Deep-Learning-Based Earthquake Detection for Fiber-Optic Distributed Acoustic Sensing

Abstract

In this paper, deep learning models trained with real seismic data are proposed and proven to detect earthquakes in fiber-optic distributed acoustic sensor (DAS) measurements. The proposed neural network architectures cover the three classical deep learning paradigms: fully connected artificial neural networks (FC-ANNs), convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Results demonstrate that training these networks with seismic waveforms measured by traditional broadband seismometers can extract and learn relevant features of earthquakes, enabling the reliable detection of seismic waves in DAS measurements. The intrinsic differences between DAS and seismograph waveforms, and eventual errors in the labelling of the DAS data, slightly reduce the performance of the models when tested with the distributed acoustic measurements. Despites of that, trained models can still reach up to 96.94% accuracy in the case of CNN and 93.86% in the case of CNN+RNN. The method and results here reported could represent an important contribution to the development of an early warning earthquake system based on DAS technology.

Full text

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1
Abstract—In this paper, deep learning models trained with real
seismic data are proposed and proven to detect earthquakes in
fiber-optic distributed acoustic sensor (DAS) measurements. The
proposed neural network architectures cover the three classical
deep learning paradigms: fully connected artificial neural
networks (FC-ANNs), convolutional neural networks (CNNs) and
recurrent neural networks (RNNs). Results demonstrate that
training these networks with seismic waveforms measured by
traditional broadband seismometers can extract and learn
relevant features of earthquakes, enabling the reliable detection of
seismic waves in DAS measurements. The intrinsic differences
between DAS and seismograph waveforms, and eventual errors in
the labelling of the DAS data, slightly reduce the performance of
the models when tested with the distributed acoustic
measurements. Despites of that, trained models can still reach up
to 96.94% accuracy in the case of CNN and 93.86% in the case of
CNN+RNN. The method and results here reported could represent
an important contribution to the development of an early warning
earthquake system based on DAS technology.
Index Terms— Distributed acoustic sensing, earthquake
detection, optical fiber sensors, machine learning
I. INTRODUCTION
IBER-OPTIC distributed acoustic sensors (DAS) are currently
taking a great deal of attention in several application fields
[1,2], becoming a relevant technology to monitor vibrations in
a distributed way in many different scenarios. They permit the
fast monitoring of oscillating mechanical (acoustic) waves that
induce a measurable amount of longitudinal strain in the optical
fiber. Their remarkable features, such as high sensitivity, fast
measurements, and capability to retrieve the entire acoustic
field of a mechanical wave, provide unique solutions for areas
like pipeline supervision [3], structural health monitoring [4],
geotechnical engineering [5], and more recently seismological
monitoring [6]. The sensing capabilities of DAS technology
have been widely enhanced in recent years by the use of
artificial intelligence approaches [7-9], permitting the detection
or classification of specific events based on the recognition of
particular features that characterize some given target events.
Manuscript received XXXXXXXX; revised XXXXXXXX and; accepted
XXXXXXX. Date of publication XXXXXXXXX; date of current version
XXXXXXX. This work was supported in part by ANID Chilean National
Agency for Research and Development, under Projects FONDECYT Regular
1200299, FONDEF IDeA I+D ID20I10089, Fondequip EQM180026 and Basal
FB0008. The work of P. D. Hernández was also supported by Dirección de
Postgrado y Programas (through Convenio PIIC: 007/2019) of Universidad
Técnica Federico Santa María.
The detection and classification of vehicles [7], intruders in
protected and restricted areas [8], and threatening situations in
pipelines [9], are some of the examples that demonstrate the
benefits that machine learning can bring to the field of
distributed acoustic sensing.
Among different applications, the use of DAS technology in
the field of seismology is nowadays growing rapidly. In an early
stage, DAS-based seismic monitoring was primarily exploited
to investigate artificially generated seismic activity, for instance
monitoring reservoirs and obtaining vertical seismic profiling
information of boreholes and wells [11,12]. More recently,
DAS has been used in the monitoring of natural seismic events,
with most scientific works focused on demonstrating the
capabilities of DAS technology for measuring isolated seismic
events under different scenarios [13-15]. This includes
earthquake detection using installed fibers in isolated areas
[16], in telecom cables under cities [6], and even using
submarine optical cables [17-19]. Compared to traditional
seismic networks, based on punctual seismographs separated by
a few tens of kilometers, DAS technology can increase the
spatial sampling of seismic waves in about three orders of
magnitude (down to a few tens of meters) [6]. This feature
provides a disruptive approach for specialists to study the
propagation of earthquakes, which combined with the
possibility of using the worldwide fiber-optic communication
infrastructure for seismic monitoring, grants DAS technology
with very promising projections to become an essential
technology in future seismic monitoring networks.
In the field of seismology, the use of deep learning to detect
earthquakes from traditional seismic measurements has been
exploited in recent years. Early works on the subject have been
based on the use of fully connected artificial neural networks
(FC-ANNs) [20,21], reporting better classification performance
than traditional methods. However, later, the ground-breaking
results obtained by convolutional neural networks (CNNs) in
computer vision and pattern recognition tasks motivated their
use on the classification of seismic signals [22,23], improving
the detection capabilities and sensitivity of its predecessors.
Other approaches based on recurrent neural networks (RNNs)
Pablo D. Hernández and Marcelo A. Soto are with the Department of
Electronic Engineering, Universidad Técnica Federico Santa María, 2390123
Valparaíso, Chile (e-mail: pablo.hernandezdo.13@sansano.usm.cl,
marcelo.sotoh@usm.cl).
Jaime A. Ramírez is with Novelcode SpA, 2580216 Viña del Mar, Chile (e-
mail: jaime@novelcode.io).
Deep-Learning-Based Earthquake Detection for
Fiber-Optic Distributed Acoustic Sensing
Pablo D. Hernández, Jaime A. Ramírez, and Marcelo A. Soto, Senior Member, OSA, Member, IEEE
F

