682 Seismological Research Letters Volume 80, Number 5 September/October 2009
doi: 10.1785/gssrl.80.5.682
INTRODUCTION
e term “earthquake early warning” (EEW) is used to describe
real-time earthquake information systems that have the poten-
tial to provide warning prior to signicant ground shaking.
is is possible by rapidly detecting the energy radiating from
an earthquake rupture and estimating the resulting ground
shaking that will occur later in time either at the same loca-
tion or some other location. Warning times range from a few
seconds to a little more than a minute and are primarily a func-
tion of the distance of the user from the earthquake epicenter.
e concept has been around for as long as we have had
electric communications (e.g., Cooper 1868), but it is only in
the last two decades that the necessary instrumentation and
methodologies have been developed (e.g., Nakamura 1988;
Espinosa-Arandaetal. 1995). e last ve years in particular
have seen a rapid acceleration in the development and imple-
mentation of EEW, fueled by a combination of seismic network
expansion, methodological development, and awareness of the
increasing threat posed by earthquakes paired with desire by
the seismological community to reduce risk.
is special issue of Seismological Research Letters is
intended to facilitate communication of EEW methodologies
and experiences in implementation. It complements the spe-
cial section of GeophysicalResearchLetters published in March
2009 (Allen, Gasparini, and Kamigaichi 2009). Together,
these collections of papers describe the science, engineering,
and societal considerations of the active warning systems in
Mexico, Japan, Taiwan, Turkey, and Romania and detail the
development and testing of methodologies in the Unites States,
Europe, and Asia (Figure 1). is introductory paper summa-
rizes this content to provide an overview of EEW status around
the world. We provide a summary of the various early warning
methodologies and then describe the active implementation of
early warning including the current users, the successes, and
the failures. e Perspectives, Misconceptions, and Challenges
section provides synthesis on the state of early warning around
the world, the progress made in the last few years, some of the
lessons learned and misconceptions uncovered, and nally the
challenges for the future.
APPROACHES TO EEW
Front Detection
All EEW systems must rst detect hazardous earthquakes and
then transmit a useful warning. e simplest approach to iden-
tifying when an earthquake is hazardous is to detect damaging
ground shaking. e rst approach to EEW was therefore to
detect strong ground shaking at one location and transmit a
warning ahead of the seismic energy. is concept, called front
detection, was initially proposed for San Francisco following
the 1868 earthquake on the Hayward fault east of the San
Francisco Bay. e radiating telegraph cables could be used to
signal coming ground shaking by ringing a characteristic bell
above the city (Cooper 1868). is system was never imple-
mented, but it captures the main concept of front detection by
installation of seismic sensors between the earthquake source
and the possible recipient of warnings. Front detection requires
good knowledge of the locations of future earthquakes.
In the late 1960s, the Japanese railway systems deployed
seismometers along their tracks that would trigger when the
ground shaking intensity exceeded some threshold and shut
o power to the trains. is approach, using “alarm-seismom-
eters,” only provides warning once serious ground shaking has
started. By deploying some instruments along the east coast of
Japan, between the large oshore earthquakes and the train
tracks, more warning time could be gained (Nakamura 1984).
All warning systems that make use of a network also use the
front detection concept by detecting an earthquake in one loca-
tion and providing warning to another.
When the only seismic sources are some distance from a
populated area, front detection can provide signicant warn-
ing times, i.e., tens of seconds. e Seismic Alert System (SAS)
for Mexico City uses front detection. Instruments along the
coast adjacent to the subduction zone trigger on earthquakes
and transmit the warning ~320 km to the city. Implemented in
1991, the SAS was the rst public warning system and contin-
ues to provide ~60s warning to Mexico City (Espinosa-Aranda
The Status of Earthquake Early Warning around
the World: An Introductory Overview
Richard M. Allen, Paolo Gasparini, Osamu Kamigaichi, and Maren Böse
Richard M. Allen,
1
Paolo Gasparini,
2
Osamu Kamigaichi,
3
andMarenBöse
4
1. University of California, Berkeley
2. University of Naples (Italy)
3. Japan Meteorological Agency
4. California Institute of Technology
Seismological Research Letters Volume 80, Number 5 September/October 2009 683
etal. 1995). e warning system for Bucharest, Romania, uses
the same concept. Seismometers in the Vrancea zone of the
southeastern Carpathians detect earthquakes and provide
~25s warning to Bucharest, which is ~160km away (Böseet
al. 2007). Istanbul also uses front detection. Ten instruments
along the northern shore of the Marmara Sea trigger based on
exceedance of an acceleration threshold at two or three sites
(Alciketal. 2009).
