GeoScienceWorld

Lithosphere
Volume 2024, Number 2, Article ID lithosphere_2024_118, 14 pages
https://doi.org/10.2113/2024/lithosphere_2024_118

Research Article
Enhancement of the Consistency and Connectivity of the Ant-
tracking Algorithm via 3D U-Net with Dual-Threshold Iteration

Seunghun Choi ,1 Min Je Lee ,1 Yonggwon Jung ,1 Juan Lee ,1 Jineon Kim ,1 and Yong‐
chae Cho 1,2
1
Department of Energy Systems Engineering, Seoul National University, Seoul, South Korea
2
Research Institute of Energy and Resources, Seoul National University, Seoul, South Korea

Correspondence should be addressed to Yongchae Cho; yc.cho@snu.ac.kr

Received 29 February 2024; Published 23 May 2024

Academic Editor: Donggeon Kim

Copyright © 2024. Choi et al. Exclusive Licensee GeoScienceWorld. Distributed under a Creative Commons Attribution License
(CC BY 4.0).

The ant-tracking algorithm is commonly used to extract faults in geological structures. However, obtaining 3D ant-tracking
data requires the calculation of various volume attributes. To obtain satisfactory data, it is necessary to iterate through
the parameters of these attributes to achieve reasonable continuity. Moreover, due to the numerous parameters involved,
the algorithm can produce different outputs with each execution. In this study, we aimed to enhance the performance
of the ant-tracking algorithm by combining it with a basic U-Net structure. The input and corresponding labels to the
model are "cubic shaped" 3D data segmented from the original 3D seismic volume to facilitate cross-validation with
distinct regions. We used the label data as the single ant-tracking result to minimize the operator’s bias by executing
the ant-tracking with several different parameters and executors and then taking the average. An evaluative comparison
of three different loss functions (MAE, RMSE, and MSE) was conducted to identify the optimal function for training
the model. Across five out of the six metrics, MSE function demonstrated predominant performance, leading to its
adoption. Apart from this, a significant number of misinterpreted faults led us to propose the post-processing algorithm
named "Dual-Threshold Iteration." It was initially used to extract fine blood vessels branching out from large vessels in
medical image segmentation and adapted in our work to ensure a high level of continuity while ignoring worthless noise.
Comparison with the F1 score and the number of 3D-connected components confirmed that the proposed method could
generate reduced bias and smoothly connected fault structures.

1. Introduction manually digitizing faults to utilizing specific algorithms.

Therefore, considerable effort has been devoted to devel‐
Fault structures are crucial patterns for tracking the oping algorithms and methods to enhance fault detection
accumulation of hydrocarbons and movement in under‐ performance.
ground reservoirs. Thus, fault mapping from three-dimen‐ Various volume attributes, including semblance [1],
sional seismic data is an essential interpretation step for variance [2], and coherency [3] are commonly employed for
evaluation. Despite the importance of faults, especially in fault detection. However, seismic attributes are vulnerable
areas with numerous layers, their quantitative interpretation to both seismic noise and complex stratigraphic features,
remains a time-consuming and challenging task. Addition‐ introducing a potential for significant misinterpretation
ally, because individual operators manually select faults, the [4]. Unlike the single volume attributes, the ant-tracking
result varies based on the evaluator. The task also consis‐ algorithm initiates with a fault attribute and utilizes "ant
tently requires expertise to ensure reliability. This emphasi‐ agents" that follow assumed fault features. The algorithm is
zes the necessity of advanced techniques that progress from usually applied to the variance attribute by a predefined set

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 2 Lithosphere

Figure 1: Ant-tracking results conducted by distinct operators with different hyperparameter combinations (cutoff value, aggressive‐
ness/passiveness, ant-agent movement, etc.). Blue clusters represent the estimated fault structures, while red dotted lines indicate the major
fault trend lines.

Figure 2: The workflow of ant-tracking algorithm employed in software. Orange boxes show the volume attributes or calculations
executed. Parameters displayed in blue boxes represent the range of values adjusted by executors.

of parameters such as orientation ranges and step sizes. This include excessive false positives and noise-sensitive seismic
application allows for the extraction of faults while limiting attributes that could have been eliminated with the manual
the influence of random noise [5]. approach, all of which highlight the need to improve the
However, despite these advantages, ant-tracking also algorithm’s efficiency [7].
has several limitations. The most serious drawback of Recently, methods based on machine learning (ML)
the ant-tracking algorithm is that it requires a considera‐ have found successful applications in the geoscience field.
ble number of seismic attributes to derive results, which Especially, several convolutional neural network (CNN)
increases the number of hyperparameters that need to be methods have been proposed to extract faults with
adjusted. Until satisfactory results are obtained, the operator multiple seismic attributes. Wu et al. [7] implemented fault
must iteratively adjust the parameters of each attribute. segmentation using a conventional U-Net trained with
Furthermore, predicting the final result during intermediate synthetic seismic images. Liu et al. [8] utilized a Resid‐
processes is difficult. This situation leads to using ant-track‐ ual U-Net (ResUNet) for more accurate fault structure
ing in a trial-and-error approach, with executors differ‐ prediction than conventional U-Net-based algorithms,
ing in their interpretations of what noise artifact is or is utilizing artificial training data. Cunha et al. [9] proposed
not. Consequently, the executor’s bias is reflected in the a transfer learning approach with a base model trained
outcomes to an increasing extent [6]. Other disadvantages on synthetic data, aiming to reduce tuning and training