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2
have leveraged on the temporal characteristics of seismic
signals for earthquake detection and seismic phase picking [24].
More recently, novel deep learning models based on
transformers and attentive mechanisms have also been applied
for earthquake detection [25]. Besides detecting the seismic
waves, CNN models have also been applied to the
characterization of other seismic features such as epicenter
location [26,27] and magnitude estimation [28-30]. Although
there has been a great progress on the use of deep learning
approaches in seismology, the limited amount of available DAS
seismic measurements has prevented the use of these
techniques with distributed seismic data. To overcome this
limitation, researchers have trained some deep learning models
with synthetic data generated by simulations. For this, a
generative adversarial network (GAN) approach has been
proposed to produce a large database with training waveforms
[31]. The generative model was trained to adjust simulated data
to become more alike true data measured in a field test. The use
of this GAN enabled a significant improvement of the trained
classifier to differentiate between footsteps, vehicle induced
vibrations and noise [31]. It is however worth noticing that the
detection of earthquakes in distributed acoustic measurements
has yet to be addressed, especially if based on real seismic DAS
measurements.
In this paper, the use of deep learning techniques trained with
real seismic data is proposed and demonstrated, for the first
time to the best of our knowledge, to detect the occurrence of
earthquakes based on DAS measurements. In particular, the
capabilities of three deep learning models to detect seismic
waves based on DAS measurements are investigated. These
three proposed models are based on fully connected artificial
neural networks, convolutional neural networks, and recurrent
neural networks, which are trained with waveforms measured
by traditional broadband seismometers. Results demonstrate
that the use of seismic timeseries obtained by traditional
seismographs allows the proper training of the proposed deep
learning models, which can learn the relevant features of
earthquakes to provide a reliable detection of earthquakes in
DAS measurements. This approach exploits existing large
databases of earthquakes obtained with traditional seismic
instrumentation, overcoming the need of a today-inexistent
large database with thousands of different DAS-based seismic
records. It is believed that the deep learning models and method
here reported are great candidates to improve earthquake
monitoring systems based on DAS technology, enabling
seismologists to study more complete earthquake catalogs.
II. PRINCIPLES AND THEORETICAL BACKGROUND
A. Distributed Acoustic Sensing
Distributed acoustic sensors are based on measuring changes
in the optical phase of the Rayleigh scattering light generated in
an optical fiber [1,2]. Mechanical vibrations reaching an optical
fiber induce a local dynamic strain that modulates the local
refractive index of the fiber. This induces an optical phase shift
of the Rayleigh backscattered light generated in the sensing
fiber when light propagates through it. Detecting local changes
in the Rayleigh optical phase can allow for the retrieval of the
amplitude, frequency, and phase of the perturbation. To obtain
spatially resolved information along a sensing fiber, there exist
two main approaches in the literature [1,2]: i) optical frequency-
domain reflectometry (OFDR), and ii) phase-sensitive optical
time-domain reflectometry (OTDR).
In the OFDR approach, a continuous-wave frequency-swept
optical signal is launched into the sensing fiber and the
Rayleigh backscattered light is combined with a delayed copy
of the input signal into a coherent detector. Making use of
Fourier transform, spatially resolved information about the
Rayleigh optical phase can be obtained with extremely sharp
spatial resolution. Whilst the spatial resolution is normally
around millimetric scale (or even sub-mm), the sensing range
in that case is normally restricted to a few tens or hundreds of
meters. However, OFDR distributed sensors with ranges of
several kilometers have also been reported but with meter-scale
spatial resolutions [32,33].
On the other hand, OTDR uses short optical pulses that are
launched into the sensing fiber to generate Rayleigh scattering.
Different detection schemes are used to dynamically obtain the
Rayleigh optical phase information. Some of the most common
detection schemes used in OTDR are based on coherent
detection (either heterodyne or homodyne detectors) and direct
detection using either interferometric schemes or chirped pulses
[1,2]. In contrast to OFDR, the spatial resolution of OTDR
sensors is normally in the range of 1 to 10 m, allowing for
sensing distances exceeding 50 km.
It is worth mentioning that the Rayleigh optical phase is only
sensitive to axial strain, and therefore DAS measurements could
contain fiber sections with poor strain response if the acoustic
(mechanical) wave reaches the fiber at specific angles. Another
reason that leads to fiber sections with low (or null) acoustic
sensitivity is the local poor strain coupling that could exist
between the acoustic propagating media (e.g., ground) and the
optical fiber. In addition, depending on the interrogating and
detection schemes, DAS measurements could be in principle
also affected by intensity fading points, which correspond to
blind fiber locations where the Rayleigh optical intensity fades
out and the local optical phase extraction becomes unreliable.
As a consequence of these three causes, each location along a
single optical fiber can have a very different sensitivity to
mechanical vibrations, resulting also in distributed acoustic
measurements with longitudinally-varying signal-to-noise ratio
(SNR) [1,2].
B. Deep Learning for seismic DAS measurements
A deep learning neural network model approximates a
function that maps a set of input values onto a desired set of
output values, such that a specific task (like regression or
classification) is accomplished [34,35]. The neural network
itself is composed of a series of computational layers, each one
performing mathematical operations on its input values and
then connecting the results to the inputs of the subsequent layer.
Every layer is further composed of a set of nodes, known as
neurons, which perform a vector to scalar operation to calculate
the layer output in parallel. The main parameters to adjust in

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
3
each node are the so-called weights of the model, which
transform the neurons inputs into their output. A special kind of
nonlinear functions, known as activation functions, are applied
to the output of each layer in the neural network model, so the
model can learn the nonlinear dynamics of data and
approximate more complex functions. There is a vast number
of specific activation functions used on these layers, such as
Sigmoid, rectifying linear units (ReLu) or hyperbolic tangent
function (Tanh) [34], and their election is a matter of the
specific practical application.
In this work, a supervised learning approach is proposed to
develop deep neural networks capable of detecting seismic
signals in DAS measurements, outputting the probability of an
input DAS signal being a seismic waveform. A large database
containing conventional seismic signals and noise measured by
traditional broadband seismometers is used, along with their
corresponding labels, to optimize the parameters of each
proposed model. Batches of seismograph measurements are
passed through the proposed neural networks to estimate the
probability of each input waveform being a seismic signal. The
model predictions and the real target labels of the examples are
used to calculate the following Binary Cross-Entropy (BCE)
cost function [35]:

  
 


(1)
where 
 is the output of the model (i.e., the probabilities of the
input measurements being seismic waves),  is the vector of
true labels for the input waveform set, and  is the number of
examples in each batch. The calculation of gradients with
respect to the complete dataset is computationally very
demanding, so an estimation is obtained with the batch
approach. Here, backpropagation [34,36] is used to calculate
the gradients of the cost function with respect to every
parameter of the network. The network weights are then
updated using the corresponding gradients and a
hyperparameter that controls their amount of change, known as
learning rate [34].
To ensure that a deep learning model correctly learns the
underlying data distribution characteristics, each timeseries of
the dataset is passed through the network more than once during
training. Every pass of the complete training dataset is known
as an epoch. Given a large enough amount of training examples
(waveforms) that capture the characteristics of the real-data
distribution, a well-designed model and the use of the right
optimizer can reach a local minimum of the cost function and
can learn the classification function. However, when the
training procedure is extended beyond a given optimal number
of epochs, the phenomenon of overfitting occurs [34,35]. In this
overfitting regime, the model has seen the training examples too
many times, so that it memorizes the corresponding outputs,
thus being incapable of generalizing the received knowledge
and failing in the proper classification of new waveforms never
seen before. To avoid this overfitting effect, a validation set of
examples completely disjoint from the training data is used to
monitor the performance of the trained models during the
training procedure. In this work, optimal model parameters are
obtained through an early stopping strategy, interrupting the
training process when the loss function has stopped improving.
In addition, regularization techniques, such as dropout and
batch normalization, have been applied to the models to
diminish overfitting [34].
To measure the performance of the trained models, a separate
test set of examples is held apart from the training and
validation data. Every example of this set is passed through the
trained neural network and the output probability of it being a
seismic waveform is compared to the target label. The
classification result of this output probability depends on a
specific threshold defined to divide the probability values that
are considered as seismic, and those considered as noise. Thus,
all metrics used to measure the performance of the model
depend on the threshold as well.
In this work, accuracy, recall, precision and F-score are used
as evaluation metrics [34]. Every metric is a function of the
number of true positives (), true negatives (), false
positives () and false negatives () in the classification.
Among these metrics, accuracy is defined as the total number
of correctly classified examples (including both true positives
and true negatives) with respect to the total number of examples
[37]:
  