Using the P Wave
Waiting for strong ground shaking to be observed at one loca-
tion before issuing a warning results in a large “blind-zone”
around the epicenter where no warning can be provided. Given
a move-out velocity of ~4 km/s for peak ground shaking fol-
lowing shallow earthquakes, every second of delay increases the
radius of the blind zone by 4km. Using the P wave to deter-
mine whether an earthquake will produce hazardous ground
shaking provides additional warning. Given that the stron-
gest ground shaking usually arrives at the time of, or aer, the
S-wave arrival, using the P wave to provide warning has the
potential to increase the warning time everywhere, reduce the
radius of the blind zone, and potentially provide warning at the
epicenter.
A variety of observational parameters have been developed
for using the P wave to assess earthquake hazard. Many are
designed to estimate the magnitude of an earthquake, which
can then be translated into expected ground shaking. ese are
discussed in this section. Others are designed to bypass mag-
nitude and to estimate the ground shaking directly; they are
discussed in the following section, Onsite Warning.
One of the rst P-wave parameters developed for early
warning was the predominant period of the rst few seconds
of the P wave (Nakamura 1988). is was found to scale with
the magnitude of an earthquake while remaining insensitive
to the epicentral distance within a few hundred kilometers of
the event. Nakamura’s original method generates a continuous
time series of predominant period from which the maximum
value, τ
p
max
, can be extracted and converted to a magnitude
estimate (Nakamura 1988; Allen and Kanamori 2003). e
advantage of this approach is that an initial magnitude esti-
mate is available rapidly, within about one second, which can
then be increased if τ
p
max
increases. Typically, the predominant
period is monitored for three or four seconds. While concern
has been expressed about the stability of the measurement
(Wolfe 2006), the use of τ
p
max
has been successfully tested
using earthquakes from the western United States, Japan, and
Taiwan (Nakamura 2004; Lockman and Allen 2007; Tsang et
al. 2007; Wurman etal. 2007; Brown etal. 2009, this issue).
e use of τ
p
max
is currently part of real-time system testing in
California (Allen, Brown, Hellweg etal. 2009), Taiwan (Hsiao
etal. 2009), and Istanbul (Fleming etal. 2009, this issue).
A slightly dierent approach to measuring the frequency
content of the P wave was formulated by Kanamori (2005). e
τ
c
method calculates the eective (average) period of the P-wave
signal over a xed time window that is commonly selected to be
three seconds. is approach shows similar scaling between τ
c
and magnitude and has been tested in California and Taiwan
Earthquake Hazard: Peak ground acceleration (ms ) with 10% probability of exceedance in 50 years
0 0.4 0.8 1.6 2.4 3.2 4.0 4.8 7.0 10.0
-2
Japan
Taiwan
Mexico
Turkey
Romania
Italy
California
Switzerland
May 2009
Allen et al - Fig 1
Pacific
Tsunami
Warning
Center
Earthquake early
warning systems
Providing warnings
Real-time testing
China
▲ Figure 1. Map of global seismic hazard showing locations where active EEW systems are providing warnings to one or more users
(blue) and where EEW systems are currently being tested as part of a real-time seismic system (green).