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 Lithosphere 3

Figure 3: The architecture of 3D U-Net used in this study. The green and blue rectangles indicate the encoder and decoder, respectively.
The numbers located on the lower left side of the rectangles indicate the number of channels.

Figure 4: (a) An input seismic data, (b) the result of ant-tracking, obtained after averaging, used as the ground truth, (c) label binarized
based on the top 10% criterion, (d) label binarized based on the top 5% criterion, and (e) label binarized based on the top 2% criterion. The
trade-off between connectivity and precision is illustrated through the label data binarized at different thresholds. Yellow circle indicates
the discontinuity of fault structures, while green circle highlights the seismic noise burst out. The colormap shown in this figure is identical
in other figures as well.

time. Similar to conventional auto-detection methods, ML synthetic data, as these data often fail to capture intricate
allows for faster execution compared to manual approaches. details [10]. Additionally, training with a large amount of
Furthermore, once a robust edge detection model is guaranteed manual data is time-consuming and challeng‐
established, results can be obtained without parameter ing.
adjustments or human evaluation. However, ML-based Our goal is to imitate the overall less-biased ant-tracking
methods are inherently dependent on the training data, and workflow by using a deep learning model. To mitigate bias,
can produce uncertain results when analyzing geological we aggregate the ant-tracking results obtained from various
structures that have not been trained before. Synthetic data executors using a combination of different parameters, each
as training data is ineffective because these data mostly following their unique standards. The initial seismic volume
lack the complexity of real geological structures, resulting in and corresponding ant-tracking results were cropped into
outcomes that can vary depending on the specific synthetic a smaller 3D volume, followed by various data augmenta‐
dataset employed. Training with manually selected data, tion processes. We employed a U-Net model with a newly
similarly, does not differ significantly from training with introduced post-processing algorithm named Dual-Thresh‐

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 4 Lithosphere

Figure 5: Pseudocode of the Dual-Threshold Iteration. Lines 6 to 12 represent the binarization process using a high threshold. Lines 14 to
26 indicate the iterative process while checking the total number of fault pixels, neighborhood pixels, and determining whether they fall
below a low threshold.

old Iteration (DTI) to enhance the continuity and model this procedure was repeated. While taking the variance,
performance. Consequently, combining the U-Net-based vertical-moving triangular shape smoothing and variance
model and post-processing algorithm demonstrates our filter size must be selected. The parameters were adjusted
comprehensive approach to fault detection in seismic aggressively in the first ant-tracking step, while in the next
volumes. step, they were adjusted passively. "Possible illegal step" is
the expanded search range when the ant agent cannot find
2. Methodology the edge in its current location. "Legal steps required" is
a necessary number of valid steps after an illegal step. If
2.1. Ant-Tracking Seismic Attribute. Supervised learning the ant agent stepped illegally to find the fault structure,
requires labeled input data. To acquire validly labeled there should be at least a 'legal steps required' amount of
images for learning, it is crucial to address the bias issue edge pixels. The larger possible illegal step allows the agent
when the ant-tracking algorithm is executed. There is no to look for the fault aggressively, and the lower legal steps
guarantee that the obtained results by a single operator required lead to less restrictive and drastic connections.
accurately mimic "True Faults" in the seismic volume, and Following this, structure smoothing was performed once
even the criteria for satisfactory results can vary depending more, and values below a certain threshold were cut off to
on the operator’s tendency. Figure 1 depicts fault volumes create the baseline value of −1.
identified by different executors and parameters, highlight‐ As we explained, ant-tracking is not a fixed procedure
ing the significant variations in fault trends. with predefined steps and formulas; rather, it involves the
The exact volume attributes that operators use to appropriate adjustment of parameters by an operator as
generate ant-tracking data are introduced in Figure 2 with they proceed. Therefore, using a single labeled ant-tracking
the adjustable parameters. Initially, structure smoothing dataset will likely result in biased outcomes. To address
was performed before calculating the variance from the this, diverse datasets, each with its standard, were construc‐
initial seismic volume data. We basically performed data ted using data from six other operators who adjusted the
smoothing by local averaging using a Gaussian filter. The parameters differently. Additionally, data considered too
variance of the Gaussian filter determines the degree of blur inaccurate were removed through quality control before
and thereby enables the removal of noise from the initial averaging the data into a single ant-tracking labeled dataset.
data. Subsequently, ant-tracking was employed to process Figure 2 presents a simple workflow of the procedure.
the smoothed structure after calculating the variance, and

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 Lithosphere 5

Figure 6: (a) A data distribution of ground truth volume and (b) data distribution of random 2D sliced ground truth image.