 (2)
where  corresponds to the threshold value defined for a
proper classification.
Recall is the number of correctly classified positive results
divided by the total amount of positive examples in the
complete dataset. A high value indicates that the number of
positive examples that the model does not detect is low. It is
defined as [37]

 (3)
Precision is the number of correctly classified positive results
divided by the total amount of examples classified as positive.
A high value indicates that most detections are correct, and thus
the triggers are trustable. This is defined as [37]

 (4)
F-score is the harmonic mean of both Recall and Precision,
so it can be written as [37]

−
  
 (5)
The F-score greatly decreases whenever the Precision or
Recall are deficient. The number of false positives and false
negatives is more important in the calculation of the F-score
than in the accuracy. In addition, F-score is better than accuracy
whenever the number of examples in every class is different.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
4
This is because accuracy can be misleading when the number
of examples in each class is unbalanced, as the correct
classification of the predominant class can hide the errors of the
other. On the other hand, for a model to have a high F-score
value it is necessary to correctly classify both unbalanced
classes, which is more precise in a general context. The model
performance as a function of the classification threshold is
better observed on a Precision-Recall (PR) curve, where the
Precision values are normally shown on the vertical axis and the
Recall values on the horizontal axis.
III. PROPOSED DEEP LEARNING APPROACH FOR SEISMIC DAS
MEASUREMENTS
A. Deep Learning Models
The neural network architectures proposed for this work
cover the three classical deep learning paradigms: fully
connected artificial neural networks, convolutional neural
networks and recurrent neural networks [34,35].
Artificial neural networks, also known as feedforward
networks or multilayer perceptrons, are the quintessential deep
learning models. Each linear layer of the network has a fixed
set of nodes (or neurons) that applies an affine vector-to-scalar
transformation to its inputs. The network defines a mapping
between the values of the input layer and output layer, so that
neurons learn the weights that best approximate a desired
function that completes the task that the model is designed for.
In the specific task of classification, the model approximates a
function    that maps an input vector  to its
corresponding category . The amount of model parameters
increases rapidly with the dimension of the input and number
of layers, because each neuron output is connected to all the
inputs of the subsequent layer. The model can approximate a
nonlinear behavior by the addition of nonlinear functions
between the computation layers.
The proposed FC-ANN model shown in Fig. 1 has an input
layer receiving the complete waveforms (of 6000 samples), two
hidden layers with 6000 neurons, and an output layer with a
single node. As an activation function, ReLu [38] is chosen
after the hidden layers, and a Sigmoid function [35] is then used
at the output layer to obtain a value between 0 and 1, which
indicates the probability of the input being a seismic waveform.
Convolutional neural networks, unlike FC-ANNs, make use
of convolutional and pooling layers to extract features from the
input data. As its name suggests, a convolutional layer applies
a convolution operation over the input data using a set of
learnable filters, known as kernels. On the training procedure,
the weights of these filters are updated so that relevant intrinsic
characteristics of the input data, known as features, are
extracted as the information passes through the network. The
calculated features reach higher levels of abstraction as the
number of convolutional layers increases. The main advantage
of this approach is that the kernel parameters are shared on
every layer, so the total number of parameters of the model is
lower compared with that necessary in an FC-ANN. This
enables the implementation of deeper models with a higher
number of layers.
The proposed CNN shown in Fig. 2 is an adaptation of the
LeNet architecture [39], firstly proposed to solve image
classification tasks and used in several areas of pattern
recognition, including seismic data processing. Most
convolutional models used for seismic detection and phase
picking are variations of this scheme [23,24]. This type of
architecture has an initial feature extraction stage composed of
a set of convolutional layers, followed by a classification stage
composed of linear feedforward layers. As the practical
Fig. 1. Architecture of the proposed fully connected artificial neural
network. Total number of parameters to train: 72,018,001.
Fig. 2. Architecture of the proposed convolutional neural network. Total
number of parameters to train: 27,241.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
5
application of deep learning suggests, deeper models normally
and in general terms obtain better results, and therefore, here an
architecture composed of eight convolutional layers with a 1x3
kernel size is proposed, followed by two linear layers with 32
and 1 output units, respectively. Every convolutional layer is
followed by a ReLu activation function and a batch
normalization layer (BatchNorm) [40] to speed up training and
improve convergence. Pooling layers (MaxPool, in Fig. 2) are
also included to reduce the dimensionality of the output of some
convolutional layers, calculating the maximum value of a fixed
sized window of features. The Sigmoid activation function is
used in the last linear layer to obtain the probability value of the
input timeseries being a seismic signal, just as in the FC-ANN
case.
On the other hand, recurrent neural networks are a family of
networks especially fitted for processing sequential data [34].
This type of architecture processes a sequence of inputs,
learning the model parameters as it retains relations among
inputs in an internal memory state. RNNs share the same
learned parameters across different parts of the sequence, and
the output at a given timestep depends on the input and the
hidden memory state. Mousavi et al. [41] proposed a neural
network architecture, named CNN-RNN Earthquake Detector
(CRED), which makes use of linear, convolutional and
recurrent layers in an efficient residual framework to detect
seismic signals and phases. The use of convolutional neural
networks altogether with recurrent neural networks enables the
extraction of relevant features from the input data in the initial
convolutional layers and learning of temporal characteristics of
data in the subsequent layers. Furthermore, the model makes
use of Batch Normalization layers to enable faster and more
stable training, and Dropout layers that prevent overfitting by
randomly setting some weights of hidden layers to zero [40].
CRED takes the short-time Fourier transform (STFT) of
