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

net/publication/321924423

An Earthquake Early Warning Algorithm Based on the P‐Wave Energy Released

in the tS–tP Interval

Article  in  Bulletin of the Seismological Society of America · December 2017

DOI: 10.1785/0120170115

CITATIONS READS
5 302

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 16 July 2018.

The user has requested enhancement of the downloaded file.

 Bulletin of the Seismological Society of America, Vol. , No. , pp. –, , doi: 10.1785/0120170115

An Earthquake Early Warning Algorithm Based on the

P-Wave Energy Released in the tS –tP Interval
by Armando Cuéllar, Gerardo Suárez, and J. M. Espinosa-Aranda

Abstract The tS –tP earthquake early warning algorithm measures the energy of the
P-wave coda on the vertical component in the tS –tP period to make a magnitude
threshold estimation. The algorithm is based on two parameters: the logarithm of
the peak ground acceleration, max
a
tS –tP  and the logarithmic cumulative accel-
eration SA
tS –tP . The model is built using a learning algorithm that iteratively pa-
rameterizes in segments the linear fit of max
a
tS –tP  and SA
tS –tP  to Mw . Training
datasets were based on 324 accelerograms from 101 earthquakes (4:8 ≤ M w ≤ 8:1) in
the Mexican subduction zone from 1985 to 2013. The algorithm is supervised to avoid
outliers in the data. The process results in a family of linear equations parameterizing
the observations to magnitude m
tS –tP  calibrated to the observed Mw . The algorithm
was successfully tested using a dataset of 28 earthquakes in the Mexican subduction
zone, from 2014 to 2017. The performance of tS –tP algorithm was tested as a warning
tool using 89 earthquakes in the Mexican subduction zone from 1985 to 2017, that met
the criterion of having at least two stations within 70 km from the epicenter. The
results show that 79 were correctly screened. The magnitude of six events was over-
estimated and four were underestimated. These earthquakes had an unfavorable sta-
tion distribution. The M w 6 South Napa, California, earthquake of 24 August 2014
was used also as a test case. The two closest stations identified it as Mw ≥ 5:8 within
2 s after the arrival of the P phase. This resulted in a lead time of 10 s in Berkeley and
12 s in San Francisco, prior to the arrival of the S waves. Thus, the tS –tP algorithm
proves to be a reliable tool for seismic early warning where hypocenters are close to
the target cities.

Electronic Supplement: Earthquake catalog and table showing performance

evaluation of the tS –tP algorithm.

Introduction
Since its inception in 1991, the Mexican Seismic Alert presented a unique opportunity to develop a seismic early
System (SASMEX) was designed to cover the southeastern warning system. If earthquakes are correctly identified and
Guerrero seismic gap. At the time, the gap was considered their magnitude calibrated, the distance of ∼350–400 km
potentially capable of generating a great M w > 8 earthquake from the subduction zone provides an enviable warning time
(McCann et al., 1979; Singh et al., 1981). The original goal of ∼60 s before the strong shaking begins. No other city in the
of the system was to warn the population of Mexico City of world offers this advantage to implement a seismic early warn-
an impending earthquake, similar to the destructive 19 Sep- ing system.
tember 1985 earthquake (Espinosa-Armanda et al., 1995). SASMEX has grown to 97 dedicated strong-motion sta-
Although the 1985 earthquake took place about 350 km from tions covering the subduction zone, with an average spacing
Mexico City, it produced a large loss of life and unprec- of ∼25km (Fig. 1). Additional seismic stations are located
edented material damage (e.g., Rosenblueth, 1986; Esteva, inland to monitor the in-slab earthquakes that occur within
1988). The soft and highly saturated soils, remnant of the the subducted Cocos plate Cuéllar et al., 2014. The original
pre-Columbian lake on which the city was built, were algorithm used by SASMEX to issue early seismic warnings
responsible for the large accelerations and the long duration to Mexico City is based on the growth of seismic energy in
of the strong motion of the ground. Therefore, Mexico City twice the observed tS –tP time. The 2
tS –tP  algorithm was

1
2 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

algorithm waits for the confirmation of

two neighboring stations that the magni-
tude threshold is exceeded to issue an alert.