2.2. U-Net Model Architecture. U-Net, a CNN-type lossMSE = N

∑iN= 1 yi − ti
1 2
(3)
architecture, is particularly effective for image segmenta‐
where yi and ti denote the output and label and N
tion tasks [11]. As shown in Figure 3, it has a distinctive
U-shaped structure: the encoder (contracting path), shown
on the left, captures the context, and the decoder (expand‐ indicates the batch size.
ing path), shown symmetrically on the right, enables the
precise localization. A noticeable feature of U-Net is the 2.3. Dual-Threshold Iteration. The label data range from 0
skip connections, which connect layers of equal resolu‐ to 1 after min-max normalization. Accordingly, the result
tion in the encoder to the decoder. These connections value of the model to be interpreted also has a value
help the network to propagate context information to between approximately 0 and 1. However, while analyz‐
higher resolution layers, facilitating precise segmentation. ing the results, the interpreter needs to determine which
However, since the ant-tracking algorithm is based on data pertain to the fault and which data do not, thus it
seismic attributes analyzed in three-dimensional context, is natural to think in terms of binary values (right or
it is more reasonable to build the model with three- wrong) rather than a continuous interpretation. In other
dimensional volume training data. Therefore, we used 3D words, although the model was trained to fit continuous
U-Net to imitate the averaged ant-tracking labels. Figure 3 values, fault extraction is considered as image segmenta‐
illustrates the comprehensive architecture of the 3D U-Net tion. Therefore, in the interpretation stage, values above a
model used in this research and shows that its depth is "certain value" are interpreted as true values and other parts
one less than the 2D version, a modification caused by the can be said to be false by binarizing them to 0 and 1.
constraints of GPU memory availability. This approach raises the question of how to determine
Considering that our label has continuous values, we the appropriate "certain value." In most binary segmenta‐
used the loss functions ordinarily used in regression rather tion tasks, data are divided into two classes based on 0.5
than cross entropy to follow the fault existence tendency. as the threshold since the label is binary, and any value
The details will be discussed when describing the evalua‐ exceeding 0.5 suggests a higher likelihood of being classified
tion metrics. Three different measures of regression loss as 1. However, the value of 1 in the ant-tracking data, being
are available: Mean Absolute Error (MAE), Mean Squared the result of normalization, does not represent a probability
Error (MSE), and Root Mean Squared Error (RMSE), which of 100%. Therefore, setting a threshold of 0.5 as a basis
are defined as follows: for division at 50% was deemed inappropriate for this
study. Instead, it was considered preferable to interpret only
lossM AE = ∑iN= 1 yi − ti
values above a "certain upper percent" as faults in the entire
N
1
(1) structure. Unfortunately, the result contains noise, making it

lossRMSE = ∑iN= 1 yi − ti
challenging to achieve a small number of false positives with
N
1 2
(2) high continuity and details.

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 6 Lithosphere

Figure 7: Comparison of the model’s output for different loss functions across each cross-validation, illustrated with a random 2D image
slice example. (a) An input seismic data, (b) ant-tracking result used as a ground truth, (c) U-Net’s output using MAE loss function,
(d) U-Net’s output using RMSE loss function, and (e) U-Net’s output using MSE loss function, and (f)–(j) and (k)–(o) show the different
cross-validation cases. The green circles highlight the enhancement of the valid signals when using the MSE loss function. The blue circles
show that MSE loss function can capture the moderate signals.

Figure 8: Comparison of the model’s output for different loss functions in second cross-validation case. These figures contain data with the
same meaning as in Figure 7. The red circles indicate the presence of residual blurred noise signals.

Figure 4 illustrates that adjustment of the percentage information but also includes noise, as marked with the
yields varying outputs. A low threshold provides detailed green circle. A high threshold case captures significant

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 Lithosphere 7

Figure 9: Comparison of the model’s output for different threshold values of DTI in the first cross-validation case, illustrated with a
random 2D image slice example. (a) An input seismic data, (b) ant-tracking result used as a ground truth, (c) binarized model’s output
at the upper 5% threshold, (d) model’s output binarized at the top 5.5% threshold, (e) model’s output binarized at the top 3% threshold,
(f) DTI result with top 3, 5.5% thresholds, (g) model’s output binarized at the top 6% threshold, (h) model’s output binarized at the top 2%
threshold, and (i) DTI result with top 2, 6% thresholds. The green circles show a reduction in noise signals.

signals but lacks continuity and subtle faults, as is marked are updated into true values. This iteration is repeated,
with the yellow circle. To overcome this trade-off, this study progressively expanding the scope of adjacent pixels, until
introduces the "Dual-Threshold Iteration" algorithm [12]. no further updates occur.
This algorithm was originally developed for extracting Figure 5 presents the algorithm’s pseudocode with some

percentage is defined as θ, we set a high bound value

fine blood vessels branching out from large vessels in retinal modifications from the original version. When the target

