IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL.

21, 2024 7501405

Earthquake Prediction Model Based on

Geomagnetic Field Data Using
Automated Machine Learning
Khairul Adib Yusof , Syamsiah Mashohor, Mardina Abdullah, Mohd Amiruddin Abd Rahman , Member, IEEE,
Nurul Shazana Abdul Hamid , Kasyful Qaedi, Khamirul Amin Matori, and Masashi Hayakawa

Abstract— The observation of geomagnetic anomalies appear- Index Terms— Asynchronous successive halving algorithm
ing prior to earthquakes (EQs) is theorized to be generated by (ASHA), automated machine learning (AutoML), earthquake
underground seismic processes. However, these pre EQ anomalies (EQ) prediction model, geomagnetic field, neural network (NN).
can only provide “postdiction” and are still inadequate for
practical applications. So, this study was conducted to pursue
the long-term quest for real EQ prediction models through I. I NTRODUCTION
the adoption of automated machine learning (AutoML), which
automates many laborious routines of model development. In this
study, more than 50 years of geomagnetic field data recorded
at 131 magnetometer observatories globally were acquired.
T HE perturbations in the geomagnetic field could be
either due to external sources like solar wind, lightning,
human activity, and magnetospheric processes as well as
Several features were extracted from them through wavelet internal sources like remanent and induced magnetization in
scattering transform (WST). The features were used as the the lithosphere [1], [2]. Numerous studies have identified
input to model optimization, of which the strategy for automatic
algorithm selection and hyperparameter tuning was performed
anomalous variations due to internal sources that are linked
based on the asynchronous successive halving algorithm (ASHA). to seismic events to detect possible EQ precursors as first
From the implementation of five classification algorithms, neural reported by [3]. The idea behind this is that underground
network (NN) yielded the best-performing model with an processes occurring during the EQ preparatory phase may
accuracy of 83.29%. The results showed that practical EQ perturb the geomagnetic field several weeks to days prior
prediction models could be achievable even for complex systems
to sizeable EQs ([2], [4] and references therein). Several
like lithospheric and seismo-induced geomagnetic processes by
employing AutoML. mechanisms have been proposed to explain the processes
like conductivity changes, electromagnetic field induction,
and electrokinetic effects, in accordance with the lithosphere-
Manuscript received 18 December 2022; revised 3 June 2023; atmosphere-ionosphere coupling mechanism [4], [5]. Past
accepted 10 January 2024. Date of publication 16 January 2024; date of researches were done using various approaches including case
current version 1 February 2024. This work was supported by the Universiti
Putra Malaysia (UPM) under Geran Putra–Inisiatif Putra Muda under Grant
and worldwide statistical studies [6], but their outputs were
