See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/326308014

A Fast Earthquake Early Warning Algorithm Based on the First 3 s of the P‐Wave
Coda

Article  in  Bulletin of the Seismological Society of America · July 2018

DOI: 10.1785/0120180079

CITATIONS READS
11 236

3 authors, including:

Gerardo Suárez
Universidad Nacional Autónoma de México
126 PUBLICATIONS   5,244 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Seismicity catalogs evaluation and detection of artificial anomalies. View project

CATAC_INETER View project

All content following this page was uploaded by Gerardo Suárez on 07 October 2019.

The user has requested enhancement of the downloaded file.

 Bulletin of the Seismological Society of America, Vol. 108, No. 4, pp. 2068–2079, August 2018, doi: 10.1785/0120180079

A Fast Earthquake Early Warning Algorithm

Based on the First 3 s of the P-Wave Coda
by Armando Cuéllar, Gerardo Suárez, and J. M. Espinosa-Aranda

Abstract A new seismic early warning algorithm is presented that estimates a mag-
nitude threshold using the first 3 s of the P-wave coda on the vertical component. The
algorithm considerably reduces the processing time compared to previous algorithms
used by the Mexican seismic early warning system (Seismic Alert System of Mexico
[SASMEX]). It was designed to alert earthquakes within the subducted Cocos
plate. The tP
3 algorithm was based on a training dataset of 76 accelerograms of
25 Mexican in-slab earthquakes, with focal depths > 40 km. The algorithm uses two
parameters based on the unfiltered vertical component of the P waves: the sum of the
cumulative quadratic acceleration, avT and θP a parameter that represents the slope of
the cumulative acceleration. The model is based on a learning machine that linearizes
piecewise the empirical relation between these two parameters and magnitude M w .
The resulting algorithm was tested on nine earthquakes that took place from 2014
to 2017, recorded in 37 strong-motion records. In addition, the algorithm was evalu-
ated in the context of the Mexican earthquake early warning, applying it to 24 in-slab
earthquakes occurring from 1995 to 2017 (5:0 < M w < 7:1). The magnitude of
19 earthquakes was properly estimated; for four of them, it was overestimated and
in one case the magnitude was underestimated. Three earthquakes M w > 6:5 that af-
fected Mexico City were included in the dataset: the Mw 6.5 event on 11 December
2011 and the destructive in-slab Tehuacán and Morelos earthquakes on 15 June 1999
(Mw 7.0) and 19 September 2017 (M w 7.1). The retrospective application of the tP
3
algorithm shows that these three earthquakes are correctly identified as M w > 6
and would activate a seismic alert. The tP
3 algorithm would have given an advance
warning of 34, 35, and 16 s respectively, before the arrival of strong motion in
Mexico City.

Electronic Supplement: Tables showing the hypocentral data of the catalog of

earthquakes and the results of the performance evaluation of the tP
3 algorithm on
earthquakes within the subducted Cocos plate.

Introduction
Early warning systems are in a constant race to discrimi- The (tS − tP ) procedure requires the arrival of the S waves,
nate as rapidly as possible the magnitude of earthquakes that whereas the 2tS − tP  needs twice this time to determine
may prove damaging to cities with a high-seismic vulnerabil- whether an alert is issued (Espinosa-Aranda et al., 1992;
ity. The challenge is to provide a warning time, as long as pos- Cuéllar et al., 2017a,b). These algorithms allow warning
sible, that enables the authorities and the population to carry times in Mexico City for earthquakes originating along the
out preventive actions before the arrival of the damaging seis- subduction zone of the Cocos plate of between 60 and 120 s.
mic waves. Several workers suggested using the first seconds Therefore, waiting for the arrival of the S wave is not an issue
of P-wave coda to estimate early warning parameters such as considering the large lead time available. In the case of cities
τc method, which uses the ratio of acceleration and velocity to located near the subduction zone, which now have a seismic
establish a warning criterion (e.g., Wu and Kanamori, 2005). alerting system, the (tS − tP ) algorithm reduces in half the
The Seismic Alert System of Mexico (SASMEX) makes time needed to process the data and make a decision on the
use of algorithms that measure acceleration of the ground in emission of an alert, increasing the warning time in those
the periods 2tS − tP  and (tS − tP ) (Cuéllar et al., 2017a,b). cities.

