Content uploaded by Richard M Allen
Author content
All content in this area was uploaded by Richard M Allen on Jun 14, 2019
Content may be subject to copyright.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Annual Review of Earth and Planetary Sciences
Earthquake Early Warning:
Advances, Scientic Challenges,
and Societal Needs
Richard M. Allen1and Diego Melgar2
1Department of Earth and Planetary Science, University of California, Berkeley,
California 94720-4760, USA; email: rallen@berkeley.edu
2Department of Earth Sciences, University of Oregon, Eugene, Oregon 97403-1272, USA;
email: dmelgarm@uoregon.edu
Annu. Rev. Earth Planet. Sci. 2019. 47:361–88
First published as a Review in Advance on
January 30, 2019
The Annual Review of Earth and Planetary Sciences is
online at earth.annualreviews.org
https://doi.org/10.1146/annurev-earth-053018-
060457
Copyright © 2019 by Annual Reviews.
All rights reserved
Keywords
earthquake early warning, hazard reduction, earthquake physics, ground
motion prediction, seismic networks, geodetic networks
Abstract
Earthquake early warning (EEW) is the delivery of ground shaking alerts or
warnings. It is distinguished from earthquake prediction in that the earth-
quake has nucleated to provide detectable ground motion when an EEW is
issued. Here we review progress in the eld in the last 10 years. We begin
with EEW users, synthesizing what we now know about who uses EEW and
what information they need and can digest. We summarize the approaches
to EEW and gather information about currently existing EEW systems im-
plemented in various countries while providing the context and stimulus for
their creation and development. We survey important advances in methods,
instrumentation, and algorithms that improve the quality and timeliness of
EEW alerts. We also discuss the development of new, potentially transfor-
mative ideas and methodologies that could change how we provide alerts in
the future.
Earthquake early warning (EEW) is the rapid detection and character-
ization of earthquakes and delivery of an alert so that protective actions
can be taken.
EEW systems now provide public alerts in Mexico, Japan, South Korea,
and Taiwan and alerts to select user groups in India, Turkey, Romania,
and the United States.
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
ShakeAlert:
earthquake early
warning system in the
United States
currently generating
alerts in California,
Oregon, and
Washington
EEW methodologies fall into three categories, point source, nite fault,and ground motion
models, and we review the advantages of each of these approaches.
The wealth of information about EEW uses and user needs must be employed to focus
future developments and improvements in EEW systems.
1. INTRODUCTION
Earthquake early warning (EEW) is the delivery of ground shaking alerts or warnings. It is dis-
tinguished from prediction in that the earthquake has nucleated to provide detectable ground
motion when an EEW is issued. The warning time available is the time between detection and
when ground motion is experienced by a user. Potential warning times are therefore seconds to
minutes. Likewise, the time available to collect and process geophysical data and deliver alerts is
seconds to minutes, and the actions of users must be possible in seconds to minutes.
The concept of EEW has been around for as long as there have been electronic communi-
cations that can outpace seismic waves. Following the 1868 earthquake on the Hayward Fault,
J.D.Cooper (1868) proposed that the new telegraph cables that radiated away from San Francisco
could be used to transmit a warning to the city and a characteristic bell would ring the alarm. While
the concept is simple, the implementation is much more complex. How do you detect an earth-
quake? How do you determine the size (magnitude or otherwise) of the event and the distance to
which ground shaking will be felt or damaging? How quickly can you do this, and how accurately?
How do you choose the right trade-off between speed and accuracy? Who should receive alerts?
How should the alerts be communicated (both message content and delivery technology) to dif-
ferent classes of users? How accurate do the alerts need to be? Since no system can be perfect,
what is the tolerance for false and missed alerts? Who should pay for the system (government
versus private sector/users)? Who is responsible for its successes and failures?
While EEW has its foundation in earthquake science, and it is earthquake scientists who have
predominantly developed the concept and been responsible for the implementation, the success of
an early warning system is dependent on many stakeholders working together to bring an EEW
system into operation in each region. For example, ShakeAlert is the US EEW system that is
currently being tested with pilot users in California, Oregon, and Washington. The phased rollout
of alerts to the entire population is now underway because the network infrastructure is complete.
Key individuals who are making this possible include (a) political leaders at the city, state, and
federal levels; (b) leadership at state and federal agencies responsible for risk reduction and disaster
mitigation; (c) leadership in the private sector representing EEW business applications; and (d)the
earthquake science community spanning geophysics, social science, and disaster mitigation. It is
important to recognize that without this broad collaboration, communication, and engagement,
EEW will achieve little.
So, what are the key components of a successful EEW system? The most important component
is a group of users who want alerts and can dene the necessary capabilities of the system. As
described in the next section, the gradual development of EEW around the world over the last
few decades provides a great deal of information about who the potential users are and what they
want and need. Next is the physical infrastructure for a system. Two physical networks are needed,
one that provides the data to detect and characterize earthquakes and a second that can deliver
the alerts. These networks include sensors, communications and telemetry,processing capabilities,
and receivers to deliver the alerts. With the growth of the Internet of Things, these could be the
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
same network. Finally, intellectual community and capacity are needed to distill the sensor data
into alerts and deliver them in a useful and usable format to users. This is, of course, a collaborative
effort between physical and social earthquake scientists.
There is also one additional characteristic of EEW that perhaps makes it unique in earthquake
science. An operational EEW system makes testable predictions about the earthquake process
and physics every day. Every time an EEW system detects or alerts on an earthquake, we test the
earthquake model used by the underlying EEW algorithms. Even more unique is the fact that
members of the public are able to evaluate the success or failure of our daily predictions when
they do, or do not, receive an alert for a felt earthquake. This is very different from other products
the earthquake science community provides. For example, earthquake forecasts of all types (e.g.,
hazard maps) provide a probability that there will be an earthquake, and correspondingly there is
a probability that there will not be a large earthquake. Whether an earthquake occurs or not, the
forecast is correct, which makes it untestable or at least difcult to understand and interpret from
the public perspective. This is perhaps at the heart of why EEW is so important in our efforts to
reduce earthquake risk: It is actionable information for every member of the public, and the public
can and will immediately assess if our information was correct or not. This provides a challenge
and an opportunity. This is why it is so important that we, the earthquake science community, get
it right.
In this review we aim to gather information about existing EEW systems around the world
and the ongoing development of methods and algorithms to improve the quality and timeliness of
EEW alerts. We start with a review of EEW users based on various studies around the world. We
then review the status of EEW systems implemented in various countries, providing the context
and stimulus for their creation and development. Next we review the development of new ideas
and methodologies for EEW algorithms. Finally, we review some new concepts for EEW that
could change how we provide alerts in the future.
1.1. Framing the Problem: User Needs
There are three broad categories of EEW users: (a) individuals receiving alerts who make per-
sonal decisions about how to respond, (b) automated response applications that typically require
institutions or companies to make decisions about how to apply and implement automated alert
responses, and (c) individuals and institutions who want rapid earthquake information for situa-
tional awareness purposes. It is also important to consider which users are more or less likely to
use a warning system. Who are the early adopters, and for whom does EEW adoption represent
a signicant expense and/or effort?
Perhaps the most important category of users is the public, broadly dened as a group of indi-
viduals who want personal alerts and will take personal protective actions. The impact of public
alerts and responses is perhaps the clearest case of the cost-benet of EEW. In the 1989 Loma Pri-
eta earthquake in the San Francisco Bay Area, more than 50% of injuries were linked to falls; in the
1994 Northridge earthquake in Southern California, more than 50% of injuries were due to non-
structural falling hazards (i.e., things falling on people rather than building collapse) (Shoaf et al.
1998). This means that if everyone got a few seconds’ warning of coming shaking and dropped,
took cover, and held on, the number of injuries in an earthquake could be halved (Strauss & Allen
2016). The estimated cost of injuries alone in the moderate M6.7 1994 Northridge earthquake
was $2–3 billion (Porter et al. 2006).
Many members of the public also fall into the category of early adopters in that they are keen to
receive the alerts as soon as possible. EEW is very popular with the public even in the face of lim-
ited or even poor performance. A survey of the public in Japan one year after the M9.1 Tohoku-Oki
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Modied Mercalli
Intensity (MMI):
scale used to describe
the shaking intensity at
a given location; varies
with distance from the
epicenter of an
earthquake
earthquake on March 11, 2011, asked the question, “Is EEW useful?” (Hoshiba 2014). Nation-
ally, 82% of those surveyed responded positively, and in the Tohoku-Oki region, 90% responded
positively. This was despite the fact that the warning for the M9.1 event was issued only in the
epicentral region and there were many alerts during the intense aftershock phase for which people
did not feel shaking due to incorrect associations and poor event locations (Hoshiba 2014). In a
less-formal survey of EEW users in Mexico City following events in September 2017, users had a
similarly positive response despite the fact that the alert was issued after most people felt shaking
in the most-damaging M7.1 September 19, 2017, Puebla event, and there had been several other
alerts in the same month for events in which most people did not feel shaking (Allen et al. 2017,
2018). In California, where there is currently no fully public warning system, a poll by Probolsky
Research in 2016 found that 88% of the sampled population supported building a statewide EEW
system, and 75% were willing to pay an additional tax to fund it. This is encouraging for EEW
developers: The public wants alerts and is accepting of imperfect warning systems.
So what type of alerts does the public want and need? We know that effective alerts must be
simple while delivering information about both the hazards and the actions to be taken (Wood
et al. 2012). While individuals may be in many different types of hazardous situations during an
earthquake, the default response message for EEW must be simple, just as it is for the response if
you feel shaking: “Drop, cover, and hold on.” What is notable is that providing information about
shaking intensity or time to shaking is neither needed nor desirable. Most people do not under-
stand the difference between intensity and magnitude, so including it causes confusion. Providing
a countdown can also delay response as a user digests the additional information and contem-
plates action. In Mexico City a simple siren sounds across the city (Cuéllar et al. 2014). In Japan
the broadcast alert is also simple: “Earthquake Early Warning.An earthquake has occurred in Area
X. Please prepare for a strong temblor” (Seki et al. 2008, p. 3).
The next question is, which area to alert? As the message contains an action to be taken, the
goal is to deliver that alert only to the people who should drop, cover, and hold on—that is, the
users who are likely to be impacted by the earthquake. Ground motion is inherently a stochastic
process. There will always be variability in the intensity of shaking from one location to the next
that can only be described statistically (e.g., Atik et al. 2010). This variability is typically a factor
of two, which corresponds to approximately one intensity level. If the goal is to alert all people
who might feel shaking, we must choose whether to alert areas where the predicted Modied
Mercalli Intensity (MMI) is II, meaning some people indoors feel the shaking, or IV, when many
feel shaking outside. When making this choice, we must consider the relative tolerance of the
public user for what is perceived as an unnecessary alert versus no alert when the user expected one.
More work is needed in this area, as there is little quantitative information as to the tolerance of
different users to false and missed alerts. However, some guidance may come from user perceptions
in Mexico City. When people were asked what they considered to be a false alert, their general
response was an alert when there was no earthquake (Allen et al. 2017, 2018). It was not an alert
with no felt shaking. This implies that users are more tolerant of unnecessary alerts than of missed
alerts.