9710900. (Corresponding author: Khairul Adib Yusof.) limited to obtaining possible EQ precursors rather than actual
Khairul Adib Yusof is with the Department of Physics, Faculty of prediction models.
Science, Universiti Putra Malaysia (UPM), Seri Kembangan 43400, Malaysia, The prospect of machine learning (ML) for training EQ
and also with the Space Science Center, Institute of Climate Change,
Universiti Kebangsaan Malaysia (UKM), Bangi 43600, Malaysia (e-mail: prediction models has been demonstrated before, mostly
[email protected]). based on seismic waveforms like [7]. Meanwhile, non-
Syamsiah Mashohor is with the Department of Computer and Communica- seismic approaches are not as common as seismic ones
tion System Engineering, Faculty of Engineering, Universiti Putra Malaysia
(UPM), Seri Kembangan 43400, Malaysia. and more inclined toward ionospheric or atmospheric
Mardina Abdullah is with the Space Science Center, Institute of Climate perturbations [8], [9]. Only a few studies have attempted
Change, Department of Electrical, Electronic and Systems Engineering, to develop geomagnetic-based ML models specifically using
Faculty of Engineering and Built Environment, Universiti Kebangsaan
Malaysia (UKM), Bangi 43600, Malaysia. convolutional neural network (NN) and gradient boosting-
Mohd Amiruddin Abd Rahman and Khamirul Amin Matori are with the based methods [10]. It is obvious that ML is a scarcely
Department of Physics, Faculty of Science, Universiti Putra Malaysia (UPM), explored avenue in achieving EQ prediction models probably
Seri Kembangan 43400, Malaysia.
Nurul Shazana Abdul Hamid is with the Space Science Center, Institute due to its inherent complexity, especially to non-experts.
of Climate Change, Department of Applied Physics, Faculty of Science and Hence, automated ML (AutoML) was introduced to automate
Technology, Universiti Kebangsaan Malaysia (UKM), Bangi 43600, Malaysia. the necessary procedures and make ML more accessible.
Kasyful Qaedi is with the Space Science Center, Institute of Climate
Change, Universiti Kebangsaan Malaysia (UKM), Bangi 43600, Malaysia.
An AutoML workflow covers automatic feature engineering,
Masashi Hayakawa is with the Hayakawa Institute of Seismo Electromag- algorithm selection, and hyperparameter tuning that are
netics Company Ltd., (Hi-SEM), UEC Alliance Center, Kojimacho 182-0026, performed iteratively and assisted by an optimizer [11]. It can
Japan, and also with the Advanced and Wireless Communications Research greatly save time and effort by removing the need for humans
Center (AWCC), University of Electro-Communications, Chofu 182-8585,
Japan. to speculate how a better model could be obtained. Despite its
Digital Object Identifier 10.1109/LGRS.2024.3354954 capabilities, we found only one study that applied AutoML to
1558-0571 © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 16,2024 at 17:21:02 UTC from IEEE Xplore. Restrictions apply.
7501405 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 21, 2024