seismogram recordings as an input and makes use of a specific
type of recurrent layer, known as long short-term memory
(LSTM) [43], in its unidirectional and bidirectional versions.
In this work, a CNN+LSTM model based on the CRED
architecture is proposed, as shown in Fig. 3. As the figure
illustrates, the two-dimensional convolutional layers used by
CRED are here changed to one-dimensional convolutional
layers, so they can process timeseries of DAS seismic data. To
compensate for the less information that the one-dimensional
representation of DAS data provides in comparison to the
original three-dimensional seismograph data used by the CRED
model, the convolutional kernel size of our CNN+LSTM model
is modified to 1x3 to obtain bigger feature maps after every
layer. In addition, compared to CRED, the last layer of our
model has only one neuron with a sigmoid activation function
that outputs the probability of the entire input waveform being
a seismic wave.
B. Seismic Datasets
The data used to train the neural network models in this work
correspond to 60 second local earthquake records measured by
conventional seismographic instruments, obtained from the
STanford Earthquake Dataset (STEAD) [44]. This dataset is a
large-scale, global collection of timeseries specially designed
for artificial intelligence research tasks. The database comprises
about 19,000 hours of seismic recordings and approximately
100,000 noise timeseries, stored as three component records of
ground motion in east-west, north-south, and vertical direction.
However, as we test our models on one dimensional data
measured by DAS arrays, only the east-west component is
extracted from the seismometer measurements for training.
Note that this dataset is characterized by being composed of
seismographic records obtained for more than 30 years from
several sources, various magnitudes, and covering many
Fig. 3. Architecture of the proposed CNN+LSTM model based on the
CRED architecture. Total number of parameters to train: 476,689.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
6
geographic locations around the world [44], ensuring a good
representation of the real distribution of seismographic events.
All signals have been detrended, band-pass filtered between 1-
45 Hz and resampled to 100 Hz, obtaining temporal signals of
60 s with 6000 samples. Noise signals have also gone through
a de-signaling process to ensure that no weak seismic waves
remain hidden within the background noise. Waveforms are
then normalized as explained in Section III.C. An extract of the
full STEAD dataset with the same amount of seismic and noise
signals is randomly chosen to train the models.
The performance of trained models is evaluated using three
different datasets with seismic signals measured by fiber-optic
DAS systems [18,45-47]. Note that these datasets contain
measurements obtained by very different DAS technologies,
being based on single-pulse coherent-detection OTDR [18,46]
or chirped-pulse direct-detection OTDR [47]. Although the
optical configuration of the DAS system determines different
features of the measurements (e.g., SNR, sensitivity, spatial
resolution, and others), the acquired strain waveforms induced
by seismic waves are, in principle, independent of the optical
DAS layout (provided the systems are well designed and show
a linear response). The first dataset corresponds to a 1.9 local
magnitude earthquake measured along a 41.5 km-long telecom
cable deployed offshore Toulon, France, using a DAS system
with an optical heterodyne detector to retrieve the phase of the
Rayleigh backscattered light [18]. The DAS array was located
at 80 to 100 km from the earthquake epicenter. The data is
available online [45] and consist of distributed strain profiles
obtained with a sampling rate of 100 Hz, and a spatial sampling
interval of 6.4 m, for a total of 6848 acoustic independent
sensing points. Every measurement has 6000 temporal samples
obtained over 60 s of ground movement. Because of the site
characteristic attenuation and radiation pattern, the propagation
of the S waves was predominant with respect to P waves. The
second dataset corresponds to measurements of a 3.4 magnitude
earthquake measured by horizontal and vertical DAS arrays at
an approximate distance of 23 km from the epicenter in Nevada,
USA [46]. For testing, only the second horizontal measurement
is here used. The DAS array has a total of 8721 acoustic sensing
points recorded by a phase coherent DAS system at a sampling
rate of 1000 Hz. All records consist of 30000 samples, for a
complete signal of 30 s. The third dataset corresponds to
measurements of teleseismic waves from the 8.2 magnitude Fiji
deep earthquake on August 2018, with a DAS array located on
Zeebrugge, Belgium [47], using a chirped-pulse direct-
detection OTDR system. Body waves arrived from an
epicentral distance higher than 16,300 km, resulting in an
extremely low signal power focused between 0.001 Hz and
1 Hz. The DAS noise dataset is built from this measurement
after applying a bandpass filter in the range 1 to 45 Hz. All three
preprocessed DAS datasets have been merged into a complete
DAS test dataset.
Seismic records obtained by conventional broadband
seismographs and DAS systems are normally very similar to
each other, as illustrated in Fig. 4. The main difference is related
to the measurement unit, being acceleration, velocity or
displacement in the case of seismometers (Fig. 4(a)) and strain
in the case of DAS (Fig. 4(b)). However, the resemblance of the
measured waveforms constitutes one of the fundamental
characteristics behind the working principle of the method here
proposed: the deep learning models can be trained with seismic
recording obtained by traditional broadband seismometers and
then tested to classify seismic measurements obtained by DAS
sensors. Note however that the longitudinal response of a DAS
sensor to earthquake activity is not uniform along the fiber
length. This can be observed in Fig. 5, which shows the
earthquake measurement of the first DAS dataset (i.e., the one
containing measurements obtained in France [18,45]). As can
be seen in the white areas of the figure, there exist fiber
locations with very poor response to the ground movement.
This could result mainly due to the combination of three
possible reasons: i) the earthquake wave arrives to the local
optical fiber section with a relative angle that induces small
longitudinal strain, ii) the optical fiber can be locally weakly
coupled to the ground, inducing deficient local strain transfer
(from ground movement) to the optical fiber, or iii) intensity
fading affecting the Rayleigh scattering measurements induce
blind positions in the fiber, where no reliable acoustic signal is
Fig. 4. Comparison of normalized seismic waveforms measured by (a) a
distributed fiber sensor and, (b) a traditional broadband seismometer.
Fig. 5. Distributed seismic measurement based on DAS technology.
Replotted from data available in [45].