A subset of this public group of alert users includes individuals who work in situations that
are more hazardous than typical ofces. Examples include construction workers on building sites,
utility workers who might be climbing high-voltage power lines, and people working with haz-
ardous chemicals or heavy machinery. While the hazard is related to the work environment and
the responsibility for providing a safe work environment may lie with the employer, it is still the
human responses of the individuals that are the most effective mitigation: stepping away from
hazards, stopping machinery, putting down chemicals, securing safety harnesses, and so on. In a
survey of potential industrial users for ShakeAlert in California, these human-response actions
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
were unanimously identied as being of the greatest value ( Johnson et al. 2016). These individual
users have the same need of a simple warning but require more specialized education and training
on how to respond to the alerts.
The public is also a good early adopter group because the daily life or work impact of adopting
EEW is minimal and has broader positive benets. If the actions to be taken are simply to drop,
cover, and hold on, or to step away from hazards at the workplace, then the cost of these actions is
minimal—just a few minutes. When this is done in a damaging earthquake, the actions can signif-
icantly reduce injuries, fatalities, and costs. Also, training people to take these actions and having
people respond to an earthquake alert even when the shaking is not that severe for them have the
benet of building a seismic culture of preparedness and thereby increasing the resiliency of a
community. This encourages people to think about earthquakes, the impacts, and their responses
and preparedness (Allen et al. 2017, 2018).
The second category of users is automated response applications. The list of potential applica-
tions is long and includes slowing and stopping trains, preventing planes from taxiing and landing,
taking elevators to the ground oor and opening doors, automatically isolating hazardous chemi-
cals, and stopping heavy and hazardous machinery. The examples of where these applications have
been implemented are more limited.
By far the most commonly implemented automated response application is slowing and stop-
ping trains. EEW has been part of the Japanese Shinkansen system since the 1960s and auto-
matically slows and stops trains when earthquakes are detected (Nakamura & Tucker 1988). The
high-speed trains in China are now also spurring the development of EEW systems. In the San
Francisco Bay Area, the Bay Area Rapid Transit (BART) train system was a very early adopter and
integrated EEW alerts into the train control system in 2012 when ShakeAlert was still a research
project (Strauss & Allen 2016). Two aspects of this application make train systems early adopters:
(a) the very serious potential consequences of earthquake shaking and (b) the ease of implemen-
tation coupled with the low costs and consequences of taking action. A potential consequence of
earthquake shaking is train derailment, which is more likely if a train is traveling at high speed and
can result in many injuries and casualties. The ease of implementation is related to the automated
nature of train control. BART automatically accelerates and decelerates trains in and out of sta-
tions. The same automated train control systems that make it easy to implement EEW also mean
that the cost of slowing and stopping a train when the shaking is not that severe is low. Trains
can be restarted as soon as the operators are ready,minimizing the impact on riders. Elevators are
another example. It is easy to have them go to the ground oor and open the doors, preventing
hundreds of people from being trapped.
The inverse of the above examples also appears to be true. EEW applications with signicant
costs of implementation and/or signicant costs or consequences of taking alert actions are un-
likely to be implemented. This is true irrespective of how high the consequences of earthquake
shaking are. Nuclear power plants are unlikely to use EEW (Cauzzi et al. 2016),as the cost of im-
plementing an emergency shutdown is signicant because it shortens the lifetime of the reactor.
The third category of users is situational awareness users. This refers to operational centers
who want and need to be aware of events that threaten infrastructure. This includes emergency
operations centers such as 911 centers and telecommunication, power, and other utility operations.
Having this information available allows them to understand possible causes of system disruption
and also reduce the impact by preventing cascading hazards. This group is also a very early adopter
of EEW, as having any additional source of information automatically streaming into these centers
is valuable.
The purpose of this summary is to draw conclusions about the types of information EEW users
want and need. First, for all the applications described, once an alert is issued, the response will
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Ground motion
prediction equation
(GMPE): estimates
the distribution of
ground shaking for an
earthquake based on
distance from the fault,
local site conditions,
and other parameters
likely be completed and the users will then wait for a period of minutes before resuming normal
activities. People will drop, cover, and hold on, and then they will wait for the shaking to stop
or wait to be sure that they will not feel shaking. Trains or elevators will stop, and operators will
wait to be sure the hazard has passed before restarting. In this sense, the alert cannot be taken
back; instead, once the alert is triggered, people will wait for an all clear. Second, because of all the
complexity and uncertainty when an earthquake is underway, the information must be reduced
to a very simple form for the alert decision to be made. All examples of known users are looking
for a binary alert: react when an alert is issued; otherwise no alert should be issued. As most users
will want to react when shaking is expected to be above some threshold, the information about
the earthquake must be reduced into a map of shaking intensity and then an alert issued to the
appropriate region for different categories of users. It is not the case that detailed earthquake
information (locations and magnitudes) can be sent to all users who will then decide whether to
react.
1.2. Approaches to Alert Generation
Now that we know who the users are and what they need, we can next consider the nature of the
alerts that are possible given the constraints of earthquake physics. Table 1 provides a summary of
the relevant distances and timescales for EEW for various magnitude earthquakes. We use MMI V
as the threshold when EEW is most useful. MMI V is described as “felt by nearly everyone; many
awakened. Some dishes, windows broken. Unstable objects overturned.” It is the lower threshold
for when we expect to see some light damage. One could argue that EEW is useful over a much
wider area—anywhere an earthquake is felt (i.e., MMI ≥II), as even people in a region where
shaking is felt but is unlikely to cause damage can benet from knowing that while they are about
to feel shaking, the hazard is minimal. Also, such alerts serve as an EEW drill and help build a
culture of seismic prevention (Allen et al. 2018).
The S-wave arrival time at the greatest distance where MMI V is expected and the approximate
end of peak shaking are included to give a sense of the total time available to provide an alert and act
in response to it. These numbers are only an approximate guide for several reasons. The shaking
intensity at any specic location can vary by a factor of two (Worden et al. 2010) compared to
the average shaking at that distance according to ground motion prediction equations (GMPEs).
Also, the strongest shaking can be signicantly later than the S-wave arrival time. A compilation
of peak shaking observations for large-magnitude events (M>6) shows that peak shaking occurs
up to 10 s after the S-wave at 50 km and up to 50 s at 400 km (Allen 2011).
Tabl e 1 Approximate estimates of relevant distances and times for earthquake early warning applications
Magnitude
Approximate
fault length
Approximate
distance from fault
where MMI ≥V
Maximum
epicentral distance
where MMI V is
expected
S-wave arrival time
at maximum
distance where
MMI V is expected
Approximate
end of peak
shaking
M5 1km 8km 10 km 4s 10 s
M6 6km 30 km 40 km 10 s 20 s
M7 50 km 100 km 200 km 40 s 60 s
M8 (crustal) 400 km 300 km 700 km 200 s 300 s
M9 (subduction) 1,000 km 400 km 1,000 km 300 s 600 s
All numbers are one signicant gure and are intended to give outer bounds to the time-space region where MMI ≥V may occur. Abbreviation: MMI,
Modied Mercalli Intensity.
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Source time function
(STF): seismic
moment release as a
function of time for a
given earthquake
Ground motion
model (GMM):
approach to
earthquake early
warning where
observations of ground
motion are directly
used to estimate future
ground motion
While the area that can be affected by a large-magnitude earthquake can be huge, most of
the damage caused by earthquakes will almost always be at smaller epicentral distances and close
to the rupturing fault. Therefore, providing an initial alert within a few seconds (Table 1)isthe
most critical objective for any EEW system so that alerts can be provided as close as possible to
the epicenter. Following this initial alert, there is time to provide better information or alert a
larger region in the tens of seconds that follow for larger events (M>7) (Table 1).
For this reason, all regional EEW systems use a point source algorithm. These typically use a
few seconds of P-wave data (0.5–4 s) from a handful of stations (two to six) close to the epicenter
to detect an earthquake and characterize location, origin time, and magnitude (Allen & Kanamori
2003). This information can be transformed into ground shaking information for users employing
an appropriate GMPE. These algorithms have the advantage of being fast; they usually provide
the earliest warning and therefore the most warning time. However, their predictions typically
saturate for ∼M7 earthquakes (Figure 1) for two reasons. First, it is difcult or impossible to
distinguish an M8 earthquake from an M7 when only information about the rst few seconds of
the source time function (STF) is known. Second, for ruptures extending hundreds of kilometers
along a fault, the shaking intensity that a user should expect is dependent on the magnitude and
distance to the fault rupture, and the lateral extent of the rupture is not provided by point source
information.
Finite fault algorithms aim to remedy the point source saturation issue by estimating the -
nite extent of the rupture. They also typically improve the magnitude estimate by reducing or
removing the magnitude saturation limitation (Colombelli et al. 2013). Both seismic and geodetic
approaches are being used to accomplish these goals. While nite fault algorithms are slower than
point source algorithms, they can predict higher intensities of shaking, and over larger areas, for
the largest earthquakes before the shaking is felt (Ruhl et al. 2019a) (Figure 1).
The last category of algorithms includes ground motion models (GMMs). These are very dif-
ferent in that they do not attempt to characterize the earthquake source at all. Instead, they use
observations of strong shaking to forward predict shaking at other locations (Hoshiba & Aoki
2015). Their advantage is that they are not susceptible to the challenges of earthquake detection,
association of seismic arrivals from multiple stations, uncertainties in location, and uncertainties
Ground motion model
Ground motion
model
Shaking intensity
Point source
Seconds after origin time
Source time
function
Source time
function
Observed ground motion
Observed
ground motion
Finite fault
Further from epicenterClose to epicenter
Fast but
saturated
Fastest
Too late to be useful Slower but
more precise
Seconds after origin time
Best but short
lead time
Figure 1
Conceptual sketch of expected earthquake early warning shaking forecasts during a moderate to large event (left) close to the epicenter
and (right) further away. The earthquake source time function (blue region) indicates the duration and character of rupture. Also shown
are the actual ground motion expected at each location (pink region) and the intensity forecasts for a point source algorithm (gold line), a
nite fault algorithm (green line), and a ground motion model (purple line).
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
SASMEX: earthquake
early warning system
in Mexico that
broadcasts alerts
including to a
loudspeaker system
that can be heard
across Mexico City
in magnitude. This can be particularly difcult during aftershock sequences, during earthquake
swarms, or for the largest magnitude events (M9). The disadvantage is that the accuracy of the
ground shaking prediction decreases as a function of the warning time. Forward predictions of
more than ∼10 s are perhaps too uncertain to be useful (Figure 1).
One last approach is the onsite method that uses data from a sensor at a site to detect an
earthquake and generate an alert at the same site. The most basic version of this method is a
simple ground motion threshold that sounds an alert at the same time the detected shaking is
unusual or reaches a damaging level. More sophisticated systems detect P-waves and trigger an
alert when the following S-wave or peak shaking is predicted to be large.
2. EARTHQUAKE EARLY WARNING IMPLEMENTATION
AROUND THE WORLD
The implementation of EEW has been driven by the advent of digital seismic instrumentation and
digital communications to collect the data and issue alerts. The availability of these technologies
has spurred the expansion of seismic networks, and the resulting data have then improved our
physical earthquake models to provide a framework for generating alerts.