Description of the tS –tP Algorithm

Detection of the P and S Waves
The detection process of the P and S
phases uses the three orthogonal compo-
nents of the strong-motion data, running
in parallel the average square input (Allen,
1978) and vertical-to-horizontal (V/H; Na-
kamura, 1996) algorithms. The real-time
detection procedure was described in de-
tail for the 2
tS –tP  algorithm (Cuéllar
et al., 2017). The phase detection process
Figure 1. Current distribution of Mexican Seismic Alert System (SASMEX) strong- takes into account empirical criteria that
motion stations (triangles) and cities where the seismic early warning system is now correct for site effects and for the charac-
operational (white circles). G, Guadalajara; M, Morelia; Mx, Mexico City; Ch, Chilpan- teristics of the seismic instrumentation
cingo; A, Acapulco; O, Oaxaca; RP, Rivera plate; CP, Cocos plate. (Cuéllar et al., 2017). In cases where the
seismic station detects abnormally high
background noise, the algorithm recalibrates its operational
thoroughly tested and discussed by Cuéllar et al. (2017). parameters. In the case of persistent levels of high noise, the
Based on the energy released, it estimates an empirical mag- seismic station is temporarily shut off the detection process
nitude range related to mb . until the background noise returns to operational levels.
In recent years, other cities were incorporated to
SASMEX (Fig. 1). Some of these cities lie close to the sub-
Estimation of Parameters max
a
tS –tP  and
duction zone or to the regions where in-slab earthquakes take SA
tS –tP 
place beneath the continent. For this reason, many of these
locations do not share the advantage of Mexico City of The algorithm uses the vertical component of three
having an almost minute-long warning time. Thus, the need orthogonal component accelerograms. The energy in the
to develop a faster algorithm to detect earthquakes and to t –t
S P time window is composed essentially of the P wave
determine a magnitude threshold. and its coda. The parameter max
a
tS –tP  is the logarithm
Here, we present an algorithm based on the seismic of the maximum value of the quadratic acceleration of the
energy released in the interval tS –tP . The algorithm initiates vertical channel within the t S –tP interval (equation 1).
from the detection of the P wave and ends with the arrival of The parameter SA
t –t
S P  is the sum of the logarithmic
the S. The acceleration of the P wave and its coda are mea- cumulative quadratic acceleration of the vertical channel ob-
sured on the vertical component. We propose a magnitude served within the same interval (equation 2). The algorithm
estimation model m
tS –tP  based on two parameters: the terminates with the detection of the S wave and proceeds to
logarithm of the maximum quadratic acceleration estimate the corresponding seismic magnitude threshold.
max
a
tS –tP  and the logarithmic cumulative quadratic max
a
tS –tP   log10 max
a
i2 tit S

1
acceleration SA
tS –tP  of the P wave. The algorithm is P
EQ-TARGET;temp:intralink-;df1;313;255

calibrated to M w using a training dataset of 101 earthquakes

occurring from 1985 to 2013 in the Mexican subduction X tS
SA
tS –tP   log10 a
i2 ;
2
zone. Only accelerograms of earthquakes M w > 4:7, located
EQ-TARGET;temp:intralink-;df2;313;229

itP
at epicentral distances < 70 km are used.
in which, a
i are the acceleration values on the vertical
The magnitude estimator m
tS –tP  is based on a family of
channel of the triaxial accelerometer and i is the sample num-
logarithmic linear regression models fit in a least-squares
ber starting from the P wave and ending with the detection of
sense, iteratively calculated using a machine learning strat- the S wave.
egy. In contrast to the two-tiered alerting protocol prescribed
by the authorities in Mexico City (Cuéllar et al., 2017), here
Correlation between M w and max
a
tS –tP  and
we propose a minimum magnitude cutoff of m
tS –tP  ≥ 5:8 to
SA
tS –tP 
activate the public earthquake early warning system. The
tS –tP algorithm neither calculates the location of the earth- To explore the correlation between the magnitude M w of
quake nor estimates potential shaking of the ground. The earthquakes in the Mexican subduction zone and parameters
An Earthquake Early Warning Algorithm Based on the P-Wave Energy Released in the tS –tP Interval 3

Figure 2. Cumulative distribution of the 1114 Racceleration records since 1985–2017 4:7 > Mw ≤ 8:1 grouped into the two categories:
max
a
tS –tP  ≤ log10
42  (open circles) and max
a
tS –tP  > log10
42  (closed diamonds) plotted with respect to (a) the tS –tP time and
(b) the epicentral distance Δ.

