This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1

Deep Learning Approach for Earthquake Parameters

Classification in Earthquake Early Warning System
Omar M. Saad , Ali G. Hafez, and M. Sami Soliman

Abstract— Magnitude determination of earthquakes is a on the window width, in which P-wave arrival and magnitude
mandatory step before an earthquake early warning (EEW) are estimated. Since using a long window causes the blind
system sends an alarm. Beneficiary users of EEW systems depend zone to be larger, then short-window analysis is required in
on how far they are located from such strong events. Therefore,
determining the locations of these shakes is an important issue
order to gain more time for taking the necessary precautions
for the tranquility of citizens as well. In light of that, this before the arrival of strong waves. Therefore, many researchers
article proposes a magnitude, location, depth, and origin time have tried to use the short-window analysis such as [5], where
categorization using earthquake Ml magnitudes between 2 and 9. a 1-s window is used to discriminate between the far and near
The dataset used is the fore and aftershocks of the great Tohoku sources. Although the accuracy of calculating the magnitude
earthquake of March 11, 2011, recorded by three stations from decreases when using a short window as discussed by Wu and
the Japanese Hi-net seismic network. The proposed algorithm
depends on a convolutional neural network (CNN) which has Zhao [6], this less accurate magnitude is enough to send the
the ability to extract significant features from waveforms that alarm signal of the EEW systems.
enabled the classifier to reach a robust performance in the A deep learning technique is currently one of the lead-
required earthquake parameters. The classification accuracies of ing techniques in the field of machine learning [7] and is
the suggested approach for magnitude, origin time, depth, and recently used in the field of seismology. Contrary to most
location are 93.67%, 89.55%, 92.54%, and 89.50%, respectively. of the machine learning approaches, deep learning does not
Index Terms— Deep learning, earthquake early warn- need preprocessing of the input data as it deals with the
ing (EEW) system, earthquake location, earthquake magnitude. raw data. It is a nonlinear technique that decomposes input
data into multiple processing layers representing data with
I. I NTRODUCTION multiple levels of abstraction and has a greater ability to
extract significant features from the unlabeled data [8]. Deep
E ARTHQUAKE early warning (EEW) system is an impor-
tant tool to protect civilians and critical applications
which could be in danger during strong shakes such as regular
learning has been proposed for earthquake detection [9]–[12],
seismic data inversion [13], and lithology prediction of seismic
data [14].
and high-speed railway trains, construction workers at height,
The main output of this letter is the classification of the
and others. An information aspect contained in this alarm
hyperparameters: location, magnitude, depth, and origin time
signal is the magnitude of the earthquake. The literature is very
of earthquakes based on convolutional neural network (CNN),
rich with articles dedicated to determine the magnitude of seis-
using only 8-s waveform, from three stations, which end 2 s
mic events. Each of these articles uses one of many hypotheses
after the latest P-wave arrival time. The proposed algorithm
explaining the nature of fault rupture [1], [2]. Olson and
is built to be used within the EEW system due to its fast
Allen [3] and Kanamori [4] have shown that the calculated
decision, ability of being evolving, and robust performance.
magnitude during the first few seconds of an earthquake is
It uses the 1970 events with locations around the Tohoku great
proportional to the magnitude calculated from the waveform
earthquake of March 11, 2011. The dataset used is collected
representing the whole rupture. This finding is important to the
from the three Japanese seismic stations N. KKWH, N. RZTH,
EEW systems since the size of the blind zone depends mainly
and N. KAKH.
Manuscript received November 27, 2019; revised February 26, 2020
and April 2, 2020; accepted May 27, 2020. (Corresponding author: II. P ROPOSED A LGORITHM
Omar M. Saad.) One of the main functions of EEW system is the determi-
Omar M. Saad is with the National Research Institute of Astronomy
and Geophysics (NRIAG), Cairo 11421, Egypt, and also with the School nation of the hyperparameters of earthquakes within the first
of Earth Sciences, Zhejiang University, Hangzhou 310027, China (e-mail: few seconds after the P-wave arrival time. CNN is proposed
[email protected]; [email protected]). to classify the earthquake hyperparameters and extract the
Ali G. Hafez is with the National Research Institute of Astronomy and
Geophysics (NRIAG), Cairo 11421, Egypt, also with the Research and significant features from 8-s waveforms, from three stations,
Development Division, LTLab, Inc., Fukuoka 814-0155, Japan, and also with which end 2 s after the latest P-wave arrival time.
the Department of Communication and Computer Engineering, Faculty of These features are fed to the softmax classifier to classify the
Engineering, Nahda University in Beni Suef, Beni Suef 65211, Egypt.
M. Sami Soliman is with the National Research Institute of Astronomy and
earthquake hyperparameters. The CNN consists of three main
Geophysics (NRIAG), Cairo 11421, Egypt, and also with the Department of layers: input, processing, and output layers. First, the input
Communication and Computer Engineering, Faculty of Engineering, Nahda layer is constructed to read and store the input data as a
University in Beni Suef, Beni Suef 65211, Egypt. 2-D tensor which represents the vertical component from three
Color versions of one or more of the figures in this letter are available
online at http://ieeexplore.ieee.org. seismic stations. The processing layer contains several types
Digital Object Identifier 10.1109/LGRS.2020.2998580 of layers: convolutional, max pooling, and activation. Several
1545-598X © 2020 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 of Exeter. Downloaded on June 13,2020 at 17:36:46 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