But the development of EEW has also been driven by several key earthquakes. The 1985 M8.1
Mexico City earthquake killed more than 20,000 people in the city and illustrated that there could
be more than 1 min between when seismic stations detected the event along the coast and when
the shaking was felt in Mexico City.Mexico City’s EEW system became operational in 1991. The
1995 M6.9 Kobe earthquake killed more than 6,000 people and led to the deployment of multiple
dense seismic networks across Japan that were then used to develop an early warning capability that
became public in 2007. The 2008 M7.9 Wenchuan earthquake in China killed 70,000 and initiated
the development of EEW in the region. One of the greatest tests of EEW systems was the 2011
M9.1 Tohoku-Oki earthquake. While an alert was successfully issued, this event and its aftershocks
highlighted potential areas for methodological improvements (see Section 3). This event in Japan
also placed the EEW research effort in the United States on a path toward public implementation.
Finally, the 2018 M7.1 Puebla earthquake in central Mexico was the most signicant test of the
Mexican EEW system and provided insights into the public response and attitude toward EEW.
Here we summarize the current operational characteristics of EEW systems around the world.
We divide the systems into three categories (Figure 2;Table 2). Japan and Mexico have public
alert distribution where the alerts are broadcast through multiple channels and are available to
all members of the public. South Korea and Taiwan also provide public alerts to cell phones and
smartphones. Several regions have limited alert distribution where alerts are distributed to some
groups of users. These often include train operators, schools, and emergency service groups. Fi-
nally, system construction is underway in many more regions where real-time testing and devel-
opment are in progress, but these alerts are not yet being issued to users beyond the earthquake
science and engineering community.
Allen et al. (2009) reviewed the history and status of EEW systems around the globe at that
time, so here we focus on the changes, improvements, and expansions over the last decade. Clinton
et al. (2016) also reviewed efforts in Europe.
Mexico’s SASMEX system operated by CIRES (Espinosa-Aranda et al. 1995) has expanded
signicantly and now covers multiple states and cities. Alerts are issued through thousands of
dedicated radio receivers deployed in schools and government ofces. In Mexico City a public
alert sounds 12,000 sirens across the city that can be heard by most residents (Cuéllar et al. 2014).
The system still uses the concept of assessing the likely magnitude of a detected earthquake at
individual stations. When two stations have detected a signicant earthquake, the alert is triggered
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Earthquake hazard: peak ground acceleration (m/s2) with 10% probability of exceedance in 50 years
0 0.4 0.8 1.6
150°W 150°E120°W 120°E
75°N
60°N
45°N
30°N
15°N
EQ
15°S
30°S
45°S
60°S
90°W 90°E60°W 60°E30°W 30°EPM
Longitude
Latitude
AMAM
2.4 3.2 4.0 4.8 7.0 10.0
July 2018
Mexico
Japan
Korea
Korea
India
India
Romania
Romania
United States
United States
Chile
Chile
Italy
Italy
Costa Rica
Nicaragua
El Salvador
El Salvador
Switzerland
Switzerland
China
China
Israel
Israel
Turkey
Turkey
Taiwan
Alerts sent to select users
Earthquake early warning systems
Alerts sent to the public
Real-time testing
GLOBAL: MyShake
Figure 2
Status of earthquake early warning systems in different regions of the globe. Shown are locations of systems that broadcast alerts to all
members of the public (purple), systems distributing alerts to select users (orange), and systems undergoing real-time development and
testing (blue). The background is the seismic hazard presented as peak ground acceleration with 10% probability of exceedance in
50 years.
in cities likely to experience shaking. The threshold for issuing alerts in Mexico City, for example,
is the detection of M>5 earthquakes at two stations as far away as the coast 300 km to the south.
The algorithm applied at each site has been improved to require less data to issue alerts faster,now
just 3 s of the P-wave (Cuéllar et al. 2018). The system issued alerts for the damaging earthquakes
occurring in September 2017, and despite some performance challenges, it is clear that the public
perception of SASMEX is very positive (Allen et al. 2017, 2018). It is notable that there are also
two independent private-sector EEW systems that are operational, SkyAlert and Grillo. Both run
Tabl e 2 Status of earthquake early warning systems in different regions of the globe
Public alert distribution Limited alert distribution System construction
Alerts broadcast to all members of the
public
Alerts distributed to selected users Real-time testing and development of
alert system
Mexico: multiple states
Japan: nationwide
South Korea: nationwide
Taiwan: nationwide
India: Roorkee region
Romania: regional
Turkey: Istanbul
United States: West Coast
Chile: nationwide
China: several regions
Costa Rica: regional
El Salvador: regional
Israel: nationwide
Italy: Irpinia region
Nicaragua: regional
Switzerland: nationwide
Global: MyShake smartphones
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
ElarmS: earthquake
early warning
algorithm using a few
seconds of P-wave data
to detect, locate, and
estimate the
magnitude of an
earthquake underway
Microelectro-
mechanical system
(MEMS): a very
compact type of
accelerometer such as
those found in
smartphones and other
devices
Finite fault detector
(FinDer): earthquake
early warning
algorithm matching
observed patterns of
ground shaking with
templates to estimate
future shaking
their own detection networks and broadcast the alerts primarily through smartphone apps and to
dedicated Internet-connected devices.
The 2011 M9.1 Tohoku-Oki earthquake was a substantial test for Japan’s public alert system
operated by the Japan Meteorological Agency ( JMA) (Hoshiba 2014). In 2011, a point source
algorithm was being used to locate earthquakes based on P-wave arrival times and estimate the
magnitude from the rst few seconds of the P-wave. An alert was issued across the most severely
affected Sendai region, where the expected ground shaking exceeded the public alert threshold of
JMA intensity 5-lower (equivalent to MMI VII). However, this massive earthquake caused signi-
cant shaking over a much larger area than predicted by the EEW algorithm, as the magnitude esti-
mate saturated at M8.1, and there was no information about the nite extent of the rupturing fault
plane, which extended 400 km to the south and caused strong shaking in Tokyo’s Kanto region as
well. Also, in the intense aftershock sequence,the algorithm incorrectly associated seismic arrivals
from separate simultaneous events, resulting in poor-quality alerts for a period of weeks (Hoshiba
2014). JMA has signicantly improved the algorithms used since (Kodera et al. 2018), as we dis-
cuss in Section 3. The public perception of the JMA system is very positive, and alerts continue to
be broadcast through multiple channels (Hoshiba 2014). These include broadcast messages that
trigger alerts on most cell phones (Seki et al. 2008); publicly available commercial smartphone
apps (e.g., Yurekuru); and TV, radio, and various other dedicated communication channels.
South Korea has been experimenting with EEW for about a decade. The effort started with the
evaluation of the ElarmS point source algorithm (Kuyuk et al. 2014, Sheen et al.2017), which the
Korean Meteorological Administration now uses to issue public alerts. Alerts are distributed using
the Cell Broadcast System, which can deliver text message–like alerts to all phones simultaneously
across the country for M>4 earthquakes. Current efforts are underway to provide more localized
and user-specic alerts that include information of what actions to take.
Taiwan is now using a total of three EEW systems. The Central Weather Bureau is the ofcial
source of EEW and uses the national seismic network to detect events and issue warnings using a
P-wave-based point source approach (Wu et al. 2014). It typically takes 15 s to generate the alert,
meaning warnings can be issued to cities more than 50 km from the epicenter, and the alerts are
available on all mobile phones. The P-alert system developed by National Taiwan University uses a
low-cost microelectromechanical system (MEMS) sensor to provide onsite warnings more rapidly
to locations close to the epicenter. It uses P-wave displacement thresholds to issue alerts within
the ∼600 buildings—most of them schools—that now have P-alert devices installed (Wu et al.
2018). Finally, the National Center for Research on Earthquake Engineering has also developed
an onsite approach. It also uses a few seconds of P-wave data to predict the coming peak shaking,
but it uses six extracted features and a support vector machine model to decide when to alert.
Its devices are currently installed in ∼30 locations, primarily schools, and it has plans to expand
rapidly (Hsu et al. 2018).
The phased rollout of a public EEW system for the West Coast of the United States (California,
Oregon, and Washington) is currently underway. The system is called ShakeAlert and is operated
by the US Geological Survey in collaboration with the University of California, Berkeley; the
California Institute of Technology; the University of Oregon; the University of Washington; and
state emergency management agencies (Kohler et al. 2017). Expansion of the seismic networks
will double the number of sensors contributing data to 1,500 over the next few years (Given et al.
2014). The system received extensive testing of multiple approaches (Cochran et al. 2017) and
uses a single point source algorithm called EPIC that is primarily based on ElarmS-3 (Chung
et al. 2019) and the nite fault detector (FinDer) source algorithm (Böse et al. 2017). Testing is
also underway of multiple algorithms that use geodetic data to both improve the magnitude and
estimate the nite extent of large ruptures. The phased public rollout in 2018 included stopping
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
PRESTo: earthquake
early warning
algorithm using
P-wave data to
estimate the location
and magnitude of an
earthquake
MyShake: global
smartphone seismic
network that detects
earthquake shaking,
records ground
acceleration time
series, and generates
earthquake alerts using
private/personal
smartphones
trains, alerts in some schools, and automated water supply management. At the beginning of 2019,
the city of Los Angeles made the ShakeAlertLA app available to the public with the intent of
delivering alerts in Los Angeles County.
In northern India a network of 84 accelerometers has been deployed by IIT Roorkee around
the main central thrust just north of Roorkee. The system takes a point source approach, using
peak displacement from 3 s of P-wave data to estimate the magnitude. When an event M6 or
greater is detected, an alert is issued that sounds sirens in student dormitories and in emergency
control rooms for all districts across Uttarakhand (Chamoli et al. 2019).
In Romania the National Institute for Earth Physics has been providing warnings in the
Vrancea region, where very deep earthquakes occur and have damaged Bucharest in the past.
Alerts go to a nuclear research facility and the Basarab Bridge, where trafc is stopped. It is us-
ing the PRESTo algorithm (Satriano et al. 2010) in addition to its own event validation approach
(M˘
armureanu et al. 2010), and alerts are now being made more widely available by private com-
panies delivering alerts from the national system (Clinton et al. 2016).
In Turkey Istanbul has a warning system based on a simple exceedance of acceleration threshold
at three stations in the network of sensors across the city. The system is operated by the Kandilli
Observatory and the Earthquake Research Institute and provides alerts to the Istanbul Gas Dis-
tribution Company and the Marmaray Tube Tunnel. Gas valves are shut off and trains are slowed
when an alert is received and conrmed by local acceleration observations (Clinton et al. 2016).
System construction is underway in many places around the world using local/regional net-
works in addition to a global smartphone system. China is beginning the implementation phase
of a national system that will initially focus on four regions. A total investment of $280 billion
has been made to install 15,000 sensors across earthquake-prone regions and deliver public alerts
(Li 2018). Also, the high-speed rail is in the process of developing independent systems (Lu et al.
2016). Chile has been expanding its seismic and geodetic networks since the M8.8 offshore of
central Chile in 2010. It has been testing the ElarmS (Kuyuk et al. 2014) and FinDer (Böse et al.