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
7
retrieved. Since this feature is not present in broadband
seismometer measurements (i.e., in the STEAD database), null
(zero-valued) traces have been added to the train dataset and
labelled as ‘noise’, allowing the classifiers to correctly identify
this type of traces in DAS seismic measurements and improving
the classification performance.
C. Signal Processing Strategy
A total of 200,000 signals from the STEAD dataset have been
selected, distributed equally between seismic and noise
waveforms. The complete extracted dataset is split in training,
validation and test datasets following an 80/10/10 split ratio.
Note that the number of signals used in the training process
corresponds only to a subset of the entire STEAD dataset; and
therefore, no data augmentation techniques are required for
training.
For a correct comparison of the classification performance
on the STEAD and DAS datasets, all signals must be equally
preprocessed. As shown in Fig. 6, all DAS signals are first
preprocessed to remove linear trends and mean values, and a
bandpass filter in the range 1-45 Hz is applied to then resample
the waveforms at 100 Hz. Every STEAD and DAS signal is
normalized to have an amplitude in the range [-1,1] to speed up
and stabilize the training procedure and to maintain the same
properties for the training and test waveforms. In the case of the
third DAS dataset (i.e., Belgium earthquake signals), 30 s of
random white noise (having the same variance as the
measurement noise) are padded in the first half of each
waveform to complete the timeseries of 60 s with 6000 samples.
The performance of the trained models is evaluated with the
fiber-optic DAS dataset as input, using the same evaluation
procedure as in the STEAD case. Note that, although these
evaluation blocks are the same for STEAD and DAS data, they
are depicted as two separate blocks in Fig. 6, as they take
different input datasets and return different sets of results. The
testing procedure gives the output probabilities, which are
compared with the ground truth labels generated with the
STA/LTA algorithm [46]. This algorithm has been widely used
by the seismological community to build collections of
earthquake signals. Its operating principle is based on the
definition of two moving windows that compute the average of
the samples they cover using a short window and a long
window. These windows capture the long- and short-time
variations of the signal. The average value of the short window
is divided by the average value of the long window and then
compared to a predefined threshold value; when the quotient
surpasses a given threshold, a detection is raised. The
evaluation metrics previously described are calculated for this
DAS dataset.
IV. TRAINING PROCEDURE AND RESULTS
The training is carried out by minimizing a loss function
defined by the binary cross entropy and updating the weights
using the Adam optimization algorithm [35]. The performance
of trained models is assessed using the loss function evaluated
with the validation dataset, as shown in Fig. 7(a). In this case
each epoch is composed of batches with 256 waveforms. The
figure shows how the trained models behave with waveforms
that the model did not see before during training. When the loss
curve reaches its lowest value and remains constant or
increases, overfitting takes place [34,35]. In that overfitting
regime, the models reduce their capabilities to generalize their
behavior in front of new datasets, and therefore this regime
must be avoided. For this, the model parameters that minimize
the validation loss function must be selected. To evaluate the
performance of the models, the F-score metric has been selected
Fig. 6. Flow diagram describing training, validation, and test procedures.
Fig. 7. (a) Validation loss and (b) F-score as a function of the training
epochs for the three model architectures.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
8
since this metric gives relevance to the false negatives and false
positives, which represents the outcomes with the highest
importance for an earthquake detection system. Fig. 7(b) shows
the F-score obtained as a function of the training epochs for the
FC-ANN, CNN and CNN+LSTM models. Curves illustrate that
a maximum F-score occurs with a similar number of batches
and weights that minimize the validation loss function, thus
confirming the optimal behavior of the models when selecting
such parameters. Note that for the sake of visualization, Fig. 7
shows the loss function and F-score as a function of the number
of epochs used in the training process, over a total of 100
epochs. Using an Intel Core i9-9980XE CPU, Nvidia Titan
RTX GPU, 128 GB RAM, and operating system Ubuntu 18.04,
the optimal models are however obtained using an early
stopping strategy and regularization techniques, leading to
training times of 1140.5 s, 186.3 s, and 1375.9 s for the FC-
ANN, CNN and CNN+LSTM models, respectively.
The classification performance of the optimal trained models
(FC-ANN, CNN, and CNN+LSTM models) is then evaluated
using the test dataset extracted from STEAD, through F-score,
accuracy, and PR metrics. This verification using STEAD data
is essential to assess the classification performance of trained
models before testing them with DAS measurements.
Fig. 8(a) shows the F-score values as a function of the
classification threshold, indicating that CNN and CNN+LTSM
models obtain better F-score results along the entire thresholds
range. The maximum F-score values achieved for each of the
models (over 0.987 in the three cases) indicate that they can
reliably discriminate the seismic waveforms from noise.
Fig. 8(b) shows the accuracy of the models as a function of the
classification threshold. Results point out that CNN obtains the
highest accuracy of 99.82% and an F-score of 0.998 for an
optimal threshold value of 0.964. CNN+LTSM obtains very
close results with an accuracy of 99.69% and an F-score of
0.997 for an optimal threshold value of 0.542. On the other
hand, FC-ANN obtains the lowest maximum accuracy of
98.71% and F-score of 0.987 for an optimal threshold value of
0.422. Fig. 8(c) shows the PR curve for the tree models,
confirming that the CNN model outperforms the other two, as
shown by the curve approaching closer to the (1,1) point. It is
worth mentioning that the FC-ANN and CNN+LSTM models
reach a classification Precision higher than Recall for the
obtained optimal thresholds, meanwhile a higher Recall
(compared to Precision) is obtained for the CNN model. This
behavior illustrates an advantage of the CNN model for
seismological applications, due to the very high and steady
Recall value near one here obtained, meaning that the model
detects almost all the seismic waveforms in the test set.
The lower performance of the FC-ANN model compared to
CNN and CNN+LSTM models is somehow expected, as this
architecture is not specially designed to automatically perform
feature extraction. Nevertheless, the results show that even
without a previous hand-designed extraction of seismic
waveform features, the FC-ANN model is capable of
classifying earthquake timeseries with high accuracy. Both
CNN and CNN+LSTM are more complex and leverage on the
automatic feature extraction of the initial convolutional layers,
justifying their better performance. Adding recurrent layers on
top of the convolutional block did not improve the classification
accuracy or F-score. It is worth to notice that the CNN+LSTM
model has a lower number of convolutional layers, and
therefore its feature extraction step is shallower. Adding more
convolutional layers to this model may slightly improve its
performance at the cost of higher computational cost.
Confusion matrices for the best classification thresholds
based on the F-score metric are illustrated in Table I, where the
best thresholds are 0.422, 0.964 and 0.542, for the FC-ANN,
