Earthquake forecasting using optimized levenberg–marquardt back-
propagation neural network
MANOJ KOLLAM
Department of Electrical and Computer Engineering
The University of West Indies
St.Augustine, TRINIDAD AND TOBAGO
AJAY JOSHI
Department of Electrical and Computer Engineering
The University of West Indies
St.Augustine, TRINIDAD AND TOBAGO
Abstract: In this study, an effective earthquake forecasting model is introduced using a hybrid metaheuristic machine learning
(ML) algorithm with CUDA-enabled parallel processing. To improve the performance and accuracy of the model, a novel
hybrid ML model is developed that utilizes parallel processing. The model consists of a Chaotic Chimp based African Vulture
Optimization Algorithm (CCAVO) for feature selection and a Hybrid Levenberg-Marquardt Back-Propagation Neural Network
(HLMt-BPNN) for prediction. The proposed model follows a four-step process: preprocessing the raw data to identify seismic
indications, extracting features from the preprocessed data, using optimized ML algorithms to forecast the earthquake and its
expected time, epicenter, and magnitude, and implementing the model using the Python platform. The model's performance is
evaluated using various criteria, including accuracy, precision, recall, F-measure, specificity, false negative ratio, false positive
ratio, negative prediction value, Matthew’s correlation coefficient, root mean square error, mean absolute error, and mean
absolute percentage error. The proposed model achieved an accuracy of 98%, which is higher than the accuracy of existing
earthquake prediction methods.
Keywords: Seismic indicators; Earthquake catalogue;Magnitude predictions; African vulture optimization; GPU,Levenberg-
Marquardt Back-Propagation Neural Network.
Received: April 27, 2022. Revised: May 25, 2023. Accepted: June 22, 2023. Published: August 3, 2023.
1. Introduction
The most frequent natural disaster is an earthquake, which
happens when tectonic plates slide past one another or
laterally. This causes significant losses in human lives and
material goods by disrupting the seas and land masses. The
Richter scale, which ranges from 0 to 9, is used to quantify the
intensity of earthquakes [1] [2] [3]. Strong earthquakes are
those with a Richter scale value greater than or equal to 6.
Furthermore, due to changes in the structure of the region that
is prone to rupture, selective release of tension, and a variety
of additional flaws, earthquakes do not occur on a regular
basis. This proves that the intervals between these seismic
occurrences must be unquestionably erratic in character [4]
[5]. One of the key factors in an earthquake's categorization is
its magnitude. The strength of the earthquake source is shown
via a logarithmic scale. Magnitude is utilized in scientific
study as well as to quickly educate the public about
earthquakes [6] [7] [8].
Numerous research have proposed several forms of
magnitude scales ever since the so-called local (ML) or
Richter scale, which is used to measure earthquake magnitude.
Although these magnitude scales may indicate fundamentally
distinct aspects of the source, they are appropriate for a variety
of magnitude of earthquakes and the distances between
epicenters despite measuring differing seismic wave
attributes. Quantity scales are often empirical. Typically, a
magnitude is calculated using a formula containing a number
of constants [9] [10] from the time and amplitude of a certain
type of seismic wave. These constants are chosen such that, at
least within a particular magnitude range, a new scale's
magnitudes match those of an existing one. On a seismogram,
the length of shaking can occasionally be used to estimate
magnitude. Because of this, there may be more than one
magnitude unit of difference between the values of the various
magnitude categories for both very large and very small
earthquakes as well as for some specific classes of seismic
source. This is due to the complicated physical mechanism
that causes an earthquake [11] [12] [13].