max
a
tS –tP  and SA
tS –tP , data were collected from max
a
tS –tP  ≤ log10
42 . The cumulative number of these
the following catalogs: the Mexican Seismological Service two groups of accelerograms was plotted relative to the
(SSN), the Global Centroid Moment Tensor (CMT) (Dzie- epicentral distance and to the tS –tP time(Fig. 2). In the case
wonski et al., 1981, 1999; Ekström et al., 2012), the strong- of records with large tS –tP times, the cumulative distribution
motion data recorded by the Instituto de Ingeniería, for max
a
tS –tP  > log10
42  decreases. In contrast,
Universidad Nacional Autónoma de México (Pérez-Yañez max
a
tS –tP  ≤ log10
42  increases as a function of time
et al., 2014) and the SASMEX network, operated by the (Fig. 2). We propose to use the intersection of these curves
Centro de Instrumentación y Registro Sísmico. A selection to empirically define a criterion of when the acceleration
criteria were applied as follows: records are above the noise level in the tS –tP time. The inter-
section of these curves corresponds to epicentral distances of
• A1. earthquakes reported in the Mexican subduction zone
~70 km and tS –tP ∼ 7 s.
in the SSN catalog with focal depths < 40 km;
Based on these results, the accelerograms selected had to
• A2. earthquakes of magnitude M w > 4:8, as reported by
the CMT catalog; and meet the following criteria:
• A3. accelerograms with clear recordings of both P and • B1. acceleration values in the vertical channel a > 4 cm=s2
S waves. during the tS –tP time;
A total of 1114 acceleration records were selected. In • B2. epicentral distances of < 70 km to the strong-motion
some cases, the accelerograms show very low accelerations site; and
• B3. acceleration records with tS –tP < 7 s.
that are close to the self-noise of the instruments. The reso-
lution of the majority of the early strong-motion SASMEX From the initial dataset of 1114 acceleration records, only
instruments was 10 or 12 bits. In those cases, the noise level 400 accelerograms recorded from 129 earthquakes ranging in
of the seismic acceleration recordings is at least 1:0 or magnitude from 4:8 ≤ Mw ≤ 8:1 met these criteria (Fig. 3 and
0:5 cm=s2 , respectively. To be well above the noise level, Ⓔ Table S1, available in the electronic supplement to this ar-
only acceleration records with values > 4 cm=s2 in the tS –tP ticle). These 400 accelerograms are divided into two sets: the
time window were utilized. This value of acceleration corre- first 324 are used as training data to develop the algorithm, and
sponds to max
a
tS –tP  > log10
42 . More recently, the in- the other 76 strong-motion records are used as test data.
struments have higher dynamic ranges using 16 or 24 bit Max
a
tS –tP  and SA
tS –tP  were computed for the
digitizers; the threshold criteria would be also acceptable selected dataset of 400 acceleration records and plotted
in these cases. against the reported magnitude M w . The correlation between
The strong-motion records were grouped in two: one max
a
tS –tP  and M w is r  0:40. In the case of the cumu-
where max
a
tS –tP  > log10
42  and the other with lative acceleration SA
tS –tP , the correlation coefficient with
4 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

The correlation between SA
tS –tP  versus M w shows an

asymptotic growth. As in the case of max
a
tS –tP , the cu-
mulative acceleration shows the presence of outliers, corre-
sponding to earthquakes Mw > ∼7. In some cases, outliers
are due to earthquakes where the directivity of the seismic
energy is radiated away from the recording seismic stations.
For example, in the case of the 1995 Copala earthquake
(Mw 7.3), the seismic energy radiated away from the closest
stations (Anderson et al., 1995). As a result, the values of
SA
tS –tP  and max
a
tS –tP  at stations GR11, COPL,
GR10, and GR09 were low at short epicentral distances
for this large magnitude earthquake (Fig. 4 and Ⓔ Table S2).

Calibration of Magnitude Based on max
a
tS –tP 

Figure 3. Map showing the 129 selected earthquakes in the and SA
tS –tP 
Mexican subduction zone from 1985 to 2017 (M w > 4:9; depth
< 40 km) used as training data (open circles). The seismic field sen- The correlation of parameters max
a
tS –tP  and
sors (triangles) are from the networks of the Instituto de Ingeniería,
Universidad Nacional Autónoma de México, and SASMEX. SA
tS –tP  with M w suggests the possibility to build an algo-
rithm relating the magnitude and the acceleration observed in
the P-wave coda (Fig. 4). To this end, we use a training set of
324 accelerograms corresponding to 101 earthquakes occur-
magnitude is r  0:50 (Fig. 4). The correlation between
ring from 1985 to 2013 that meet the selection criteria de-
max
a
tS –tP  and SA
tS –tP  is r  0:88, indicating that
scribed above (Ⓔ Table S1).
these two parameters are not completely independent. The relationships between magnitude and the parame-
Although the plot of max
a
tS –tP  versus M w indicates ters reflecting the energy released are estimated by a learning
that there is a linear correlation between these two parameters, machine algorithm exemplified in Figure 5. The input data
there are several outliers, particularly in the case of large earth- to the learning machine is vector xj ; y j  in which
quakes M w > ∼7 (Fig. 4a). The maximum acceleration of the x j 
SA
tS –tP j ; max
a
tS –tP j , formed from the 324
ground observed in the near field in the first few seconds does acceleration records, and y j is the target data: y j  Mwj .
not reflect necessarily the resulting magnitude of the earth- Before it is introduced into the learning machine process,
quake. A case in point is the 19 September 1985 earthquake the data are ordered so that SA
tS –tP j < SA
tS –tP j1 .
(M w 8.1) (Fig. 4a and Ⓔ Table S2). The low acceleration ob- The process solves the overdetermined set of equations
served in station CALE was attributed to the fact that the nu- log10 Mwj  αSA
tS –tP j  β max
a
tS –tP j .
cleation of the 1985 earthquake was located immediately Considering that the cumulative energy released
beneath that station (Mendez and Anderson, 1991). SA
tS –tP  behaves in an asymptotic manner, the process is

Figure 4. Correlation between parameters (a) max
a
tS –tP  and (b) SA
tS –tP  with Mw . The dataset shows the strong-motion records of
324 accelerograms recorded from 1985 to 2013.
An Earthquake Early Warning Algorithm Based on the P-Wave Energy Released in the tS –tP Interval 5

Figure 5. Flow chart showing the process followed by the learning procedure to determine the linear fit in k segments.