Fig. 1. CNN general topology.

Fig. 2. Proposed CNN topology.
kernels are used in the convolutional layer to extract the
significant features. Each convolutional layer extracts different seismic stations. The inputs for the CNN are s1 [t0 − 6s :
significant feature maps, where each feature map output, t0 +2s], s2 [t0 −6s : t0 +2s], and s3 [t0 −6s : t0 +2s]. Comparing
out[i, j ], can be obtained as follows: the arrival times of the three stations has two benefits. First,

N 
M the mutual difference between arrival times is a key point in
out[i, j ] = (x[i + n, j + m] ∗ k[n, m]), (1) estimating the travel times and consequently the location of
n=0 m=0 the earthquake. Second, these mutual differences will decide
where x is the input data, k is the filter coefficients (kernel), N if the event is far or near to these three stations. In case
and M are the filter order (kernel size), and i and j vary from the event is near to these assigned stations (the differences
zero to le with step size of stride, where le is the size of the are less than 6 s), then these three stations will be used in
input data, and the stride is the number of shifted sample for event parameters’ classification. In case these differences are
the next convolutional process. For squashing the network and greater than 6 s, it means that the recorded event is far, and the
nonlinearity purpose, rectified linear unit (ReLU) activation other three stations should be used in classification process.
function is used, whose output f [i, j ] is obtained using the In this study, as the sampling rate is 100 samples/S, and the
following formula: input is a 2-D matrix with a size of 3 × 800 samples. The
proposed algorithm has four identical architectures of CNN
f [i, j ] = max (0, out [i, j ]). (2) topologies, and each topology obtains different earthquake
hyperparameters. Therefore, the input is copied four times
Next, the maxpooling layer is utilized to reduce the network to be fed to the four CNN topologies. In order to reach the
dimension to give the flexibility of expanding the number best performance of the proposed algorithm, the number of
of feature maps and allowing the network to be deeper. The convolutional layers, the number of feature maps, the kernel
maxpooling layer output, out1[i, j ], is then determined by size, the dropout rate, and the type of activation function are
out1[i, j ] = max ( f [i + z, j + r ]), (3) tuned using the following procedure.
1) Fixing the network parameters and varying the number
where z and r are the dimensions of the maxpooling window. of convolutional layers.
The processing layer is the core of the CNN, and consecutive 2) Using the optimum network parameters obtained by the
processing layers can be utilized to construct deep network previous step, varying the number of feature maps in
architecture. Finally, the extracted features are fed to the each layer.
softmax classifier to categorize the hyperparameters of the 3) Varying the kernel size using the optimum network
earthquake. The classification process is done in a supervised parameters obtained by the previous steps.
way, where the softmax classifier tries to match the K-mean 4) Varying the type of activation function, using the opti-
derived clusters. The softmax output, Pc , can be obtained as mum network parameters obtained by the previous steps.
follows: 5) Varying the dropout rate, using the optimum network
exp(E T ∗ W j ) parameters obtained by the previous steps.
Pc = L L . (4) In each step, the accuracy of the proposed algorithm is
Q=1 exp(E ∗ W Q )
T
determined, and the highest accuracy refers to the best archi-
Here, E is the extracted features, W is the softmax classifier tecture. Hence, five convolutional layers with 16, 32, 64, 128,
weights, Pc is the probability of the c class, and Q is the and 256 feature maps are enough for a robust classification
class number which varies from 1 to L L. The general CNN performance. Each feature map has a kernel size of 3 × 3.
topology is shown in Fig. 1. The convolution operation is done by sliding this filter over
The proposed algorithm will function as follows: first, the the input. The stride of the convolutional filter is one sample
P-wave arrival time of the event is estimated manually by the in each step. For the output layer, the fully connected layer is
Japan Meteorological Agency (JMA), where this step will be used to wrap up the CNN topology. In order to train the net-
made in real-time detection by one of the algorithms developed work faster and add kind of regularization, batch normalization
in [15]–[18]. Second, if s1 (t), s2 (t), and s3 (t) are the seismic layer is used after each convolutional layer. The dropout layer
data from three different seismic stations, then the arrival times with a rate of 0.2 is used after each pooling layer to avoid the
of the three seismic stations are compared with the latest overfitting problem. Finally, the extracted features are fed to
arrival time which is set as t0 . Third, 6 s are extracted before the softmax classifier to obtain the final output. The proposed
the time t0 , and 2 s are stored after that time for the three CNN topology is shown in Figs. 2 and 3.