The development of an awareness system utilizing ML has
been a growing area of research across all sectors of
engineering and science as a result of the losses brought on by
an earthquake. Numerous studies have advanced in this
approach. Geologists and earthquake specialists now have a
new, creative technique to assess seismic risk and trigger
future earthquakes that exceeds the traditional, established
ways they had previously used. Earthquake projections can be
divided into two categories: forecast predictions and short-
term predictions [14] [15]. In contrast to long-term estimates,
which are made months to years before it happened, short-
term earthquake predictions are created hours to days
beforehand. The main goal of this research is to use various
ML methods to forecast whether a significant earthquake
would be labelled as a negative or positive event. The model
cannot be solved perfectly using ML alone. A new ML model
is created in parallel to improve the model's accuracy and
performance. Since the parallelism is naturally supplied by
employing the architecture for constructing GPU utilizing
computational techniques, known as the Compute Unified
Device Architecture, the shortcomings of ML using a central
processing unit (CPU) may be solved by Graphic Processing
unit (GPU) implementation (CUDA). The implementation of
hybrid state vector machine (HSVM) algorithm using parallel
processing through CUDA is used to forecast earthquakes.
The foremost contribution of the paper is as follows,
• Chaotic Chimp based African Vulture Optimization
Algorithm (CCAVO) is used for feature selection.
• The CUDA model is used for train the extracted data.
The CUDA model will process the data parallelly. This is the
main advantage of CUDA model.
• Then the prediction is performed with the help of
HLMt-BPNN model. The accuracy of the prediction model is
improved using the ISOA.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
90
Volume 22, 2023
• By comparing the CPU and GPU's respective
computation times, the performance of the proposed model is
compared to that of the present model.
This essay's remaining sections are organized as
follows: Section 2 discussed the literature reviews that were
completed by earthquake forecasting and the history of the
ideas employed in this article. The methodology of the
proposed models was explained in Section 3. the outcomes of
all the approaches are provided, and the best method is
determined by comparison with a few other tried-and-true
techniques are described in Section 4. The study was
concluded in Section 5. The formatter will need to create
these components, incorporating the applicable criteria that
follow.
2. Proposed Methodology
In this article, an HLMt-BPNN algorithm is created for use
with GPUs and the CUDA programming environment. The
computation speed and performance of the forecasting model
were increased by the method by adding the GPU to machine
learning, which further enhanced resilience. Preprocessing,
feature extraction, feature selection, model training, and
predictions are eventually made on an unobserved portion of
the dataset are other procedures that are involved. The
effectiveness of the prediction model is ultimately assessed,
and comparisons are made. Fig.1 depicts the overall
architecture diagram.
The The eight seismic parameters described in Section 3.1
have been subjected to a variety of ML techniques. The
earthquake is classified by considering the threshold value
specifically based on magnitude i.e., Magnitude>5.5
illustrates that the earthquake is occurred. Conversely, the
Magnitude<5.5 shows that the earthquake is non occurred.
This can be stated as 0 or 1 problem which means 0 denotes
non-occurred and the 1 denotes occurred. The major
indention of the proposed model is to deal with the binary
classification issue. After these approaches have been trained,
output on unknown data parameters is generated, and
performance is then assessed in Section 4. The preprocessing
is significant step for enhancing the prediction performance.
The subsequent section describes the feature selection step in
detail.
2.1 Feature Selection
To select the best optimal features from the extracted
features, a Chaotic Chimp based African Vulture
Optimization Algorithm (CCAVO) is used.
1) Chaotic Chimp based African Vulture Optimization
Algorithm
The AVOA (African Vulture Optimization Algorithm) is a
nature-inspired metaheuristic algorithm that was developed
as a tool for optimization. It is based on the observed behavior
of African vultures, which are known for their ability to find
food in a wide range of environments. One potential
advantage of AVOA is its ability to effectively search for
solutions in a wide range of optimization problems, including
those with many variables and complex constraints. It is also
relatively simple to implement, as it only requires a few
parameters to be set by the user. AVOA has been applied to
various optimization problems and has shown to be effective
at finding good solutions. AVOA has been tested on a variety
of optimization problems and has demonstrated its ability to
find high-quality solutions.
Fig. 1. Block diagram of the proposed methodology.