CNN, and CNN+LSTM models, respectively. Results point out
that the number of true positives and true negatives are
consistently high, with low percentage of false positives and
false negatives. The table shows that the amount of correctly
classified timeseries reaches 98.71% for the FC-ANN, 99.82%
for the CNN and 99.69% for the CNN+LSTM model. The
Fig. 8. Test results for the trained models using STEAD data. (a) F-score
vs threshold, (b) Accuracy vs threshold, and (c) Precision-Recall curve.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
9
overall smaller number of false positives and false negatives for
the CNN verifies the better performance of this model, which
also shows a classification F-score of 0.9981, with a precision
and recall of 99.94% and 99.69% respectively.
These results indicate that the three proposed models are
successfully trained and can efficiently learn the fundamental
differences between seismic and noise signals, classifying with
high confidence earthquake waveforms obtained by classical
seismographs. This is also observed from the very high and low
output probabilities obtained by the models using the STEAD
waveforms.
V. CLASSIFICATION OF REAL SEISMIC DAS MEASUREMENTS
The classification performance of the optimal trained models
(same ones tested in Section IV) has been tested using real DAS
acoustic waveforms, obtained after preprocessing the timeseries
in the selected DAS datasets. Considering the high similarity
between seismometer and DAS waveforms, no domain
adaptation technique is needed to further improve the trained
models. The output probabilities of the models are compared to
the ground truth target labels generated with the STA/LTA
algorithm. The same metrics, i.e., F-score, accuracy, recall and
precision, are calculated for the optimal FC-ANN, CNN, and
CNN+LSTM models using DAS data. Fig. 9(a) and 9(b) show
the F-score and accuracy obtained as a function of the
classification threshold, for the three optimal models. Results
point out that the CNN model can achieve the best performance
among all three models when using DAS measurements, similar
to the behavior reported in Section IV when testing the models
with STEAD waveforms. This outcome can be verified by the
higher accuracy and F-score obtained by CNN, being also the
most robust model with an F-score above 0.9 over almost the
entire threshold range. The CNN+LSTM model also shows a
good performance, achieving an F-score also above 0.9 over all
the threshold range below 0.95. On the other hand, the F-score
for the FC-ANN model slightly reduces as the threshold value
increases, reaching the lowest F-score among the three models.
The accuracy in the three cases behaves similarly to the F-score
metric, reaching maximum values of 90.3%, 97.1%, and 94.1%
for the FC-ANN, CNN, and CNN+LSTM, respectively. Fig.
9(c) reports the PR curve using DAS data, showing a consistent
behavior with respect to the one previously obtained in
Fig. 8(c).
Table II presents the confusion matrices with the
performance of every optimal model when tested with DAS
seismic data. The table shows that the amount of correctly
classified waveforms reaches 90.17% for the FC-ANN, 96.94%
for the CNN and 93.86% for the CNN+LSTM model. Note that
these values are obtained with the optimal thresholds defined
during testing with STEAD data and are slightly lower than the
maximum values shown in Fig. 9(b). Compared to the
classification obtained with test STEAD waveforms, the FC-
ANN model has the largest decrease in the number of correctly
classified waveforms, while the CNN model shows the lowest
TABLE I
CONFUSION MATRICES FOR THE THREE MODELS EVALUATED WITH
CONVENTIONAL SEISMIC WAVEFORMS FROM THE STEAD DATASET
Model
Recognized
Class
Real Class
Seismic
Noise
FC-ANN
(th: 0.422)
Seismic
49.00%
(9799)
0.28%
(57)
Noise
1.00%
(201)
49.72%
(9943)
CNN
(th: 0.964)
Seismic
49.85%
(9969)
0.03%
(6)
Noise
0.15%
(31)
49.97%
(9994)
CNN + LSTM
(th: 0.542)
Seismic
49.76%
(9952)
0.07%
(15)
Noise
0.24%
(48)
49.93%
(9985)
The values indicated between brackets represent the number of waveforms
recognized in the respective class.
Fig. 9. Test results for the trained models using DAS data. (a) F-score vs
threshold, (b) Accuracy vs threshold, and (c) Precision-Recall curve.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
10
reduction. This implies that the CNN model has the highest
potential (among the analyzed models) to better classify DAS
seismic waveforms and eventually develop future novel
improvements. In addition, note that the classification
performance of the CNN and CNN+LSTM models on the
seismic traces of the DAS dataset is almost the same, which is
reflected in the number of true positives and false negatives
obtained in the two cases. However, the slightly worst
performance of the CNN+LSTM model relies on the higher
number of noise traces incorrectly classified as seismic, i.e., the
false positives.
These results verify that the learning parameters optimized
during training using conventional seismic waveforms (e.g.,
STEAD timeseries) can be reliably applied to the classification
of distributed seismic waves obtained by a DAS system.
However, comparing the analyzed metrics obtained for the
three models using DAS measurements with respect to those
obtained with traditional seismic waveforms, a small reduction
in the performance of the models is observed. This could be
justified by the intrinsic differences between seismic
waveforms obtained by DAS and traditional seismographs, and
eventual errors in the labelling of the measured DAS signals.
To exemplify the performance of the models using DAS
waveforms, Fig. 10 shows the output probability given by all
three models in the case of the first DAS dataset under analysis.
In particular, Fig. 10(a) shows the DAS strain recording as a
function of time and fiber position. Fiber sections with no
acoustic signal (white colored sections) are clearly observed,
which could have resulted from poor local strain transfer from
ground movement to the optical fiber. The behavior is correctly
classified by all three models, as shows the output probability
of the trained the FC-ANN (Fig. 10(b)), CNN (Fig. 10(c)), and
CNN+LSTM (Fig. 10(d)) models. Most of the differences in the
output probabilities given by the models are found over optical
fiber zones with low measurement SNR (i.e., from 17.92 km to
18.27 km, 19.8 km to 19.96 km, and 20.71 km to 21.12 km). It
is clearly observed that within those fiber sections, the
classification provided by the CNN and CNN+LSTM models
outperforms the FC-ANN model. On the other hand, a near-one
probability output is obtained for the high-level signal zone
between 18.27 km and 19.8 km, whereas the poorly detected
zone between 19.96 km and 20.71 km is classified with high
confidence as noise by all models. This behavior of the models
can also be observed in Fig. 11, which shows two random
seismic waveforms correctly classified as ‘seismic’ (Figs. 11(a)
and 11(b)), and two random noisy waveforms classified as
‘noise’ (Figs. 11(c) and 11(d)) by all three models.
TABLE II
CONFUSION MATRICES FOR THE THREE MODELS TESTED WITH SEISMIC
WAVEFORMS MEASURED BY DAS SYSTEMS
Model
Recognized
Class
Real Class
Seismic
Noise
FC-ANN
(th: 0.422)
Seismic
44.10%
(12239)
3.93%
(1092)
Noise
5.90%
(1637)
46.07%
(12784)
CNN
(th: 0.964)
Seismic
47.85%
(13280)
0.91%
(253)
Noise
2.15%
(596)
49.09%
(13623)
CNN + LSTM
(th: 0.542)
Seismic
47.58%
(13205)
3.72%
(1031)
Noise
2.42%
(671)
46.28%
(12845)
The values indicated between brackets represent the number of waveforms
recognized in the respective class.
Fig. 10. Output probability of the trained models in the case of DAS seismic
measurements with fiber sections showing poor or no strain sensitivity (due
to coupling or fading issues). (a) Zoom-in of DAS seismic measurements,
and model outputs for (b) FC-ANN, (c) CNN and (d) CNN+LSTM models.