684 Seismological Research Letters Volume 80, Number 5 September/October 2009
(Wu and Kanamori 2005a; Wu etal. 2007). While the τ
p
max
and τ
c
methods aim to characterize the same property of the P
wave, i.e., the frequency content, it has been shown that there is
an advantage to using both simultaneously (Shieh etal. 2008).
e τ
c
parameter is currently part of real-time system tests in
California (Böse, Hauksson, Solanki, Kanamori, and Heaton
2009), Taiwan (Hsiao etal. 2009), Japan, China, and at the
Pacic Tsunami Warning Center in Hawaii.
e amplitude of the P wave is also a useful parameter for
estimating earthquake hazard, provided a correction for the
epicentral distance can be made. e use of the peak displace-
ment, velocity, or acceleration of the rst few seconds, typi-
cally three seconds, of the P wave has been shown to scale with
magnitude (Wu and Kanamori 2005b; Zolloetal. 2006; Wu
and Kanamori 2008a) and ground motion (Böseetal. 2007;
Wu and Kanamori 2005a). e peak displacement, referred
to as P
d
, is usually found to be the most robust amplitude
parameter provided that integration or double integration of
the waveform at a station is possible. For instruments where
robust calculation of P
d
is not possible due to, for example,
noisy accelerometers, P
v
has also been shown to be useful
(Wurman etal. 2007). P
d
is currently in use by the real-time
systems in California (Allen, Brown, Hellweg et al. 2009;
Böse, Hauksson, Solanki, Kanamori, and Heaton 2009),
Taiwan (Hsiao etal. 2009), Istanbul (Fleming etal. 2009, this
issue), Japan, China, the Pacic Tsunami Warning Center in
Hawaii, and the Campania region of southern Italy (Lancieri
and Zollo 2008).
In Japan, the amplitude of the seismic waveform is used
in a continuous fashion to estimate the magnitude of an earth-
quake. No limit on the number of seconds of P wave is imposed.
Instead, the continuously monitored maximum amplitude of
the P-wave displacement (vector summation of the three com-
ponents) is used until the S wave arrives at a station. en, a
similar scaling between the amplitude of the S-wave displace-
ment and magnitude is used (Kamigaichi 2004). e epicentral
distance, which is necessary to translate the maximum ampli-
tude value into magnitude, is estimated by using the rst two
seconds of the log-transferred P-wave envelope slope (B-delta
method; Odaka etal. 2003).
A similar envelope function is used by several methodolo-
gies to parameterize the seismic waveforms. e peak accelera-
tion, velocity, and displacement is calculated every second and
used as an input. e Virtual Seismologist uses this data to trig-
ger on an event and estimate the magnitude using a linear com-
bination of the peak values (Cua and Heaton 2007). While
this approach uses amplitude, a frequency dependence is intro-
duced by the combination of acceleration, velocity, and dis-
placement. ere is therefore a similarity between this ampli-
tude-based approach and the frequency-based methodologies
described above. e neural-network based PreSEIS approach
makes use either of the cumulative absolute velocity (CAV;
Böseetal. 2008) or of other envelope parameters (Köhler etal.
2009, this issue) to predict seismic source and ground-motion
parameters. e Virtual Seismologist and PreSEIS are cur-
rently being tested in California and Switzerland (Cua etal.
2009, this issue; Köhler etal. 2009, this issue).
Finally, the cumulative energy of the seismogram can also
be used. In Mexico, the energy and growth rate are matched
to similar observations from previous earthquakes and used to
estimate the magnitude (Espinosa-Arandaetal. 1995).
One of the concerns when using the rst few seconds of the
P wave to estimate the magnitude of an earthquake is that the
magnitude estimate may saturate for large magnitude events.
If four seconds of P-wave data is being used, and the duration
of an M 6.5 earthquake is four seconds, can P-wave parameters
distinguish an M 7.5 earthquake from an M 6.5 event? ere is
clear evidence that the amplitude of the P wave within the rst
few seconds does saturate for M > ~7 earthquakes (Zollo etal.
2006; Wurman etal. 2007; Murphy and Nielsen 2009; Brown
etal. 2009, this issue). To try to reduce this problem, several
methodologies use longer time windows of the P wave and/or
the S wave to update magnitude estimates (Zolloetal. 2006;
Kamigaichi etal. 2009, this issue). When using the frequency
content of the P wave there is less empirical evidence for satura-
tion (Kanamori 2005; Olson and Allen 2005; Lewis and Ben-
Zion 2008; Brownetal. 2009, this issue), although this conclu-
sion is controversial (Rydelek and Horiuchi 2006; Yamada and
Ide 2008), and a satisfactory physical explanation remains elu-
sive. It may be sucient to know that an event is M 6.5 or larger
and broadcast a warning. However, methodologies to map the
nite ruptures of these large (M > 7) earthquakes could also be
developed and would enhance system performance (Yamada et
al. 2007; Yamada and Heaton 2008; Zollo etal. 2009).