2017) algorithms as well as experimenting with the deployment of modied smartphones to col-
lect geodetic data (Minson et al. 2015). Costa Rica,El Salvador,and Nicaragua have been building
capacity in EEW and are testing the Virtual Seismologist (VS) (Cua & Heaton 2007) and FinDer
(Böse et al. 2017) algorithms. Israel is currently constructing an improved national seismic net-
work with the specic goal of delivering alerts. It is currently testing the ElarmS algorithm on the
growing network (Nof & Allen 2016). The Irpinia region of Italy is running the PRESTo point
source algorithm with the goal of delivering alerts in Naples and the surrounding region (Satriano
et al. 2010). Switzerland has a high-quality national seismic network and is continuing testing and
development of the VS (Cua & Heaton 2007) and FinDer (Böse et al. 2017) algorithms.
Finally, the MyShake project (https://myshake.berkeley.edu/) has developed the capability
to detect earthquakes using private/personal smartphones with the goal of collecting earthquake
data from their global smartphone network and delivering alerts to users (Kong et al. 2016b).
The use of this nontraditional sensor network makes it possible to provide alerts in earthquake-
prone regions where there are smartphones (i.e., wherever there are people). With over 300,000
downloads to date, the system has detected over 800 earthquakes and demonstrated the end-to-
end capability to create alerts (Kong et al. 2016a). To issue alerts in multiple regions around the
world, the number of users will need to increase signicantly, but the hope is that once the system
starts issuing alerts, many more people will download the app and participate.
3. RECENT ADVANCES
As previously noted, the bulwark of EEW for the last 30 years has been the point source algorithm.
Point source algorithms continue to be improved upon, and new approaches such as nite faulting
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Gutenberg
algorithm: uses
frequency content of
the P-wave at a single
site to estimate the
likely distance to the
earthquake source and
its magnitude
Integrated particle
lter (IPF): method
using information
about stations
triggering on a seismic
arrival to determine if
one or more
earthquakes are
underway
and ground motion methods have matured. In this section we review these and other relevant
methodological advances.
3.1. Improvements to Point Source Algorithms
Timely alerts are the most important attribute of point source methods. For fast alerts network
density exerts a rst-order control on how quickly an earthquake will be detected and thus the size
of the blind zone where no alert is possible (Kuyuk & Allen 2013). However, as noted in Section 1.2
and Figure 1, there is a trade-off, introduced by the details of the algorithm, between magnitude
uncertainty (and thus ground motion uncertainty) and speed of the alert. Broadly speaking, al-
gorithms that require several stations to trigger before issuing an alert will experience fewer false
alerts and produce more robust source and ground motion estimates. However, they will be slower.
Single station algorithms, conversely, will be faster but will be more error prone both by having
larger uncertainty in their estimate of magnitude and by being more susceptible to false alerts.
Several heuristics have been proposed (e.g., Böse et al. 2009) to improve the performance of sin-
gle station event detection with modest success. A more robust approach was proposed by Meier
et al. (2015), who used a novel lterbank technique. This signicantly reduces the uncertainty by
exploiting more features of the early onset waveforms. The lterbank outputs are used for joint
Bayesian estimation of the magnitude and source-to-station distance. By inferring these parame-
ters jointly,this proposed Gutenberg algorithm relies on the notion that large amplitudes at high
frequencies will be observed only if the source-to-station distance is short. Similarly, high ampli-
tudes at long periods should be observed only if the magnitude is large. The Gutenberg algorithm
produces better estimates with only one to two stations than traditional onsite methods and thus
can speed up the rst alert and reduce the blind zone substantially.
Meier et al. (2015) also suggested that prior information such as proximity to known fault
structures or areas of recent seismicity could be used as a benecial constraint; Yin et al. (2018)
showed exactly that. They combined traditional waveform features with epidemic-type aftershock
sequence seismicity forecast in a Bayesian framework and showed that misidentication of non-
earthquake signals was greatly reduced, especially during periods of substantial activity such as
during swarms or mainshock-aftershock sequences. Suboptimal performance of EEW systems
during aftershock sequences is an important open problem. In the days following the M9 Tohoku-
oki earthquake, due to the rich aftershock sequence, 63% of the 70 warnings issued were false
warnings where the intensity was overestimated by two intensity levels (Liu & Yamada 2014).
This is due to events occurring close together in time being misidentied as a single event. Liu &
Yamada (2014) proposed a novel integrated particle lter (IPF) method to separate such events.
The IPF algorithm is a Bayesian estimator that uses information from both triggered and non-
triggered stations. Data from nontriggered stations, often referred to as not-yet-arrived data, are
critical for the success of the IPF algorithm. During the 2016 M7.2 Kumamoto earthquake, the
IPF algorithm showed good ofine performance, and it is being phased into production by JMA
(Kodera et al. 2016).
Other impacts of the M9 Tohoku-oki earthquake relate to the well-documented magnitude
saturation experienced by the point source algorithm (Hoshiba & Ozaki 2014), which estimated
the event at M8.1. Substantial effort has been invested in approaches that allow more prompt and
unsaturated magnitude estimation of large events. A core improvement revolves around develop-
ing better approaches for using the evolving features of a waveform rather than xed-length time
windows. Noda & Ellsworth (2016, 2017) studied events in Japan and suggested that by using
varying window lengths between 0.5 s and 4 s, convergence to nal magnitudes could be sub-
stantially accelerated for moderate magnitude events. However, those studies, echoing Hoshiba
& Ozaki (2014), found no improvement for the largest magnitude events.
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Peak ground
displacement (PGD):
maximum
displacement that
occurs at a site during
the passage of seismic
waves
High-rate global
positioning system
(HR-GPS): high-rate
geodetic time series
having sample rates of
one sample per second
or more, providing
information about
ground displacement
This contrasts with Colombelli et al. (2014) and Colombelli & Zollo (2015). They found more
positive results for the same Japanese data set of strong motion recordings. In their approach, in
addition to using progressively expanding P-wave windows, they averaged many waveforms over
many distances and azimuths together. They found that the growth of this average displacement
waveform event exhibited markedly different character for events as large as M9. For events in
the M8–8.5 range, such a method would yield useful source parameters 30–40 s after origin time,
although for the M9 Tohoku-oki event, this method still underestimated the earthquake at M8.4.
This averaging approach is tricky to apply in real time. As the P-wave time window expands,
some stations will not be useful after a certain point because S-wave energy will begin to leak into
the parameter estimation. Kodera (2018) proposed an interesting solution to this, showing that
through ground motion polarization analysis of borehole sites, it is possible to build a P-detector to
measure P-waves on the vertical channels even after the rst S-wave onsets. For large earthquakes
in Japan, these late-arriving P-waves can be used to track an evolving rupture and provide better
intensity estimates with more lead time. One potential limitation is that the P-detector method is
likely useful only for borehole sites since free-eld stations will be affected by soil response that
will degrade the performance.
Other approaches for obtaining timelier unsaturated magnitudes of large events focus on alter-
native features of ground motion recordings. Noda et al. (2016) showed that magnitude correlates
well to the time difference between the S-wave onset and the arrival of the peak high-frequency
amplitude in an accelerogram. Importantly, this time is shorter than the source duration. For the
M9 Tohoku-oki earthquake, such an approach would have produced an unsaturated magnitude
estimate 120 s after origin time. However, by using high-frequency data, the method is suscepti-
ble to complexities in the source such as the location of strong motion-generating areas (Asano &
Iwata 2012), which can produce anomalously long apparent durations.
Similarly, several researchers have studied the behavior of peak ground displacement (PGD)
as measured by a high-rate global positioning system (HR-GPS). Crowell et al. (2013) rst noted
that for the 2003 M8.3 Tokachi-oki and M9 Tohoku-oki earthquakes, PGD could be a reliable
magnitude estimator. Melgar et al. (2015) then expanded the data set to include a number of
M8+events in Chile and M7+events worldwide and proposed an algorithm for rapid PGD
magnitude estimation. PGD does not exhibit magnitude saturation, and it can be used to estimate
nal magnitudes, depending on network conguration, between one-half and two-thirds of the
way through the source process. For example, nal magnitudes were obtained after 60 s and 100 s
for the M8.8 Maule and M9.0 Tohoku-oki earthquakes. Crowell et al. (2016, 2018b) produced
operational prototypes of the PGD magnitude algorithm as part of the ShakeAlert EEW system
and for Chile as well.
3.2. Finite Fault Algorithms
One of the other open challenges in EEW is unsaturated magnitude and ground motion estima-
tion of large events. One potential solution is to explicitly quantify fault niteness in real time
from seismic and geodetic measurements. This is important because a particular location can be
far from the hypocenter but close to a strong motion-generating area of a fault, in which cases
point source algorithms, when converted to ground motion, will lead to large misestimations of
the actual hazard. For example, during the 2011 M9 Tohoku-oki earthquake, shaking in the Kanto
region around Tokyo, 350 km from the epicenter, was seriously underpredicted (Hoshiba & Ozaki
2014).
The FinDer algorithm (Böse et al. 2012) estimates fault rupture extent and strike by analyzing
the spatial distribution of ground motions in real time. At any given instant FinDer interpolates
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
G-larmS: earthquake
early warning
algorithm using
HR-GPS to estimate
the nite extent of the
fault rupture and the
magnitude
the observed maximum ground motion to produce an image of the maximum observed shaking.
FinDer simplies a potentially complex distribution of shaking by using ground motion thresholds
to divide sites into near and far from the rupture. This classication produces a binary image of
maximum-observed shaking at a given point in time. FinDer then compares this binary image
to previously computed templates of shaking distributions obtained from GMPEs and nds the
best-matching template. Each template has an associated fault length, and strike and magnitude
are then determined from fault length scaling relationships.
Unlike the point source algorithms, FinDer is not a predictive algorithm. It does not estimate
whether an event will grow; rather, it is a real-time ground motion assessment. Although FinDer
was initially conceived for large events, modications to the ground motion thresholds and other
parameters have made it suitable for small-magnitude events as well (Böse et al. 2015). With re-
gards to timeliness, if a network is dense enough, FinDer is surprisingly fast,with rst alerts often
only a few seconds behind point source algorithms (Böse et al. 2017). Indeed, recent demonstra-
tions of FinDer performance for large crustal events such as the M7 Kumamoto earthquake (Böse
et al. 2017) showcased the utility of the algorithm. Because it relies on template images of the
distribution of ground shaking, FinDer performs at its best when an earthquake occurs within
the footprint of the seismic network providing it with data. Böse et al. (2015) added asymmetric
or one-sided templates to allow the algorithm to deal with subduction zone earthquakes occur-
ring out of the network. Ofine testing during the M9.0 Tohoku-oki earthquake was encouraging.
FinDer estimated the event to be M8.5 and 270 km long 160 s after origin time. More rigorous
testing is necessary with more large subduction zone events that have one-sided distributions of
shaking.