Authorized licensed use limited to: University of Exeter. Downloaded on June 13,2020 at 17:36:46 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SAAD et al.: DEEP LEARNING APPROACH FOR EARTHQUAKE PARAMETERS CLASSIFICATION 3

Fig. 3. (a) Architecture of Block A. (b) Architecture of Block B, note that

the number of feature maps is doubled each Block B, such as the first Block
B has 16 feature maps and the last Block B has 256 feature maps.
Fig. 4. Events’ map for the Japanese dataset used in this article.
For training process, each CNN topology, Block A,
is trained separately using Adam optimizer [19] with a learning of an event. K-means clustering algorithm is a method that can
rate of 0.001. The dataset is split randomly to 70% for training be used for seismic zonation depending on observed seismicity
and 30% for testing. The number of epochs is 120, and the instead of gathering geophysical parameters of the region of
batch size is 16. interest such as source parameters, path, and site effects. This
letter proposes a novel approach for partitioning a catalog of
III. DATASET
earthquake spontaneously by using K-means on the mutual
The data from the Japanese Hi-net seismic network has time differences between P-wave arrival times, which repre-
been used in this letter to provide the datasets for training sents the impact of all geophysical parameters, which could be
and testing. The earthquake catalog used in this letter is applied for any different dataset without prior knowledge about
produced by the JMA [20]. The catalog is based on seismic this specific region. Using the K-means algorithm, the dataset
data provided by the National Research Institute for Earth has been clustered into three classes based on these mutual
Science and Disaster Resilience, the JMA, and various insti- time differences. These clusters are shown in Fig. 5(a). These
tutions. Python scripts were used for creating the dataset. cluster identifiers (ids) are the labels for the proposed location
“WIN32” waveform data and arrival times were downloaded classifier. The accuracy of the training dataset reached 99.93%.
and converted using the HinetPy [21]. Karakuwa, N. KKWH, The proposed algorithm accuracy using 30% of the dataset
Rikuzentakata N. RZTH, and Kahoku N. KAKH stations are to test the proposed location classifier reached 94.57%. The
selected to construct the used dataset because they are the most location confusion matrix (CM L ) for the proposed algorithm
used picks of phase arrivals according to the JMA bulletin [22], is
and accordingly, 1970 events during the period from January 1, ⎡ ⎤
176 1 13
2010 to December 31, 2011, have been used with a total
CM L = ⎣ 0 160 11 ⎦. (5)
of 5910 traces. All selected events occur within the region
4 3 222
having 141◦ E < Longitude < 144.5◦ E and 37.4◦ N <
Latitude < 39.2◦ N. The events’ map is shown in Fig. 4. Each For further evaluation of the proposed location classifier
dataset record consists of the waveform data segment from performance, the precision, recall, and F1-score are obtained.
the vertical component for the three stations with 8 s length, These parameters can be obtained as follows:
2 s after the latest P-wave arrival time, and 6 s before that
arrival. Given that the sampling rate is 100 samples per second, TP
therefore, every data record is 2400 (800 × 3) samples long, precision(Pr) =
TP + FP
and the P phase time is extracted from the downloaded manual TP
picked arrival times. Each dataset record is labeled with the recall(Rc) =
TP + FN
matched label from the catalog information according to the precision ∗ recall
required classifications: location, magnitude, and depth. F1 − score(Fsc) = 2 ∗ , (6)
precision + recall
IV. R ESULTS
where TP, FP, and FN are the true positive, false positive,
The proposed system classifies the detected events accord- and false negative, respectively. For the proposed location
ing to their locations, magnitudes, depths, and origin times. classifier, the precision, recall, and F1-score are 94.79%,
The state of the art of this letter is that multiple parameters had 94.58%, and 94.60%, respectively. In order to reduce the span
been obtained with acceptable accuracy. The real-time EEW between the clusters, we used six clusters instead of three
system should have an array picking algorithm functioning to clusters which can significantly decrease the span between
detect the P-wave arrival times. Once these times are declared, the clusters to be less than 200 km as shown in Fig. 5(b).
the CNN will be activated to calculate the required event The same procedure is done to train the proposed location
parameters. classifier with 70% of the dataset and then tested with the
remaining 30% of the dataset. Despite that, the proposed
A. Location Classification location classifier showed a robust classification performance.
In traditional location algorithms, the difference between That is, the precision, recall, and F1-score have been slightly
P-wave arrivals is the key factor for determining the location decreased to 90.67%, 90.33%, and 90.40%, respectively.