linearized in m sections of the curve, in which for each section through coefficients αk and βk in the m segments of the proc-
k, the system of equations is solved with the (r
k − r
k − 1) ess. As a result, these nine models are grouped in a family of
elements, in which r
k represents the position of the ordered linear equation parameterized by the γ-values:
input data and its value corresponds to xj ; y j  (Fig. 5). The 8
solution to the overdetermined set of equations determines the >
>
10SA
tS –tP  α1 ×
10max
a
tS –tP  β1 if SA
tS –tP >γ 1
>
>
value of coefficients αk ; βk  and the error in the linear fit of the >
>
10SA
tS –tP  α2 ×
10max
a
tS –tP  β2 if SA
tS –tP >γ 2
>
>
kth segment. Errork is the mean absolute error jMwk − Yek j; >
10SA
tS –tP  α3 ×
10max
a
tS –tP  β3
> if SA
tS –tP >γ 3
>
>
in which M wk is the reported magnitude and Yek is the result- >
10SA
tS –tP  α4 ×
10max
a
tS –tP  β4
< if SA
tS –tP >γ 4
ing magnitude of the inversion process in section k. m
tS –tP 
10SA
tS –tP  α5 ×
10max
a
tS –tP  β5 if SA
tS –tP >γ 5 ;
>
>
EQ-TARGET;temp:intralink-;df3;313;435

The process starts by building a solution to the first >

>
10SA
tS –tP  α6 ×
10max
a
tS –tP  β6 if SA
tS –tP >γ 6
segment k  1. When the error for that segment is >
>
>
>
10SA
tS –tP  α7 ×
10max
a
tS –tP  β7 if SA
tS –tP >γ 7
Errork < tolerance, the process recycles, adds the next data > SA
tS –tP  α8
>
>
>
10  ×
10max
a
tS –tP  β8 if SA
tS –tP >γ 8
point xj1 ; y j1 , and solves again the overdetermined set >
:
10SA
tS –tP  α9 ×
10max
a
tS –tP  β9 if SA
tS –tP >γ 9
of equations (Fig. 5). This iterative process continues until
Errork > tolerance. After several tests, we settled for an error
3
tolerance of 5% in the least-squares fit for each segment.
At that point, the algorithm moves on to solve the next section in which
k  1 and sets the upper bound γ of section k, as
α1  23:570 × 102 ; β1  −29:573 × 103 ; γ 1  3:230;
EQ-TARGET;temp:intralink-;;313;273

γ k  log10
0:5
10SA
tS –tP j  10SA
tS –tP j1 : Thus, the last
data point in section k is the jth term xj ; y j . Subsequently, α2  23:556 × 102 ; β2  −56:052 × 103 ; γ 2  3:380;
the first data point in section k  1 will be the j  1th
α3  21:214 × 10 ; 2
β3  −33:404 × 10 ; 3
γ 3  3:690;
item xj1 ; y j1 .
The correlations between SA
tS –tP  and max
a
tS –tP  α4  21:214 × 102 ; β4  −44:066 × 103 ; γ 4  4:016;
with M w show that the accelerograms of large magnitude α5  18:176 × 10 ; 2
β5  −8:438 × 10 ; 3
γ 5  4:149;
earthquakes in the near field often exhibit relatively low values
of peak ground acceleration and cumulative energy (Fig. 4). α6  21:492 × 102 ; β6  −73:543 × 103 ; γ 6  4:278;
These outliers introduce a nonlinear behavior and increase the α7  22:792 × 10 ; 2
β7  −102:763 × 10 ; 3
γ 7  4:637;
value of Errork . Thus, the inversion process is supervised at
α8  10:767 × 102 ; β8  −47:098 × 103 ; γ 8  5:394;
each iteration. If a data point xj ; y j  is identified as a clear
outlier in model k, the data point is manually eliminated α9  18:747 × 102 ; β9  −62:704 × 103 ; γ 9  5:609;
by an active supervision of the learning algorithm (Fig. 5).
The resulting tS –tP algorithm, based on the 324 training Based on the expressions (4) and (5), described below,
acceleration records, is defined by nine linear equations that the mean absolute error and its standard deviation are
empirically relate max
a
tS –tP  and SA
tS –tP  with m
tS –tP  estimated
6 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

Figure 6. The estimated magnitude m
tS –tP  plotted versus the

observed M w of 324 strong-motion records of 101 earthquakes
Mw < 4:8 recorded from 1985 to 2013 in the Mexican subduction
Figure 7. The estimated magnitude m
tS –tP  plotted versus the
observed Mw for the test data composed of 76 strong-motion re-
zone.
cords of 28 earthquakes M w > 4:8 recorded from 2014 to 2017
in the Mexican subduction zone.
1X N
merror  jM − m
tS –tP j j
4
N j1 wj
EQ-TARGET;temp:intralink-;df4;55;471

estimation, the success rate was 70% and 93%, respectively.