Stage 1: Vulture Group Formation
In the first phase of the CCAVO method, the initial
population of vultures is created and the fitness of all
solutions is evaluated. The vulture corresponding to the best
solution is identified as the first vulture, the vulture
corresponding to the second-best solution is identified as the
second-best vulture using the Eq. (11), and all the other
vultures are assigned to the third group according to the
second criteria. This phase sets the foundation for the
subsequent phases of the foraging stage, in which the
vultures' positions are updated and their fitness values are
reevaluated.
(1)
In this phase, the variables and represent the best
and second-best vultures, respectively and r_1 and r_2are two
random values in the range [0,1] such that their sum is 1. The
value of is determined using the roulette-wheel technique
as shown in Eq. (2).
(2)
In this phase, , which represents the fitness of the first
and second groups of vultures, and n, which represents the
combined number of vultures in both groups, are used.
Stage 2: Vulture Starvation Level
The CCAVO algorithm uses the hunger level of vultures, as
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
91
Volume 22, 2023
calculated by Eq. (3), to determine their exploration and
exploitation behavior. When vultures are not hungry, they
have the energy and endurance to fly long distances in search
of food. However, if they are hungry, they will not be able to
sustain flight for as long and may act aggressively in their
search for food. The hunger level of the vultures,
represented by (), serves as an indicator of their transition
from exploitation to exploration. In this way, the CCAVO
algorithm is able to simulate the real-world behavior of
vultures in the search for food and apply it to the optimization
process.
(3)
Where denotes that the vultures have consumed all of
the available food, is a random variable with a value
between 0 and 1, is a random number with a range of [1,1]
that varies with each iteration, and t is determined by Eq. (4).
(4)
Where the value c determines the likelihood that the vulture
will execute the exploitation stage. In addition, stands for
the current iteration number, for the total number
of iterations, and for a random number between -2 and 2.
When the value of is greater than 1, the vultures begin
the exploration phase and look for new food sources in
diverse places. In the absence of this, vultures enter the stage
of exploitation and search the nearby area for better food.
Stage 3: Search Stage
Vultures can swiftly seek food and identify dead animals
because of their excellent vision in the natural world. But
because they spend a lot of time scanning their surroundings
before taking off, vultures can have trouble finding food. A
long way in search of nourishment. A parameter labelled
in the range [0,1] is used to select which of two different
techniques vultures in the CCAVO can use to check
numerous random sites.
A random number randp1 between 0 and 1 is used to select
one of the strategies during the exploration phase.
represents the position of the vulture in the next iteration of
the optimization process
(5)
-+
(6)
Where, is a random integer between 0 and 1,
is one of the best vultures selected in the current iteration,
is the current iteration's rate of vulture satiation derived
using Eq. (6), and and are the variables' lower and
upper bounds, respectively. is used to give a high
random coefficient at the search environment scale,
increasing diversity and the search for different search space
areas. Eq. (7) calculates , which stands for the separation
between the vulture and the currently optimal one.
(7)
Here, ‘A’ is a randomly chosen number between 0 and 2,
and denotes the location of the ith vulture.
Stage 4: First Exploitation Stage
The efficiency stage of the CCAVO is investigated at this
point. If value is less than 1, the CCAVO initiates the
first phase of exploitation. The selected approach is
determined by the parameter in the interval [0,1]. A
random integer between 0 and 1 is generated at the beginning
of this phase, . If this is greater than or equal
to parameter , the siege-fight tactic is employed gradually.
If not, the circular flying method is employed. As per Eq. (8),
(8)
Where represents the distance between the vulture and
one of the two groups' top vultures, as determined by Eq. (9),
and is a random number between 0 and 1.
(9)
(10)
(11)
(12)
(13)
and are random numbers between 0 and 1.
Eq. (10) and Eq. (11) are used to determine saturated vulture
one and saturated vulture two , and is the chaotic
vector based on chimp optimization.