2068
A Fast Earthquake Early Warning Algorithm Based on the First 3 s of the P-Wave Coda 2069

The situation is different, however, for earthquakes accelerations > 20 cm=s2 . Although the earthquake caused
beneath central Mexico that take place within the subducted no substantial damage, there were power outages in ∼30
Cocos plate. Hereafter, we refer to these events as in-slab districts in Mexico City and the temporary suspension of
earthquakes. Because of the contorted nature of the sub- telephone and mobile communications. At that time, SAS-
ducting Cocos plate (Suárez et al., 1990; Pardo and Suárez, MEX had no coverage in this inland region of Mexico. The
1993, 1995), in-slab earthquakes range in depth from 50 km in monitoring stations were concentrated on the Pacific coast,
central Mexico to 180 km, beneath the Isthmus of Tehuan- covering only the subduction zone. Therefore, SASMEX
tepec. Although in-slab events are generally smaller and less was unable to issue an early warning in Mexico City.
frequent than subduction earthquakes, they occur beneath the This experience showed that the coverage of SASMEX
more highly populated areas of Mexico, frequently causing monitoring stations needed to be expanded to cover the in-
damage and loss of life. In these cases, algorithms that require land regions of the country affected by in-slab earthquakes.
the arrival of the S waves to issue an alert are of little use be- Furthermore, it became clear that a faster algorithm was nec-
cause the arrival of the S waves coincides with the strong essary, that need not wait for the arrival of the S waves, as the
ground motion. other algorithms traditionally used by SASMEX (Cuéllar
The need of seismic early warning systems in cities et al., 2017a,b). This becomes clear when the (tS − tP ) and
located above in-slab earthquakes stimulated the develop- the 2tS − tP  algorithms are used in a post facto analysis
ment of a fast algorithm based only on the P-wave coda. based on the two closest strong-motion stations that recorded
We present here an algorithm that uses the unfiltered vertical the 2011 Zumpango earthquake: TNLP-Tonalapa and
ground acceleration records of the first 3 s of the P-wave COMD-La Comunidad (Fig. 1). These strong-motion stations,
coda. The algorithm is designed for earthquakes with hypo- operated by the Instituto de Ingeniería of the National Autono-
central depths > 40 km, where strong-motion instruments mous University of Mexico (II-UNAM), are located at epicen-
are located at distances of a few tens of kilometers from the tral distances of 52 and 65 km. The application of the (tS − tP )
epicenter. The algorithm called tP
3 is correlated to M w and and 2tS − tP  algorithms on these two seismic records would
makes use of two parameters to determine a magnitude thresh- have resulted in warning times of 27 and 16 s, prior to the
old to issue a seismic early warning: avT , the cumulative arrival of the S waves observed at station SCT2-Secretaría
quadratic acceleration in the tP
3 s period, and a parameter de Comunicaciones in Mexico City (Fig. 1). However, the
called θP that parameterizes the shape of the cumulative quad- population would have probably felt the ground motion pro-
ratic acceleration in the first 3 s of P-wave coda. duced by the coda of the P waves prior to the alert signal.
A magnitude model mtP
3 , is constructed using an ac- The experience of the Zumpango earthquake demon-
tive and supervised learning machine method. The algorithm strated that for seismic foci within the subducted Cocos plate
was developed using a training dataset of 76 strong-motion that take place close to major cities in central Mexico, it is
records recorded from 25 earthquakes from 1997 to 2013, possible to issue a seismic early warning that provides a useful
with a magnitude range of 5:0 < M w < 7:1. The resulting warning time. Moreover, an algorithm that substantially re-
algorithm was tested on nine recent earthquakes that took duces the time needed to discriminate the magnitude threshold,
place between 2014 and 2017, recorded in 37 strong-motion would offer a much longer warning time to take civil protec-
records. These earthquakes are all in-slab events with hypo- tion measures prior to the arrival of strong ground shaking.
central depths > 40 km. In addition to the test earthquakes,
three of the largest in-slab events are used to analyze the
retrospective performance of the proposed algorithm. These Estimation of Parameters avT and θP
are the M w 7.0 Tehuacán earthquake on 15 June 1999, the of the tP
3 Algorithm
Mw 6.5 Zumpango del Río earthquake on 11 December
Cumulative Quadratic Acceleration avT
2011, and the destructive earthquake that occurred near
Mexico City on 19 September 2017 (M w 7.1). The results The records of free-field strong-motion stations located
on these test earthquakes indicate that the tP
3 algorithm above the hypocenter of in-slab earthquakes exhibit high ver-
is a robust and reliable early warning algorithm that offers a tical accelerations of the P waves compared to the horizontal
substantial time advantage for preparedness purposes, com- components. This facilitates the identification of the P phase
pared with algorithms that wait for the arrival of the S phase. on the vertical channel of triaxial accelerometers. Several
works proposed earthquake early warning algorithms that
Background on the Development use the first seconds after the detection of P waves to esti-
of the tP
3 Algorithm mate seismic magnitude (e.g., Wu and Kanamori, 2005; Wu
et al., 2007). Others propose to calculate in 2 or 3 s the epi-
On 11 December 2011, an earthquake M w 6.5 occurred central distance to the sensor and to use the growth shape of
near the town of Zumpango, Guerrero, at a depth of 65 km. an empirical function to estimate magnitude (Odaka, et al.,
This earthquake was felt strongly in Mexico City, located at 2003). Here, we propose the parameter avT, which mea-
an epicentral distance of 200 km (Fig. 1). The Mexico City sures the cumulative growth of seismic energy in the vertical
strong-motion network (RACM) recorded peak ground component during the interval tP
3 s. The cumulative
2070 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

Figure 1. Post facto analysis of the 11 December 2011 Mw 6.5 Zumpango earthquake with a focal depth of 65 km. The (tS − tP ) and
the 2tS − tP  algorithm would have issued an alert in Mexico City ∼27 and 16 s before the arrival of the S waves.