A second class of nite fault algorithm relies on HR-GPS data. Automated static slip inver-
sion was rst demonstrated by Crowell et al. (2009), and research into it has been facilitated pre-
dominantly by two signicant events recorded in real time. First, the M7.2 El Mayor-Cucapah
earthquake in northern Mexico was recorded across a large network of GPS stations in Southern
California. From these data Allen & Ziv (2011) and Crowell et al. (2012) made the rst concrete
demonstrations and recommendations of what an operational GPS-enhanced warning would en-
tail. The second signicant event was, of course, the 2011 M9 Tohoku-oki earthquake, which was
recorded across more than 1,000 HR-GPS stations in Japan. From these data a number of workers
noted and demonstrated that had HR-GPS data been used for simplied inversion, then the mag-
nitude saturation problem would have been resolved (Ohta et al. 2012, Wright et al. 2012, Melgar
et al. 2013). Knowledge of fault niteness improves ground motion estimates by allowing one to
use more physically realistic estimates of site-to-fault distance. Indeed, for the GPS approaches,
Colombelli et al. (2013) showed that not only were magnitude estimates from simplied GPS slip
inversions reliable across a large range but also the ground motion estimates from such source
models were a substantial improvement over point source–driven calculations. Specically, in the
M9 Tohoku-oki case, the GPS models were available quickly enough to be useful for better ground
motion estimates in the Kanto region around Tokyo.
Because of these ndings, many algorithms have been proposed and are being tested. Notably,
in the United States, three algorithms are undergoing testing and being considered for ShakeAlert
(Murray et al. 2018). All of them rely on event notications or triggers from the seismic system.
This is preferred because GPS data can be noisy, and event detection on these waveforms can lead
to many false alerts. The rst of these three is the G-larmS algorithm (Grapenthin et al. 2014a,
2014b), which, following an event trigger, calculates static slip inversions on a series of predened
geometries that correspond to tectonic domains. For example, within Northern California, it
considers slip to be possible on San Andreas fault parallel and conjugate vertical strike slip
geometries as well as on blind thrust faults with strike and dip similar to the well-known Mt.
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
G-FAST: earthquake
early warning
algorithm using
HR-GPS to estimate
the seismic moment
tensor and then
inverting for slip on
the fault plane
BEFORES:
earthquake early
warning algorithm
using HR-GPS to
simultaneously
estimate the most
likely fault plane and
slip distribution
Propagation of local
undamped motion
(PLUM) algorithm:
algorithm using a
dense network to
measure the ground
motion and forward
predict likely motion
at surrounding sites
Diablo fault. Another algorithm, G-FAST (Crowell et al. 2016), rst computes a moment tensor
from the static offsets using the Melgar et al. (2012) method and then attempts the slip inversion
on the two nodal planes from the moment tensor and determines that the one that best ts the
data is the preferred solution. A third algorithm, BEFORES (Minson et al. 2014), simultaneously
estimates slip and the most likely fault geometry using a simplied Bayesian formulation.
Outside the United States, implementation of GPS nite faulting algorithms has begun in
Chile, where the G-FAST algorithm is undergoing testing (Crowell et al. 2018b). Similar efforts
are underway in Japan. Following the M9 Tohoku-oki earthquake, Ohta et al. (2012) proposed
an inversion algorithm that inverts for fault geometry and extent with homogenous slip. They
demonstrated convergence to M8.8 within ∼150 s of origin time. This detection and inversion
algorithm, called REGARD, has begun operational testing with some modications (Kawamoto
et al. 2016, 2017). REGARD consists of two simultaneous inversions. First, a nonlinear inversion
of a single rectangular fault loosely constrained to an a priori model of allowed fault orientations
is carried out. Concurrently, a linear inversion of a slip distribution model xed to the assumed
subducting plate boundary is performed. The algorithm has been successfully tested for several
crustal and subduction zone events.
Testing the performance of GPS algorithms under a variety of circumstances remains challeng-
ing. GPS is somewhat insensitive and typically can measure ground motions only for earthquakes
larger than M6. As a result, because only a few events occur within the footprint of a given net-
work, retrospective real-time replays of events elsewhere in the world are common. For example,
Crowell et al. (2018a) showed that for the very complex M7.8 Kaikoura earthquake, the simplied
G-FAST solution was useful for ground motion prediction.Ruhl et al. (2019b) have made available
a database of more than 3,000 HR-GPS recordings for 29 large events that can be used by algo-
rithm developers for benchmarks and testing. Indeed, in a follow-up study Ruhl et al.(2019a) used
these data and publicly available strong motion recordings to show that the G-larmS algorithm,
while slower to converge to the nal magnitude, is still fast enough to forecast strong motions.
This is because of the protracted nature of the source process and the ability of the algorithm to
estimate fault niteness. For large earthquakes that are expected to occur but for which no example
data are available, one possibility is to use scenarios and simulated data. For example,the Cascadia
subduction zone is known to produce events up to M9; however, because no instrumental data of
any signicant megathrust event are available, Melgar et al. (2016) proposed a simulation approach
that synthesizes HR-GPS data from kinematic rupture scenarios. Ruhl et al. (2017) used these data
to assess the performance of the G-larmS algorithm in the Cascadia subduction zone and, because
of that analysis, were able to propose a set of improvements and modications to the algorithm.
Overall, nite fault algorithms provide a more complete characterization of the source and do
not saturate, which results in better ground motion estimates. However, they need observations
at many sites and for longer times and thus will always be slower than point source solutions. As a
result, they are of limited use at short epicentral distances and can provide actionable information
only some distance away from the earthquake origin (Figure 1).
3.3. Ground Motion–Driven Approaches
A new generation of algorithms has emerged (Hoshiba 2013, Hoshiba & Aoki 2015, Kodera et al.
2018) that avoids estimation of source parameters. These algorithms use physics-driven data as-
similation techniques. The core idea is to use the present state of the ground shaking and knowl-
edge of propagation physics to forecast the likely evolution of intensity some short time (<20 s) in
the future. Specically, the propagation of local undamped motion (PLUM) algorithm (Hoshiba
& Aoki 2015) has undergone substantial renement and testing in Japan.
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
0 100 200 300
km
0 100 200 300
km
Seconds after origin time
0
100
200
300
400
500
600
kmJMA intensity in Kanto region
0 100 200
Kanto
Kanto
300
km
1
2
3
4
5
6
JMA
intensity
Observed Data assimilation
Prediction:
10 second lead time
Prediction:
20 second lead time
1
2
3
4
5
6
7
0 50 100
Observed
150 200
0 100 200 300
km
5 seconds
Lead time predictions
10 seconds
20 seconds
Figure 3
Retrospective performance of the PLUM algorithm during the M9 Tohoku-oki earthquake. Shown are the ground motion forecasts
available 10 s and 20 s before the onsets of shaking. Abbreviations: JMA, Japan Meteorological Agency; PLUM, propagation of local
undamped motion. Figure adapted with permission from Hoshiba & Aoki (2015).
In the PLUM method, rst a spatially dense image of the present distribution of intensity is
formed. Through data assimilation techniques, this image is forecast into the future using radiative
transfer theory.This is a high-frequency ray theoretical approximation that neglects details of the
waveeld and is, rather, an estimate of energy propagation. The approximation is chosen because
it is far faster than numerical approaches that propagate the full waveeld, so it is more suitable for
real-time applications. An example of this procedure for the M9 Tohoku-oki earthquake is shown
in Figure 3. After an image of the present waveeld is formed, it can be forecast 10 s and 20 s into
the future.
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
The main strength of PLUM is that it does not require any knowledge of the source; it simply
assumes that the physics is known to forecast the observed intensities a short time into the future.
Like FinDer, PLUM is not a predictive algorithm and cannot anticipate whether a rupture will
keep growing. However, by its very nature, it will capture fault niteness effects if they have been
observed in the distribution of ground shaking. Additionally, it can easily handle multiple events
in quick succession, such as during early aftershock sequences. It is possible to empirically add
site-effect and path-effect corrections, and thus, with short 10- to 20-s lead times, PLUM can
provide very accurate intensity forecasts.
Development of the algorithm was driven by the low ground motion forecasts in the Kanto re-
gion around Tokyo during the M9 Tohoku-oki earthquake and the many missed alerts in the after-
shock sequence. Retrospective testing of this event has shown the desired improved performance
(Hoshiba & Aoki 2015). Furthermore, during the more recent 2016 M7 Kumamoto earthquake,
PLUM performed well during the mainshock and also during the vigorous aftershock sequence
(Kodera et al. 2016). The algorithm is now in real-time operations in Japan (Kodera et al. 2018).
3.4. Combining Algorithms
Individual algorithms—point source, nite fault, ground motion driven, and others—all have dif-
ferent strengths and weaknesses. Thus, to achieve the best performance in all possible situations,
it is desirable to simultaneously operate several algorithms. The challenge is how to then syn-
thesize, or combine, information from all these disparate sources. Combining earthquake source
information from a point source and a nite fault algorithm is not desirable; rather, it is prefer-
able to combine ground motion forecasts. A simple solution proposed by Kodera et al. (2016) is
to take the largest intensity estimate from any given algorithm for a particular location. Indeed,
in the operational system in Japan, the maximum ground motion from the PLUM and the con-
ventional point source algorithm is taken at any point in time (Kodera et al. 2018). Minson et al.
(2017) proposed a more elegant solution. Using a Bayesian formulation, they designed a central
decision module that provides a single estimate of shaking by assigning likelihoods to the fore-
casts of different algorithms by comparing predicted and observed waveform envelopes. In this
framework the predictions from many algorithms, even when they have physically incompatible
source models such as point source or nite faulting, can be combined (e.g., Ruhl et al. 2019a).
3.5. Timeliness and Accuracy of Ground Motion Forecasts
There has been a recent push to systematically evaluate how well shaking intensities of certain
levels can be forecast and how large the prediction errors are as a function of time (e.g., Hsu et al.
2016, Meier 2017, Kodera et al. 2018, Minson et al. 2018). Critical to these analyses of ground
motion is the concept of an alert threshold, which signies a user-dened ground motion level at
which an action will be triggered. Different thresholds will produce different warning times.
Once a threshold is selected for any particular location, alerts are classied as true or false pos-
itives and true or false negatives if the alert threshold was correctly or incorrectly forecast to be
exceeded or not. Using this approach Meier (2017) analyzed several thousand waveforms employ-
ing a P-wave displacement point source method and a theoretical idealized nite fault algorithm.
Meanwhile, Minson et al. (2018) studied point source and nite fault algorithms from theoretical
considerations. Both studies concluded that for a high MMI threshold, fewer alerts—both false
and correct—will be issued, but the alerts that are issued can have a large proportion of false alarms
due to the inherent difculty in forecasting large intensity ground motions. Ruhl et al. (2019a) ex-
tended this approach and conducted a retrospective test of 32 6 <M<9 events worldwide. They
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
replayed seismic and geodetic data through both point source and nite fault algorithms and quan-
tied the accuracy of ground motion prediction. The study provided the rst systematic test of a
system on a wide array of real and varied geophysical data. It effectively showed that the addition
of the global navigation satellite system (GNSS) nite faulting approach substantially improves
the ground motion prediction ability of the system. As the point source algorithm saturates, the
GNSS algorithm begins to provide useful information for users who, while further aeld from
the hypocenter,are still close to the nite fault and experience substantial shaking. Similarly, a low
MMI threshold means that alerts are issued more often and the probability that any individual
alert is correct is much higher; however, cumulatively, more false alerts are issued.