γℎ corresponding to the high threshold’s percentage θℎ .

extraction. When images of vessels are extracted from the

Similarly, we set a low bound value γl and low threshold’s

dataset, thick vessels are well observed, but fine vessels

percentage θl . The algorithm initializes values to 1 for

located next to the main vessel may not be adequately

pixels that exceed the high threshold γℎ . While iterating,

connected or extracted. To address this issue, DTI is applied
iteratively extending the image. Initially, a high threshold
is applied to eliminate noise signals that may confuse the
bound value γl are transformed to true values until they
the pixels adjacent to true values and in excess of the low
model. By doing so, only the main vessels are left with‐
reach the target percentage θ. If there are no more pixels
out noise and subsequently, the pixels that are adjacent to
true values and have values higher than the low threshold

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 8 Lithosphere

Figure 10: Comparison of the model’s output for different threshold values of DTI in second cross-validation case, illustrated with a
random 2D image slice example. These figures contain data with the same meaning as in Figure 9. The green circles show a reduction in
noise signals.

to update even though the target percentage is not satisfied, only to secure GPU storage space, but also to set distinct
they expand the range of neighboring pixels generously to regions to facilitate cross-validation.
find additional fault candidates. Given the limited quantity of possible training data,
these data were considered insufficient to train the model.
3. ML-Based Fault Extraction Workflow Hence, data augmentations were performed to increase the
diversity of the datasets. Despite the existence of var‐
3.1. Data Preparation and Implementation. We used a ious augmentation schemes such as shearing and blur‐
seismic volume (14.905 km × 12.155 km × 1844 ms) with ring, only a few of these schemes were utilized in this
55 m spacing and 4 ms sampling intervals (271 [inline] × study. Usually, data augmentation during training helps
221 [crossline] × 461 [vertical]) in this study. Because of to achieve variety; however, in this case, the input was
the lack of GPU resources, the model used in this study seismic data, which contains stratigraphic features, thus
received cropped seismic data as input. First, the 3D volume the neural network was prevented from learning irrelevant
cube was transformed into sixteen small cubes by cutting and unrealistic patterns. Consequently, the dataset was
it in the middle of the inline and crossline directions and augmented exclusively through mirror flipping and rotation
dividing it into four parts in the vertical direction (135 × around the vertical axis, with the number of randomly
110 × 115). Cropping into small cubes was necessary not cropped cubes expanding from 16 to 128 as a result of

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 Lithosphere 9

Figure 11: Comparison of the model’s output for different threshold values of DTI in third cross-validation case, illustrated with a random
2D image slice example. These figures contain data with the same meaning as in Figure 9. The green circles show a reduction in noise
signals.

the image augmentation process. Additionally, we further Despite the implementation of data augmentation
randomly cropped the seismic data with a side length of strategies, it was determined that there was insufficient
ninety-six pixels in the data loader since several binary training data to assess the model’s fitness. Hence, the
pooling sequences are meant to be conducted afterward. cross-validation was employed. Initially, from sixteen
It also enhanced the dataset’s randomness to mitigate distinct areas segmented into 135 × 110 × 115, one area was
overfitting. excluded for validation and another for testing, using the

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 10 Lithosphere

R2score y, t =
remaining fourteen areas as the total training set. Consid‐ ∑iN= 1 yi − t
−2
N
∑i = 1 ti − t
ering data augmentation, 112 3D volumes were used for −2 (7)
training, while 8 volumes each were allocated for valida‐
F1score ybin, tbin =
tion and testing sets. Cross-validation was conducted three 2TP
(8)
times, each time training the 3D U-Net model with a 2TP + TN + FP
different combination of areas to evaluate its performance.
ODSscore y, tbin = ∑iN= 1 F1score yi, γ, tbin
The experiments consist of two parts: 1) loss func‐
N
1
γ
max (9)
tion comparison between MAE, RMSE, and MSE, and 2)
application of DTI to the output of the trained model. The

OISscore y, tbin = ∑iN= 1 max F1score yi, γi, tbin

training processes were executed for up to 100 epochs using
N
1
γi
multiple NVIDIA GeForce RTX 3090s by using distrib‐ (10)
uted data parallel package, each equipped with 24 GB of
where yi and ti denote the output and target and N
memory. We selected the model with the lowest validation

indicates the batch size. The value of t in the R2 score

loss as the final model. The batch size was set to 8 for −
all datasets. The learning rates were set to 0.0005, and the
AdamW optimizer with the CosineAnnealingWarmRestart means the mean value of the target data. An R2 score of
learning rate scheduler was utilized [13, 14]. 1 means that every output pixel matches the target, and a
score of 0 indicates that the performance is the same as
predicting all pixels as the average of the target.