quadratic acceleration is calculated beginning at the initial The Shape of the Seismic Acceleration Curve:
detection of the P wave, as in the following equation: Parameter θP
tX
P
T The growth of the cumulative acceleration in the period
EQ-TARGET;temp:intralink-;df1;55;299 avT  xUD i2 ; 1 tP
3 s shows three very different growth shapes, depend-
itP ing on the magnitude of the earthquake: linear, sigmoidal,
in which xUD is the acceleration on the vertical channel, tP is the and exponential (Fig. 2). Thus, the shape of the growth rate
arrival time of the P phase, and T is the 3 s processing window. of the quadratic acceleration appears to be a good indicator
As an example, parameter avT is calculated on stations to estimate magnitude.
GR04-Cacalutla, GR18-Joluta, and GR13-Petatlán, of the Based on equations (2) and (3), the slope of the growth of
SASMEX network, for three earthquakes: 6 October 2013 the P-wave quadratic acceleration is divided into two segments:
(Mw 5.1; focal depth  21 km); the 21 April 2013 (Mw 6.1; avT1 − avT0
depth  27 km); and the 18 April 2014 (M w 7.2; mvT1  2
avT1
EQ-TARGET;temp:intralink-;df2;313;197

depth  21 km) (Fig. 2). The recording strong-motion sta-

tions lie at similar epicentral distances to these three earth-
quakes. Although these earthquakes are not in-slab events,
the exercise highlights the difference in the cumulative quad- avT2 − avT1
ratic acceleration avT of the P wave during the first seconds
mvT2  ; 3
avT2 
EQ-TARGET;temp:intralink-;df3;313;152

for earthquakes with different magnitudes. The cumulative

growth of energy, measured by parameter avT, shows values in which avT0 is the cumulative sum of the acceleration from
of 102 , 104 , and 105 cm2 =s4 for earthquakes with magni- tP to tP
0:5 s; avT1 is the cumulative sum of the quadratic
tudes 5.1, 6.1, and to 7.2, respectively (Fig. 2). acceleration from tP
0:5 s to tP
1:75 s; and avT2 is the
A Fast Earthquake Early Warning Algorithm Based on the First 3 s of the P-Wave Coda 2071

 
mvT2
θP  arctan : 4
mvT1
EQ-TARGET;temp:intralink-;df4;385;733

The parameter θP is an angular meas-

urement of the rate of change of the quad-
ratic acceleration on the vertical component
avT , in the first 3 s after the P arrival. When
θP ≈ π=4, the rate mvT2 does not have a
significant change with respect to mvT1 .
In this case, the growth rate is constant
and the P-wave growth is linear. When
θP < π=4, mvT1 is larger than mvT2 and
the P-wave growth is sigmoidal. Finally,
when θP > π=4, mvT2 is larger than
mvT1 and the P-wave growth is expo-
nential.

Selection Criteria of Strong-Motion

Records
Data from 349 strong seismic records
from 57 earthquakes were selected to assess
the usefulness of parameters log10 avT
and θP to estimate M w in an earthquake
early warning system. The networks of the
II-UNAM (Pérez-Yáñez et al., 2014) and
SASMEX recorded these accelerograms.
The sampling rate of the acceleration re-
cords was normalized to 100 Hz.
In some cases, the accelerograms of
small earthquakes or records at distant sta-
tions show very low accelerations that are
close to the self-noise of the instruments.
The resolution of the majority of the early
strong-motion SASMEX instruments was
10 or 12 bits. In those cases, the noise level
of the seismic acceleration recordings is
at least 1:0 or 0:5 cm=s2, respectively.
To be above the noise level, only accele-
ration records with values > 4 cm=s2 in
the tP
3 s time window were utilized.
This value of acceleration corresponds to
Figure 2. Cumulative quadratic acceleration observed in the first 3 s of the P-wave maxxUD tP
3 > log10 42 .
coda for three earthquakes at similar epicentral distances, but with different magnitudes M w . The strong-motion records were
grouped in two: one where xUD tP
3 >
cumulative sum of the quadratic acceleration from tP
1:75 s log10 42  and the other with xUD tP
3 ≤ log10 42 . The
to tP
3 s. cumulative number of these two groups of accelerograms was
In the case of intermediate-depth earthquakes, P waves plotted relative to the epicentral distance (Fig. 3). In the case
arrive almost vertically at seismic stations near the epicenter, of accelerograms recorded at long distances, the cumulative
producing large vertical accelerations. Thus, we propose a distribution for xUD tP
3 > log10 42  decreases. In con-
0.5 s waiting period to stabilize subsequent measurements. trast, xUD tP
3 ≤ log10 42  increases as a function of time
The normalized rates mvT1 and mvT2 define the growth (Fig. 3). The intersection of these curves corresponds to epi-
rate with duration of 1.25 s, in the interval between 0.5 and central distances of ∼110 km. We propose to use this empirical
3 s, after the arrival of the P wave. Thus, the parameter express- criterion in the selection data to develop the algorithm.
ing the shape of the cumulative energy growth may be Out of this dataset, 113 acceleration records of 34 earth-
expressed as quakes were selected based on the following criteria (Fig. 4
2072 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

and Ⓔ Table S1, available in the elec-

tronic supplement to this article):
1. earthquakes Mw > 4:6, occurring from
1985 to 2017, based on the centroid
moment tensor catalog (Ekström et al.,
2012);
2. focal depth > 40 km located in the in-
traplate region, based on the catalog of
the Seismological Service of Mexico;
3. accelerograms with minimum mean
accelerations of 4 cm=s2 , considered
as the acceptable floor of self-noise of
the strong-motion instruments (Cuéllar
et al., 2017b); and
4. strong-motion stations located at epi-
central distances < 110 km.