Fig. 11. DAS measurements containing traces classified as (a)-(b) seismic,
and as (c)-(d) noise, by the three deep learning models.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
11
VI. CONCLUSION
In this paper, the classification of seismic waves based on
deep learning models applied to distributed acoustic sensing
measurements is demonstrated. Results validate the proposed
classification and training strategies based on the use of seismic
timeseries obtained by conventional seismographs (e.g.,
STEAD database). This approach has permitted us to train
models based on FC-ANN, CNN, and CNN+LSTM with
existing seismic databases, and then use them to classify DAS
measurements. Note that, this strategy does not need large DAS
datasets to achieve a successful training. Evaluating metrics like
F-score and accuracy show that convolutional models are more
suitable to learn the features of seismic waveforms, and
particularly CNN has proven to be the best model, among the
studied ones, to classify DAS seismic waveforms.
Note that the analyzed evaluating metrics using DAS
measurements are slightly lower than the ones obtained when
testing the models with traditional seismic waves. This behavior
might be explained by the intrinsic differences between seismic
waveforms measured by DAS systems and conventional
seismographs, caused by combination of the three following
reasons: i) the uneven fiber coupling to the ground along the
cable and different soil properties, which results in different
strain transfer efficiencies along the sensing fiber, ii) the angle
of arrival of the earthquake wave with respect to the optical
cable, which might induce different local longitudinal strain in
the fiber, and iii) the existence of amplitude fading points,
which lead to unreliable phase retrieval and short blind fiber
sections. These situations lead to seismic DAS measurements
with very different SNR and temporal waveforms along a given
optical fiber cable, affecting the classification and labeling of
the timeseries.
The method here proposed could represent a first step to
develop an early warning earthquake system based on DAS
technology. The performance of the classification models could
be further improved by including DAS recordings on the
seismograph training dataset. Such an approach would include
specific features of DAS measurements into the learning
process, while also leveraging on the great availability of
conventional seismic recordings. In addition, models pre-
trained with conventional seismic waveforms could also make
use of a fine-tuning training stage, on which additional DAS
training dataset could be used to further optimize the classifier
parameters. These approaches can be especially applied to the
CNN architecture here proposed, since this kind of model
shows greater capability to adapt to seismic waveforms not seen
during training. Further performance enhancements could also
include the improvement of the CNN architecture and a training
stage using heterogeneous databases to discriminate seismic
waveforms from other sources of mechanical vibrations. This
could be of particular importance when installed telecom
optical fibers (e.g., in urban areas or near highways) are
exploited for distributed seismic measurements.
Finally, it must be also pointed out that all these alternatives
of improvement could additionally benefit from the proper
labelling of DAS measurements, which may be required point
by point along the sensing optical fiber. Labelling seismic DAS
data is actually essential to achieve more reliable training and
testing when using distributed seismic measurements. This
would represent a challenging task, in which the involvement
of specialist in seismology is crucial to obtain accurate target
labels.
REFERENCES
[1] A.H. Hartog, An Introduction to Distributed Optical Fibre Sensors; CRC
Press: Boca Raton, FL, USA, 2017.
[2] Z. He and Q. Liu, “Optical Fiber Distributed Acoustic Sensors: A Review,”
J. Lightw. Technol., vol. 39, no. 12, pp. 3671-3686, 2021
[3] P. Stajanca, S. Chruscicki, T. Homann, S. Seifert, D. Schmidt, and A.
Habib, “Detection of Leak-Induced Pipeline Vibrations Using Fiber—
Optic Distributed Acoustic Sensing,” Sensors, vol 18, no 9, pp. 2841, 2018.
[4] P. G. Hubbard, J. Xu, S. Zhang, et al. “Dynamic structural health
monitoring of a model wind turbine tower using distributed acoustic
sensing (DAS).” J Civil Struct Health Monit., vol. 11, pp. 833–849, 2021.
[5] L. Schenato, L. Palmieri, M. Camporese, et al. “Distributed optical fibre
sensing for early detection of shallow landslides triggering.” Sci. Rep., vol.
7, pp, 14686, 2017.
[6] M. R. Fernández-Ruiz, M. A. Soto, E. F. Williams, S. Martin-Lopez, Z.
Zhan, M. Gonzalez-Herraez, and H. F. Martins, “Distributed acoustic
sensing for seismic activity monitoring,” APL Photonics, vol. 5, pp.
030901, 2020.
[7] H. Liu, J. Ma, T. Xu, W. Yan, L. Ma, X. Zhang, “Vehicle Detection and
Classification Using Distributed Fiber Optic Acoustic Sensing,” IEEE T.
Veh. Technol., vol. 69, no. 2, pp. 1363-1374, Feb. 2020
[8] Z. Li, J, Zhang, M. Wang, Y. Zhong, and F. Peng, “Fiber distributed
acoustic sensing using convolutional long short-term memory network: a
field test on high-speed railway intrusion detection,” Opt. Express, vol. 28,
no. 3m pp. 2925-2938, 2020.
[9] J. Tejedor, J. Macias-Guarasa, H. F. Martins, J. Pastor-Graells, P.
Corredera, and S. Martin-Lopez, “Machine Learning Methods for Pipeline
Surveillance Systems Based on Distributed Acoustic Sensing: A Review”
Appl. Sci., vol. 7, no. 8, pp. 841, 2017.
[10] A. Mateeva, J. Lopez, H. Potters, J. Mestayer, B. Cox, D. Kiyashchenko,
P. Wills, S. Grandi, K. Hornman, B. Kuvshinov, W. Berlang, Z. Yang and
R. Detomo, “Distributed acoustic sensing for reservoir monitoring with
vertical seismic profiling,” Geophys. Prospect., vol. 62, no 4, pp. 679-692,
2014.
[11] A. Hartog, B. Frignet, D. Mackie and M. Clark, “Vertical seismic optical
profiling on wireline logging cable,” Geophys. Prospect., vol. 62, no. 4, pp.
693-701, 2014.
[12] N. J. Lindsey, E. R. Martin, D. S. Dreger, B. Freifeld, S. Cole, S. R. James,
B. L. Biondi and J. B. Ajo-Franklin. “Fiber‐optic network observations of
earthquake wavefields,” Geophys. Res. Lett., vol. 44, no 23, pp. 11,792-
11,799, 2017.
[13] H. F. Wang, X. Zeng, D. E. Miller, D. Fratta, K. L. Feigl, C. H. Thurber
and R. J. Mellors, “Ground motion response to an ML 4.3 earthquake using
co-located distributed acoustic sensing and seismometer arrays,” Geophys.
J. Int., vol. 213, no 3, pp. 2020-2036, 2018.
[14] Z. Li and Z. Zhan, “Pushing the limit of earthquake detection with
distributed acoustic sensing and template matching: A case study at the
Brady geothermal field,” Geophys. J. Int., vol. 215, no 3, pp. 1583-1593,
2018.
[15] C. Yu, Z. Zhan, N. J. Lindsey, J. B. Ajo-Franklin, M. Robertson, “The
Potential of DAS in Teleseismic Studies: Insights From the Goldstone
Experiment,” Geophys. Res. Lett., vol. 46, no.3, pp. 1320-1328, 2019.
[16] P. Jousset, T. Reinsch, T. Ryberg, H. Blanck, A. Clarke, R. Aghayev, G. P.
Hersir, J. Henninges, M. Weber and C. M. Krawczyck, “Dynamic strain
determination using fibre-optic cables allows imaging of seismological and
structural features,” Nat. Commun., vol. 9, p. 2509, 2018.
[17] G. Marra, C. Clivati, R. Luckett, A Tampellini, J. Kronjäger, L. Wright, A.
Mura, F. Levi, S. Robinson, A. Xuereb, B. Baptie and D. Calonico,
“Ultrastable laser interferometry for earthquake detection with terrestrial
and submarine cables,” Science, vol. 361, no. 6401, pp. 486-490, 2018.
[18] A. Sladen, D. Rivet, J. P. Ampuero, L. De Barros, Y. Hello, G. Calbris and
P. Lamare, “Distributed sensing of earthquakes and ocean-solid Earth
interactions on seafloor telecom cables,” Nat. Commun., vol. 10, p. 5777,
2019.
[19] E. F. Williams, M. R. Fernández-Ruiz, R. Magalhaes, R. Vantillho, Z.
Zhan, M. González-Herráez and H. F. Martins, “Distributed sensing of