The magnitude estimated by the tS –tP algorithm of the test
 X 1=2 data is compared with the catalog magnitude in Figure 7 and
1 N
σ error 
m
tS –tP j − merror 2

5 Ⓔ Table S2.
N j1
EQ-TARGET;temp:intralink-;df5;55;439

In the case of the two largest earthquakes in the test dataset,

the 18 April 2014 (Mw 7.3) and the 8 May 2014 (Mw 6.5), the
The mean absolute error of the estimated magnitude of the two strong-motion stations closest to the epicenter estimated the
324 accelerograms corresponding to the training set was magnitude within an error of 0:5. The largest errors in mag-
merror  0:32, with a standard deviation of σ error  0:30. nitude estimation are in stations that are far from the epicenter.
This mean absolute error in magnitude is within the tolerance However, these other stations, although they are not at an ideal
accepted in the algorithm training process. Assuming an er- epicentral distance for the algorithm, they lie within the estab-
ror tolerance in magnitude of 0:5 and 1:0 as considered lished distance range. Thus, they were included for complete-
in Japan (Hoshiba et al., 2008), the algorithm accurately pre- ness and transparency in the tests.
dicts the magnitude with a success rate of 85% and 96%,
respectively. The results of the proposed algorithm, compar- Performance Evaluation of the tS –tP Algorithm
ing the estimated magnitude m
tS –tP  versus the catalog Mw is on Mexican Subduction Earthquakes M w > 4:7
shown in Figure 6. from 1985 to 2017

Magnitude Estimation m
tS –tP  on Independent Test Iglesias et al. (2007) and Suárez et al. (2009) pointed
Data from 2014 to 2017 out the difficulties of the 2
tS –tP  algorithm of correctly
identifying the magnitude in the narrow bins prescribed
To test the resulting algorithm, a test dataset was selected by the Mexico City authorities to issue preventive alerts
composed of 76 acceleration records of 28 earthquakes (5:5 < M w < 6). To avoid this problem of estimating magni-
occurring in the Mexican subduction zone from 2014 to tude within very narrow bins, we propose a single earthquake
2017. The magnitude of these 28 earthquakes ranges from early warning criterion: Mw ≥ 5:8.
4:8 < M w < 7:3. Because of the short time span of the test The activation criterion traditionally used by SASMEX
data available, there are only three earthquakes with M w > 6. to minimize spurious signals (Espinosa-Aranda et al., 1995)
Based on the expressions (4) and (5), the mean absolute requires that two nearby sensors confirm that the magnitude
error of the estimated magnitude was merror  0:5, with a estimate on both is above the prescribed threshold to issue an
standard deviation σ error  0:38. Thus, the resulting algo- alert. To test the alerting robustness of the tS –tP algorithm
rithm predicted the magnitude m
tS –tP  of the test earthquakes using a dataset with a broader range of magnitudes, we use
within the prescribed error tolerance of ±0.5. Considering the 89 earthquakes from 1985 to 2017 that meet the criterion
an error tolerance of 0:5 and 1:0 in the magnitude of having accelerograms within a epicentral distances of
An Earthquake Early Warning Algorithm Based on the P-Wave Energy Released in the tS –tP Interval 7