Stage 5: Second Exploitation Stage (Chaotic Chimp based
Enhancement in AVO) (proposed)
The chaotic maps listed in Table 1 are used to enhance the
performance of CCAVO. These deterministic processes can
also produce random behavior. The update process is
modeled as follows as per Eq. (12)
(14)
where, is the random number in [0,1].
To summarize, the CCAVO algorithm begins by generating
a random population of "vultures" (candidate solutions). Each
vulture then updates its coefficients using its own group's
strategy. During the iteration, the attacker, barrier, chaser,
and driver all estimate the possible locations of the prey. The
candidate solutions also update their distance from the prey.
The adaptive tuning of the and parameters help to avoid
local optima and improve the convergence rate. Additionally,
the value of is reduced from 2.5 to 0 to enhance the
exploitation process. If the inequality is satisfied, the
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
92
Volume 22, 2023
chimps diverge from the prey, otherwise they eventually
converge towards it. Finally, the chaotic maps help to speed
up convergence without getting stuck in local minima.
Table 1: Chaotic Maps
S. No
Name
Chaotic Map
Range
1
Bernoulli
(0,1)
2
Quadratic
(0,1)
3
Iterative
(-1,1)
2.2 Earthquake prediction
The eight seismic parameters have been subjected to
various ML techniques. With earthquakes of magnitude 5.5
and bigger being classified as Yes or 1 and earthquakes of
lesser magnitude as No or 0, the prediction job is approached
as a binary classification issue. These strategies produce
results on unknown data parameters after training.
1) Levenberg–Marquardt backpropagation (LMA)
The Levenberg-Marquardt method is implemented here
using the usual backpropagation approach. The algorithm
bears the names of the researchers who developed it. It is
taught how to train feedforward networks using the
Levenberg-Marquardt method, and it is made clear how much
better neural networks compute when they use this algorithm
rather than backpropagation as is often done. Because of this,
attempts were made to change the LMA-based
backpropagation learning algorithm, which is noteworthy
from the perspective of a contribution. The writers of this
work described every mathematical formulation and function
used to modify the conventional backpropagation along the
lines of LMA. The construction of a "Hessian" matrix using
this approach has the benefit of using initial derivatives with
regard to network weights, which are conveniently handled
by the usual backpropagation. So, the algorithm's overall
computing complexity decreases. The algorithm is
specifically made to reduce the total squared mistakes. A
Taylor series can be used to expand the error vector to first
order if there is little difference between the old and new
weight vectors. The error function can therefore be provided
as per Eq. (15),
(15)
is an error vector and is its element,
and are new and previous weight vector
respectively. When the aforementioned function is minimised
with regard to the new weight vector as per Eq. (16)
(16)
The formula is based on linear approximation, which is
another factor. In order to guarantee the validity of the linear
approximation, the step size is kept small in the LMA while
the error function is reduced. The error function is somewhat
adjusted as a result:
+
(17)
In this, determines the step size. Similar to this, reducing
the error now in relation to results in
(18)
2) Levenberg–Marquardt backpropagation (LMA)
The topology of the BP neural network is shown in Fig. 2.
The network's input layer, hidden layer, and output layer's
corresponding node counts are represented by the letters n, T,
and m. and are used to represent connection weights.
The BP neural network's input value is represented by the
letters , , , and , while the predicted value is
represented by the letters , , , and . The BP neural
network is trained using the following method: the neural
network should be started. According to the criteria for real
prediction, the values are chosen, and the hidden layer
threshold a and output layer threshold b are initialized.
Following that, the neural network's learning rate and the
neuron's excitation function are calculated.
1. The chosen implicit layer excitation function in this
work, f, is:
(19)
Fig 2: Topological structure of NN
2. Determine the buried layer's output. Given that a, ,
and p are known, it is possible to calculate the hidden layer's
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
93
Volume 22, 2023
output :
(20)
The buried layer's N nodes are what make up the formula.