Recognizing the inherent uncertainty in ground motion estimation, Kodera et al. (2018) pro-
posed that, rather than study the classication on a station-by-station basis, it is preferable to
study the accuracy of ground motion prediction over small, spatially averaged geographical areas.
In their analysis of events in Japan using the hybrid point source–PLUM method, Kodera et al.
(2018) found that PLUM achieves a high prediction score of 90%. This score is given as the pro-
portion of correctly classied true positives and negatives and false positives and negatives for an
alert threshold of 3.5 on the JMA intensity scale. The algorithm had warning times between 6 and
35 s.
These quantitative efforts at assessing algorithm performance are important because they are
an objective way to measure improvements (or degradation) of a system’s performance and to ad-
vocate for one algorithm over another.In the United States in particular,a testing and certication
platform based on these metrics, and others, is in the processof being standardized (Cochran et al.
2017). However, recall from Section 1 that the kinds of users of an EEW system are varied, and an
open problem is that no quantitative research exists on the tolerance of any kind of user to false or
missed alerts. So, while we can quantify the seismological performance of a system and set goals
for what is and is not acceptable from a system performance standpoint, we lack the information
to combine these ndings with quantitative assessments of user tolerance.
3.6. The Question of Determinism
The question of source determinism, which has been debated since modern EEW was proposed,
continues to be debated today. Determinism means when, within a potentially minutes-long rup-
ture process, a very large earthquake can be distinguished from a large one. For EEW it is an
important issue because it denes the minimum theoretical time at which the hazard (ground
shaking) can be characterized. One end member is strong determinism, where the nucleation
process is different between earthquakes of eventually different nal magnitudes (e.g., Ellsworth
& Beroza 1995). In this view a few seconds of observation can be used to identify events (e.g.,
Olson & Allen 2005). Colombelli et al. (2014) suggested that the evolution of displacements im-
aged by dense strong motion networks in Japan was strong evidence of this. However, Meier et al.
(2016), also from an analysis of strong motion data, found that nucleation is likely a universal pro-
cess and thus independent of nal magnitude. Hoshiba & Iwakiri (2011) analyzed strong motion
records from the rst 30 s of the M9 Tohoku-oki earthquake and found no difference between it
and events of smaller magnitude originating in the same region, echoing previous ndings from
Rydelek & Horiuchi (2006). Meier et al.(2017) further analyzed several large databases of teleseis-
mically determined STFs and contended that, at least in the rst third, there are no differences
in the STFs of large and very large events. From an analysis of the same STFs and near-eld
GPS, a more nuanced view was proposed by Melgar & Hayes (2017), who suggested that there is
weak determinism. In this view nucleation is a magnitude-independent process, but shortly there-
after (∼10 s), a self-similar slip pulse develops whose properties (rise time and pulse width) are
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
diagnostic of the eventual nal magnitude. Goldberg et al. (2018) analyzed many HR-GPS records
in Japan and reached similar conclusions.
4. FUTURE OUTLOOK
Throughout Section 3 we described the incremental improvements made in the last ∼10 years to
EEW systems. Despite these advances, the basic concepts of the EEW paradigm demonstrated by
the rst operational systems (e.g., Nakamura 1988, Espinosa-Aranda et al. 1995) regarding both
sensors and algorithms have remained largely unchanged. In this section we highlight research
that could impact, at its core, how EEW is realized.
EEW has relied strictly on onshore networks. However, in many subduction zone environ-
ments, nucleation of large, damaging events occurs predominantly offshore.Thus, there has been
a push for real-time telemetered ocean bottom networks. The largest effort to date is that of the
S-net network in Japan (Kanazawa et al. 2016), which consists of more than 800 km of ber-optic
cable covering the Japan trench with 150 nodes, or observatories, spaced roughly every 30 km.
Each node contains absolute-pressure, strong motion, broadband, and short-period sensors. Al-
though smaller in scope, similar real-time cables exist at the Cascadia subduction zone (Figure 4)
on both the Canadian (Barnes et al. 2011) and the US (Tréhu et al. 2018) portions of the system.
Such real-time cabled observatories have the potential to speed up detection of unfolding
events and thus warning; however, they come at a cost that is signicantly greater than their on-
shore counterparts’. Techniques to deploy and maintain infrastructure like this are still evolving.
The ocean bottom is a challenging environment. Both the cables and the observing nodes need
to be buried or made trawl resistant to protect them from shing activities. Additionally, there
are numerous noise sources such as bottom currents and internal waves that can make the data
difcult to use, particularly at long periods (Webb 1998). However, research has shown that it is
possible to correct and account for the noise (e.g., Bell et al. 2014), and thus it is becoming possi-
ble to reliably use real-time seaoor data. Offshore networks have intrinsic value not just for these
hazards applications but also for basic science exploration, so it is likely that they will continue
to expand in the coming decades and allow for better and more timely alerts, especially for the
largest events.
Onshore, it has been shown that a signicant portion of the delay between detection and char-
acterization of an event and issuance of the rst alert is directly related to the density of the sens-
ing network (e.g., Kuyuk & Allen 2013, Ruhl et al. 2017). Simply put, with a sparser network,
the blind zone is larger. Observatory-grade equipment is expensive and thus limits the growth of
most networks. One potential approach is to supplement traditional sensor networks with low-
cost accelerometers. Substantial progress has been made in MEMSs, and their noise characteristics
and sensitivity have been well studied. They are suitable for local monitoring (Evans et al. 2014,
Saunders et al. 2016). MEMS accelerometers are a fraction of the cost of traditional inertial seis-
mometers and thus could be deployed to augment a network’s density. Chung et al. (2015) showed
that the sensors can be used by currently existing point source algorithms, and Clayton et al. (2015)
deployed more than 500 MEMS devices in the Los Angeles region. In Taiwan 543 MEMS sensors
supplement the traditional network (Chen et al. 2015).
Despite these improvements, large MEMS deployments still face some of the same challenges
as traditional seismic stations. Permits and permissions need to be secured to deploy devices, and
telemetry paths need to be built. Thus, while MEMSs are cheap, their maintenance might not be
any less costly than that of a traditional network. Perhaps the most exciting development in low-
cost sensing comes from the potential to use mobile phones as seismometers. Most smartphones
today carry onboard a three-component MEMS accelerometer of sufcient quality to detect
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Newport
Portl and
Gigapop
Midplate
Axial
seamount
Slope
base
Blanco
Fracture Zone
Seattle
Gigapop
Ocean Networks
Canada
Pacic City
Vancouver
Vancouver
Seattle
Seattle
Portland
Portland
Depth (m)
3,500
0
Main sensor node
Redding
Redding
Eureka
Eureka
Crescent City
Crescent City
Coos Bay
Coos Bay
Branch sensor node
Substation
Shore station SHR
Juan de Fuca
Plate
Pacic
Plate
Moorings
Potential
expansion
sites
Figure 4
(To p ) Current state of seaoor networks in the Cascadia subduction zone. (Bottom) Conceptual sketch of a potential expansion to the
seaoor infrastructure. Abbreviation: SHR, Southern Hydrate Ridge. Figure adapted with permission from original by Will Wilcock
and Hunter Hadaway, University of Washington.
moderate and large events at local to regional distances (Kong et al. 2016b). Indeed, if the mobile
phone is at rest, the onboard sensors are good enough to observe P-waves and thus are suitable for
EEW. Kong et al. (2016a,b) demonstrated the ability to collect and analyze data from the phones
in real time with minimal impact to the user’s phone in terms of battery consumption or pro-
cessor use. Additionally, Kong et al. (2016b) demonstrated with an articial neural network that
it is possible to reliably separate human signals from earthquake signals recorded on the phone.
Kong et al. (2016a) further demonstrated how the data can be used for EEW. In related work,
Minson et al. (2015) showed through simulations of large events that the GPS chips in modern
smartphones could also be useful for warning. Because there are more than 2 billion smartphone
users around the world, the ability to harness them for warning is an exciting prospect.
By and large, however, EEW systems continue to rely on traditional inertial seismometers.
There is potential for contributions from other geophysical sensors. Barbour & Crowell (2017)
showed that strainmeters, which measure elastodynamic strain over a very broad frequency range,
could be reliably used for rapid source characterization. While they conducted their analysis on
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
onshore instruments, their ndings are interesting because it is possible that strain will be one
of the preferred geodetic measurements in future offshore networks. Another interesting result
relating to geodesy was described by Montagner et al. (2016). They found that a gravity signal
preceding the P-wave was detected by the Kamioka superconducting gravimeter during the M9
Tohoku-oki earthquake. During the rupture process, slip on a fault redistributes mass within the
crust, as do the radiated elastic waves, and thus a gravity perturbation is expected.Montagner et al.
(2016) and Vallée et al. (2017) produced a theoretical formulation for such transient elastogravity
signals. For the M9 Tohoku-oki earthquake, Vallée et al.(2017) showed that the postulated gravity
perturbations preceding the P-wave arrivals could indeed be observed on broadband seismometers
(Figure 5). These ndings are encouraging because gravity perturbations propagate at the speed
of light, so a system that uses such measurements operationally would be substantially faster by
eliminating the travel-time delay from P-waves traveling from source to site. However, even for
the M9 Tohoku-oki earthquake, the perturbations were quite small (∼1 nm/s2), and thus, only
the largest of events will generate signals that could be measured with the current generation of
broadband sensors.
−100 0 100 200 300
427
587
Seconds after origin time
Distance of
gravity observation
from epicenter (km)
673
1,284
1,390
1,398
1,897
2,276
2,438
3,041
3,044
Tohoku observation
Tohoku simulation
M8.5 simulation
1 nm/s 2
Figure 5
Agreement between observed and simulated elastogravity acceleration signals for the M9 Tohoku-oki
earthquake. The simulation for a ctitious M8.5 earthquake shows large amplitude differences, directly
illustrating the magnitude determination potential existing in these signals. The waveforms are truncated
before P-wave arrivals. Figure adapted with permission from Vallée et al. (2017).
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Machine learning
(ML) algorithm:
algorithm designed to
recognize patterns in
data to classify them or
predict future actions
or observations
without a physical
model
Finally, another broad and exciting avenue of research for future EEW systems comes from
modern data science and specically from machine learning (ML). Modern seismic equipment
will sense other environmental signals, both anthropogenic and otherwise. From time to time,
one of these nonearthquake signals will trigger an individual station. If there are many stations,
especially in noisy urban environments, the potential for several of these false triggers to occur
close to each other in space and time and be associated into a false alarm is non-negligible. One
solution is to introduce new algorithms for signal discrimination. Perol et al. (2018) demonstrated
that a convolutional neural network (CNN) algorithm, often referred to as deep learning, could
be used for detection and location. Ross et al. (2018) used the CNN algorithm to identify P-waves,
S-waves, and nonevent signals in continuous data. The performance is far better than with tradi-
tional triggering and detection algorithms and, if applied to real-time EEW networks, will sub-
stantially reduce false alerts. It is easy to envision other as-yet-to-be-explored ML applications.
ML algorithms could be trained to recognize magnitude, location, and other source features di-
rectly from the waveforms themselves, forgoing the need for a magnitude regression. Additionally,
as in the PLUM method, images of ground shaking can be formed and used to train ML algo-
rithms in a predictive sense such that when the next earthquake occurs, an ML algorithm can
forecast the most likely evolution of the present waveeld some time into the future.