of the confusion matrix after binarizing with threshold γ

3.2. Evaluation Metrics. The ultimate purpose of this study Terms from equation. 8 to 10 are illustrated in terms

as yi, γ means output after binarizing. True Positive (TP)

was to emulate the "reliable" overall process of ant-track‐
ing which produces a value approximately ranging from
0 to 1 by using a deep learning model. Additionally, for and True Negative (TN) denote the number of correctly
this reason, there was a need for the evaluation metrics to segmented fault pixels and nonfault pixels, respectively.
become slightly more complex. In the case of ML-based False Positive (FP) and False Negative (FN) denote
models trained on synthetic data or manually picked fault the numbers of incorrect pixels predicted as faults and
data, the label values are binary. In such cases, evaluating nonfaults. The F1 score is calculated as the harmonic mean
the model’s performance is straightforward: the logits will of the precision and recall, where an F1 score reaches its
be binarized based on 0.5, and the model can be assessed best value at 1 (perfect precision and recall) and worst at
using metrics like Accuracy, Precision, or F1 score. 0. It provides a balanced measure of a model’s performance
However, one minor drawback of ant-tracking is that in class imbalance situations or excessive outliers. ODS and
the result of ant-tracking does not directly represent the OIS are standard methods to evaluate the edge detection
probability of a fault’s presence in that space. Thus, as briefly performance with imbalanced classes [15]. By using a
mentioned in Section 2.2, it should be interpreted as an binarized target, these two metrics evaluate the perform‐

a different threshold. tbin denotes the upper 5% binarized

image regression problem that mimics continuous labels. ance of the model by binarizing the model’s output with
However, as noted in Section 2.3, ultimately, it is necessary
to determine whether the structure is a fault, thus requir‐ target ant-tracking label. The reason for binarizing the
ing interpretation as an image segmentation problem as target label based on the top 5% is not due to any other
well. Consequently, this study suggests that it is essential to factor, but because it visually appears to be the point that
evaluate both image regression (in cases where the label is best distinguishes between fault and nonfault terrains This
continuous) and image segmentation (after binarizing the value needs to be determined empirically according to the
label). given data environment. It should be noted that this 5%
For the numerical analysis of the performances, we criterion may vary depending on the executor’s perception
used seven different metrics. We first compared the MAE, or target dataset. Figure 4 shows the difference between
RMSE, Root Mean Squared Log Error (RMSLE), R2 score various percentage criteria.
before binarizing the target and the F1, Optimal Dataset
Scale (ODS), and Optimal Image Scale (OIS) scores were 4. Numerical Examples
measured on the binarized target. The detailed formulas for
the other newly introduced metrics are as follows: 4.1. Determination of Loss Functions. We compared

M AE y, t = ∑iN= 1 yi − ti
models trained on the following three loss functions: MAE,
N
1
(4) RMSE, and MSE. The metrics measured for each of the
three cross-validation cases and their averages are organized

RMSE y, t = ∑iN= 1 yi − ti
in Table 1. Excluding the MAE as an evaluation index, it
N
1 2
(5) was observed that the models using MSE loss outperformed
in most of the cases. Models trained with MAE loss were
RMSLE y, t = ∑iN= 1 log yi + 1 − log ti + 1
interpreted to have a particularly lower MAE value as
N
1 2
(6) they were obviously trained to reduce this loss. However,
the other metrics that were used to measure the overall
performance of the model did not improve upon these

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 Lithosphere 11

Table 1: Model performance comparison by different loss functions. MAE, RMSE, and RMSLE are indicators where a smaller value is
better. In contrast, R2, ODS, and OIS scores are metrics where a larger value is considered better. Except for one case, MSE shows better
performance than other loss functions in the averaged case.

Cross val Loss MAE RMSE RMSLE R2 ODS OIS

1 MAE 1.71.E-02 5.01.E-02 4.22.E-02 0.47 0.56 0.57
1 RMSE 2.01.E-02 4.58.E-02 3.88.E-02 0.52 0.59 0.59
1 MSE 1.49.E-02 3.02.E-02 2.60.E-02 0.77 0.76 0.76
2 MAE 1.62.E-02 4.75.E-02 3.99.E-02 0.54 0.69 0.69
2 RMSE 1.65.E-02 4.03.E-02 3.40.E-02 0.78 0.73 0.72
2 MSE 1.85.E-02 4.04.E-02 3.42.E-02 0.75 0.72 0.72
3 MAE 2.46.E-02 6.94.E-02 5.75.E-02 0.37 0.50 0.49
3 RMSE 2.68.E-02 6.37.E-02 5.28.E-02 0.40 0.51 0.50
3 MSE 2.72.E-02 6.46.E-02 5.35.E-02 0.50 0.51 0.50
AVG MAE 1.93.E-02 5.56.E-02 4.66.E-02 0.46 0.58 0.58
AVG RMSE 2.11.E-02 4.99.E-02 4.19.E-02 0.57 0.61 0.60
AVG MSE 2.02.E-02 4.51.E-02 3.79.E-02 0.67 0.66 0.66