3. Determine what the output layer should be. The
expected output value may be determined from Hl, ,
and y.
(21)
4. Determine the inaccuracy in the forecast. By
deducting from the anticipated output b, one may
derive the model prediction error .
(22)
5. Update the weights. Update and in
accordance with . The following are the expressions:
(23)
where
(24)
where
where, h is the learning rate.
6. Update threshold. Update according to
(25)
where
where (26)
7. Determine if the model has converged, and if not,
return to step 2 to continue the iteration.
3. Result and Discussion
In this section, the performance of the proposed model is
compared with the existing models by using the performance
metrics. The computation time for the CPU is compared with
the GPU-CUDA model. The comparison table is shown in
table 2.
3.1 Dataset description
Modern automatic phase pickers have been compared
using the dataset, a global collection of more than a million
seismic waveforms labelled with both P- and S-arrival
(Mousavi et al., 2020). To evaluate how successfully the
recommended technique identified phases, we used the same
test set (120,000 waveforms). The phase selecting networks
were used to choose P- and S-phase arrivals from the
waveforms after the feature extraction network processed the
waveforms. In order to identify phases and pinpoint arrival
timings, we chose the peaks from the predicted activation
sequences that were higher than a threshold of 0.5. True
positives are those projected selections that are within 0.5
seconds of the manual labelling. The remainder are regarded
as false positives.
3.2 Performance Metrics
Define Utilizing performance metrics including accuracy,
precision, recall, F-measure, RMSE, MAE, and MAPE, the
suggested model's performance is assessed.
Table 2: Comparison of the performance metrics between
the proposed and existing techniques
Method
Accu
racy
Preci
sion
Recall
F-
measu
re
specif
icity
FNR
Proposed
_HLMt_B
PNN
98.00
95.84
97.95
94.57
95.73
0.04
Existing_
LSTM
95.80
92.38
96.58
91.36
92.72
0.07
Existing_
GRU
90.12
90.58
90.99
87.58
90.22
0.07
Existing_
CNN
89.77
86.40
90.10
85.91
86.40
0.09
Existing
ANN
85.80
84.50
88.52
84.25
87.00
0.10
Method
FPR
MCC
NPV
RMSE
MAE
MAPE
Proposed
_HLMt_
BPNN
0.01
96.21
93.74
0.25
0.41
0.36
Existing_
LSTM
0.05
93.78
91.60
5.84
18.87
13.58
Existing_
GRU
0.08
90.68
90.66
57.90
49.96
46.90
Existing_
CNN
0.08
86.63
88.69
32.69
38.87
34.71
Existing
ANN
0.09
83.99
87.95
236.59
183.95
175.78
The tabulated values are shown in the form of graphs. The
performance metrics for the proposed model are higher than
the existing models which are explained separately.
The accuracy values for the proposed and the existing
techniques like LSTM, GRU, CNN, and ANN are 98.00,
95.80, 90.12, 89.77, and 85.80 respectively. The proposed
model produces higher accuracy than the other existing
techniques.
The precision values for the proposed and the existing
techniques, such as LSTM, GRU, CNN, and ANN, are,
respectively, 95.84, 92.38, 90.58, 86.40, and 84.50. The
suggested model has higher precision than the other methods
that are already in use.
The values for recall for the suggested and existing
techniques, such as LSTM, GRU, CNN, and ANN, are 97.95,
96.58, 90.99, 90.10, and 88.52, respectively. In comparison
to other methodologies, the suggested model produces results
with higher recall.
iv) F-Measure
The Proposed HLMt-BPNN's F-measure is compared to
the accuracy of existing models such as the LSTM, GRU,
CNN, and ANN. Fig. 6 displays a graphical comparison of
the F-measure rates.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
94
Volume 22, 2023
Fig 6: Comparison of F-measure between proposed and existing papers
The values for F-measure for the suggested and the
existing techniques, such as LSTM, GRU, CNN, and ANN,
are 94.57, 91.36, 87.58, 85.91and 84.25, respectively. In
comparison to other methodologies, the suggested model
produces results with higher F-measure.