predict casualty rates and economic losses due to EQs [12], month [4]. In this study, extending the time window longer
but none used AutoML for the development of a prediction than seven days was not ideal as it would considerably reduce
model. available datasets due to more observation period overlaps.
Prior reports on pre EQ geomagnetic anomalies and the Based on the criteria, 3762 geomagnetic field datasets were
accessibility of AutoML motivated this study in the pursuit obtained and labeled “seismic.” Binary classification is the aim
of actual prediction models. Vast global geomagnetic field of this study; therefore, datasets that belong to a contrary class
data measured between 1970 and 2021 were collected and that would be labeled “non-seismic” were constructed.
underwent wavelet scattering transform (WST) for feature For each seismic dataset, a non-seismic dataset was searched
extraction. These features were input into an AutoML from the same station by shifting the seismic period earlier
framework, which was optimized by an asynchronous or later by a year, month, or both, done one step at a
successive halving algorithm (ASHA) that tried various time until a period that was not followed by any EQs was
classification algorithms. The performance of the models found. The search was not always successful due to the
produced was evaluated using standard metrics. In addition occurrence of other nearby EQs during the seven-day period
to obtaining good models, we have thoroughly discussed the for some datasets. Ultimately, 3686 non-seismic datasets
operational strategy of AutoML with ASHA. were obtained to aggregate reasonably class-balanced datasets
(n = 3762 + 3686 = 7448). Each of these 7448 datasets is
II. DATA AND M ETHODOLOGY a 10 080-by-3 matrix that corresponds to the number of data
A. Data Description points (7 days × 24 h × 60 min) and the 3 geomagnetic
components.
Large volumes of geomagnetic field data (double-precision
floating-point format, 1-min sampling period) measured C. Wavelet Scattering Transform
globally were acquired from the SuperMAG open-access Possible seismic-related anomalies were previously found
database (www.supermag.jhuapl.edu) as the primary dataset mostly in the vertical component [2], and a few in the
in this study. The data were measured between the years horizontal components of the geomagnetic field [6], [16].
1970 through 2021 at 131 magnetometer stations worldwide, In terms of frequency, anomalies were detected in the ultra and
although the operational periods of certain stations were
extremely low-frequency ranges [2], [17], besides as slowly
discontinuous or commenced much later than 1970. The
varying diurnal fluctuations [15]. Hence, it was essential
magnetometers are either induction-coil or fluxgate types [13].
to extract these anomalies from various components and
Nevertheless, the measured field data are all in a common
frequency ranges simultaneously. This could be achieved
coordinate system comprising geomagnetic northward (N ),
through WST which constructs a scattering network consisting
eastward (E), and downward vertical (Z ) components [6].
of Gabor wavelets {ψ j,k } and scaling functions, φ J . Through
Additionally, the baselines of the data were removed by the
the network, the input data, f was processed in multiple
data provider based on a three-step process (i.e., daily, yearly,
layers where each layer comprises three operational steps:
and residual offset [13]) to ensure comparability between data
convolution (wavelets), nonlinearity (modulus), and averaging
from different locations.
(scaling functions) [18].
The geomagnetic field data were supplemented by EQ cat-
In the zeroth layer, the input data are averaged ( f ∗ φ J )
alog data that were downloaded from the United States Geo-
with the wavelet low-pass filter that captures only the low-
logical Survey online database (www.earthquake.usgs.gov).
frequency component. The high-frequency component is lost
The catalog includes all substantial global EQ events that
during the filtering but is recovered in the first layer (i = 1)
occurred during the same period, defined to have hypocentral
by performing a continuous wavelet transform ( f ∗ ψi,k ) to
depths of 200 km or less and magnitudes of M5.0 or higher.
obtain a set of scalogram coefficients (k = 1, 2, . . . , kn ).
We ensure that the EQ catalog is complete under the selected
A modulus filter is then applied to the coefficients before
parameters. Several procedures were implemented on both
the output is filtered with the wavelet low-pass filter to
acquired datasets to prepare them for the subsequent model
yield the first layer’s scattering coefficients, | f ∗ ψi,k | ∗ φ J .
development.
The same set of operations are repeated in the subsequent
layers, i = 2, 3, . . . , i n , but only i n = 2 is required
B. Data Preparation as the scattering coefficients in layers beyond that possess
The EQ catalog was further declustered to remove fore- less than 1% of energy [19]. In this study, the scattering
and aftershocks (i.e., dependent EQs). An EQ is considered coefficients (or features) were initially collated as a 219 ×
dependent if it occurs within 10 km from a mainshock 10 × 7448 × 3 tensor (i.e., features × time windows ×
epicenter and a 10-day time window of the mainshock datasets × components). Some dimensions of the tensor were
occurrence. It must also have a magnitude of M ≤ Mms − 1, then merged, reshaping it into a 74 480 × 657 matrix to be
where Mms is the mainshock magnitude [14]. The final EQ the input in the model development.
filtration process was performed based on the distance between
the EQ epicenter and the nearest magnetometer station where D. Model Development Using AutoML
they must not be more than 200 km apart [15]. Additionally, The input data, which contains 74 480 data points and
the station data must be available from seven days prior to the 657 features, were first split into 75% training and 25%
EQ up until the time of occurrence, considering that short-term testing datasets. A set of algorithms namely ensemble (ES;
EQ prediction should have a lead time of several days to one weak learners: k-nearest neighbor, discriminant analysis and