< 70 km. (Ⓔ Tables S1 and S2). Therefore, this test evalu-

ates the robustness of the proposed algorithm of discriminat-
ing earthquakes above or below the M w 5.8 threshold and not
only the magnitude estimation based on individual accel-
erograms.
Out of these 89 earthquakes, in 79 cases the seismic alert
based on the tS –tP algorithm would have been correctly
broadcast. In four cases, even though the magnitude of these
earthquakes reported in the seismic catalogs was M w > 5:8,
the magnitude m
tS –tP  was underestimated and no alert would
have been issued by the tS –tP algorithm. In the case of six
events, the algorithm overestimated the magnitude. This
means that 89% of the earthquakes would have been correctly
classified based on the pre-established warning criteria.
Earthquakes that are larger than the prescribed magni- Figure 8. Location of the Napa earthquake of 24 August 2014
tude threshold, but for which no alert is issued, represent (Mw 6) and of the four closest Center for Engineering Strong Mo-
tion Data (CESMD) seismic stations. NHC and station 68150 are
the possibility that a damaging earthquake may not be
located at distances of less than 10 km from the epicenter. The sta-
alerted. Thus, it is worth reviewing the four earthquakes tions C007 and JO56, in Berkeley and San Francisco, respectively,
Mw > 5:8 for which the magnitude m
tS –tP  was underesti- are used to determine the arrival of the S waves at these locations.
mated. In the case of the 31 May 1990 earthquake
(Mw 5.8), one out of four accelerograms (LLAV) estimated
the magnitude as m
tS –tP  6.4. The other three records yield an
low acceleration observed at station CALE, located immedi-
estimate of m
tS –tP  < 5:8. The second earthquake is the
ately above the hypocenter of the 1985 earthquake, is the
Mw 6.0 earthquake on 1 January 2004. In this case, the only
result of the nucleation process. From these results, it may
two stations available estimated m
tS –tP  < 5:8 (Ⓔ Table S2).
be concluded that, even under unfavorable distribution of
Both strong-motion stations lie at the limit of the design cri-
strong-motion stations, the tS –tP algorithm is a reliable tool
teria of the algorithm, with tS –tP times of 5.75 and 6.80 s,
to issue alerts in a short time span based only on the P-wave
respectively. The third case is the 14 June 2004 M w 5.9 earth-
coda. It is clear that with better seismic coverage, the tS –tP
quake. It was recorded by five acceleration stations. Station
algorithm would have a better performance than the one re-
PNTP estimated m
tS –tP  6.3. The next closest station OX04
ported here.
with tS –tP  3:01 s, estimated m
tS –tP  5.5 (Ⓔ Table S2).
Finally, the 2 April 2012 M w 6.0 earthquake, in this case
two stations selected estimated m
tS –tP  5.7.
Performance of M 6 South Napa Earthquake
It is worth pointing out that these earthquakes occurred
prior to the installation of the new SASMEX stations and, The South Napa earthquake is an ideal example of a
therefore, had a very unfavorable distribution of strong- shallow crustal earthquake occurring close to important
motion records. Furthermore, they correspond to cases where urban areas on which to test the tS –tP algorithm. Several
the reported magnitude is very close to the prescribed thresh- instruments from the Center for Engineering Strong Motion
old of M w 5.8. Data (CESMD) recorded the South Napa earthquake of
It is also interesting to evaluate the performance of the 24 August 2014 (M w 6) at relatively close distances. The
tS –tP algorithm on the eight earthquakes Mw > 7 in the data- NHC-Huichica Creek and the 68150-Napa College strong-
set (Ⓔ Table S2). Three of these eight earthquakes, the 14 motion stations are the closest to the epicenter and were
September 1995 Copala (Mw 7.3), the 20 March 2012 the first to record the earthquake. They are located 3.9
(Mw 7.5), and the 18 April 2014 (M w 7.3), had at least and 6.4 km away from the epicenter respectively (Fig. 8).
two strong-motion records. Although the true magnitude was The application of the tS –tP algorithm on the
underestimated, the tS –tP algorithm identified them as large NHC-Huichica Creek seismic record results in a magnitude
magnitude events and would have issued a correct early m
tS –tP  6.2 with a processing time of 2.02 s. The second seis-
warning in all cases (Ⓔ Table S2). mic record in 68150-Napa College processed the algorithm
In the case of the other five earthquakes M w > 7, only parameters in 2.08 s and estimated a magnitude m
tS –tP  6.5
one accelerogram was available. Thus, we could not evaluate (Ⓔ Table S3). Based on these two stations, which are the
them using the criteria of two confirming stations. Never- closest to the epicenter, an earthquake early warning would
theless, except for the 1985 M w 8.1 Michoacán earthquake have been broadcast ∼2 s after the arrival of the P wave
with magnitude estimation m
tS –tP  5.7, the accelerograms of (Fig. 9). Considering that the strong shaking coincides
the other four earthquakes were identified as having a mag- generally with the arrival of the S waves, the warning time
nitude M w > 5:8 and would have been alerted. As mentioned in Berkeley would have been ∼10 s, prior to the arrival of the
before, in the case of the 1985 Michoacán earthquake, the S wave, as measured on the C007-Forest Lane strong-motion
8 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

Figure 9. Graphical representation of the performance of tS –tP algorithm applied to the 24 August 2014 Mw 6 South Napa, California,
earthquake.

record (Figs. 7 and 8). In the case of San Francisco, the mic early warning system, would shorten the warning time.
warning time would have been 12 s prior to the arrival of The purpose of deriving a fast algorithm is to rapidly issue
strong shaking, based on the S-wave arrival at the J056- early warnings to cities that lie at distances of a few tens of
Golden Gate Avenue station (Fig. 8). kilometers from the epicenter. Although the warning time
It is interesting to compare the warning times reported may be of only a few seconds, it may help in establishing
by ShakeAlert to those obtained by the tS –tP algorithm. The rapid and automated civil protection measures.
warning time reported for the Napa earthquake based on The proposed tS –tP algorithm reduces in half the time
ShakeAlert was 5 s prior to the arrival of the S waves in needed to estimate the magnitude, making use of two param-
Berkeley (Allen et al., 2014; Grapenthin et al., 2014a,b). eters measured from the vertical component: the logarithm of
As a result of this process, the tS –tP algorithm would almost the maximum acceleration max
a
tS –tP  and the logarith-
double the early warning time. A similar result is obtained in mic cumulative quadratic acceleration SA
tS –tP . Based
San Francisco, where a warning time of 12 s would allow the on a linear regression model that maps SA
tS –tP  and
implementation of basic civil protection procedures. max
a
tS –tP  to Mw , a magnitude m
tS –tP  is estimated.
The construction of the algorithm makes use of a supervised
Summary and Conclusions machine learning process that linearizes piecewise the rela-
tion between maximum and cumulative acceleration and the
The recent expansion of SASMEX created the demand magnitude.
from cities that lie close to the seismic zones for the possibil- The algorithm was derived using a training dataset com-
ity to issue early earthquake warnings. In these cases, the posed of 324 accelerograms that correspond to 101 earth-
2
tS –tP  algorithm, traditionally used by the Mexican seis- quakes occurring between 1985 and 2013 (Ⓔ Table S1).
An Earthquake Early Warning Algorithm Based on the P-Wave Energy Released in the tS –tP Interval 9

