI methods have been
studied in the medical elds to understand the needs of healthcare
practitioners (Panigutti et al., 2022; Du et al., 2022b), similarly, there is
a need for user studies with geology experts to explore XAI in fault
interpretation.
In addition to the focus on explainability of DL models, future re-
searchers could also prioritize the design of suitable loss functions for
addressing the uncertainty challenge in fault interpretation. Guillon
et al. (2020) have discussed their plans for developing an uncertainty-
aware loss function in their future work. This approach holds promise
in effectively tackling the uncertainties associated with fault
interpretation.
5.4.4. Data-related
Despite the prevalence of data-related challenges discussed in the 56
collected articles, with considerable efforts made to address them, our
analysis in Section 5.2 indicates a lack of high-quality, open-source,
large-scale and high-diversity seismic datasets in the current domain. In
computer vision, there exists a large-scale image classication dataset
known as ImageNet, which is frequently regarded as a central testbed for
pioneering achievements in DL (Beyer et al., 2020). Future work could
consider building a similar dataset in fault interpretation, which could
greatly facilitate related research. Ideally, this highly anticipated open
dataset should undergo meticulous annotation by multiple expert in-
terpreters, ensuring its substantial size and incorporation of seismic data
from diverse geolocations to achieve a high level of diversity. Kuang
et al. (2021) noted the high priority of having proper data management
regulation, including unifying data labelling, promoting data interop-
erability, and sharing.
Overall, seismic fault interpretation is a complex task that involves
multidisciplinary efforts, including but not limited to data acquisition
experts, geophysicists, fault experts, AI experts and enterprise software
specialists. Close collaboration among all stakeholders is required to
address all the challenges fully. Future research could focus on under-
standing what users really need and customising specic expert systems.
One potential direction is to incorporate the human-in-the-loop
approach, a methodology that synergistically combines the expertise
of human labelers with the power of DL models (Mosqueira-Rey et al.,
2023). A specic class of the human-in-the-loop approach that shows
promise is active learning, which entails a learning algorithm interac-
tively querying a human labeller to annotate the most informative data
points with desired outputs (Ma et al., 2023). By leveraging this tech-
nique, geologists can actively supervise and iteratively enhance the fault
interpretation capabilities of DL models on new seismic datasets.
Consequently, this boosts geologists’ condence in DL-based fault in-
terpreters and helps alleviate any skepticism they may harbor towards
black box DL models. Moreover, expert systems can integrate features
such as exporting fault sticks in a format that seamlessly integrates with
geologists’ preferred software, thereby facilitating a smooth transition
between the expert system and their established workows.
6. Conclusion
This review presents the rst systematic literature review on DL-
based seismic fault interpretation. The ndings reveal that the current
focus of DL algorithms in this domain is primarily on convolutional
neural networks (CNNs), while the exploration of other neural network
types remains limited. Furthermore, despite the inclusion of 56 DL
methods and 73 seismic datasets, the availability of code and data for
public access is scarce. To address these limitations, the establishment of
an open community and standardized benchmark evaluation standards
is recommended. Additionally, the challenges associated with data, DL
methodologies, evaluation, and practical use must be addressed to fully
harness the potential of DL methods in seismic fault interpretation.
Future research could concentrate on practical aspects such as devel-
oping user-friendly interfaces and incorporating geological constraints
into DL decision-making processes to enhance the application of DL
methods in this eld.
CRediT authorship contribution statement
Yu An: Conceptualization, Methodology, Software, Investigation,
Resources, Project administration, Data curation, Formal analysis,
Validation, Visualization, Writing – original draft, Writing – review &
editing. Haiwen Du: Investigation, Data curation, Formal analysis,
Visualization, Writing – original draft, Writing – review & editing.
Siteng Ma: Investigation, Data curation, Formal analysis, Writing –
original draft, Writing – review & editing. Yingjie Niu: Investigation,
Data curation, Formal analysis, Validation, Writing – review & editing.
Dairui Liu: Investigation, Data curation, Formal analysis, Validation,
Writing – review & editing. Jing Wang: Investigation, Data curation,
Formal analysis, Validation, Writing – review & editing. Yuhan Du:
Conceptualization, Methodology, Writing – review & editing. Conrad
Childs: Supervision, Writing – review & editing. John Walsh: Super-
vision, Writing – review & editing. Ruihai Dong: Supervision, Project
administration, Funding acquisition, Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Data availability
No data was used for the research described in the article.
References
Ahmad, A.B., Tsuji, T., 2021. Machine learning for automatic slump identication from
3D seismic data at convergent plate margins. Mar. Pet. Geol. 133, 105290. https://
doi.org/10.1016/j.marpetgeo.2021.105290. URL:https://linkinghub.elsevier.
com/retrieve/pii/S0264817221003937.
Alcalde, J., Bond, C.E., Johnson, G., Kloppenburg, A., Ferrer, O., Bell, R., Ayarza, P.,
2019. Fault interpretation in seismic reection data: an experiment analysing the
impact of conceptual model anchoring and vertical exaggeration. Solid Earth 10,
1651–1662. https://doi.org/10.5194/se-10-1651-2019.