Onsite Warning
e principle of onsite or single-station warning is to detect
seismic energy at a single location and provide warning of
coming ground shaking at the same location, i.e., detect the
P wave and predict the peak shaking. is is possible using a
combination of the P-wave parameters described above. e
simplest approach is to look for a scaling relation between the
P-wave amplitude and the peak ground-shaking (e.g., Wu and
Kanamori 2005b). However, small magnitude earthquakes
may have very large amplitude but high frequency spikes.
Combining the amplitude with frequency information is
therefore a more robust approach. If large amplitudes are also
associated with low frequencies, i.e., larger magnitudes, then a
warning should be issued.
UrEDAS is the grandfather of earthquake early warning
systems in general and onsite warning specically (Nakamura
1988). UrEDAS uses three seconds of the P wave to estimate
source parameters. Predominant period is used to estimate
magnitude, and then the P-wave amplitude and magnitude
provide an estimate of epicentral distance. e particle motion
is used to estimate the event azimuth and depth. All this is
achieved with a single three-component seismometer and with
remarkable accuracy. e complete point source description
can then be used to trigger an onsite alert based on predened
criteria. e Compact UrEDAS uses the same principles but
only one second of the P wave to provide more rapid warn-
Seismological Research Letters Volume 80, Number 5 September/October 2009 685
ings (Nakamura and Saita 2007b). UrEDAS continues to be
used along the rail systems in Japan (Nakamura 1996, 2004;
Nakamura and Saita 2007b) and a mobile unit named FREQL
is also in use by emergency response teams (Nakamura and
Saita 2007a).
e τ
c
–P
d
method is another onsite method that com-
bines two of the P-wave parameters described above (Wu and
Kanamori 2005a; Wu and Kanamori 2005b; Wuetal. 2007;
Wu and Kanamori 2008a, 2008b). When a station triggers on
the P-wave arrival, τ
c
is used to estimate magnitude and P
d
to
estimate peak ground velocity (PGV). rough careful selec-
tion of the appropriate combinations of τ
c
and P
d
, as well as
through the introduction of station corrections, the system can
be tuned to alert only when strong shaking is expected (Böse,
Hauksson, Solanki, Kanamori, Wu, and Heaton 2009). is
methodology is currently being tested in California (Böse,
Hauksson, Solanki, Kanamori, and Heaton2009), Taiwan
(Hsiao etal. 2009), Japan, China, and at the Pacic Tsunami
Warning Center in Hawaii.
Regional Warning
Regional warning makes use of a seismic network and typi-
cally combines many or all of the components described above.
Historically, seismic networks have been designed to stream
waveform data back to a network processing center, where
they are processed to detect, assess, and report on earthquakes.
e development of EEW has followed this route, installing
new processing modules at the network centers to generate
earthquake alerts and then distributing the warning to users.
“Onsite” processing algorithms are still used in a network set-
ting to provide the most rapid warnings, i.e., warnings based
on a P wave at a single station. In addition, networks also allow
data from multiple stations to be combined. is gives more
accurate predictions of the distribution of ground shaking
across the aected region by providing the earthquake location
and reduced uncertainties in earthquake magnitude estimates.
Once a communications network is in use both to col-
lect the seismic data and to give warning to users, the front
detection concept also provides for increased warning times.
Seismometers close to the epicenter are used to detect the event
and assess the hazard, and the communications provide warn-
ing to users at greater distances. In the case of large magnitude
earthquakes (M > 6.5), this approach can provide tens of sec-
onds of warning to areas that can expect damage. But there
will also likely be a blind zone around the epicenter where no
warning is available. is is due to the time lost transmitting
data to a processing center, processing the data, and sending
out a warning. A true onsite approach, where a seismometer is
installed at the user’s location and provides a warning at that
location, removes these telemetry delays, but with increased
probability of false or missed alarms.