Table 2: ODS, OIS performance comparison by different target binarization threshold. Even if the percentage of the binarized target
slightly varies, the ODS and OIS results are almost the same as the original criteria case.

tbin 5% (Original) 2% 8%
Cross val Loss ODS OIS ODS OIS ODS OIS
1 MAE 0.56 0.57 0.54 0.54 0.56 0.57
1 RMSE 0.59 0.59 0.60 0.59 0.57 0.57
1 MSE 0.76 0.76 0.77 0.77 0.74 0.74
2 MAE 0.69 0.69 0.67 0.66 0.68 0.68
2 RMSE 0.73 0.72 0.70 0.70 0.72 0.72
2 MSE 0.72 0.72 0.71 0.71 0.71 0.71
3 MAE 0.50 0.49 0.54 0.53 0.45 0.45
3 RMSE 0.51 0.50 0.57 0.56 0.46 0.46
3 MSE 0.51 0.50 0.54 0.54 0.47 0.46
AVG MAE 0.58 0.58 0.58 0.58 0.57 0.57
AVG RMSE 0.61 0.60 0.62 0.62 0.58 0.58
AVG MSE 0.66 0.66 0.68 0.67 0.64 0.64

results compared to the other loss functions. RMSE loss loss functions are sensitive to outliers, the loss function is
function showed finer performance in several cases, but on relatively affected by the error introduced by the outliers.
average, MSE was the most dominant. A drawback of using Therefore, the model is generalized by considering outliers.
deep learning models for fault extraction is the occurrence Nevertheless, this class of generalization should not be
of structures that resemble noisy dots, which are recog‐ considered to have been inappropriate in all cases. In some
nized as faults despite not being faults, with the numerical cases, it is essential to ignore outliers with MAE loss, and
interpretation resulting in numerous false positives [11]. in other cases, it is necessary to generalize with MSE or
From this perspective, an improvement in the ODS and OIS RMSE loss by considering outliers [16]. Figure 6 shows the
score that yields improved values with fewer false positives data distribution of the ant-tracking result for the entire
suggests that the model is converging in a positive direction. volume and a randomly selected sliced image. Despite both
MAE is robust against outliers because it allocates an having a log scale on the y-axis, the linear decrease observed
equal weight to each error. This may help to prevent as the value increases suggests an exponential decrease
overfitting; learning on outliers would cause the model to in the actual linear scale. In fact, using the ant-tracking
perform poorly on unseen data. In other words, if there are result directly as a training label is challenging because of
few outliers, it is appropriate to use MAE when attempting the low percentage of meaningful faults. Fault structures
to create a model that is less affected by those outliers. usually occupy a small proportion of the entire 3D volume,
However, in the case of MSE and RMSE, because these making it rigorous for the model to learn patterns based

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 12 Lithosphere

Table 3: DTI performance comparison by different threshold with the top 5% value as a target percentage. The experi‐
combinations. ment compares two scenarios of DTI application: one with
the top 3% and 5.5% as the low and high thresholds and
F1 score another with the top 2% and 6%. The selection of the exact
Cross Val 1 Cross Val 2 Cross Val 3 number of threshold percentages is determined empirically,
5% 0.305 0.321 0.227 influenced by the visually assessed 5% criterion. Here is
a simple guideline for choosing the threshold percentages:
DTI 5.5%, 3% 0.750 0.717 0.478
Employing a more robust high threshold initially focuses
DTI 6%, 2% 0.728 0.706 0.481 on the most reliable fault structures. This approach can
effectively reduce the noise in the intermediate stages.
However, this strategy might terminate the iteration loop
Table 4: Indirect assessment of DTI’s impact through comparison
of 3D-connected components. before achieving the target percentage. Similarly, starting
with a generous low threshold can easily find the additional
3D-Connected Components true pixels, but it may terminate the iteration loop too
quickly, possibly within one or two times. This understand‐
Cross Val 1 Cross Val 2 Cross Val 3 ing informs the strategy of combining thresholds, such as
5% 31 15 106 6/2 and 5.5/3, around the benchmark of 5, where extreme
DTI 5.5%, 3% 15 9 45 values are paired to balance the detection process.
DTI 6%, 2% 9 5 23 Table 3 presents a comparison across these two DTI
cases and the simple upper 5% binarized output using F1
score. Generally, we observed that the application of DTI
on seismic data. Therefore, it can be inferred that effective resulted in an F1 score that was nearly double compared
learning with MSE loss was achieved on ant-tracking data to scenarios where it was not applied. Based on equation.
with a small proportion of significant values in the overall 8, the change in F1 score indicates an improvement in
structure. RMSE also showed notable performance relative precision, resulting from a significant decrease in false
to MSE since it also considered the outlier patterns, but in positive cases, addressing the main challenge of automated
terms of the average case, MSE was superior to RMSE fault detection.
performance. Additionally, to understand the extent to which DTI
One instance of each of the input, ground truth, and enhanced the connectivity while reducing the amount of
results of the model with the MAE, RMSE, and MSE loss negligible noise, the number of 3D-connected components
for different cross-validation cases is plotted in Figure 7. was compared. This measure counts the total number of
The figure shows a single slice of the 3D cube data, where a fault clusters that can be considered as one group within
clearer extraction of fault structures can be seen in the MSE the binarized 3D cube, revealing that the number of fault
case. We see that MSE loss function improves continuities clusters was reduced by more than half in the cases where
of significant signals, as indicated by the green circles, and DTI was applied, as seen in Table 4. This suggests that using
yields more similar to the label. Blue circles in cross-valida‐ DTI somewhat compensated for the phenomenon where
tion case 2 show the MSE loss function can capture the structures that likely would have been connected in the
intermediate-level signals, but the others cannot. In Figure output were interrupted. Overall, cases in which DTI was
8, however, some blurred seismic noise, which is indicated applied delivered improved performance compared to those
by the red circles, still remains in MSE loss function cases. that did not.
It seems to stem from the characteristic of MSE, which Figures 9–11 display plots for each of the three cross-val‐
emphasizes outliers, as is mentioned above. Such noise idation cases, including the seismic input, label, model’s
necessitates alternative methods for its removal. output, high threshold application cases, low threshold
As mentioned earlier, the 5% criterion was selected based application cases, and DTI results. Note that there are
on visual appearances. Therefore, if the ODS and OIS two significantly different semantic objectives. One is