Alfarhan, M., Deriche, M., Maalej, A., AlRegib, G., Al-Marzouqi, H., 2020a. Multiple
Events Detection In Seismic Structures Using A Novel U-Net Variant. In: 2020 IEEE
International Conference on Image Processing (ICIP). IEEE, pp. 2900–2904. https://
doi.org/10.1109/ICIP40778.2020.9190682.
Alfarhan, M., Maalej, A., Deriche, M., 2020b. Concurrent Detection of Salt Domes and
Faults using ResNet with U-Net. In: Proceedings - 2020 6th Conference on Data
Science and Machine Learning Applications, CDMA 2020, pp. 118–122. doi: 10.11
09/CDMA47397.2020.00026.
Alzubaidi, L., Zhang, J., Humaidi, A.J., Al-Dujaili, A., Duan, Y., Al-Shamma, O.,
Santamaría, J., Fadhel, M.A., Al-Amidie, M., Farhan, L., 2021. Review of deep
learning: Concepts, cnn architectures, challenges, applications, future directions.
J. Big Data 8, 1–74.
An, Y., Dong, R., 2023. Understanding the effect of different prior knowledge on cnn
fault interpreter. IEEE Access 11, 15058–15068. https://doi.org/10.1109/
ACCESS.2023.3244683.
An, Y., Guo, J., Ye, Q., Childs, C., Walsh, J., Dong, R., 2021. Deep convolutional neural
network for automatic fault recognition from 3D seismic datasets. Comput. Geosci.
153 https://doi.org/10.1016/j.cageo.2021.104776.
An, Y., Ye, Q., Guo, J., Dong, R., 2020. Overlap training to mitigate inconsistencies
caused by image tiling in CNNs. Lecture Notes in Computer Science (including
subseries Lecture Notes in Articial Intelligence and Lecture Notes in Bioinformatics)
12498 LNAI, pp. 35–48. doi: 10.1007/978-3-030-63799-6_3.
Ao, Y., Lu, W., Jiang, B., Monkam, P., 2021. Seismic Structural Curvature Volume
Extraction With Convolutional Neural Networks. IEEE Trans. Geosci. Remote Sens.
59, 7370–7384. https://doi.org/10.1109/TGRS.2020.3042098. URL:https:
//ieeexplore.ieee.org/document/9301201/.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
17
Ao, Y., Lu, W., Xu, P., Jiang, B., 2022. Seismic Dip Estimation With a Domain Knowledge
Constrained Transfer Learning Approach. IEEE Trans. Geosci. Remote Sens. 60,
1–16. https://doi.org/10.1109/TGRS.2021.3061438. URL:https://ieeexplore.ieee.or
g/document/9374570/.
Aribido, O.J., AlRegib, G., Alaudah, Y., 2021. Self-supervised delineation of geologic
structures using orthogonal latent space projection. GEOPHYSICS 86, V497–V508.
https://doi.org/10.1190/geo2020-0541.1.
Aribido, O.J., AlRegib, G., Deriche, M., 2020. Self-Supervised Annotation of Seismic
Images Using Latent Space Factorization. In: 2020 IEEE International Conference on
Image Processing (ICIP). IEEE, pp. 2421–2425. https://doi.org/10.1109/
ICIP40778.2020.9190698. URL:https://ieeexplore.ieee.org/document/9190698/.
Badley, M., Freeman, B., Roberts, A., Thatcher, J., Walsh, J.J., Watterson, J., Yielding, G.,
1991. Fault interpretation during seismic interpretation and reservoir evaluation.
The integration of geology, geophysics, petrophysics and petroleum engineering in
reservoir delineation, description and management. In: Proc. conference, Houston,
1990, pp. 224–241.
Baraniuk, R.G., Steeghs, P., 2017. Compressive sensing: A new approach to seismic data
acquisition. Lead. Edge 36, 642–645.
Bengio, Y., Courville, A., Vincent, P., 2013. Representation learning: A review and new
perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35, 1798–1828.
Beyer, L., H´
enaff, O.J., Kolesnikov, A., Zhai, X., Oord, A.v.d., 2020. Are we done with
imagenet? arXiv preprint arXiv: 2006.07159.
Bi, Z., Wu, X., Geng, Z., Li, H., 2021. Deep Relative Geologic Time: A Deep Learning
Method for Simultaneously Interpreting 3-D Seismic Horizons and Faults.
J. Geophys. Res.: Solid Earth 126, 1–24. https://doi.org/10.1029/2021JB021882.
Bond, C.E., 2015. Uncertainty in structural interpretation: Lessons to be learnt. J. Struct.
Geol. 74, 185–200.
Borghesi, A., Baldo, F., Milano, M., 2020. Improving deep learning models via constraint-
based domain knowledge: a brief survey. arXiv preprint arXiv: 2005.10691.
Bugge, A.J., Clark, S.R., Lie, J.E., Faleide, J.I., 2018. A case study on semiautomatic
seismic interpretation of unconformities and faults in the southwestern barents sea.
Interpretation 6, SD29–SD40.