Authorized licensed use limited to: University of Exeter. Downloaded on June 13,2020 at 17:36:46 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

Fig. 5. (a) K-mean location clustering (three clusters). (b) K-mean location
clustering (six clusters).

Fig. 7. (a) Border of the three magnitude clusters. (b) Border of the two
depth clusters. (c) Border of the three origin time clusters. (d) Border of
the six location clusters. The colored star represents the event class number
according to the proposed algorithm.

depth distribution of the dataset is shown in Fig. 6(b). The

proposed depth classifier is trained with 70% of the dataset
with an accuracy of 99.56% and tested with the remaining 30%
reaching an accuracy of 92.54%. Meanwhile, the accuracy of
precision, recall, and F1-score of the proposed depth classifier
Fig. 6. (a) Histogram of dataset magnitude. (b) Histogram of dataset depth. is 92.55%, 92.48%, and 92.52%, respectively. The depth
(c) Histogram of dataset time difference between origin and arrival times. confusion matrix (CM D ) of the proposed depth classifier is
B. Magnitude Classification

124 32
CM D = . (8)
The proposed magnitude classifier classifies the events into 13 422
three categories according to Ml, 2 Ml to 3.49 Ml, 3.5 Ml to
4.5 Ml, and greater than 4.5 Ml, and is tested with 30% of the
dataset. The magnitude distribution of the dataset is shown in D. Origin Time Classification
Fig. 6(a) and its magnitude confusion matrix (CM M ) is Time difference between expected origin time and the
⎡ ⎤
474 5 2 earliest P-wave arrival from the three stations is denoted as td .
CM M = ⎣ 22 60 8 ⎦. (7) This value is divided by the K-means into three clusters: less
0 4 15 than 15.5 s, 15.5–22.6 s, and greater than 22.6 s as shown in
the histogram of Fig. 6(c). Then, the proposed origin time clas-
The training accuracy of the proposed magnitude algorithm sifier is trained with 70% of the dataset and tested with the rest
is 98.94%, and the accuracy of the proposed magnitude of the data. The training accuracy reached to 99.06%, whereas
classifier using dataset of testing is 93.67%. The accuracies the accuracy of the test dataset is 89.55%. The accuracies of
of precision, recall, and F1-score are 92.81%, 91.77%, and precision, recall, and F1-score of the proposed origin time
92.29%, respectively. The previous studies showed that nor- classifier are 91.51%, 89.51%, and 90.05%, respectively. The
malizing the waveform could be used such as [23]. This step obtained origin time confusion matrix (CM O ) for the origin
should be optimized, where the normalization process will time classifier is
add a step in calculations which will add a small delay in ⎡ ⎤
the decision of the EEW system. In this regard, the input 366 23 3
waveforms were normalized using the maximum value which CM O = ⎣ 5 56 12 ⎦. (9)
is added as an extra input for the proposed network. The clas- 6 14 107
sifier accuracy improved slightly, e.g., the magnitude classifier
accuracy enhanced to 93.72% instead of 93.67%. Therefore,
we decided to keep the algorithm without using normalization V. A NALYSIS AND R ESULT C OMPARISON
step. For each event, we plot the output magnitude, depth, origin
time, and location clusters corresponding to the proposed algo-
C. Depth Classification rithm as well as the border of each cluster as shown in Fig. 7.
Two categories assigned to the proposed depth classifier are Accordingly, most of the miss-classified magnitude, depth,
less than or equal to 20 km and higher than 20 km. The location, and origin time clusters are near to the border of