(tbin), it could compromise the reliability of these met‐

metrics rely too heavily on the target binarization threshold "reduction of noise," and the other is "similarity to the
label." Table 3 compares the F1 score, which is related
rics. To address this concern, we conducted an additional to "similarity to the label" as considering false-positive
experiment using different criteria (top 8% and 2%) to pixels. Table 4 compares the number of 3D-connected
demonstrate the effectiveness of the ODS and OIS metrics. components as an indirect indicator of "reduction of noise."
The results, as presented in Table 2, empirically show that Figures 9–11 can show both objectives via 2D sliced
varying the criterion does not significantly alter the trend of components for each cross-validation case. While both
the scores. metrics are important, achieving the maximum index for
both objectives is impossible since the label contains noise,
4.2. Dual-Threshold Iteration Application. We used DTI as shown in Figures 9–11, parts (b) and (c). A result leaving
to reduce the noise while enhancing the connectivity of only the main stem with zero noise would not resemble
fault structures. We applied DTI to the case with the MSE the label, while a perfectly imitated result would inevitably
loss function, which showed the best performance overall. have noise. Thus, achieving high but not overwhelming
To implement DTI, lower and higher thresholds were set values in both dimensions is essential since both objectives

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 Lithosphere 13

are valuable. Although this paper aims to "mimic less- Conflicts of Interest
biased ant-tracking data," extracting faults while excluding
abnormal noise is also critical. Regarding this dilemma, it The authors declare that there is no conflicts of interest
is not a suitable way to determine a single case of DTI regarding the publication of the paper.
percentage combination. If DTI proceeded moderately, it
would converge towards the label while reducing noise Acknowledgments
adequately. If DTI was applied aggressively, it could mitigate
the noise significantly but slightly differ from the label. We thank SLB for providing Petrel license. This research
was partially supported by Korea National Oil Company.
5. Conclusions This work was supported by a grant from the Human
Resources Development program (No. 20204010600250)
The aim of our study was to enhance the performance of the Korea Institute of Energy Technology Evaluation
of the existing ant-tracking method, which has two major and Planning (KETEP), funded by the Ministry of Trade,
drawbacks: first, the application of numerous attributes Industry, and Energy of the Korean Government. Also,
individually led to an excessive inclusion of the operator’s this work was partially supported by the Korea Institute of
bias in the output and made the process overly com‐ Energy Technology Evaluation and Planning (KETEP) and
plex. Second, an excessive number of FPs were obtained, the Ministry of Trade, Industry & Energy of the Republic of
which means that structures were interpreted as seismic
Korea (No. 20224000000120).
faults when they were not. To overcome these drawbacks,
we constructed a 3D U-Net model trained on a single
ground truth, which was an average of various ant-track‐ References
ing results that could arise from an operator’s actions,
thereby minimizing bias as much as possible. This approach [1] K. J. Marfurt, R. L. Kirlin, S. L. Farmer, and M. S. Bahorich,
“3-D seismic attributes using a semblance-based coherency
eliminated unnecessary tasks and enabled the acquisition
algorithm,” Geophysics, vol. 63, no. 4, pp. 1150–1165, 1998.
of less-operator-biased fault structures. We built this model
by adopting the MSE loss function, which showed greater [2] P. P. Bemmel, R. E. Pepper, and Van, Seismic signal process‐
ing method and apparatus for generating a cube of variance
performance across various metrics in cross-validation
values: U.S. Patent. 6,151,555, issued 2000,
tests. Furthermore, to further lower the number of FPs,
[3] K. J. Marfurt, V. Sudhaker, A. Gersztenkorn, K. D. Crawford,
we applied the DTI algorithm to decrease intermediate-
and S. E. Nissen, “Coherency calculations in the presence of
level noise and enhance the connectivity of faults. The structural dip,” Geophysics, vol. 64, no. 1, pp. 104–111, 1999.