Bugge, A.J., Erik Lie, J., Evensen, A.K., Faleide, J.I., Clark, S., 2019. Automatic extraction
of dislocated horizons from 3d seismic data using nonlocal trace matching.
Geophysics 84, IM77–IM86.
Chang, D., Yong, X., Yang, W., Wang, Y., Guo, T., 2019. U_net & residual neural networks
for seismic fault interpretation. In: 81st EAGE Conference and Exhibition 2019.
European Association of Geoscientists & Engineers, pp. 1–5.
Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H., 2018. Encoder-decoder with
atrous separable convolution for semantic image segmentation.arXiv:1802.02611.
Cui, L., Wu, K., Liu, Q., Wang, D., Guo, W., Liu, Y., Xu, G., 2022. Enhanced interpretation
of strike-slip faults using hybrid attributes: Advanced insights into fault geometry
and relationship with hydrocarbon accumulation in Jurassic formations of the
Junggar Basin. J. Petrol. Sci. Eng. 208 https://doi.org/10.1016/j.
petrol.2021.109630. URL:https://linkinghub.elsevier.com/retrieve/pii/S0920410
52101264X.
Cunha, A., Pochet, A., Lopes, H., Gattass, M., 2020. Seismic fault detection in real data
using transfer learning from a convolutional neural network pre-trained with
synthetic seismic data. Comput. Geosci. 135, 104344. https://doi.org/10.1016/j.
cageo.2019.104344. URL:https://linkinghub.elsevier.com/retrieve/pii/S0098300
418307040.
Egorov, D., 2019. Automatic fault interpretation from seismic data via convolutional
neural networks. In: EAGE Subsurface Intelligence Workshop. European Association
of Geoscientists & Engineers, pp. 1–5. https://doi.org/10.3997/2214-
4609.2019X610105.
Di, H., Li, C., Smith, S., Li, Z., Abubakar, A., 2021. Imposing interpretational constraints
on a seismic interpretation convolutional neural network. GEOPHYSICS 86,
IM63–IM71. https://doi.org/10.1190/geo2020-0449.1.
Di, H., Shaq, M.A., AlRegib, G., 2017. Seismic-fault detection based on multiattribute
support vector machine analysis. In: 2017 SEG International Exposition and Annual
Meeting. OnePetro.
Di, H., Shaq, M.A., Wang, Z., AlRegib, G., 2019. Improving seismic fault detection by
super-attribute-based classication. Interpretation 7, SE251–SE267. https://doi.org/
10.1190/INT-2018-0188.1.
Di, H., Truelove, L., Li, C., Abubakar, A., 2020. Accelerating seismic fault and
stratigraphy interpretation with deep CNNs: A case study of the Taranaki Basin, New
Zealand. Lead. Edge 39, 727–733. https://doi.org/10.1190/tle39100727.1.
Di, H., Wang, Z., AlRegib, G., 2018a. Seismic Fault Detection from Post-Stack Amplitude
by Convolutional Neural Networks. In: 80th EAGE Conference & Exhibition 2018.
URL:http://www.earthdoc.org/publication/publicationdetails/?publicatio
n=92119, doi: 10.3997/2214-4609.201800733.
Di, H., Wang, Z., AlRegib, G., 2018b. Why using CNN for seismic interpretation? An
investigation. In: SEG Technical Program Expanded Abstracts 2018. Society of
Exploration Geophysicists, pp. 2216–2220. https://doi.org/10.1190/segam2018-
2997155.1.
Dimililer, K., Dindar, H., Al-Turjman, F., 2021. Deep learning, machine learning and
internet of things in geophysical engineering applications: An overview.
Microprocess. Microsyst. 80, 103613 https://doi.org/10.1016/j.
micpro.2020.103613. URL:https://linkinghub.elsevier.com/retrieve/pii/S0141933
120307602.
Dou, Y., Li, K., Zhu, J., Li, X., Xi, Y., 2022. Attention-Based 3-D Seismic Fault
Segmentation Training by a Few 2-D Slice Labels. IEEE Trans. Geosci. Remote Sens.
60, 1–15. https://doi.org/10.1109/TGRS.2021.3113676. URL:https://ieeexplore.
ieee.org/document/9559897/,arXiv:2105.03857.
Du, H., An, Y., Ye, Q., Guo, J., Liu, L., Zhu, D., Childs, C., Walsh, J., Dong, R., 2022a.
Disentangling noise patterns from seismic images: Noise reduction and style transfer.
IEEE Trans. Geosci. Remote Sens. 60, 1–14.
Du, Y., Antoniadi, A.M., McNestry, C., McAuliffe, F.M., Mooney, C., 2022b. The role of
xai in advice-taking from a clinical decision support system: A comparative user
study of feature contribution-based and example-based explanations. Appl. Sci. 12
https://doi.org/10.3390/app122010323. URL:https://www.mdpi.com/20
76-3417/12/20/10323.