The specificity values for the suggested and the currently
employed techniques, such as LSTM, GRU, CNN, and ANN,
are 95.73, 92.72, 90.22, 86.40, and 87.00, respectively.
Comparing the specificity of the suggested model to other
methods currently in use.
The FNR values for the suggested and existing techniques,
such as LSTM, GRU, CNN, and ANN, are 0.04, 0.07, 0.07,
0.09, and 0.10 correspondingly. Compared to other methods
already in use, the suggested model produces results with low
FNR.
The recommended and current techniques, such as LSTM,
GRU, CNN, and ANN, have FPR values of 0.01, 0.05, 0.08,
0.09, and 0.09, respectively. The proposed approach yields
outcomes with lower FPR compared to existing techniques.
The MCC values for the proposed and existing techniques,
such as LSTM, GRU, CNN, and ANN, are, respectively,
96.21, 93.78, 90.68, 86.63, and 83.99. The suggested model
generates results with higher MCC compared to other
techniques already in use.
The NPV values for the proposed and existing techniques,
such as LSTM, GRU, CNN, and ANN, are, respectively,
98.00, 95.80, 90.12, 89.77, and 85.80. The suggested model
generates results with higher NPV compared to other
techniques already in use.
The FRR values for the proposed and the existing
techniques, such as LSTM, GRU, CNN, and ANN, are 98.00,
95.80, 90.12, 89.77, and 85.80, respectively. The proposed
model generates results with lower FRR compared to other
existing approaches.
The recommended and current techniques, such as LSTM,
GRU, CNN, and ANN, have RMSE values of 0.25, 5.84,
57.90, 32.69, and 236.59, respectively. The proposed
approach yields outcomes with higher RMSE compared to
existing techniques.
The values for F-measure for the suggested and the
existing techniques, such as LSTM, GRU, CNN, and ANN,
are 94.57, 91.36, 87.58, 85.91and 84.25, respectively. In
comparison to other methodologies, the suggested model
produces results with higher F-measure.
The MAPE values for the proposed and the currently used
techniques, including LSTM, GRU, CNN, and ANN, are
0.36, 13.58, 46.90, 34.71, and 175.78, respectively. The
suggested model yields low-loss findings when compared to
other techniques currently in use.
In order to determine the proposed model's processing
performance on CPU and GPU, 4000 epochs are taken into
account because the model's correctness is constant after
2000 epochs. The proposed model is run on an Intel Core i7,
8 GB of RAM, and a GT 1050Ti GPU with 4 GB of RAM
and 768 CUDA cores. The training of 4000 epochs for the
PSVR model using GPU took 200 seconds, but the same task
on the CPU took roughly 900 seconds. As can be observed
the proposed model significantly outperformed the CPU in
terms of computing speed when training the seismic catalog
model.
Table 3: Exploring the Differences in Computation Time between CPU
and GPU
Method
CPU
(ms)
GPU
(ms)
Proposed_HLMt_BPNN
44526
21102
Existing_LSTM
71421
48903
Existing_GRU
74234
55274
Existing_CNN
88706
59001
Existing_ANN
99934
77392
Table 3 shows the computation time differences between
CPU and GPU models. The computation time of the CPU is
high because the data is processed serially. But in the GPU
model, the data is processed parallelly, so the computation
time is low. This is the main advantage of the proposed model
for forecasting an earthquake.
4. Conclusion
In this study, an effective earthquake forecasting model
was presented that employs a hybrid metaheuristic machine
learning algorithm with CUDA-enabled parallel processing.