reduction of noise signals was illustrated by comparing
[4] Y. Cho, “Stochastic discrete fracture network modeling in
F1 scores and 3D-connected components. Consequently,
shale reservoirs via integration of seismic attributes and
we demonstrated that an ML-based model could function petrophysical data,” Interpretation, vol. 9, no. 4, pp. SG47–
effectively as a fault detector without training on numer‐ SG58, 2021.
ous simple synthetic data. To replace synthetic data, we [5] H. Jun and Y. Cho, “Repeatability enhancement of time-lapse
used stacked ant-tracking results with minimized operator seismic data via a convolutional autoencoder,” Geophysical
bias. To improve precision, we utilized DTI, allowing us to Journal International, vol. 228, no. 2, pp. 1150–1170, 2021.
reduce the quantity of excessive false positives. [6] M. C. Lorentzen, K. BredesenSmitF. W. Het al., “Mapping
In terms of the general applications of our study, directly cretaceous faults using a convolutional neural network – a
applying our trained model may not be suitable because field example from the Danish North sea,” Bulletin of the
its training was from a single area without validation Geological Society of Denmark, vol. 71, pp. 31–50, 2022.
across different regions. We suggest that a more beneficial [7] X. Wu, L. Liang, Y. Shi, and S. Fomel, “FaultSeg3D: using
approach would be to adopt the model’s architecture and synthetic data SETS to train an end-to-end convolutional
our method of generating complex labels through ant-track‐ neural network for 3D seismic fault segmentation,”
ing. Conversely, our DTI algorithm is versatile and can be Geophysics, vol. 84, no. 3, pp. IM35–IM45, 2019.
integrated into any fault extraction methodology during [8] L. N. Liu Naihao, H. T. He Tao, T. Y. Tian Yajun, W. B. Wu
the post-processing phase. Further research is needed to Bangyu, G. J. Gao Jinghuai, and X. Z. Xu Zongben, “Com‐
employ the model training with additional distinctive data mon-azimuth seismic data fault analysis using residual
and apply the trained model when seismic data from UNet,” Interpretation (Tulsa), vol. 8, no. 3, pp. SM25–SM37,
entirely different regions are used as input. Additionally, 2020.
beyond the use of the ant-tracking algorithm, other seismic [9] A. Cunha, A. Pochet, H. Lopes, M. Gattass, and Elsevier Ltd,
attributes like chaos or semblance could also serve as labels “Seismic fault detection in real data using transfer learning
for training the model, offering the potential for a broader from a convolutional neural network pre-trained with
range of applications. synthetic seismic data,” Computers & Geosciences, vol. 135, p.
104344, 2020.
[10] Y. An, J. Guo, Q. Ye, C. Childs, J. Walsh, and R. Dong, “Deep
Data Availability convolutional neural network for automatic fault recognition
Authors do not have permission to share the original data. from 3D seismic datasets,” Computers & Geosciences, vol.
153, p. 104776, 2021.

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest
 14 Lithosphere

[11] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: convolu‐

tional networks for BIOMEDICAL image segmentation,” in
Lecture Notes in Computer Science (including Subseries
Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics), 9351:pp. 234–241, Springer International
Publishing AG, Switzerland, 2015.
[12] W. Liu, H. Yang, T. Tian, et al., “Full-resolution network and
dual-threshold iteration for retinal vessel and coronary
angiograph segmentation,” IEEE Journal of Biomedical and
Health Informatics, vol. 26, no. 9, pp. 4623–4634, 2022.
[13] I. Loshchilov and F. Hutter, “SGDR: stochastic gradient
descent with warm restarts,” arXiv, 2016.
[14] I. Loshchilov and F. Hutter, “Decoupled weight decay
regularization,” arXiv, 2017.
[15] P. Arbeláez, M. Maire, C. Fowlkes, and J. Malik, “Contour
detection and hierarchical image segmentation,” IEEE
Transactions on Pattern Analysis and Machine Intelligence,
vol. 33, no. 5, pp. 898–916, 2011.
[16] Q. Wang, Y. Ma, K. Zhao, and Y. Tian, “A comprehensive
survey of loss functions in machine learning,” Annals of Data
Science, vol. 9, no. 2, pp. 187–212, 2022.

Downloaded from http://pubs.geoscienceworld.org/gsw/lithosphere/article-pdf/doi/10.2113/2024/lithosphere_2024_118/6441552/lithosphere_2024_118.pdf

by guest