Mexico. e operational network-based warning system
in Mexico consists of two components: the SAS system for
Mexico City and Guerrero and the SASO system for Oaxaca.
e SAS system uses a network of 12 accelerometers along the
coast of Guerrero above the subduction zone to provide warn-
ing to Mexico City ~320 km to the north (Espinosa-Aranda et
al. 1995; Espinosa-Aranda etal. 2009, this issue). e system
1) integrates energy at each station in a time interval starting
at the P-wave arrival and equal to twice the S-minus-P time
and 2) measures the growth rate. Empirical scaling relations
are used to determine whether the earthquakes are “strong”
or “moderate,” corresponding approximately to M ≥ 6 and
6>M≥5 respectively. When two or more stations report a
strong earthquake, a “public” warning is issued; when two or
more report a moderate earthquake, a “preventive” warning is
issued. Although the system waits until well aer the S-wave
arrival before issuing any alert, ~60s warning is still available
to Mexico City due to the distance. Ground-shaking is still
high in the city, and much stronger than is typically the case
at such distances, due to amplication by factors of 100–500
caused by the basin sediments on which the city is built (Suárez
etal. 2009, this issue).
e SASO system uses a network of 36 seismic stations
distributed across Oaxaca. e subduction zone earthquakes
are distributed over a wider area than the adjacent Guerrero
section of the subduction zone (see Figure 9 in Espinosa-
Aranda etal. 2009, this issue). Because the distances between
the earthquake sources and population centers are shorter,
algorithms requiring shorter time-windows of data must be
used. Empirical relations use parameters measured in two time
windows. During the interval from the P- to the S-arrival, the
dominant period, peak acceleration, and energy are measured.
When the S-minus-P time is more than three seconds, the
dominant period of the rst three seconds of the P wave is also
measured. ese parameters are used to determine whether the
earthquake is strong or moderate. Public or preventive warn-
ings are then issued, as with the SAS system.
Japan. e operational warning system implemented by
the Japan Meteorological Agency (JMA) combines both an
alert-seismograph concept and a network-based approach
(Kamigaichi etal. 2009, this issue). e system makes use of
~1,000 seismic instruments across Japan, 200 operated by
JMA and 800 by the National Research Institute for Earth
Science and Disaster Prevention (NIED), and integrates
methodologies developed by JMA (Hoshibaetal. 2008) and
NIED (Nakamuraetal. 2009). When a single station observes
ground-shaking above 100 cm/sec
2
, an alert is triggered
(the alert-seismograph approach). In addition, the network
approach is used. e source is characterized based on single
and multiple P-wave detections. First, the location is estimated.
Using a single P-wave detection, the slope of the onset is trans-
lated to epicentral distance, and the azimuth is estimated by
tting an ellipsoid to the particle motion (Odakaetal. 2003).
In addition, when one or two stations have detected a P wave,
a “territory” region is dened where the event must have
occurred based on the fact that other stations have not yet trig-
gered (Horiuchietal. 2005). e centroid of the territory is
used as the epicenter. e depth is xed at 10 km. Once three
or more stations have triggered, a grid search for the optimal
686 Seismological Research Letters Volume 80, Number 5 September/October 2009
location that minimizes mist with the observed arrival times
is used (Kamigaichi 2004).
e magnitude is estimated using scaling between the
P-wave amplitude and magnitude, correcting for the epicen-
tral distance. e vector summation of the three-component
waveform is monitored continuously and the magnitude esti-
mate updated as the amplitude increases. Once the S wave
arrives, a new scaling relation is used. Given the location and
magnitude, peak ground motion is obtained by estimating the
distribution of peak velocity on bedrock, applying site ampli-
cation factors, and converting peak velocity to JMA intensity
(Kamigaichi 2004), essentially the same approach as is used for
ShakeMaps aer an earthquake. e system provides an esti-
mate of the intensity and time until shaking for each subpre-
fecture (Kamigaichietal.2009, this issue).