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 16,2024 at 17:21:02 UTC from IEEE Xplore. Restrictions apply.
YUSOF et al.: EQ PREDICTION MODEL BASED ON GEOMAGNETIC FIELD DATA USING AutoML 7501405

Fig. 2. Confusion matrices for the three models. The percentage of each
Fig. 1. (Left) Validation losses, L of model optimization with the algorithms class over the total observations is also shown.
with value drops are marked. (Right) The number of cumulative iterations is
also shown.

decision tree), NN, Naïve Bayes (NB), support vector machine

(SVM), and binary decision tree (BDT) classification models
were iteratively trained with various hyperparameter values
that were optimized by ASHA. ASHA works by first selecting
an algorithm randomly and training it with a subset of
the training data. The performance of the trained model
was computed using cross-validation classification error so
that ASHA could decide whether to promote the model by
retraining it using a larger training dataset or select a new
model and its hyperparameter values using random search.
The cross-validation is also generally effective in preventing
model overfitting [20].
At the end of the optimization process, the function returns
the best model that could be tested with various performance
metrics, namely accuracy, precision, recall, F1 -score, and Fig. 3. ROC curve for the three models with AUC values shown in legend.
specificity, given as follows: The markers indicate OOP and labeled with the classification threshold for
each model.
TP + TN
Accuracy = (1) TABLE I
TP + FP + TN + FN
VALIDATION L OSS , L S TATISTICS , N UMBER OF I TERATIONS , AND M EAN
TP O PTIMIZATION T IME FOR E ACH A LGORITHM . T HE D ESIRED VALUE
Precision = (2)
TP + FP OF E ACH PARAMETER OR ATTRIBUTE I S E MBOLDENED
TP
Recall = (3)
TP + FN
TP
F1 = (4)
TP + 12 (FP + FN)
TN
Specificity = (5)
TN + FP
where TP, FP, TN, and FN refer to true and false positives
and negatives, respectively. Since failure to predict upcoming
EQs as well as broadcasting false alarms to the public are
both equally detrimental, we stipulated that the proportion of The value of ρ M = 1 implies perfect predictions of both
correct EQ (i.e., high precision, recall, and F1 -score) and non- classes, even if one of the classes is under or overrepre-
EQ (i.e., high specificity) predictions must be balanced for our sented [21]. We normalized ρ M (i.e., ρ̃ M = (ρ M + 1)/2)
models [21]. to convert the coefficient range from [−1, 1] to [0, 1]
Hence, we also computed the Matthews correlation like other metrics for fair comparisons.
coefficient (ρ M ) that is given by
TP · TN − FP · FN III. R ESULTS AND D ISCUSSION
ρM = √ .
(TP + FP) · (TP + FN) · (TN + FP) · (TN + FN) AutoML-based model optimization was performed for the
(6) first time with the inclusion of all five classification algorithms

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 16,2024 at 17:21:02 UTC from IEEE Xplore. Restrictions apply.
7501405 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 21, 2024