5. CONCLUSIONS
Earthquakes can be major catastrophic events and disrupt people’s lives in very signicant, sudden,
and uncontrollable ways. EEW is one relatively new tool to help reduce an earthquake’s impact
and provide people with an ability to take back some control by reacting to the warning. Public
interest in EEW has driven the rapid development of methodologies over the last decade, which
is in turn driving a deeper understanding of earthquake physics as we develop improved models
of the earthquake process to better predict ground shaking. At the same time, EEW leads to the
deployment of new geophysical networks providing the data for EEW and further research and
development. This process can be observed in the countries that currently deliver public alerts
(Mexico, Japan, South Korea, and Taiwan), the countries delivering limited alerts (India, Romania,
Turkey, and the United States), and the many additional regions developing systems.
The algorithms in use fall into four categories. Point source algorithms are in use in all regional
systems and provide earthquake information in a few seconds. In larger earthquakes (M>7), there
is time for nite source algorithms to provide unsaturated magnitude estimates and describe the
fault geometry, both of which lead to more accurate alerts. GMMs provide the most accurate
estimate of forthcoming shaking by forward predicting observed shaking, but they are limited to
a few seconds of warning. Finally, onsite approaches are the simplest in that they use a sensor at
one location to warn the same location, but they have limited accuracy and warning time.
EEW systems several decades from now will evolve out of what exists today but will likely be
substantially different. Systems will be amphibious and cross shorelines. More diverse geophysical
observations will exist, including not just HR-GPS but also perhaps strainmeters and pressure or
acoustic sensors on the seaoor. Observatory-grade sensors will form a backbone network sup-
plemented by MEMS accelerometers in key areas, perhaps surrounding major fault lines, and by
millions of mobile phones in urban environments. ML algorithms will more promptly discrimi-
nate between myriad environmental signals and correctly identify the onsets of signicant events
from all these disparate data. Other ML algorithms will classify, from these onset signals, the
characteristics of the event and combine that with information of the present state of the wave-
eld to make forecasts of the likely intensities of shaking some time into the future. Alerts will be
tailored according to the level of shaking a geographic region is likely to experience and will be
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
disseminated through multiple communication channels to millions of people with minimal lag
and with simple, easy-to-digest messages. Automated systems will act to secure critical infras-
tructure, lifelines, and transportation systems, substantially reducing the risk of post-earthquake
hazards such as re. In such a future the benets of EEW will extend beyond the immediate af-
termath of a signicant earthquake. A modern EEW system will increase resiliency and allow a
society to bounce back to its pre-event condition far faster than it would have otherwise.
DISCLOSURE STATEMENT
R.M.A. has received research funding from the Gordon and Betty Moore Foundation and state and
federal agencies for EEW, holds a patent on the use of smartphones for EEW, and has advocated
for the implementation of EEW around the world. D.M. has received research funding from state
and federal agencies for EEW and advocated for implementation.
ACKNOWLEDGMENTS
Much assistance was provided from colleagues around the world in compiling this review. In par-
ticular we thank John Clinton, Brendan Crowell, Mitsuyuki Hoshiba, Qingkai Kong, Ashok Ku-
mar, Young-Woo Kwon, Yih-Min Wu, and Masumi Yamada. Original versions of gures were
provided by Martin Vallée, Will Wilcock, and Hunter Hadaway. Funding was provided for R.M.A.
by the Gordon and Betty Moore Foundation through grants GBMF3024 and GBMF5230 to UC
Berkeley and by US Geological Survey grant G17AC00346.
LITERATURE CITED
Allen RM. 2011. Earthquakes, early and strong motion warning. In Encyclopedia of Solid Earth Geophysics,ed.
HK Gupta, pp. 226–33. Boston: Springer
Allen RM, Cochran ES, Huggins T, Miles S, Otegui D. 2017. Quake warnings, seismic culture. Science
358:1111
Allen RM, Cochran ES, Huggins T, Miles S, Otegui D. 2018. Lessons from Mexico’s earthquake early warning
system. Eos Trans. AGU 99
Allen RM, Gasparini P, Kamigaichi O, Böse M. 2009. The status of earthquake early warning around the
world: an introductory overview. Seismol. Res. Lett. 80(5):682–93
Allen RM, Kanamori H. 2003. The potential for earthquake early warning in southern California. Science
300:786–89
Allen RM, Ziv A. 2011. Application of real-time GPS to earthquake early warning. Geophys. Res. Lett.
38(16):L16310
Asano K, Iwata T. 2012. Source model for strong ground motion generation in the frequency range 0.1–10 Hz
during the 2011 Tohoku earthquake. Earth Planets Space 64:6
Atik LA, Abrahamson N, Bommer JJ, Scherbaum F, Cotton F, Kuehn N. 2010. The variability of ground-
motion prediction models and its components. Seismol. Res. Lett. 81(5):794–801
Barbour AJ, Crowell BW. 2017. Dynamic strains for earthquake source characterization. Seismol. Res. Lett.
88(2A):354–70
Barnes CR, Best MM, Pautet L, Pirenne B. 2011. Understanding Earth–ocean processes using real-time data
from NEPTUNE, Canada’s widely distributed sensor networks, northeast Pacic. Geosci. Can. 38(1):21–
30
Bell SW, Forsyth DW, Ruan Y. 2014. Removing noise from the vertical component records of ocean-bottom
seismometers: results from year one of the Cascadia Initiative. Bull. Seismol. Soc. Am. 105(1):300–13
Böse M, Felizardo C, Heaton TH. 2015. Finite-fault rupture detector (FinDer): going real-time in Californian
ShakeAlert warning system. Seismol. Res. Lett. 86(6):1692–704
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Böse M, Hauksson E, Solanki K, Kanamori H, Heaton T. 2009. A new trigger criterion for improved real-time
performance of onsite earthquake early warning in southern California. Bull. Seismol. Soc. Am. 99:897–905
Böse M, Heaton TH, Hauksson E. 2012. Real-time nite fault rupture detector (FinDer) for large earthquakes.
Geophys. J. Int. 191(2):803–12
Böse M, Smith DE, Felizardo C, Meier MA, Heaton TH, Clinton JF. 2017. FinDer v. 2: improved real-
time ground-motion predictions for M2–M9 with seismic nite-source characterization. Geophys. J. Int.
212(1):725–42
Cauzzi C, Behr Y, Guenan TL, Douglas J, Auclair S, et al. 2016. Earthquake early warning and operational
earthquake forecasting as real-time hazard information to mitigate seismic risk at nuclear facilities. Bull.
Earthq. Eng. 14:2495–512
Chamoli BP, Kumar A, Chen D-Y, Gairola A, Jakka RS, et al. 2019. A prototype earthquake early warning
system for northern India. J. Earthq. Eng. In press
Chen DY, Wu YM, Chin TL. 2015. Incorporating low-cost seismometers into the Central Weather Bureau
seismic network for earthquake early warning in Taiwan. Terr. Atmos. Ocean. Sci. 26(5):503–13
Chung AI, Cochran ES, Kaiser AE, Christensen CM, Yildirim B, Lawrence JF. 2015. Improved rapid magni-
tude estimation for a community-based, low-cost MEMS accelerometer network. Bull. Seismol. Soc. Am.
105(3):1314–23
Chung AI, Henson I, Allen RM.2019. Optimizing earthquake early warning performance: ElarmS-3. Seismol.
Res. Lett. 90(2A):727–43
Clayton RW, Heaton T, Kohler M, Chandy M, Guy R, Bunn J. 2015. Community seismic network: a dense
array to sense earthquake strong motion. Seismol. Res. Lett. 86(5):1354–63
Clinton J, Zollo A, M ˘
armureanu A, Zulkar C, Parolai S. 2016. State-of-the art and future of earthquake early
warning in the European region. Bull. Earthq. Eng. 14(9):2441–58
Cochran ES, Kohler MD, Given DD, Guiwits S, Andrews J, et al. 2017. Earthquake early warning ShakeAlert
system: testing and certication platform. Seismol. Res. Lett. 89(1):108–17
Colombelli S, Allen RM, Zollo A.2013. Application of real-time GPS to earthquake early warning in subduc-
tion and strike-slip environments. J. Geophys. Res. Solid Earth 118(7):3448–61
Colombelli S, Zollo A. 2015. Fast determination of earthquake magnitude and fault extent from real-time
P-wave recordings. Geophys. J. Int. 202(2):1158–63
Colombelli S, Zollo A, Festa G,Picozzi M. 2014. Evidence for a difference in rupture initiation between small
and large earthquakes. Nat. Commun. 5:3958
Cooper JD. 1868. Earthquake indicator. San Franc. Bull. San Franc. Publ. Co., San Francisco, CA
Crowell BW, Bock Y, Melgar D. 2012. Real-time inversion of GPS data for nite fault modeling and rapid
hazard assessment. Geophys. Res. Lett. 39(9):L09305
Crowell BW, Bock Y, Squibb MB. 2009. Demonstration of earthquake early warning using total displacement
waveforms from real-time GPS networks. Seismol. Res. Lett. 80(5):772–82
Crowell BW, Melgar D, Bock Y, Haase JS, Geng J. 2013. Earthquake magnitude scaling using seismogeodetic
data. Geophys. Res. Lett. 40(23):6089–94
Crowell BW, Melgar D, Geng J. 2018a. Hypothetical real-time GNSS modeling of the 2016 Mw7.8 Kaik ¯
oura
earthquake: perspectives from ground motion and tsunami inundation prediction. Bull. Seismol. Soc. Am.
108:1736–45
Crowell BW, Schmidt DA, Bodin P, Vidale JE, Baker B, et al. 2018b. G-FAST earthquake early warning
potential for great earthquakes in Chile. Seismol. Res. Lett. 89(2A):542–56
Crowell BW, Schmidt DA, Bodin P, Vidale JE, Gomberg J, et al. 2016. Demonstration of the Cascadia
G-FAST geodetic earthquake early warning system for the Nisqually, Washington, earthquake. Seismol.
Res. Lett. 87(4):930–43
Cua G, Heaton T. 2007. The virtual seismologist (VS) method: a Bayesian approach to earthquake early warn-
ing. In Earthquake Early Warning Systems, ed. P Gasparini, G Manfredi, J Zschau, pp. 85–132. Berlin:
Springer
Cuéllar A, Espinosa-Aranda JM, Suarez R, Ibarrola G, Uribe A, et al. 2014. The Mexican seismic alert sys-
tem (SASMEX): its alert signals, broadcast results and performance during the M 7.4 Punta Maldonado
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
earthquake of March 20th, 2012. In Early Warning for Geological Disasters, ed. F Wenzel, Z Zschau,
pp. 71–87. Berlin: Springer-Verlag
Cuéllar A, Suarez G, Espinosa-Aranda J. 2018. A fast earthquake early warning algorithm based on the rst
3softheP-wavecoda.Bull. Seismol. Soc. Am. 108:2068–79
Ellsworth WL, Beroza GC. 1995. Seismic evidence for an earthquake nucleation phase. Science 268(5212):851–
55
Espinosa-Aranda JM, Jimenez A, Ibarrola G, Alcantar F, Aguilar A, et al. 1995. Mexico City seismic alert
system. Seismol. Res. Lett. 66:42–52
Evans JR, Allen RM, Chung AI, Cochran ES, Guy R, et al. 2014. Performance of several low-cost accelerom-
eters. Seismol. Res. Lett. 85(1):147–58
Given DD, Cochran ES, Heaton T, Hauksson E, Allen RM, et al. 2014. Technical implementation plan for the
ShakeAlert production system—an earthquake early warning system for the West Coast of the United States.