Correlation between Parameters

avT and θP versus M w
Figure 3. Cumulative distribution of 349 acceleration records of in-slab earthquakes
The correlation between log10 avT from 1985 to 2017 5:0 ≥ M w ≤ 7:1 grouped into the two categories: maxxUD tP
3 ≤
versus Mw and θP versus M w based on the log10 42  (open circles) and maxxUD tP
3 > log10 42  (closed diamonds). The data
selected 113 acceleration records is defined are plotted with respect to the epicentral distance Δ.
as:
Pn
i1 θPi −θP M wi −M w  algorithm. The magnitude estimation mtP
3 is based on a su-
r2  q
Pn q
Pn
EQ-TARGET;temp:intralink-;df5;55;437 5
θ
i1 Pi −θ P  2
i1 M wi −M w  2 pervised learning model that linearizes piecewise the behavior
of the avT and θP parameters as a function of magnitude Mw .
Pn The data used by the learning machine are xj ; y j , in which
i1 log a
10 vT −log a
10 vT M wi −M w  x j  avT ; θP  is the input data and y j is the target data,
r1  q
Pn q
P 
y j  M wj . The training data used to construct the algorithm
EQ-TARGET;temp:intralink-;df6;55;398

n
i1 log10 avT −log10 avT  i1 M wi −M w 
2 2
consists of 76 acceleration records corresponding to 25 earth-
6 quakes occurring from 1997 to 2013 and is expressed as the

(Pearson, 1895), in which T is the time

window of 3 s, starting at the arrival of
the P waves; n is the number of elements,
in our case 113; and θP , log10 avT , and M w
are the average values of these parameters.
The correlation between log10 avT
and θP versus M w of the 113 acceleration
records correspond to values of r1  0:42
and r2  0:53, respectively (Fig. 5). These
correlations show that it is feasible to con-
struct a model relating the seismic magni-
tude M w with these parameters. On other
hand, the correlation between log10 avT
and θP is r  0:1, indicating that these
parameters are linearly independent.

Estimation of Magnitude Based on

Parameters avT and θP
Figure 4. Location of the 34 in-slab earthquakes included in the training and testing
A procedure similar to the one used datasets, based on the criteria described in the Selection Criteria of Strong-Motion
to develop the (tS − tP ) algorithm (Cuéllar Records section. Open circles represent the epicenters of the earthquakes and closed
et al., 2017b) is used to calibrate the tP
3 triangles the location of the strong-motion stations.
A Fast Earthquake Early Warning Algorithm Based on the First 3 s of the P-Wave Coda 2073

Figure 5. (a) Plot of log10 avT versus Mw for the selected accelerograms; (b) plot of parameter θP versus Mw .

8
vector xj ; y j . Prior to initiating the process, the data are >
>
>
> aαvT
1
× θβP1 if avT ≥ γ 1
ordered as avTj < avTj
1 . The flow chart of the learning >
> α
machine is shown in Figure 6.
>
>
> avT
2
× θβP2 if avT ≥ γ 2
>
>
The process solves the overdetermined set of equations: > α3
> avT × θβP3 if avT ≥ γ 3
<
log10 Mwk  αk log10 avTk
βk log10 θPk . The procedure is mtp
3  aαvT
4
× θβP4 if avT ≥ γ 4 ; 7
>
>
EQ-TARGET;temp:intralink-;df7;313;499

linearized piecewise in m number of sections of the curve, >

> aαvT
5
× θβP5 if avT ≥ γ 5
where for each k section the system of equations is solved >
>
>
>
with the elements (rk − rk − 1), in which rk is the po- > aαvT
>
6
× θβP6 if avT ≥ γ 6
>
>
>
:a α7
× θ β7 if avT ≥ γ 7
sition of the ordered input data and its value corresponds to vT P
xj ; y j . The solution of the system of equations determines the
value of coefficients αk ; βk . Errork is the absolute mean error
jM wk − Yek j; in which M wk is the reported magnitude and Yek in which
is the resulting magnitude of the inversion process in section k.
The process starts by estimating a solution for the
α1  253:30 × 10−3 ; β1  −48:18 × 10−3 ;
EQ-TARGET;temp:intralink-;;313;363