TABLE II was the most computationally efficient with the fastest mean
T RAINING S ET S IZE , n AND H YPERPARAMETERS OF THE B EST M ODEL optimization time of 10.05 s.
FOR THE T HREE A LGORITHMS O PTIMIZED I NDIVIDUALLY
Optimization was re-performed on the three algorithms
individually with 2048 iterations for each, which was higher
than iterations earned by any algorithm as described in Table I.
The intention of the individual optimization was to allocate
more iterations and presumably obtain the best possible model
for each algorithm. Table II summarizes the training set
size, n and hyperparameters that were automatically tuned
by ASHA to obtain the best model for each algorithm.
A lower L was obtained only for ES (i.e., 0.3104 versus
0.3684), meanwhile, a negligible increase was observed for
TABLE III NN (i.e., 0.1859 versus 0.1681). However, the results for
P ERFORMANCE M ETRICS FOR THE T HREE M ODELS . T HE H IGHEST SVM contradicted our initial assumption as the individual
VALUE FOR E ACH M ETRIC I S E MBOLDENED optimization led to a significantly higher L (i.e., 0.4877 versus
0.2110). The results could be attributed to the starkly smaller
n used by ASHA for the SVM model in comparison with the
other two algorithms. This demonstrated the dependence of
ASHA strategy and, by extension, AutoML on the available
algorithms to be selected from that could eventually lead to
different optimization outcomes. We also comment on the
difference in total training time for each algorithm, where NN
and SVM took 69.8 and 55.4 h, respectively. Whereas ES
maintained its efficiency by taking only 22.5 h for the same
for 4096 iterations. Fig. 1 shows the progress of validation number of iterations. Considering that ES’ performance was
loss, L during training (black line) where the algorithms moderately good, its efficiency could be more favorable for
that achieved value drops are indicated by different markers cases of limited computational infrastructure.
on the left axis. Additionally, the number of cumulative The confusion matrices of the models are summarized in
iterations for each algorithm is shown by colored lines on Fig. 2 where the box colors show that ES and NN models
the right axis. At the beginning (i.e., the first 1000 iterations), correctly (blue) and incorrectly (red) predicted both 1 (seismic)
ES (blue line) was the least frequently selected by ASHA and 0 (non-seismic) classes in proportion. In contrast, SVM
shown by its cumulative iterations being the least. However, overly predicted the seismic class resulting in more than half
ES proved to be the best performing in this early stage as of the predictions being incorrect.
it contributed to six consecutive drops (blue diamonds) of The models were evaluated on the test dataset. The receiver
L before being succeeded by SVM (purple upright triangle). operating characteristic (ROC) curve was plotted and the
This observation suggests that ES is an efficient algorithm area under the curve (AUC) was calculated to examine their
to decrease L quickly, although its performance is rather performance at all classification thresholds (Fig. 3). It can be
stagnant in the later stage. In contrast, NN maintained to seen that SVM’s curve (yellow line) is below the reference
be the most preferred algorithm most of the time throughout line (dashed black line) indicating that the model is worse
the optimization (yellow line) although it contributed to three than a random predictor. In contrast, this was not the case for
drops only (yellow left-pointing triangles). The results show ES (orange line) and NN (blue line) models. Their optimal
that NN is a steadily well-performing algorithm without many operating points (OOP) are indicated by the diamond and
drastic improvements but still offered the lowest L ultimately. circle markers, respectively, where they achieved a balance
The temporal evolution of model optimization illustrated in between true and false positive rates (TPR and FPR) at around
Fig. 1 can be summarized in Table I which shows statistics 0.5000 classification threshold. At OOP, the TPR of NN
of L, the total number of iterations, and the mean time of was expectedly higher than ES (i.e., 0.8403 versus 0.7619),
model optimization for each algorithm. In terms of statistics, whereas its FPR was lower (i.e., 0.1740 versus 0.3443).
we are most interested in algorithms that recorded the lowest Parametric performance evaluations of the models are
mean and minimum losses (i.e., L̄ and L min ) as they indicate elucidated in Table III using standard metrics. NN gave the
higher accuracy models. We also looked at highly fluctuating highest accuracy of 0.8329 implying that its predictions are
L (i.e., high standard deviation, L σ ), which represent models correct 83.29% of the time, followed by ES (0.7076) and
that improve the most frequently using hyperparameters that SVM (0.4619). In terms of precision, sensitivity, F1 -score,
changed more volatilely. and specificity, it is apparent that NN surpassed the other
From the table, ES, SVM, and NN satisfy the conditions two models with values of 0.8275–0.8382, indicating it is
with the lowest mean (i.e., 0.4664), highest SD (i.e., 0.0508), the most balanced in achieving both TP and TN predictions.
and lowest minimum (i.e., 0.1681), respectively. Hence, further Finally, NN trumps other models with an ρ̃ M of 0.8328, thus
discussions are focused only on these three algorithms and the ultimately confirming its superiority. This shows that NN is
models that they yielded. As also shown in the table, NN was advantageous for non-linear, complex time series data as has
the most pursued by ASHA with 1041 iterations, whereas ES been reported many times before [22]. As comparisons, our

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 16,2024 at 17:21:02 UTC from IEEE Xplore. Restrictions apply.
YUSOF et al.: EQ PREDICTION MODEL BASED ON GEOMAGNETIC FIELD DATA USING AutoML 7501405