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JLT.2021.3138724, Journal of
Lightwave Technology
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
12
microseisms and teleseisms with submarine dark fibers,” Nat. Commun.,
vol. 10, p. 5778, 2019.
[20] J. Wang, T.-L. Teng, “Artificial neural network-based seismic detector,”
Bulletin of the Seismological Society of America, vol. 85, no. 1, pp. 308–
319, 1995.
[21] H. Dai, C. MacBeth, “Automatic picking of seismic arrivals in local
earthquake data using an artificial neural network,” Geophys. J. Int., vol.
120, no. 3, pp. 758–774, 1995.
[22] Y. Wu, Y. Lin, Z. Zhou, D. C. Bolton, J. Liu and P. Johnson, “DeepDetect:
A Cascaded Region-Based Densely Connected Network for Seismic Event
Detection,” IEEE Trans. Geosci. Remote Sens, vol 57, no 1, pp 62-75, 2019
[23] A. Lomax, A. Michelini and D. Jozinović, “An investigation of rapid
earthquake characterization using single‐station waveforms and a
convolutional neural network,” Seismol. Res. Lett., vol. 90, no 2A, pp. 517-
529, 2019.
[24] M. Meier, Z. E. Ross, A, Ramachandran, A. Balakrishna, S. Nair, P.
Kundzicz, Z. Li, J. Andrews, E. Hauksson and Y. Yue, “Reliable real‐time
seismic signal/noise discrimination with machine learning,” J. Geophys.
Res. Solid Earth, vol. 124, no 1, pp. 788-800, 2019.
[25] S. M. Mousavi, W. S. Ellsworth, W. Zhu, L. Y. Chuang and G. C. Beroza,
“Earthquake transformer—an attentive deep-learning model for
simultaneous earthquake detection and phase picking,” Nat. Commun., vol.
11, p. 3952, 2020
[26] T. Perol, M. Gharbi and D. Marine, “Convolutional neural network for
earthquake detection and location,” Sci. Adv., vol. 4, no 2, p. e1700578,
2018
[27] X. Zhang, J. Zhang, C. Yuang, S. Liu, Z. Chen and W. Li, “Locating
induced earthquakes with a network of seismic stations in Oklahoma via a
deep learning method,” Sci. Rep., vol. 10, p. 1941, 2020.
[28] S. M. Mousavi, G. C. Beroza, “A machine‐learning approach for
earthquake magnitude estimation,” Geophys. Res. Lett., vol. 47, no 1, p.
e2019GL085976, 2020.
[29] N. C. Ristea and A. Radoi, “Complex Neural Networks for Estimating
Epicentral Distance, Depth, and Magnitude of Seismic Waves,” IEEE
Geosci. Remote. Sens. Lett., Early Access, 2021.
[30] D. Jozinović, A. Lomax, I. Štajduhar and A. Michelini, “Rapid prediction
of earthquake ground shaking intensity using raw waveform data and a
convolutional neural network,” Geophys. J. Int., vol. 222, no 2, pp. 1379-
1389, 2020.
[31] L. Shiloh, A. Eyal and R. Giryes, “Efficient processing of distributed
acoustic sensing data using a deep learning approach,” J. Lightw. Technol.,
vol. 37, no 18, pp. 4755-4762, 2019.
[32] D. Arbel and A. Eyal, “Dynamic optical frequency domain
reflectometry,” Opt. Express, vol. 22, no. 8, pp. 8823-8830, 2014.
[33] Z. Zhang, X. Fan and Z. He, “Long-Range Distributed Static Strain
Sensing With <100 Nano-Strain Resolution Realized Using OFDR,” J.
Lightw. Technol., vol. 37, no. 18, pp. 4590-4596, 2019.
[34] I. Goodfellow, Y. Bengio and A. Courville, “Deep learning,” MIT press,
2016.
[35] F. Chollet. “Deep learning with Python,” Manning Publications, 1st
edition, 2017
[36] D. E. Rumelhart, F. E. Hinton and R. J. Williams, “Learning
representations by back-propagating errors,” Nature, vol. 323, no 6088, pp.
533-536, 1986.
[37] A. Fernández, S. García, M. Galar, R. C. Prati and B. Krawczyk, “Learning
from imbalanced data sets,” Berlin: Springer, 2018.
[38] V. Nair and G. F. Hinton, “Rectified linear units improve restricted
boltzmann machines,” Proc. Int. Conf. Machine Learning, 2010.
[39] Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, “Gradient-based learning
applied to document recognition,” Proc. of IEEE, vol. 86, no. 11, pp.
2278-2324, 1998.
[40] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” Int. Conf. on
Machine Learning, PMLR, pp. 448-4,56, 2015.
[41] S. M. Mousavi, W. Zhu, Y. Zheng and G. C. Beroza, “CRED: A deep
residual network of convolutional and recurrent units for earthquake
signal detection,” Sci. Rep., vol. 9, no 1, pp. 1-14, 2019.
[42] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R.
Salakhutdinov, “Dropout: a simple way to prevent neural networks from
overfitting,” J. Mach. Learn. Res., vol. 15, no 1, pp. 1929-1958, 2014.
[43] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no 8, pp. 1735-1780, 1997.
[44] S. M. Mousavi, Y. Sheng, W. Zhu and G. C. Beroza, “STanford
EArthquake Dataset (STEAD): A Global Data Set of Seismic Signals for
AI,” IEEE Access, vol. 7, pp. 179464-179476, 2019,
[45] A. Sladen, A, “MEUST-NUMerEnv/KM3NeT DAS experiment Feb.
2018,” [Online] Available. https://doi.org/10.17605/OSF.IO/X6AWB,
December 2019
[46] University of Wisconsin, K. Feigl, “Brady's Geothermal Field DAS
Earthquake Data [data set]” 2016, [Online] Available
https://dx.doi.org/10.15121/1334285.
[47] Williams, E. F., Fernandez-Ruiz, M. R., Magalhaes, R., Vanthillo, R.,
Zhan, Z., Gonzalez-Herraez, M., & Martins, H. F. (2019). Belgium
Distributed Acoustic Sensing Array Raw Data (Version 1.0) [Data set].
CaltechDATA. [Online] Available https://doi.org/10.22002/D1.1296
[48] R. V. Allen, “Automatic earthquake recognition and timing from single
traces,” Bulletin of the Seismological Society of America, vol. 68, no. 5,
pp. 1521-1532, Oct. 1978.
Pablo D. Hernández was born in Viña del Mar, Chile, in 1995.
He is currently a Master student of Electronic Engineering at
Universidad Técnica Federico Santa María, Valparaíso, Chile.
His research interests include optical sensing technology,
computer vision, data science and applications of deep learning
and neural networks.
Jaime A. Ramírez received the M.Sc. degree in Electronic
Engineering from Universidad Técnica Federico Santa María,
Valparaíso, Chile, in 2007.
From 2006 to 2015, he worked as a specialist engineer and
project manager in the fields of sensors, data science and
computer vision for the health and mining sectors. Between
2015 and 2021, he was the leader of the applied R&D and
technology transfer groups at the Advanced Center of Electrical
and Electronic Engineering of Universidad Técnica Federico
Santa María. He is currently the founder and head data scientist
of the company Novelcode, dedicated to computer vision,
artificial intelligence and instrumentation for industrial and
mining sectors.
He is author or co-author of several scientific publications
and 7 patents in the field of image processing applied to
industrial processes.
Marcelo A. Soto (M’20) received the M.Sc. degree in
Electronic Engineering from Universidad Técnica Federico
Santa María, Valparaíso, Chile, in 2005, and the Ph.D. degree
in Telecommunications from the Scuola Superiore Sant’Anna,
Pisa, Italy, in 2011.
During 2010–2011, he was a Research Fellow at Scuola
Sant’Anna, where he worked on distributed optical fiber
sensors based on Raman and Brillouin scattering. Later, he was
a Postdoctoral Researcher at the EPFL Swiss Federal Institute
of Technology of Lausanne, Switzerland, where he worked on
high-performance Brillouin and Rayleigh distributed fiber
sensing, nonlinear fiber optics, optical signal processing, and
optical Nyquist pulse generation. Since March 2018, he is a
Tenure-Track Assistant Professor at Universidad Técnica
Federico Santa María, Valparaíso, Chile. He also has an invited
position as one of the “100 distinguished invited professors” at
Guangzhou University, in China. He is author or coauthor of
over 180 scientific publications in international refereed
journals and conferences, 3 book chapters and 8 patents in the
fields of optical communications and optical fiber sensing.
Dr. Soto is senior member of the Optical Society of America
(OSA), and he is in the Board of Reviewers of major
international journals in photonics.