The mean absolute error is merror  0:32, within the accepted Data and Resources
error tolerance of 0:5. In other words, considering an error
Accelerograms from the Mexican Seismic Alert System
tolerance of 0:5 and 1:0 in the magnitude estimation, the
(SASMEX) stations were provided by the Centro de Instru-
success rate was 85% and 96%, respectively.
mentación y Registro Sísmico (CIRES) with the authoriza-
The test dataset was based on 28 earthquakes occurring
tion of the Institutopara la Seguridad en las Construcciones
in the subduction zone from 2014 to 2017. The results of
del Distrito Federal in Mexico City and the Coordinación
this test on 76 accelerograms results in the absolute mean
Estatal de Protección Civil de Oaxaca, Oaxaca state.
error of 0.5. The test dataset had never been seen by the Strong-motion data for the Mexican subduction earthquakes
algorithm. Assuming error tolerances of 0:5 and 1:0 are from the webpage of the Instituto de Ingeniería of the
in the magnitude estimation, the success rate was 70% Universidad Nacional Autónoma de México (UNAM)
and 93%, respectively. https://aplicaciones.iingen.unam.mx/AcelerogramasRSM/
A test on the performance of tS –tP algorithm as a warn- (last accessed September 2017). Data from the Global Cent-
ing tool was conducted using 89 earthquakes in the Mexican roid Moment Tensor Project are from http://www.globalcmt.
subduction zone that met the criteria of having two strong- org/CMTsearch.html (last accessed September 2017).
motion records of nearby stations to issue a seismic warning
to avoid false alerts. The results show that 79 earthquakes,
out of the 89 tested, were correctly classified as being above Acknowledgments
or below the magnitude threshold M w 5.8. The magnitude of
The authors express their gratitude to the institutions that provided the
six events was overestimated and of four was underesti- strong-motion data. The governments of Mexico City and of the state of
mated. This represents a success rate of 89%. Oaxaca allowed us use the waveform database of Mexican Seismic Alert
Admittedly, if the detection of the S phase is late, the System (SASMEX). Thanks are also due to the Instituto de Ingeniería of
value of max
a
tS –tP  would increase and, as a result, the National University in Mexico (UNAM) for accelerograms in Mexico,
and to the Center for Engineering Strong Motion Data for the Napa
the magnitude would be overestimated. However, earthquake records. G. S. acknowledges support of the Consejo Nacional
max
a
tS –tP  is measured on the vertical channel, where de Ciencia y Tecnología (CONACYT) under Grant 082821. Special thanks
normally the amplitude of the S waves is smaller relative to are due to Guadalupe Rico, Sandra Ramos, Marisela Palomino, and Cecilia
the horizontal components, which are not incorporated in the Hernández for their invaluable support in collecting and organizing the data
presented in this article. The article was substantially improved by the valu-
algorithm. Nevertheless, using two stations to issue the alert able and constructive comments and corrections made by three anonymous
would help avoid this situation, as shown in Ⓔ Table S2. reviewers and by Associate Editor D. Melgar.
Clearly, now that the seismic coverage is substantially
better along the subduction zone (Fig. 1), the magnitude
estimations will tend to improve. It is important to underline References
that in the implementation of this algorithm in the earthquake Allen, R. V. (1978). Automatic earthquake recognition and timing from
early warning system, new relevant data will be systemati- single traces, Bull. Seismol. Soc. Am. 68, no. 5, 1521–1532.
cally incorporated as part of the process of an effective learn- Allen, R. V., D. Given, T. H. Heaton, and J. Vidale (2014). Successful
ing machine. ShakeAlert performance for the Napa quake, 2014 AGU Fall Meeting,
San Francisco, California, S44D-01.
The tS –tP algorithm was also tested on the 24 August Anderson, J., R. Quaas, S. K. Singh, J. M. Espinosa, A. Jimenez,
2014 Napa earthquake. The results for a hypothetical earth- J. Lermoand, and S. Alcocer (1995). The Copala, Guerrero, Mexico
quake early warning for the Napa earthquake show that the earthquake of September 14, 1995 (M w  7:4): A preliminary report,
tS –tP algorithm would have identified the event as having Seismol. Res. Lett. 66, no. 6, 11–39.
Cuéllar, A., J. M. Espinosa-Aranda, G. Suárez, G. Ibarrola, A. Uribe, F. H.
Mw > 5:8 at both stations within 2 s after the arrival of the Rodríguez, and B. Frontana (2014). The Mexican Seismic Alert Sys-
P phase. This would have given a warning time of ∼10 s in tem (SASMEX): Its alert signals, broadcast results and performance
the city of Berkeley prior to the arrival of the S waves. In the during the M 7.4 Punta Maldonado earthquake of March 20th,
case of San Francisco, the warning time would have 2012, in Early Warning for Geological Disasters, Springer, Berlin,
Germany, 71–87.
been 12 s.
Cuéllar, A., G. Suárez, and J. M. Espinosa-Aranda (2017). Performance
Early seismic warning algorithms face the challenge to evaluation of the earthquake detection and classification algorithm
reduce the processing time needed to identify earthquakes 2
tS –tP  of the Seismic Alert System of Mexico (SASMEX), Bull.
and to discriminate a magnitude threshold with a high degree Seismol. Soc. Am. 107, no. 3, 1451, doi: 10.1785/0120150330.
Dziewonski, A. M., T. A. Chou, and J. H. Woodhouse (1981). Determination
of confidence to make a decision on issuing a seismic alert.
of earthquake source parameters from waveform data for studies of
In this respect, the results of the tS –tP algorithm proposed global and regional seismicity, J. Geophys. Res. 86, 2825–2852,
here are encouraging for regions with large population cen- doi: 10.1029/JB086iB04p02825.
ters that are close to seismic zones. Although the warning Dziewonski, A. M., G. Ekström, and N. N. Maternovskaya (1999). Centroid-
moment tensor solutions for April–June, 1998, Phys. Earth Planet. In.
time would be only of a few seconds, well-designed proto-
112, no. 1, 11–19.
cols could activate a civil protection response designed to Ekström, G., M. Nettles, and A. M. Dziewonski (2012). The global
mitigate damage and loss of life. CMT project 2004–2010: Centroid-moment tensors for 13,017 earth-
 10 A. Cuéllar, G. Suárez, and J. M. Espinosa-Aranda