first segment k  1. When the error for that segment is

Errork < tol, the process iterates, adding the next data point γ 1  0:4 × 103 cm2 =s4 ; α2  241:32 × 10−3 ;
xj
1 ; y j
1  and solves the new overdetermined set of equa- β2  30:42 × 10−3 ; γ 2  1:0 × 103 cm2 =s4 ;
tions (Fig. 6). This iterative process is followed until
Errork > tol. At that point, the algorithm moves on to solve α3  223:08 × 10−3 ; β3  79:81 × 10−3 ;
for the next section k
1 and sets the upper border γ of sec- γ 3  2:3 × 103 cm2 =s4 ; α4  219:70 × 10−3 ;
tion k, as the value γ k  0:5avTj
avTj
1 , in which the
β4  109:14 × 10−3 ; γ 4  4:6 × 103 cm2 =s4 ;
j
1th item caused the Errork t > tol. Thus, the last data
point of section k is the jth item xj ; y j , and the first datum α5  199:21 × 10−3 ; β5  61:87 × 10−3 ;
of section k
1 will be the j
1th pair xj
1 ; y j
1 . γ 5  6:3 × 103 cm2 =s4 ; α6  189:50 × 10−3 ;
The correlation between avT and M w shows that in
some cases, the accelerograms of large magnitude earth- β6  56:48 × 10−3 ; γ 6  9:8 × 103 cm2 =s4 ;
quakes exhibit relatively low acceleration values (Fig. 5). α7  171:69 × 10−3 ; β7  89:83 × 10−3 ;
These outliers increase the value of Errork , introducing a
γ 7  3:8 × 104 cm2 =s4 :
nonlinear behavior. Thus, the inversion process is supervised
at each iteration. If a data point xj ; y j  is identified as a clear
outlier in the k model, the data point is manually eliminated Additionally, in the construction of the algorithm, it is
by an active supervision of the learning algorithm (Fig. 6). necessary to establish two criteria in the model to eliminate
Using the training set of 76 acceleration records from the data inputs of the smallest and the greatest parameter val-
25 earthquakes the resulting model relating the magnitude ues used to calibrate the algorithm. In this case, based on the
mtP
3 to parameters avT and θP is defined by the following small number of training dataset, we propose these empirical
seven nonlinear equations: ranges:
2074 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

Figure 6. Schematic diagram showing the active supervised learning machine using least-squares method to construct a split model to
map M w .


< 5:0 if avT < 0:4 × 103 cm2 =s4 Magnitude Estimation mtP
3 on Independent
mtP
3 : 8
> 7:0 if avT > 1:0 × 105 cm2 =s4 Testing Data
EQ-TARGET;temp:intralink-;df8;55;438

To verify the robustness of the algorithm, a testing data-

set was used based on 37 acceleration records, stemming
As the number of training data grows, the criteria
from nine in-slab earthquakes occurring from 2014 to 2017
expressed in equations (7) and (8) must be modified in a new
(5:0 ≤ M w ≤ 7:1). This more recent dataset was used to test
calibration process.
the resulting algorithm considering that the seismic coverage
Based on the expressions (9) and (10), the mean abso-
in this period had a substantial improvement and is similar
lute error and its standard deviation are estimated as
to the current conditions. The test data includes the M w 7.1
Morelos earthquake that occurred on 19 September 2017.
1X N
merror  jM − mtP
3j j: 9 It is worth emphasizing that these data were not used in the
N j1 wj
EQ-TARGET;temp:intralink-;df9;55;330

construction of the algorithm.

Based on expressions (9) and (10), the mean absolute
error of the estimated magnitude was merror  0:5, with a
standard deviation σ error  0:20. Considering an error toler-
 X 1=2 ance of 0:5 and 1:0 in the magnitude estimation, the suc-
1 N
σ error  mtP
3j − merror 2 : 10 cess rate was 89% and 97%, respectively. Thus, the estimated
N j1
EQ-TARGET;temp:intralink-;df10;55;265

magnitude mtP
3 resulting from the test accelerograms com-

pares favorably with the catalog magnitude M w (Fig. 8).

The mean absolute error of the estimated magnitude Earthquake Early Warning Performance Evaluation
based on the 76 accelerograms of the training set was of the tP
3 Algorithm on Mexican In-Slab
merror  0:30 and a standard deviation of σ error  0:22. Earthquakes from 1997 to 2017
The resulting error lies within the tolerance in magnitude
of 0:5 and 1:0. This last tolerance error is accepted by The warning criterion used by SASMEX requires that
the earthquake early warning system in Japan (Hoshiba et al., two seismic sensors confirm the magnitude threshold to issue
2008). Based on these tolerance criteria, the algorithm accu- an alert. Thus to test the performance of the tP
3 algorithm
rately predicts the magnitude with a success rate of 82% as an actual warning tool, it was tested on 24 earthquakes
and 100%, respectively. A comparison between the estimated with magnitudes ranging from 5 ≤ M w ≤ 7:1. These 24
magnitude mtP
3 and the catalog M w shows a good correla- earthquakes met the criteria of having at least two strong-
tion (Fig. 7 and Ⓔ Table S2). motion stations that meet the specifications described above.
A Fast Earthquake Early Warning Algorithm Based on the First 3 s of the P-Wave Coda 2075