El Zini, J., Rizk, Y., Awad, M., 2020. A Deep Transfer Learning Framework for Seismic
Data Analysis: A Case Study on Bright Spot Detection. IEEE Trans. Geosci. Remote
Sens. 58, 3202–3212. https://doi.org/10.1109/TGRS.2019.2950888. URL:https:
//ieeexplore.ieee.org/document/8933067/.
Faleide, T.S., Braathen, A., Lecomte, I., Mulrooney, M.J., Midtkandal, I., Bugge, A.J.,
Planke, S., 2021a. Impacts of seismic resolution on fault interpretation: Insights from
seismic modelling. Tectonophysics 816, 229008.
Faleide, T.S., Braathen, A., Lecomte, I., Mulrooney, M.J., Midtkandal, I., Bugge, A.J.,
Planke, S., 2021b. Impacts of seismic resolution on fault interpretation: Insights from
seismic modelling. Tectonophysics 816, 229008. https://doi.org/10.1016/j.
tecto.2021.229008. URL:https://www.sciencedirect.com/science/article/pii/
S0040195121002900https://linkinghub.elsevier.com/retrieve/pii/S0040195
121002900.
Feng, R., Grana, D., Balling, N., 2021. Uncertainty quantication in fault detection using
convolutional neural networks. GEOPHYSICS 86, M41–M48. https://doi.org/
10.1190/geo2020-0424.1.
Fossen, H., 2010. Structural Geology. Cambridge University Press. https://doi.org/
10.1017/CBO9780511777806.
Gao, K., Huang, L., Zheng, Y., 2022a. Fault Detection on Seismic Structural Images Using
a Nested Residual U-Net. IEEE Trans. Geosci. Remote Sens. 60, 1–15. https://doi.
org/10.1109/TGRS.2021.3073840. URL:https://ieeexplore.ieee.org/document
/9418625/.
Garg, A., Bagga, S., Singh, Y., Anand, S., 2022. Hiermatch: Leveraging label hierarchies
for improving semi-supervised learning. In: Proceedings of the IEEE/CVF Winter
Conference on Applications of Computer Vision, pp. 1015–1024.
Gao, K., Huang, L., Zheng, Y., Lin, R., Hu, H., Cladohous, T., 2022b. Automatic fault
detection on seismic images using a multiscale attention convolutional neural
network. GEOPHYSICS 87, N13–N29. https://doi.org/10.1190/geo2020-0945.1.
Gibbs, A.D., 1984. Structural evolution of extensional basin margins. J. Geol. Soc. 141,
609–620. https://doi.org/10.1144/gsjgs.141.4.0609. URL:https://jgs.lyellcoll
ection.org/content/141/4/609.
Goodfellow, I., Bengio, Y., Courville, A., 2016. Deep learning. MIT press.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S.,
Courville, A., Bengio, Y., 2014. Generative adversarial nets. Adv. Neural Inf. Process.
Syst. 27.
Gou, J., Yu, B., Maybank, S.J., Tao, D., 2021. Knowledge distillation: A survey. Int. J.
Comput. Vis. 129, 1789–1819.
Guillon, S., Joncour, F., Barrallon, P.E., Castani´
e, L., 2020. Ground-truth uncertainty-
aware metrics for machine learning applications on seismic image interpretation:
Application to faults and horizon extraction. Lead. Edge 39, 734–741. https://doi.
org/10.1190/tle39100734.1.
Guitton, A., 2018. 3D Convolutional Neural Networks for Fault Interpretation. In: 80th
EAGE Conference & Exhibition 2018. URL:http://www.earthdoc.org/publication/pu
blicationdetails/?publication=92118, doi: 10.3997/2214-4609.201800732.
Guitton, A., Wang, H., Trainor-Guitton, W., 2017. Statistical imaging of faults in 3d
seismic volumes using a machine learning approach. In: SEG Technical Program
Expanded Abstracts 2017. Society of Exploration Geophysicists, pp. 2045–2049.
Gunning, D., Stek, M., Choi, J., Miller, T., Stumpf, S., Yang, G.Z., 2019.
Xai—explainable articial intelligence. Sci. Robot. 4, eaay7120.
Hale, D., 2013. Methods to compute fault images, extract fault surfaces, and estimate
fault throws from 3d seismic images. Geophysics 78, O33–O43.
Haroon, S., Alyamkin, S., Shenoy, R., 2018. Big Data-Driven Advanced Analytics:
Application of Convolutional and Deep Neural Networks for GPU Based Seismic
Interpretations. In: Abu Dhabi International Petroleum Exhibition & Conference.
SPE, D021S035R003. https://doi.org/10.2118/193259-MS. URL:https://onepetro.
org/SPEADIP/proceedings/18ADIP/2-18ADIP/AbuDhabi, UAE/213071.
Hu, G., Hu, Z., Liu, J., Cheng, F., Peng, D., 2022. Seismic Fault Interpretation Using Deep
Learning-Based Semantic Segmentation Method. IEEE Geosci. Remote Sens. Lett. 19,
1–5. https://doi.org/10.1109/LGRS.2020.3041301. URL:https://ieeexplore.ieee.or
g/document/9293123/.