NN model performed better than an AdaBoost model trained [2] K. A. Yusof, M. Abdullah, N. S. A. Hamid, S. Ahadi, and A. Yoshikawa,
on atmospheric data (ρ M = 0.6581) [9], but slightly worse “Correlations between earthquake properties and characteristics of
possible ULF geomagnetic precursor over multiple earthquakes,”
than a LightGBM model trained on DEMETER satellite data Universe, vol. 7, no. 1, p. 20, Jan. 2021, doi: 10.3390/universe7010020.
(mean accuracy = 0.8687) [8]. [3] A. C. Fraser-Smith, A. Bernardi, P. R. McGill, M. E. Ladd,
Through the Minimum Redundancy Maximum Relevance R. A. Helliwell, and O. G. Villard Jr., “Low-frequency magnetic field
measurements near the epicenter of the Ms 7.1 Loma Prieta earthquake,”
algorithm, we ranked the features based on their importance. Geophys. Res. Lett., vol. 17, no. 9, pp. 1465–1468, 1990.
Among the 657 features, two were determined to be at least [4] M. Hayakawa, Earthquake Prediction With Radio Technique. Singapore:
2.5 times more important than the others, where they were Wiley, 2015.
extracted from the N and Z components, respectively. The [5] D. Ouzounov, S. Pulinets, K. Hattori, and P. Taylor, Pre-Earthquake
Processes: A Multidisciplinary Approach To Earthquake Predication
precursory property of the Z component is already a well- Studies. Hoboken, NJ, USA: Wiley, 2018.
established theory to explain pre EQ anomalies by most [6] K. A. Yusof, M. Abdullah, N. S. A. Hamid, S. Ahadi, and
past studies, especially those that adopted the polarization E. Ghamry, “Statistical global investigation of pre-earthquake anomalous
(Z /H ) method [2]. Whereas the N component was recently geomagnetic diurnal variation using superposed epoch analysis,” IEEE
Trans. Geosci. Remote Sens., vol. 60, 2022, Art. no. 2001413, doi:
discovered to be effective at dispersing anomalies over a great 10.1109/TGRS.2021.3093555.
distance, possibly through the ionosphere [15]. Hence, it is [7] O. M. Saad, A. G. Hafez, and M. S. Soliman, “Deep learning approach
concluded that both components might contain information for earthquake parameters classification in earthquake early warning
system,” IEEE Geosci. Remote Sens. Lett., vol. 18, no. 7, pp. 1293–1297,
related to underground EQ generation. Jul. 2021, doi: 10.1109/LGRS.2020.2998580.
[8] P. Xiong, C. Long, H. Zhou, R. Battiston, X. Zhang, and X. Shen,
IV. C ONCLUSION “Identification of electromagnetic pre-earthquake perturbations from the
In this letter, we implemented AutoML to develop EQ DEMETER data by machine learning,” Remote Sens., vol. 12, no. 21,
p. 3643, Nov. 2020, doi: 10.3390/rs12213643.
prediction models based on geomagnetic field data. Vast [9] P. Xiong et al., “Towards advancing the earthquake forecasting by
data (year 1970–2021) were utilized where relevant features machine learning of satellite data,” Sci. Total Environ., vol. 771,
were extracted using WST to capture information in both Jun. 2021, Art. no. 145256, doi: 10.1016/j.scitotenv.2021.145256.
time- and frequency-domains. The extracted features were [10] L. Petrescu and I.-A. Moldovan, “Prospective neural network model for
seismic precursory signal detection in geomagnetic field records,” Mach.
input for model training and hyperparameter tuning that were Learn. Knowl. Extraction, vol. 4, no. 4, pp. 912–923, Oct. 2022, doi:
automatically optimized based on the ASHA algorithm. It was 10.3390/make4040046.
found that SVM performed second best (L = 0.2110) but [11] A. Masood and A. Sherif, Automated Machine Learning, 1st ed.
Birmingham, U.K.: Packt Publishing, 2021.
this was the case only when it was optimized alongside
[12] W. Chen and L. Zhang, “An automated machine learning approach
other algorithms. Meanwhile, ES was only moderately good for earthquake casualty rate and economic loss prediction,”