parameters are sent in real time via redundant communica-

tion to the processing centers located in the cities where the
early warning system is operational. The data package is
small and in this manner avoids delays and bottlenecks that
would potentially take place if the full data stream were
transmitted to a central location.
Of the 24 earthquakes analyzed, 19 were correctly clas-
sified considering the magnitude threshold M w ≥ 5:8. In four
cases, the magnitude was overestimated and in only one
case it was underestimated. The only earthquake for which
the algorithm underestimated the magnitude was the Mw 5.8
earthquake on 28 April 2008 (Ⓔ Table S2). The strong-
motion stations COMD, TNLP, and ATYC estimated
magnitudes mtP
3 of 5.8, 5.5, and 5.7, respectively. TNLP
estimated a lower magnitude due to the low acceleration
recorded. Because of the large hypocentral distance to ATYC
(112 km), the magnitude estimated was slightly below the
Figure 7. The magnitude estimation mtP
3 is plotted versus
the catalog magnitude M w for the training dataset composed of 76 threshold and did not confirm the seismic alert. A more
accelerograms of 25 Mexican in-slab earthquakes recorded from densely distributed network of monitoring stations probably
1997 to 2013. would have estimated the correct magnitude threshold.
The four earthquakes that were overestimated have
relatively small magnitudes 5.5, 5.7, 5.3, and 5.5. In all cases,
the epicentral distance was greater than 110 km. The algo-
rithm is not designed for these large epicentral distances. As
discussed before, a denser coverage of strong-motion station,
as exists today in central Mexico, probably would yield
better results.

Performance of the tP
3 Algorithm on Large

Mexican In-Slab Earthquakes (M w > 6:5)
To further evaluate the performance of the tP
3
algorithm as an earthquake early warning tool, three large
magnitude in-slab earthquakes Mw > 6:5 that occurred
relatively close to Mexico City are analyzed in detail.

M w 7.0 Tehuacán Earthquake on 15 June 1999

The M w 7.0 Tehuacán earthquake on 15 June 1999 had a
Figure 8. Magnitude estimation mtP
3 compared with the cata-
log magnitude Mw of the test data consisting of 37 accelerograms focal depth of 61 km. It is one of the more damaging in-slab
of nine Mexican in-slab earthquakes occurring from 2014 to 2017. recent earthquakes in Mexico (Singh et al., 1999; García
The estimated magnitude for the M w 7.1 Morelos earthquake was et al., 2005). Forty people lost their lives and the earthquake
Mw > 7:0. caused damage to many private and public buildings in more
than 600 localities in central México, including Mexico City.
The earthquake was well recorded by strong-motion stations
Iglesias et al. (2007) and Suárez et al. (2009) pointed out the of Instituto de Ingeniería, UNAM: the closest stations were
difficulties of the 2tS − tP  algorithm of correctly identifying CHFL-Chila de las Flores and CSER-Ciudad Serdán, which
the magnitude in the narrow bins prescribed by the Mexico City are located 50 km and 83 from the epicenter, respectively.
authorities to issue preventive alerts (5:5 < M w < 6). To avoid The estimated magnitudes mtP
3 were 6.3 and 6.8, respec-
this problem of estimating magnitude within very narrow bins, tively (Ⓔ Table S2). Therefore, a warning would have been
we propose a single earthquake early warning criterion: issued based on these two accelerograms. The resulting
Mw ≥ 5:8. warning time in Mexico City based on the tP
3 algorithm
It is worth pointing out that the parameters used by the would have been 34 s (Fig. 9). In contrast, the 2tS − tP  and
SASMEX algorithms are calculated in real time at each of (tS − tP ) algorithms would have given a warning time of only
the strong-motion stations of the SASMEX network. The 11 and 24 s, respectively.
2076 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

Figure 9. Post facto analysis of the Mw 7.0 Tehuacán earthquake comparing an earthquake early warning system using the tP
3
algorithm and the (tS − tP ) and 2tS − tP  algorithms. The horizontal axis corresponds to Greenwich mean time (GMT) beginning at
the origin time of the earthquake. The red dot on the map shows the location of the Tehuacán earthquake.