Authorized licensed use limited to: University of Exeter. Downloaded on June 13,2020 at 17:36:46 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SAAD et al.: DEEP LEARNING APPROACH FOR EARTHQUAKE PARAMETERS CLASSIFICATION 5

its corresponding cluster; however, few miss-classified clusters the seismic data used in this research. They would also like
are far away from its corresponding cluster border. to thank Dr. Aly Sherif, Nahda University, who made English
For further evaluation of the proposed algorithm perfor- editing to the letter and the anonymous reviewers whose com-
mance, k-fold cross-validation has been done. The dataset is ments had significantly improved the letter. The code of this
split into K partitions, K − 1 is used for training, whereas letter is available at https://github.com/omarmohamed15/Deep-
one is used for testing. This process is repeated K times, and learning-for-earthquake-parameters-classification-in-EEW.
each time, a different partition is used. Finally, the accuracy
from all K test sets are averaged where K is set to be R EFERENCES
10. As a result, the average test accuracies of the 10 folds [1] J. N. Brune, “Implications of earthquake triggering and rupture prop-
are 89.85%, 91.72%, 92.49%, and 91.52% for the location, agation for earthquake prediction based on premonitory phenomena,”
magnitude, depth, and origin time classifier, respectively. Note J. Geophys. Res., Solid Earth, vol. 84, no. B5, pp. 2195–2198, 1979.
[2] Y. Fukao and M. Furumoto, “Hierarchy in earthquake size distribution,”
that, the reported accuracy for the location classier is related to Phys. Earth Planet. Interiors, vol. 37, nos. 2–3, pp. 149–168, Feb. 1985.
the six clusters’ model. Accordingly, the results demonstrate [3] E. L. Olson and R. M. Allen, “The deterministic nature of earthquake
the robust performance of the proposed algorithm. rupture,” Nature, vol. 438, no. 7065, p. 212, 2005.
[4] H. Kanamori, “Real-time seismology and earthquake damage miti-
The use of deep learning EEW system is a new era in gation,” Annu. Rev. Earth Planet. Sci., vol. 33, no. 1, pp. 195–214,
the seismology field. The long short-term memory (LSTM) May 2005.
approach has been proposed to classify earthquakes to near [5] H. S. Kuyuk and O. Susumu, “Real-time classification of earthquake
source, less than 17 km, or far source, greater than 17 km [5]. using deep learning,” Procedia Comput. Sci., vol. 140, pp. 298–305,
Jan. 2018.
However, the proposed algorithm is more comprehensive since [6] Y.-M. Wu and L. Zhao, “Magnitude estimation using the first three
it classifies the earthquake location, magnitude, depth, and seconds P-wave amplitude in earthquake early warning,” Geophys. Res.
origin time to release the alarm in a more accurate way because Lett., vol. 33, no. 16, pp. 1–4, 2006.
[7] J. Schmidhuber, “Deep learning in neural networks: An overview,”
not only it depends on the information of far source or near Neural Netw., vol. 61, pp. 85–117, Jan. 2015.
source but also on several earthquake parameters. [8] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
On the other hand, one of the robust location classifier no. 7553, pp. 436–444, May 2015.
[9] T. Perol, M. Gharbi, and M. Denolle, “Convolutional neural network for
deep learning networks is ConvNetQuake [9]. It shows the earthquake detection and location,” Sci. Adv., vol. 4, no. 2, Feb. 2018,
ability of deep learning to classify the earthquake location in Art. no. e1700578.
a good manner. However, the input of ConvNetQuake is 10-s [10] O. M. Saad, K. Inoue, A. Shalaby, L. Samy, and M. S. Sayed, “Automatic
waveform which does not support EEW systems. In this letter, arrival time detection for earthquakes based on stacked denoising autoen-
coder,” IEEE Geosci. Remote Sens. Lett., vol. 15, no. 11, pp. 1687–1691,
high accuracy could be obtained using only 8-s waveforms, Nov. 2018.