Huang, L., Dong, X., Clee, T.E., 2017. A scalable deep learning platform for identifying
geologic features from seismic attributes. Lead. Edge 36, 249–256.
Jiang, F., Norlund, P., 2020. Super Resolution of Fault Plane Prediction by a Generative
Adversarial Network. In: First EAGE Digitalization Conference and Exhibition.
European Association of Geoscientists & Engineers, pp. 1–5. https://doi.org/
10.3997/2214-4609.202032011.
Keele, S., et al., 2007. Guidelines for performing systematic literature reviews in software
engineering. Technical Report. Technical report, Ver. 2.3 EBSE Technical Report.
EBSE.
Kenny, E.M., Ford, C., Quinn, M., Keane, M.T., 2021. Explaining black-box classiers
using post-hoc explanations-by-example: The effect of explanations and error-rates
in xai user studies. Artif. Intell. 294, 103459.
Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead,
S., Berg, A.C., Lo, W.Y., et al., 2023. Segment anything. arXiv preprint arXiv:
2304.02643.
Kozhenkov, A., Naeini, E., Prindle, K., 2019. Machine Learning and Learning from
Machines. In: Progress’19. European Association of Geoscientists & Engineers,
pp. 1–5. https://doi.org/10.3997/2214-4609.201953052.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
18
Kuang, L., Liu, H., Ren, Y., Luo, K., Shi, M., Su, J., Li, X., 2021. Application and
development trend of articial intelligence in petroleum exploration and
development. Pet. Explor. Dev. 48, 1–14. https://doi.org/10.1016/S1876-3804(21)
60001-0. URL:https://www.sciencedirect.com/science/article/pii/S1876380
421600010.
Kumar, A., Bi, L., Kim, J., Feng, D.D., 2020. Chapter ve - machine learning in medical
imaging. In: Feng, D.D. (Ed.), Biomedical Information Technology, second ed.
Academic Press, Biomedical Engineering, pp. 167–196. https://doi.org/10.1016/
B978-0-12-816034-3.00005-5. URL:https://www.sciencedirect.com/science/arti
cle/pii/B9780128160343000055.
Kumar, P.C., Mandal, A., 2018. Enhancement of fault interpretation using multi-attribute
analysis and articial neural network (ANN) approach: a case study from Taranaki
Basin, New Zealand. Explor. Geophys. 49, 409–424. https://doi.org/10.1071/
EG16072.
Kumar, P.C., Omosanya, K.O., Alves, T.M., Sain, K., 2019. A neural network approach for
elucidating uid leakage along hard-linked normal faults. Mar. Pet. Geol. 110,
518–538. https://doi.org/10.1016/j.marpetgeo.2019.07.042. URL:https://linkingh
ub.elsevier.com/retrieve/pii/S0264817219303526.
Lapteva, A., Loginov, G., Duchkov, A., Alyamkin, S., 2019. The comparison of
convolution neural network for localized capturing detection of faults on seismic
images. In: 81st EAGE Conference and Exhibition 2019. European Association of
Geoscientists & Engineers, pp. 1–5.
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521, 436–444.
Li, F., Lyu, B., Qi, J., Verma, S., Zhang, B., 2021. Seismic Coherence for Discontinuity
Interpretation. Surv. Geophys. 42, 1229–1280. https://doi.org/10.1007/s10712-
021-09670-4.
Li, J., Wu, X., Hu, Z., 2022a. Deep Learning for Simultaneous Seismic Image Super-
Resolution and Denoising. IEEE Trans. Geosci. Remote Sens. 60, 1–11. https://doi.
org/10.1109/TGRS.2021.3057857. URL:https://ieeexplore.ieee.org/document
/9364884/.
Li, S., Liu, N., Li, F., Gao, J., Ding, J., 2022b. Automatic Fault Delineation in 3-D Seismic
Images With Deep Learning: Data Augmentation or Ensemble Learning? IEEE Trans.
Geosci. Remote Sens. 60, 1–14. https://doi.org/10.1109/TGRS.2022.3150353. URL:
https://ieeexplore.ieee.org/document/9709337/.
Lin, L., Zhong, Z., Cai, Z., Sun, A.Y., Li, C., 2022. Automatic geologic fault identication
from seismic data using 2.5 d channel attention u-net. Geophysics 87, IM111–IM124.
Liu, N., He, T., Tian, Y., Wu, B., Gao, J., Xu, Z., 2020a. Common-azimuth seismic data
fault analysis using residual UNet. Interpretation 8, SM25–SM37. https://doi.org/
10.1190/INT-2019-0173.1.
Liu, Y., Cheng, M.M., Hu, X., Wang, K., Bai, X., 2016. Richer convolutional features for
edge detection.arXiv:1612.02103.
Liu, Z., Zhou, C., Hu, G., Song, C., 2020b. Interpretability-Guided Convolutional Neural
Networks for Seismic Fault Segmentation. In: ICASSP 2020 - 2020 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE,
pp. 4312–4316. https://doi.org/10.1109/ICASSP40776.2020.9053472. URL:https:
//ieeexplore.ieee.org/document/9053472/.