ElarmS. e ElarmS EEW methodology (
http://www.ElarmS.
org)
is currently being tested statewide on the real-time seismic
systems in California (Allen, Brown, Hellweg etal. 2009) and
has been modied for oine testing using earthquakes from
Japan (Brown etal. 2009, this issue). e approach processes
seismic waveforms individually to generate P-wave trigger
times, P
d
, τ
p
max
, and signal-to-noise ratios. e peak ampli-
tude every second is also determined. ese parameters stream
continuously into an event monitor module that associates
triggers with detected earthquakes, locates earthquakes based
on observed arrival times, estimates the magnitude using P
d
and τ
p
max
relations, and predicts the distribution of ground-
shaking using attenuation relations and site corrections. e
system updates every second, providing an “AlertMap” of the
predicted ground-shaking distribution (Wurman etal. 2007).
Approximately 600 seismic instruments at ~400 sites cur-
rently stream into the system in California. At the time of this
writing, the system only provides warning to members of the
research group.
Virtual Seismologist. e Virtual Seismologist methodology
is also undergoing real-time testing in California, using the
same ~600 instruments at ~400 sites (Cua etal. 2009, this
issue), and in Switzerland. e approach uses the peak accelera-
tion, velocity, and displacement every second to detect earth-
quakes, locate them, and estimate the magnitude. e meth-
odology uses a Bayesian approach to predict the likelihood of
a given magnitude and source location using prior information
such as past seismicity and the Gutenburg-Richter relation
(Cua and Heaton 2007), though these are not yet incorporated
into the current real-time test version of this system. Virtual
Seismologist has also developed a sophisticated scheme for l-
tering false events. At the time of this writing, the system only
provides warning to members of the research group.
Presto. is is a probabilistic evolutionary approach to EEW
that is currently undergoing real-time testing using the 28-sta-
tion Irpinia Seismic Network (ISNet) in southern Italy (Weber
etal. 2007). It uses the P-wave arrival times, along with the
information that some stations have not yet triggered, to
identify the 3D region where the earthquake origin could be
located (Satriano et al. 2008). e low-frequency amplitude
of the P wave and also S wave is used to estimate magnitude
(Lancieri and Zollo 2008). e uncertainty in the EEW alerts
has been modeled extensively and shows that the largest source
of uncertainty in PGA estimates is the inherent uncertainty
in the ground motion prediction equation and not the rapid
magnitude and location estimates (Iervolinoetal. 2009). e
performance of the system in large-magnitude events has also
been studied and indicates a signicant eect of source nite-
ness (Zolloetal. 2009).
PreSEIS. PreSEIS is a neural-network based approach to early
warning using P-wave arrival times and ground motions at one
or more stations to locate earthquake hypocenters, estimate the
magnitude and expected ground motions, and predict the nal
expansion of the evolving seismic rupture for large magnitude
earthquakes (Böse etal. 2008). In order to train and test the
neural networks, the PreSEIS approach has been applied to
the station distributions in Istanbul and southern California
using synthetic seismograms for nite-fault scenarios (Böseet
al. 2008) and envelope functions of real earthquakes (Köhleret
al.2009, this issue), respectively.
SOSEWIN. e Self-organizing Seismic Early Warning
Information Network is a network of 20 stations installed in
the Ataköy district of Istanbul in June 2008, which is currently
undergoing real-time testing (Flemingetal. 2009, this issue).
is system is dierent from those described above in that
there is no central network processing center. Instead, each
station combines a sensor, onsite processing, and wireless com-
munications to the adjacent stations. Soware at each location
triggers on seismic arrivals and calculates a range of parameters
including arrival time, peak amplitudes, predominant period,
etc. Adjacent stations share this information, and alerts can be
generated based on a single-station detection, multiple adjacent
stations, or multiple components of the array. e current test-
ing is focused on communications, operation reliability, and
identifying trigger thresholds. In principle, any of the meth-
odologies described above could be applied to a self-organizing
wireless network like this.
Geodetic Networks
Most early warning system development to date has focused
on the use of seismological networks. Historically, it is seismic
networks that have provided real-time and post-earthquake
information, and these networks are very eective at detecting
earthquakes, locating earthquakes, and rapidly estimating the
magnitude for small and moderate earthquakes. e ability of
these P-wave-based methods to rapidly and accurately detect
large earthquakes (M > 7) is controversial as discussed above.
Real-time, high-sample-rate GPS networks are now becoming
more prevalent and could provide real-time constraints. e
real-time accuracies are currently at the subcentimeter level,
meaning that they could provide constraints for large magni-
tude events.