from three stations, which end 2 s after the latest P-wave [11] Y. Chen, G. Zhang, M. Bai, S. Zu, Z. Guan, and M. Zhang, “Automatic
arrival time. waveform classification and arrival picking based on convolutional
neural network,” Earth Space Sci., vol. 6, no. 7, pp. 1244–1261,
Jul. 2019.
VI. C ONCLUSION [12] R. M. H. Dokht, H. Kao, R. Visser, and B. Smith, “Seismic event and
Along this letter, the magnitude, location, depth, and ori- phase detection using time–frequency representation and convolutional
neural networks,” Seismol. Res. Lett., vol. 90, no. 2A, pp. 481–490,
gin time of earthquakes had been classified using features Mar. 2019.
extracted by the CNN. This categorization can be used in [13] B. Liu et al., “Deep learning inversion of electrical resistivity data,”
an EEW system due to the fast decision and high accu- IEEE Trans. Geosci. Remote Sens., early access, Feb. 11, 2020, doi:
10.1109/TGRS.2020.2969040.
racy of the proposed classifier. The performance of this [14] G. Zhang, Z. Wang, and Y. Chen, “Deep learning for seismic lithology
algorithm achieved an accuracy of 93.67%, 89.55%, 92.54%, prediction,” Geophys. J. Int., vol. 215, no. 2, pp. 1368–1387, Aug. 2018.
and 89.50% for magnitude, origin time, depth, and location [15] A. G. Hafez, M. T. A. Khan, and T. Kohda, “Clear P-wave arrival
classifier, respectively. It is worth to say that the obtained of weak events and automatic onset determination using wavelet filter
banks,” Digit. Signal Process., vol. 20, no. 3, pp. 715–723, May 2010.
model will work on earthquakes coming from near locations [16] A. G. Hafez, M. Rabie, and T. Kohda, “Seismic noise study for accurate
of these events used in the selected dataset. This approach P-wave arrival detection via MODWT,” Comput. Geosci., vol. 54,
can be extended to any other area by training the model using pp. 148–159, Apr. 2013.
[17] O. M. Saad, A. Shalaby, L. Samy, and M. S. Sayed, “Automatic arrival
events from the new area. This system can be implemented time detection for earthquakes based on modified Laplacian of Gaussian
with reasonable dataset and computation power for any other filter,” Comput. Geosci., vol. 113, pp. 43–53, Apr. 2018.
area and nearby recording stations with high accuracy of the [18] A. G. Hafez, A. A. Azim, M. S. Soliman, and H. Yayama, “Real-
time P-wave picking for earthquake early warning system using discrete
classifiers as had been proved for the existing case study, wavelet transform,” NRIAG J. Astron. Geophys., vol. 9, no. 1, pp. 1–6,
where the execution training time is 400 s. It is necessary Jan. 2020.
to highlight that this letter did not depend on the geological [19] D. P. Kingma and J. Ba, “Adam: A method for stochastic opti-
mization,” 2014, arXiv:1412.6980. [Online]. Available: http://arxiv.
parameters of this area, but it used the waveform to extract the org/abs/1412.6980
necessary features to make its classification. The simulation [20] (2019). Japan Meteorological Agency. Accessed: Dec. 27, 2019.
has been done using a PC of Intel Processor Core i7-7700HQ [Online]. Available: https://www.jma.go.jp/jma
CPU at 2.80 GHz, Ram of 16 Gb, Nvidia Geforce 1050 GPU, [21] (2019). HinetPy. Accessed: Dec. 27, 2019. [Online]. Available:
https://pypi.org/project/HinetPy/
and windows-10, 64-bit Operating System. [22] (2020). Japan Meteorological Agency Bulletin. Accessed: Feb. 25, 2020.
[Online]. Available: https://www.data.jma.go.jp/svd/eqev/data/bulletin/
ACKNOWLEDGMENT [23] A. Lomax, A. Michelini, and D. Jozinović, “An investigation of
rapid earthquake characterization using single-station waveforms and
The authors would like to thank the Japan Meteorological a convolutional neural network,” Seismol. Res. Lett., vol. 90, no. 2A,
Agency (JMA), Japanese Hi-net seismic network, for sharing pp. 517–529, Mar. 2019.

Authorized licensed use limited to: University of Exeter. Downloaded on June 13,2020 at 17:36:46 UTC from IEEE Xplore. Restrictions apply.