Lu, P., Morris, M., Brazell, S., Comiskey, C., Xiao, Y., 2018. Using generative adversarial
networks to improve deep-learning fault interpretation networks. Lead. Edge 37,
578–583. https://doi.org/10.1190/tle37080578.1.
Ma, S., An, Y., Wang, J., Lawlor, A., Dong, R., 2023. Adaptive adversarial samples based
active learning for medical image classication. In: Proceedings of the 12th
International Conference on Pattern Recognition Applications and Methods.
SCITEPRESS - Science and Technology Publications. https://doi.org/10.5220/
0011622100003411.
Ma, Z., Li, Y., 2021. Rotated-UNet: A seismic fault identication network based on
inverse sampling block construction. In: 2021 3rd International Academic Exchange
Conference on Science and Technology Innovation (IAECST). IEEE, pp. 740–744.
https://doi.org/10.1109/IAECST54258.2021.9695612. URL:https://ieeexplore.
ieee.org/document/9695612/.
Manral, S., 2020. Enhancing Fault Interpretation Efciency and Accuracy with Deep
Convolutional Neural Network and Elastic Cloud Compute. In: First EAGE
Digitalization Conference and Exhibition. European Association of Geoscientists &
Engineers, pp. 1–5. https://doi.org/10.3997/2214-4609.202032100.
Marfurt, K.J., Kirlin, R.L., Farmer, S.L., Bahorich, M.S., 1998. 3-d seismic attributes using
a semblance-based coherency algorithm. GEOPHYSICS 63, 1150–1165. https://doi.
org/10.1190/1.1444415.
Marfurt, K.J., Sudhaker, V., Gersztenkorn, A., Crawford, K.D., Nissen, S.E., 1999.
Coherency calculations in the presence of structural dip. GEOPHYSICS 64, 104–111.
https://doi.org/10.1190/1.1444508.
Michie, E.A.H., Alaei, B., Braathen, A., 2022. Assessing the accuracy of fault
interpretation using machine-learning techniques when risking faults for CO 2
storage site assessment. Interpretation 10, T73–T93. https://doi.org/10.1190/INT-
2021-0077.1.
Mosqueira-Rey, E., Hern´
andez-Pereira, E., Alonso-Ríos, D., Bobes-Bascar´
an, J.,
Fern´
andez-Leal, ´
A., 2023. Human-in-the-loop machine learning: A state of the art.
Artif. Intell. Rev. 56, 3005–3054.
Mosser, L., Purves, S., Naeini, E., 2020. Deep Bayesian Neural Networks for Fault
Identication and Uncertainty Quantication. In: First EAGE Digitalization
Conference and Exhibition. European Association of Geoscientists & Engineers,
pp. 1–5. https://doi.org/10.3997/2214-4609.202032036.
Nashaat, M., Ghosh, A., Miller, J., Quader, S., 2021. Semi-supervised ensemble learning
for dealing with inaccurate and incomplete supervision. ACM Trans. Knowl. Discov.
Data 16. https://doi.org/10.1145/3473910.
Ouali, Y., Hudelot, C., Tami, M., 2020. An overview of deep semi-supervised learning.
arXiv preprint arXiv: 2006.05278.
Panigutti, C., Beretta, A., Giannotti, F., Pedreschi, D., 2022. Understanding the impact of
explanations on advice-taking: a user study for ai-based clinical decision support
systems. In: CHI Conference on Human Factors in Computing Systems, pp. 1–9.
Pochet, A., Diniz, P.H., Lopes, H., Gattass, M., 2019. Seismic Fault Detection Using
Convolutional Neural Networks Trained on Synthetic Poststacked Amplitude Maps.
IEEE Geosci. Remote Sens. Lett. 16, 352–356. https://doi.org/10.1109/
LGRS.2018.2875836.
Qi, J., Lyu, B., Wu, X., Marfurt, K., 2020. Comparing convolutional neural networking
and image processing seismic fault detection methods. In: SEG International
Exposition and Annual Meeting. OnePetro.
Randle, C.H., Bond, C.E., Lark, R.M., Monaghan, A.A., 2019. Uncertainty in geological
interpretations: Effectiveness of expert elicitations. Geosphere 15, 108–118.
Raza, K., Singh, N.K., 2021. A tour of unsupervised deep learning for medical image
analysis. Curr. Med. Imag. 17, 1059–1077.
Redmon, J., Divvala, S., Girshick, R., Farhadi, A., 2015. You only look once: Unied,
real-time object detection arXiv:1506.02640.
Reilly, J.M., Aharchaou, M., Neelamani, R., 2023. A brief overview of seismic resolution
in applied geophysics. Lead. Edge 42, 8–15.
Roberts, A., 2001. Curvature attributes and their application to 3d interpreted horizons.
First Break 19, 85–100. https://doi.org/10.1046/j.0263-5046.2001.00142.x.
Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for
biomedical image segmentation.arXiv:1505.04597.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy,
A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L., 2014. Imagenet large scale visual
recognition challenge.arXiv:1409.0575.
Schmidhuber, J., 2015. Deep learning in neural networks: An overview. Neural Netw. 61,
85–117.