A novel hybrid ML model was developed to improve model
performance and accuracy, using Chaotic Chimp based
African Vulture Optimization Algorithm (CCAVO) for
feature selection and a Levenberg-Marquardt Back-
Propagation Neural Network for prediction. The Seagull
Optimization Algorithm was also utilized to further enhance
prediction accuracy. The model follows a four-step process
involving preprocessing raw data, extracting features, using
optimized ML algorithms to predict earthquakes, and
implementing the model using the Python platform. The
performance of the proposed model was evaluated using a
variety of performance criteria, and the model achieved an
accuracy of 98%, outperforming existing earthquake
prediction methods. The use of parallel processing in the
model's design enables efficient and fast prediction, making
it suitable for real-time applications. These findings suggest
that the proposed model could be a valuable tool for
predicting earthquakes and potentially mitigating their
impact.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
95
Volume 22, 2023
References
[1] Wu, Y., Hou, G. and Chen, S., 2021. Post-
earthquake resilience assessment and long-term restoration
prioritization of transportation network. Reliability
Engineering & System Safety, 211, p.107612.
[2] Hammad, A. and Moustafa, M.A., 2021. Numerical
analysis of special concentric braced frames using
experimentally-validated fatigue and fracture model under
short and long duration earthquakes. Bulletin of Earthquake
Engineering, 19(1), pp.287-316.
[3] Cremen, G., Velazquez, O., Orihuela, B. and
Galasso, C., 2021. Predicting approximate seismic responses
in multistory buildings from real-time earthquake source
information, for earthquake early warning applications.
Bulletin of Earthquake Engineering, 19(12), pp.4865-4885.
[4] Tena-Colunga, A., 2021. Conditions of structural
irregularity. Relationships with observed earthquake damage
in Mexico City in 2017. Soil Dynamics and Earthquake
Engineering, 143, p.106630.
[5] Triantafyllou, I., Papadopoulos, G.A. and Lekkas,
E., 2020. Impact on built and natural environment of the
strong earthquakes of April 23, 1933, and July 20, 2017, in
the southeast Aegean Sea, eastern Mediterranean. Natural
Hazards, 100(2), pp.671-695.
[6] Zhang, X., Zhang, M. and Tian, X., 2021. Real‐time
earthquake early warning with deep learning: Application to
the 2016 M 6.0 Central Apennines, Italy earthquake.
Geophysical Research Letters, 48(5), p.2020GL089394.
[7] Tzouvaras, M., Kouhartsiouk, D., Agapiou, A.,
Danezis, C. and Hadjimitsis, D.G., 2019. The use of Sentinel-
1 synthetic aperture radar (SAR) images and open-source
software for cultural heritage: An example from Paphos area
in Cyprus for mapping landscape changes after a 5.6
magnitude earthquake. Remote Sensing, 11(15), p.1766.
[8] Khalilpourazari, S. and Arshadi Khamseh, A., 2019.
Bi-objective emergency blood supply chain network design
in earthquake considering earthquake magnitude: a
comprehensive study with real world application. Annals of
Operations Research, 283(1), pp.355-393.
[9] Shi, Y., Liao, X., Zhang, D. and Liu, C.P., 2019.
Seismic waves could decrease the permeability of the shallow
crust. Geophysical Research Letters, 46(12), pp.6371-6377.
[10] Trugman, D.T., Chu, S.X. and Tsai, V.C., 2021.
Earthquake Source Complexity Controls the Frequency
Dependence of Near‐Source Radiation Patterns. Geophysical
Research Letters, 48(17), p.e2021GL095022.
[11] Tsai, V.C., Hirth, G., Trugman, D.T. and Chu, S.X.,
2021. Impact versus frictional earthquake models for high‐
frequency radiation in complex fault zones. Journal of
Geophysical Research: Solid Earth, 126(8),
p.e2021JB022313.
[12] Hutchison, A.A., 2020. Inter‐episodic tremor and slip
event episodes of quasi‐spatiotemporally discrete tremor and
very low frequency earthquakes in Cascadia suggestive of a
connective underlying, heterogeneous process. Geophysical
Research Letters, 47(3), p.e2019GL086798.