quakes, Phys. Earth Planet. In. 200/201, 1–9, doi: 10.1016/j.pepi Perez-Yañez, C., L. Ramirez-Guzman, G. A. L. Ruiz, D. R. Delgado, A.
.2012.04.002. Macias C Marco, G. H. Sandoval, N. L. Alcantara, and R. A. Quiroz
Espinosa-Aranda, J. M., A. Jimenez, G. Ibarrola, F. Alcantara, A. Aguilar, (2014). Strong ground motion database system for the Mexican
M. Inostroza, and S. Maldonado (1995). Mexico City seismic alert seismic network, AGU Fall Meeting, San Francisco, California,
system, Seismol. Res. Lett. 66, no. 6, 42–53. 14–19 December.
Esteva, L. (1988). The Mexico earthquake of September 19, 1985— Rosenblueth, E. (1986). The 1985 earthquake: Causes and effects in Mexico
Consequences, lessons, and impact on research and practice, Earthq. City, J. Am. Concr. Inst. 8, no. 5, 23–24.
Spectra 4, no. 3, 413–426. Singh, S. K., L. Astiz, and J. Havskov (1981). Seismic gaps and recurrence
Grapenthin, R., I. Johanson, and R. M. Allen (2014a). Operational real-time periods of large earthquakes along the Mexican subduction zone: A
GPS-enhanced earthquake early warning, J. Geophys. Res. 119, no. 10, reexamination, Bull. Seismol. Soc. Am. 71, no. 3, 827–843.
7944–7965. Suárez, G., D. Novelo, and E. Mansilla (2009). Performance evaluation of
Grapenthin, R., I. Johanson, and R. M. Allen (2014b). The 2014 M w 6.0 the seismic alert system (SAS) in Mexico City: A seismological and a
Napa earthquake, California: Observations from real-time GPS- social perspective, Seismol. Res. Lett. 80, no. 5, 707–716.
enhanced earthquake early warning, Geophys. Res. Lett. 41, no. 23,
8269–8276.
Hoshiba, M., O. Kamigaichi, M. Saito, S. Y. Tsukada, and N. Hamada
(2008). Earthquake early warning starts nationwide in Japan, Eos Centro de Instrumentación y Registro Sísmico (CIRES)
Anaxagoras 814
Trans. AGU 89, no. 8, 73–74.
Colonia Narvarte
Iglesias, A., S. K. Singh, M. Ordaz, M. A. Santoyo, and J. Pacheco (2007). Mexico City, Mexico 03020
The seismic alert system for Mexico City: An evaluation of its perfor- [email protected]
mance and a strategy for its improvement, Bull. Seismol. Soc. Am. 97, [email protected]
no. 5, 1718–1729. (A.C., J.M.E.-A.)
McCann, W. R., S. P. Nishenko, L. R. Sykes, and J. Krause (1979). Seismic
gaps and plate tectonics: Seismic potential for major boundaries, in
Earthquake Prediction and Seismicity Patterns, Birkhäuser, Basel,
Switzerland, 1082–1147. Instituto de Geofísica
Mendez, A. J., and J. G. Anderson (1991). The temporal and spatial evo- Universidad Nacional Autónoma de México
lution of the 19 September 1985 Michoacan earthquake as inferred Mexico City, Mexico 04510
[email protected]
from near-source ground-motion records, Bull. Seismol. Soc. Am.
(G.S.)
81, no. 3, 844–861.
Nakamura, Y. (1996). Real-time information systems for seismic hazards
mitigation UrEDAS, Q. Rep. Railway Tech. Res. Inst. 37, 112–127. Manuscript received 6 April 2017

View publication stats