Shapiro, L.G., Stockman, G.C., et al., 2001. Computer vision, vol. 3. Prentice Hall New
Jersey.
Shi, Y., Wu, X., Fomel, S., 2021. Interactively tracking seismic geobodies with a deep-
learning ood-lling network. GEOPHYSICS 86, A1–A5. https://doi.org/10.1190/
geo2020-0042.1.
da Silva, M., Charl´
ety, J., Fraysse, A., Pesquet, J.C., 2021. A small-scale network for
seismic patterns classication. In: 2020 28th European Signal Processing Conference
(EUSIPCO). IEEE, pp. 1402–1406.
Simonyan, K., Zisserman, A., 2014. Very deep convolutional networks for large-scale
image recognition arXiv:1409.1556.
Smith, S., Zimina, O., Manral, S., Nickel, M., 2022. Machine-learning assisted
interpretation: Integrated fault prediction and extraction case study from the
Groningen gas eld, Netherlands. Interpretation 10, SC17–SC30. https://doi.org/
10.1190/INT-2021-0137.1.
Soubaras, R., Dowle, R., Sablon, R., 2012. Broadseis: Enhancing interpretation and
inversion with broadband marine seismic. CSEG Record. 37, 40–46.
Tariq, Z., Aljawad, M.S., Hasan, A., Murtaza, M., Mohammed, E., El-Husseiny, A.,
Alari, S.A., Mahmoud, M., Abdulraheem, A., 2021. A systematic review of data
science and machine learning applications to the oil and gas industry. J. Pet. Explor.
Prod. Technol. 11, 4339–4374. https://doi.org/10.1007/s13202-021-01302-2.
van der Velden, B.H., Kuijf, H.J., Gilhuijs, K.G., Viergever, M.A., 2022. Explainable
articial intelligence (xai) in deep learning-based medical image analysis. Med.
Image Anal. 79, 102470 https://doi.org/10.1016/j.media.2022.102470. URL:
https://www.sciencedirect.com/science/article/pii/S1361841522001177.
Voulodimos, A., Doulamis, N., Doulamis, A., Protopapadakis, E., 2018. Deep learning for
computer vision: A brief review. Comput. Intell. Neurosci. 2018.
Walsh, J.J., Watterson, J., Childs, C., Nicol, A., 1996. Ductile strain effects in the analysis
of seismic interpretations of normal fault systems. Geol. Soci. Lond. Spec. Publ. 99,
27–40. https://doi.org/10.1144/GSL.SP.1996.099.01.04.
Wang, B., Ma, J., 2020. CNN Based Fault Recognition with Multi-Scale Fusion Attention
Mechanism. IOP Conf. Ser.: Earth Environ. Sci. 474, 042007 https://doi.org/
10.1088/1755-1315/474/4/042007.
Wang, S., Liu, W., Wu, J., Cao, L., Meng, Q., Kennedy, P.J., 2016. Training deep neural
networks on imbalanced data sets. In: 2016 international joint conference on neural
networks (IJCNN). IEEE, pp. 4368–4374.
Wang, J., Zhang, J.H., Zhang, J.L., Lu, F.M., Meng, R.G., Wang, Z., 2021a. Research on
fault recognition method combining 3D Res-UNet and knowledge distillation. Appl.
Geophys. 18, 199–212. https://doi.org/10.1007/s11770-021-0894-2.
Wang, X., Zhao, Y., Pourpanah, F., 2020. Recent advances in deep learning. Int. J. Mach.
Learn. Cybern. 11, 747–750.
Wang, Z., Di, H., Shaq, M.A., Alaudah, Y., AlRegib, G., 2018. Successful leveraging of
image processing and machine learning in seismic structural interpretation: A
review. Lead. Edge 37, 451–461. https://doi.org/10.1190/tle37060451.1.
Wang, P., Fan, E., Wang, P., 2021b. Comparative analysis of image classication
algorithms based on traditional machine learning and deep learning. Pattern Recogn.
Lett. 141, 61–67.
Wang, Z., Li, B., Liu, N., Wu, B., Zhu, X., 2022. Distilling Knowledge From an Ensemble of
Convolutional Neural Networks for Seismic Fault Detection. IEEE Geosci. Remote
Sens. Lett. 19, 1–5. https://doi.org/10.1109/LGRS.2020.3034960. URL:https:
//ieeexplore.ieee.org/document/9269391/.
Wei, X.L., Zhang, C.X., Kim, S.W., Jing, K.L., Wang, Y.J., Xu, S., Xie, Z.Z., 2022. Seismic
fault detection using convolutional neural networks with focal loss. Comput. Geosci.
158, 104968. https://doi.org/10.1016/j.cageo.2021.104968. URL:https://linkingh
ub.elsevier.com/retrieve/pii/S0098300421002533.
Wrona, T., Pan, I., Bell, R.E., Gawthorpe, R.L., Fossen, H., Brune, S., 2021. 3D seismic
interpretation with deep learning: A brief introduction. Lead. Edge 40, 524–532.
https://doi.org/10.1190/tle40070524.1.