[13] Mignan, A., Ouillon, G., Sornette, D. and Freund, F.,
2021. Global earthquake forecasting system (GEFS): The
challenges ahead. The European Physical Journal Special
Topics, 230(1), pp.473-490.
[14] Nandan, S., Kamer, Y., Ouillon, G., Hiemer, S. and
Sornette, D., 2021. Global models for short-term earthquake
forecasting and predictive skill assessment. The European
Physical Journal Special Topics, 230(1), pp.425-449.
[15] Wikelski, M., Mueller, U., Scocco, P., Catorci, A.,
Desinov, L.V., Belyaev, M.Y., Keim, D., Pohlmeier, W.,
Fechteler, G. and Martin Mai, P., 2020. Potential short‐term
earthquake forecasting by farm animal monitoring. Ethology,
126(9), pp.931-941.
[16] Xiong, P., Tong, L., Zhang, K., Shen, X., Battiston,
R., Ouzounov, D., Iuppa, R., Crookes, D., Long, C. and Zhou,
H., 2021. Towards advancing the earthquake forecasting by
machine learning of satellite data. Science of The Total
Environment, 771, p.145256.
[17] Gitis, V.G. and Derendyaev, A.B., 2019. Machine
learning methods for seismic hazards forecast. Geosciences,
9(7), p.308.
[18] Jena, R., Pradhan, B., Beydoun, G., Alamri, A.M. and
Sofyan, H., 2020. Earthquake hazard and risk assessment
using machine learning approaches at Palu, Indonesia.
Science of the total environment, 749, p.141582.
[19] Mousavi, S.M. and Beroza, G.C., 2020. A machine‐
learning approach for earthquake magnitude estimation.
Geophysical Research Letters, 47(1), p.e2019GL085976.
[20] Rundle, J.B., Donnellan, A., Fox, G., Crutchfield, J.P.
and Granat, R., 2021. Nowcasting earthquakes: imaging the
earthquake cycle in California with machine learning. Earth
and Space Science, 8(12), p.e2021EA001757.
[21] Aslam, B., Zafar, A., Khalil, U. and Azam, U., 2021.
Seismic activity prediction of the northern part of Pakistan
from novel machine learning technique. Journal of
Seismology, 25(2), pp.639-652.
[22] Asim, K.M., Moustafa, S.S., Niaz, I.A., Elawadi,
E.A., Iqbal, T. and Martínez-Álvarez, F., 2020. Seismicity
analysis and machine learning models for short-term low
magnitude seismic activity predictions in Cyprus. Soil
Dynamics and Earthquake Engineering, 130, p.105932.
[23] Xiong, P., Long, C., Zhou, H., Battiston, R., Zhang,
X. and Shen, X., 2020. Identification of electromagnetic pre-
earthquake perturbations from the DEMETER data by
machine learning. Remote Sensing, 12(21), p.3643.
[24] Cui, S., Yin, Y., Wang, D., Li, Z. and Wang, Y., 2021.
A stacking-based ensemble learning method for earthquake
casualty prediction. Applied Soft Computing, 101, p.107038.
[25] Zhang, Y., Burton, H.V., Sun, H. and Shokrabadi, M.,
2018. A machine learning framework for assessing post-
earthquake structural safety. Structural safety, 72, pp.1-16.
[26] Zhu, W., Tai, K.S., Mousavi, S.M., Bailis, P. and
Beroza, G.C., 2022. An End‐To‐End Earthquake Detection
Method for Joint Phase Picking and Association Using Deep
Learning. Journal of Geophysical Research: Solid Earth,
127(3), p.e2021JB023283.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
96
Volume 22, 2023
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.11
Manoj Kollam, Ajay Joshi
E-ISSN: 2224-2872
97
Volume 22, 2023