M w 6.5 Zumpango del Río Earthquake on 11 M w 7.1 Morelos Earthquake on 19 September 2017
December 2011
The 19 September 2017 (Mw 7.1) took place a couple of
The Zumpango del Río, Guerrero, earthquake occurred hours after the annual national drill, which annually sounds
on 11 December 2011, at a depth of 53 km. The distance the seismic early warning alert to commemorate the destruc-
from the epicenter to Mexico City was 200 km. In Mexico tive earthquakes of September 1985 and to remind the pop-
City, some high-rise buildings swayed for more than 1 min, ulation of the existing seismic hazard. The 2017 earthquake
causing damage in the electrical system that affected nearly caused 380 casualties and major damage in Mexico City and
82,000 people and the disruption of the phone service. Land- in surrounding locations. More than 11,500 buildings were
slides blocked a toll road between Acapulco and Mexico City damaged and 36 collapsed. Thanks to the expanded coverage
in several places. of the SASMEX network, the nearest strong-motion instru-
At the time, the strong-motion stations closest to the epi- ment was located just 17 km from the epicenter. Based on the
center were TNLP-Tonalapa and COMD-La Comunidad, at (tS − tP ) algorithm, an alert was issued in Mexico City.
epicentral distances of 52 and 65 km, respectively. The mag- Although the alert was heard ∼11 s prior to the arrival of the
nitude estimation mtP
3 in both strong-motion records was S waves, most people felt strongly the incoming P waves,
6.2. Here again, the tP
3 algorithm would have issued a almost simultaneously with the sound of the seismic early
public alert in Mexico City with a lead warning time of 35 s, warning alert (Fig. 11).
prior to the beginning of the strong shaking (Fig. 10). In At the time, the tP
3 algorithm was still under testing
contrast, the 2tS − tP  and (tS − tP ) algorithms would have and calibration and had not been operationally implemented
provided a warning time of 16 and 27 s, respectively. by SASMEX. However, it is interesting to compare the
A Fast Earthquake Early Warning Algorithm Based on the First 3 s of the P-Wave Coda 2077

Figure 10. Post facto analysis of Zumpango del Río earthquake of 11 December 2011 (M w 6.5) (centroid moment tensor catalog),
comparing an earthquake early warning system using the tP
3 algorithm with the 2tS − tP  and (tS − tP ) algorithms. The horizontal
axis corresponds to GMT beginning at the origin time of the earthquake. The red dot on the map shows the location of the Zumpango
earthquake.

theoretical performance of this new method, designed spe- The algorithm was constructed using a learning machine
cifically for these types of earthquakes, with the (tS − tP ) based on training data of 76 strong-motion records produced
algorithm that was operational at the time. The two strong- by 25 earthquakes, occurring between 1997 and 2013.
motion stations closest to the epicenter were PB01-San Juan The tP
3 algorithm was tested post facto as a seismic
Pilcaya and PB02-Tehuitzingo, at epicentral distances of 17 early warning tool on 24 earthquakes with magnitudes
and 51 km, respectively. The magnitude estimation for PB01 5 ≤ M w ≤ 7:1. We propose a simple confirmation criterion as
was saturated due to the very large value of avT , which a warning trigger of M w ≥ 5:8. Nineteen events were correctly
is notoriously larger than those used in the training data. classified; in four cases, the magnitude was overestimated and
By default, the magnitude at this station would have been in only one case was underestimated. The performance of the
fixed to mtP
3 > 7:5. PB02 estimated a magnitude mtP
3 tP
3 algorithm was also analyzed on the three largest and
6.7. Under these conditions, the tP
3 algorithm would more recent earthquakes in central Mexico. Thus, the magni-
have issued a public alert ∼16 s prior to the S-wave arrival tude mtP
3 would have correctly screened the M w 7.0 Tehua-
and ∼5 s prior to the P waves (Fig. 11).
cán, Mw 6.5 Zumpango, and Mw 7.1 Morelos earthquakes.
An early warning alert would have been activated with warning
Summary and Conclusions
times substantially longer than the algorithms that require the
The tP
3 algorithm developed here makes use of the arrival of the S waves. These results show that the tP
3
first 3 s of the P-wave acceleration observed on the vertical algorithm is a robust and reliable method to rapidly estimate
component to estimate the magnitude threshold of earthquakes the magnitude threshold of earthquakes within the subducted
and to make a decision whether an alert should be issued. in-slab in central Mexico.
2078 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

Figure 11. Post facto comparison of the performance of the tP
3 algorithms with the 2tS − tP  and (tS − tP ) algorithms for the Mw 7.1
Morelos earthquake on 19 September 2017. The horizontal axis corresponds to GMT starting 10 s before the origin time of the earthquake.
The red dot on the map shows the location of the Morelos earthquake.