(L = 0.3684), but its significantly fast training time is worth Rel. Eng. Syst. Saf., vol. 225, Sep. 2022, Art. no. 108645, doi:
mentioning, which could be advantageous if computational 10.1016/j.ress.2022.108645.
[13] J. W. Gjerloev, “The SuperMAG data processing technique,” J. Geophys.
cost is an issue. Ultimately, NN was determined to consistently Research: Space Phys., vol. 117, no. A9, pp. 1–19, Sep. 2012, doi:
output the best models with the lowest L of 0.1859 during 10.1029/2012ja017683.
optimization and highest values for all performance metrics [14] A. De Santis et al., “Precursory worldwide signatures of earthquake
during testing, for instance, 83.29% accuracy. As shown in this occurrences on swarm satellite data,” Sci. Rep., vol. 9, no. 1, p. 20287,
Dec. 2019, doi: 10.1038/s41598-019-56599-1.
study, most algorithms exhibited better performances when [15] K. A. Yusof, M. Abdullah, N. S. A. Hamid, S. Ahadi, and E. Ghamry,
they were optimized together, hence, this is the recommended “Statistical significance of geomagnetic diurnal variation anomalies prior
AutoML implementation to obtain the best possible model. to worldwide earthquakes,” Geografia, vol. 17, no. 4, pp. 366–377, 2021,
doi: 10.17576/geo-2021-1704-25.
Individual optimization, on the other hand, can be useful [16] X. Liu, K. Hattori, P. Han, H. Chen, Y. Chie, and X. Zhao, “Possible
when the best algorithm is already known, and thus AutoML anomalous changes in solar quiet daily geomagnetic variation (Sq)
is used only for hyperparameter tuning. Additionally, the related to the 2011 off the Pacific coast of Tohoku earthquake (Mw
9.0),” Pure Appl. Geophysics, vol. 177, no. 1, pp. 333–346, Jan. 2020,
results supported the existence of geomagnetic seismo-induced doi: 10.1007/s00024-018-02086-z.
signals statistically as a global phenomenon before the [17] A. Schekotov, J. Izutsu, and M. Hayakawa, “On precursory ULF/ELF
M5.0+ EQs. In conclusion, this study demonstrated the utility electromagnetic signatures for the Kobe earthquake on April 12,
of AutoML and its great potential in obtaining reliable EQ 2013,” J. Asian Earth Sci., vol. 114, pp. 305–311, Dec. 2015, doi:
10.1016/j.jseaes.2015.02.019.
prediction models. Development of a near real-time prediction [18] L. Liu, J. Wu, D. Li, L. Senhadji, and H. Shu, “Fractional wavelet
system is a highly prospective work to accomplish, given that scattering network and applications,” IEEE Trans. Biomed. Eng., vol. 66,
the current model accuracies could be further improved. no. 2, pp. 553–563, Feb. 2019, doi: 10.1109/TBME.2018.2850356.
[19] B. Soro and C. Lee, “A wavelet scattering feature extraction approach for
ACKNOWLEDGMENT deep neural network based indoor fingerprinting localization,” Sensors,
vol. 19, no. 8, p. 1790, 2019, doi: 10.3390/s19081790.
The magnetometer data were contributed by the SuperMAG [20] L. Li et al., “A system for massively parallel hyperparameter tuning,”
and its collaborators, whereas the EQ data were by the USGS 2018, arXiv:1810.05934.
EQ Hazards Program. [21] D. Chicco and G. Jurman, “The advantages of the Matthews correlation
coefficient (MCC) over F1 score and accuracy in binary classification
evaluation,” BMC Genomics, vol. 21, no. 1, p. 6, Dec. 2020, doi:
R EFERENCES 10.1186/s12864-019-6413-7.
[1] N. S. A. Hamid, N. I. M. Rosli, W. N. I. Ismail, and A. Yoshikawa, [22] H. I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller,
“Effects of solar activity on ionospheric current system in the Southeast “Deep learning for time series classification: A review,” Data Mining
Asia region,” Indian J. Phys., vol. 95, no. 4, pp. 543–550, 2021, doi: Knowl. Discovery, vol. 33, no. 4, pp. 917–963, Jul. 2019, doi:
10.1007/s12648-020-01734-2. 10.1007/s10618-019-00619-1.

Authorized licensed use limited to: UNIVERSITY PUTRA MALAYSIA. Downloaded on April 16,2024 at 17:21:02 UTC from IEEE Xplore. Restrictions apply.