Y. An et al.

Earth-Science Reviews 243 (2023) 104509
19
Wu, J., Liu, B., Zhang, H., He, S., Yang, Q., 2021. Fault Detection Based on Fully
Convolutional Networks (FCN). J. Mar. Sci. Eng. 9, 259. https://doi.org/10.3390/
jmse9030259. URL:https://www.mdpi.com/2077-1312/9/3/259.
Wu, X., Geng, Z., Shi, Y., Pham, N., Fomel, S., Caumon, G., 2020. Building realistic
structure models to train convolutional neural networks for seismic structural
interpretation. GEOPHYSICS 85, WA27–WA39. https://doi.org/10.1190/geo2019-
0375.1.
Wu, X., Liang, L., Shi, Y., Fomel, S., 2019a. FaultSeg3D: Using synthetic data sets to train
an end-to-end convolutional neural network for 3D seismic fault segmentation.
GEOPHYSICS 84, IM35–IM45. https://doi.org/10.1190/geo2018-0646.1.
Wu, X., Liang, L., Shi, Y., Geng, Z., Fomel, S., 2019b. Deep learning for local seismic
image processing: Fault detection, structure-oriented smoothing with edge-
preserving, and slope estimation by using a single convolutional neural network. In:
Seg technical program expanded abstracts 2019. Society of Exploration
Geophysicists, pp. 2222–2226.
Wu, X., Liang, L., Shi, Y., Geng, Z., Fomel, S., 2019c. Multitask learning for local seismic
image processing: fault detection, structure-oriented smoothing with edge-
preserving, and seismic normal estimation by using a single convolutional neural
network. Geophys. J. Int. 219, 2097–2109. https://doi.org/10.1093/gji/ggz418.
URL:https://academic.oup.com/gji/article/219/3/2097/5571094.
Xie, S., Tu, Z., 2015. Holistically-nested edge detection.arXiv:1504.06375.
Wu, X., Shi, Y., Fomel, S., Liang, L., Zhang, Q., Yusifov, A.Z., 2019d. FaultNet3D:
Predicting Fault Probabilities, Strikes, and Dips With a Single Convolutional Neural
Network. IEEE Trans. Geosci. Remote Sens. 57, 9138–9155. https://doi.org/
10.1109/TGRS.2019.2925003. URL:https://ieeexplore.ieee.org/document/
8772081/.
Xiong, W., Ji, X., Ma, Y., Wang, Y., AlBinHassan, N.M., Ali, M.N., Luo, Y., 2018. Seismic
fault detection with convolutional neural network. GEOPHYSICS 83, O97–O103.
https://doi.org/10.1190/geo2017-0666.1.
Yan, Z., Zhang, Z., Liu, S., 2021. Improving Performance of Seismic Fault Detection by
Fine-Tuning the Convolutional Neural Network Pre-Trained with Synthetic Samples.
Energies 14, 3650. https://doi.org/10.3390/en14123650. URL:https://www.mdpi.
com/1996-1073/14/12/3650.
Yuan, Z., Huang, H., Jiang, Y., Tang, J., Li, J., 2019. An enhanced fault-detection method
based on adaptive spectral decomposition and super-resolution deep learning.
Interpretation 7, T713–T725. https://doi.org/10.1190/INT-2018-0180.1.
Zhang, Y., Yang, Q., 2018. An overview of multi-task learning. Natl. Sci. Rev. 5, 30–43.
Zhang, Z.Y., Zhao, P., Jiang, Y., Zhou, Z.H., 2021. Learning from incomplete and
inaccurate supervision. IEEE Trans. Knowl. Data Eng. 1–1 https://doi.org/10.1109/
TKDE.2021.3061215.
Zhao, T., Mukhopadhyay, P., 2018. A fault detection workow using deep learning and
image processing. In: 2018 SEG international exposition and annual meeting.
OnePetro.
Zheng, Y., Zhang, Q., Yusifov, A., Shi, Y., 2019. Applications of supervised deep learning
for seismic interpretation and inversion. Lead. Edge 38, 526–533. https://doi.org/
10.1190/tle38070526.1.
Zhou, R., Cai, Y., Zong, J., Yao, X., Yu, F., Hu, G., 2020. Automatic fault instance
segmentation based on mask propagation neural network. Articial Intelligence.
Geosciences 1, 31–35. https://doi.org/10.1016/j.aiig.2020.12.001. URL:https://li
nkinghub.elsevier.com/retrieve/pii/S2666544120300022.
Zhou, R., Yao, X., Wang, Y., Hu, G., Yu, F., 2021. Seismic fault detection with progressive
transfer learning. Acta Geophys. 69, 2187–2203. https://doi.org/10.1007/s11600-
021-00668-5.
Zhou, Z.H., 2018. A brief introduction to weakly supervised learning. Natl. Sci. Rev. 5,
44–53.
Zhuang, F., Qi, Z., Duan, K., Xi, D., Zhu, Y., Zhu, H., Xiong, H., He, Q., 2020.
A comprehensive survey on transfer learning. Proc. IEEE 109, 43–76.
Y. An et al.