Data and Resources Engineering Strong Motion Data. A. C. acknowledges support of Centro
de Instrumentación y Registro Sísmico (CIRES). Special thanks are due
Accelerograms from the Seismic Alert System of to Guadalupe Rico Zetina for her invaluable support in collecting and organ-
izing the data presented in this article. Valuable comments by Shunta Noda
Mexico (SASMEX) stations were provided by the Centro de
and an anonymous reviewer helped improve the original article.
Instrumentación y Registro Sísmico (CIRES) with the authori-
zation of the Instituto para la Seguridad en las Construcciones
del Distrito Federal in Mexico City and the Coordinación
References
Estatal de Protección Civil de Oaxaca. Data from the Instituto
de Ingeniería of the Universidad Nacional Autónoma de Cuéllar, A., G. Suárez, and J. M. Espinosa-Aranda (2017a). Performance
México (UNAM) are in the strong-motion database (https:// evaluation of the earthquake detection and classification algorithm
aplicaciones.iingen.unam.mx/AcelerogramasRSM/Default. 2tS − tP  of the Seismic Alert System of Mexico (SASMEX), Bull.
Seismol. Soc. Am. 107, no. 3, 1451, doi: 10.1785/0120150330.
aspx, last accessed May 2018). Data from the Global Centroid Cuéllar, A., G. Suárez, and J. M. Espinosa-Aranda (2017b). An earthquake
Moment Tensor Project are from http://www.globalcmt.org/ early warning algorithm based on the P-wave energy released in the
CMTsearch.html (last accessed May 2018). tS − tP interval, Bull. Seismol. Soc. Am. 108, no. 1, 238–247, doi:
10.1785/0120170115.
Acknowledgments Ekström, G., M. Nettles, and A. M. Dziewonski (2012). The global CMT
project 2004–2010: Centroid-moment tensors for 13,017 earthquakes,
The authors express their gratitude to the institutions that provided Phys. Earth Planet. In. 200, 1–9, doi: 10.1016/j.pepi.2012.04.002.
strong-motion data. The Mexico City and Oaxaca governments allowed Espinosa-Aranda, J. M., A. Jimenez, O. Contreras, G. Ibarrola, and R. Or-
us to use the waveform database of Seismic Alert System of Mexico tega (1992). Mexico City Seismic Alert System, International Sympo-
(SASMEX) stations and the Instituto de Ingeniería of the National Autono- sium on Earthquake Disaster Prevention, Mexico, 18–21 May 1992,
mous University of Mexico (UNAM) provided data from the Center for Proceedings Centro Nacional de Prevención de Desastres and Japan
 A Fast Earthquake Early Warning Algorithm Based on the First 3 s of the P-Wave Coda 2079

International Cooperation Agency CENAPRED-JICA, Mexico I, Suárez, G., T. Monfret, G. Wittlinger, and C. David (1990). Geometry of
1992, 315–324. subduction and depth of the seismogenic zone in the Guerrero gap,
García, D., S. K. Singh, M. Herráiz, M. Ordaz, and J. F. Pacheco (2005). In- Mexico, Nature 345, no. 6273, 336–338.
slab earthquakes of central Mexico: Peak ground-motion parameters Suárez, G., D. Novelo, and E. Mansilla (2009). Performance evaluation of
and response spectra, Bull. Seismol. Soc. Am. 95, no. 6, 2272–2282. the seismic alert system (SAS) in Mexico City: A seismological and a
Hoshiba, M., O. Kamigaichi, M. Saito, S. Y. Tsukada, and N. Hamada social perspective, Seismol. Res. Lett. 80, no. 5, 707–716.
(2008). Earthquake early warning starts nationwide in Japan, Eos Wu, Y. M., and H. Kanamori (2005). Experiment on an onsite early warning
Trans. AGU 89, no. 8, 73–74. method for the Taiwan early warning system, Bull. Seismol. Soc. Am.
Iglesias, A., S. K. Singh, M. Ordaz, M. A. Santoyo, and J. Pacheco (2007). 95, no. 1, 347–353.
The seismic alert system for Mexico City: An evaluation of its perfor- Wu, Y. M., H. Kanamori, R. M. Allen, and E. Hauksson (2007). Experiment
mance and a strategy for its improvement, Bull. Seismol. Soc. Am. 97, using the τc and Pd method for earthquake early warning in southern
no. 5, 1718–1729. California, Geophys. J. Int. 170, 711–717.
Odaka, T., K. Ashiya, S. Y. Tsukada, S. Sato, K. Ohtake, and D. Nozaka
(2003). A new method of quickly estimating epicentral distance
and magnitude from a single seismic record, Bull. Seismol. Soc. Am.
93, no. 1, 526–532. Centro de Instrumentación y Registro Sísmico (CIRES)
Pardo, M., and G. Suárez (1993). Steep subduction geometry of the Rivera Anaxágoras 814
Colonia Narvarte
plate beneath the Jalisco block in western Mexico, Geophys. Res. Lett.
Mexico City 03020
20, no. 21, 2391–2394. Mexico
Pardo, M., and G. Suárez (1995). Shape of the subducted Rivera and (A.C., J.M.E.-A.)
Cocos plates in southern Mexico: Seismic and tectonic implications,
J. Geophys. Res. 100, 12,357–12,373.
Pearson, K. (1895). Note on regression and inheritance in the case of two
parents, Proc. Roy. Soc. Lond. 58, 240–242. Instituto de Geofísica
Pérez-Yáñez, C., L. Ramirez-Guzman, A. L. Ruiz, D. R. Delgado, C. Universidad Nacional Autónoma de México
Macias, A. Marco, G. H. Sandoval, N. L. Alcantara, and R. A. Quiroz Ciudad Universitaria
(2014). Strong ground motion database system for the Mexican Seis- Mexico City 04510
Mexico
mic Network, AGU Fall Meeting, San Francisco, California, 14–19
[email protected]
December. (G.S.)
Singh, S. K., M. Ordaz, J. F. Pacheco, R. Quaas, L. Alcántara, S. Alcocer,
and E. Ovando (1999). A preliminary report on the Tehuacan, Mexico
earthquake of June 15, 1999 (Mw 7.0), Seismol. Res. Lett. 70, no. 5, Manuscript received 15 March 2018;
489–504. Published Online 10 July 2018

View publication stats