Open-File Rep. 2014-1097, US Geol. Surv., Reston, VA. https://doi.org/10.3133/ofr20141097
Goldberg DE, Melgar D, Bock Y, Allen RM. 2018. Geodetic observations of weak determinism in rupture
evolution of large earthquakes. J. Geophys. Res. Solid Earth 123:9950–62
Grapenthin R, Johanson IA, Allen RM. 2014a. Operational real-time GPS-enhanced earthquake early warn-
ing. J. Geophys. Res. Solid Earth 119(10):7944–65
Grapenthin R, Johanson IA, Allen RM. 2014b. The 2014 Mw6.0 Napa earthquake, California: observations
from real-time GPS-enhanced earthquake early warning. Geophys. Res. Lett. 41(23):8269–76
Hoshiba M. 2013. Real-time prediction of ground motion by Kirchhoff Fresnel boundary integral equa-
tion method: extended front detection method for earthquake early warning. J. Geophys. Res. Solid Earth
118:1038–50
Hoshiba M. 2014. Review of the nationwide earthquake early warning in Japan during its rst ve years. In
Earthquake Hazard, Risk, and Disasters, ed. JF Shroder, M Wyss, pp. 505–29. Waltham, MA: Academic
Hoshiba M, Aoki S. 2015. Numerical shake prediction for earthquake early warning: data assimilation, real-
time shake mapping, and simulation of wave propagation. Bull. Seismol. Soc. Am. 105:1324–38
Hoshiba M, Iwakiri K. 2011. Initial 30 seconds of the 2011 off the Pacic coast of Tohoku earthquake (Mw
9.0)—amplitude and τcfor magnitude estimation for earthquake early warning. Earth Planets Space
63:553–57
Hoshiba M, Ozaki T. 2014. Earthquake early warning and tsunami warning of the Japan Meteorological
Agency, and their performance in the 2011 off the Pacic coast of Tohoku earthquake (M9.0). In Early
Warning for Geological Disasters, ed. F Wenzel, J Zschau, pp. 1–28. Berlin: Springer-Verlag
Hsu TY, Lin PY, Wang HH, Chiang HW, Chang YW, et al. 2018. Comparing the performance of the
NEEWS earthquake early warning system against the CWB system during the 6 February 2018 Mw
6.2 Hualien earthquake. Geophys. Res. Lett. 45:6001–7
Hsu TY, Wang HH, Lin PY, Lin CM, Kuo CH, Wen KL. 2016. Performance of the NCREE’s on-site
warning system during the 5 February 2016 Mw6.53 Meinong earthquake. Geophys. Res. Lett. 43:8954–
59
Johnson L, Rabinovici S, Kang G, Mahin SA. 2016. California earthquake early warning system benet study.
CSSC Publ.16-04 PEER Rep. 2016/06, Pac. Earthq. Eng. Res. Cent., Berkeley, CA
Kanazawa T, Uehira K, Mochizuki M, Shinbo T, Fujimoto H, et al. 2016. S-NET project, cabled observation
network for earthquakes and tsunamis.Paper presented at the 9th Conference in the SubOptic Series, Dubai,
Apr. 18–21
Kawamoto S, Hiyama Y, Ohta Y, Nishimura T. 2016. First result from the GEONET real-time analysis
system (REGARD): the case of the 2016 Kumamoto earthquakes. Earth Planets Space 68(1):190
Kawamoto S, Ohta Y, Hiyama Y, Todoriki M, Nishimura T, et al. 2017. REGARD: a new GNSS-based real-
time nite fault modeling system for GEONET. J. Geophys. Res. Solid Earth 122(2):1324–49
Kodera Y. 2018. Real-time detection of rupture development: earthquake early warning using Pwaves from
growing ruptures. Geophys. Res. Lett. 45(1):156–65
Kodera Y, Saitou J, Hayashimoto N, Adachi S, Morimoto M, et al. 2016. Earthquake early warning for the 2016
Kumamoto earthquake: performance evaluation of the current system and the next-generation methods
of the Japan Meteorological Agency. Earth Planets Space 68(1):202
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Kodera Y, Yamada Y, Hirano K, Tamaribuchi K, Adachi S, et al. 2018. The propagation of local undamped
motion (PLUM) method: a simple and robust seismic waveeld estimation approach for earthquake early
warning. Bull. Seismol. Soc. Am. 108(2):983–1003
Kohler M, Cochran E, Given D, Guiwits S, Neuhauser D, et al. 2017. Earthquake early warning ShakeAlert
system: West Coast wide production prototype.Seismol. Res. Lett. 89(1):99–107
Kong Q, Allen RM, Schreier L. 2016a. MyShake: initial observations from a global smartphone seismic net-
work. Geophys. Res. Lett. 43(18):9588–94
Kong Q, Allen RM, Schreier L, Kwon YW. 2016b. MyShake: a smartphone seismic network for earthquake
early warning and beyond. Sci. Adv. 2(2):e1501055
Kuyuk HS, Allen RM. 2013.Optimal seismic network density for earthquake early warning: a case study from
California. Seismol. Res. Lett. 84(6):946–54
Kuyuk HS, Allen RM, Brown H, Hellweg M, Henson I, Neuhauser D. 2014. Designing a network-based
earthquake early warning algorithm for California: ElarmS-2. Bull. Seismol. Soc. Am. 104:162–73
Li S. 2018. Approaching earthquake early-warning. Overv. Disaster Prev. 2:14–24
Liu A, Yamada M. 2014. Bayesian approach for identication of multiple events in an early warning system.
Bull. Seismol. Soc. Am. 104(3):1111–21
Lu C, Zhou L, Zhang Z. 2016. Research and test on China high-speed railway earthquake early-warning
system. Sci. Technol. Rev. 34(18):258–64
M˘
armureanu A, Ionescu C, Cioan CO. 2010. Advanced real-time acquisition of the Vrancea earthquake early
warning system. Soil Dyn. Earthq. Eng. 31:163–69
Meier MA. 2017. How “good” are real-time ground motion predictions from earthquake early warning sys-
tems? J. Geophys. Res. Solid Earth 122(7):5561–77
Meier MA, Ampuero JP, Heaton TH. 2017. The hidden simplicity of subduction megathrust earthquakes.
Science 357(6357):1277–81
Meier MA, Heaton T, Clinton J. 2015. The Gutenberg algorithm: evolutionary Bayesian magnitude estimates
for earthquake early warning with a lter bank. Bull. Seismol. Soc. Am. 105(5):2774–86
Meier MA, Heaton T, Clinton J. 2016. Evidence for universal earthquake rupture initiation behavior. Geophys.
Res. Lett. 43(15):7991–96
Melgar D, Bock Y, Crowell BW. 2012. Real-time centroid moment tensor determination for large earthquakes
from local and regional displacement records. Geophys. J. Int. 188(2):703–18
Melgar D, Crowell BW, Bock Y, Haase JS. 2013. Rapid modeling of the 2011 Mw 9.0 Tohoku-Oki earthquake
with seismogeodesy. Geophys. Res. Lett. 40(12):2963–68
Melgar D, Crowell BW, Geng J, Allen RM, Bock Y, et al. 2015. Earthquake magnitude calculation without
saturation from the scaling of peak ground displacement. Geophys. Res. Lett. 42(13):5197–205
Melgar D, Hayes GP. 2017. Systematic observations of the slip pulse properties of large earthquake ruptures.
Geophys. Res. Lett. 44(19):9691–98
Melgar D, LeVeque RJ, Dreger DS, Allen RM. 2016. Kinematic rupture scenarios and synthetic displacement
data: an example application to the Cascadia subduction zone. J. Geophys. Res. Solid Earth 121(9):6658–
74
Minson SE, Brooks BA, Glennie CL, Murray JR, Langbein JO, et al. 2015. Crowdsourced earthquake early
warning. Sci. Adv. 1(3):e1500036
Minson SE, Meier MA, Baltay AS, Hanks TC, Cochran ES. 2018. The limits of earthquake early warning:
timeliness of ground motion estimates. Sci. Adv. 4(3):eaaq0504
Minson SE, Murray JR, Langbein JO, Gomberg JS. 2014. Real-time inversions for nite fault slip models and
rupture geometry based on high-rate GPS data. J. Geophys. Res. Solid Earth 119(4):3201–31
Minson SE, Wu S, Beck JL, Heaton TH. 2017. Combining multiple earthquake models in real time for earth-
quake early warning. Bull. Seismol. Soc. Am. 107(4):1868–82
Montagner JP, Juhel K, Barsuglia M, Ampuero JP, Chassande-Mottin E, et al. 2016. Prompt gravity signal
induced by the 2011 Tohoku-Oki earthquake. Nat. Commun. 7:13349
Murray JR, Crowell BW, Grapenthin R, Hodgkinson K, Langbein JO, et al. 2018. Development of a geodetic
component for the US West Coast earthquake early warning system. Seismol. Res. Lett. 89:2322–36
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Nakamura Y. 1988. On the urgent earthquake detection and alarm system (UrEDAS). In Proceedings of the 9th
World Conference on Earthquake Engineering, Vol. 7, pp. 673–78. Tokyo-Kyoto, Japan: Jpn. Assoc. Earthq.
Disaster Prev.
Nakamura Y, Tucker B. 1988. Japan’s earthquake early warning system: Should it be imported to California?
Calif. Geol. 41:33–40
Noda S, Ellsworth WL. 2016.Scaling relation between earthquake magnitude and the departure time from P
wave similar growth. Geophys. Res. Lett. 43(17):9053–60
Noda S, Ellsworth WL. 2017. Determination of earthquake magnitude for early warning from the time de-
pendence of P-wave amplitudes. Bull. Seismol. Soc. Am. 107(4):1860–67
Noda S, Yamamoto S, Ellsworth WL. 2016. Rapid estimation of earthquake magnitude from the arrival time
of the peak high-frequency amplitude. Bull. Seismol. Soc. Am. 106(1):232–41
Nof RN, Allen RM. 2016. Implementing the ElarmS earthquake early warning algorithm on the Israeli Seis-
mic Network. Bull. Seismol. Soc. Am. 106:2332–44
Ohta Y, Kobayashi T, Tsushima H, Miura S, Hino R, et al. 2012. Quasi real-time fault model estimation
for near-eld tsunami forecasting based on RTK-GPS analysis: application to the 2011 Tohoku-Oki
earthquake (Mw9.0). J. Geophys. Res. Solid Earth 117(B2):B02311
Olson EL, Allen RM. 2005. The deterministic nature of earthquake rupture. Nature 438(7065):212–15
Perol T, Gharbi M, Denolle M. 2018. Convolutional neural network for earthquake detection and location.
Sci. Adv. 4(2):e1700578
Porter K, Shoaf K, Seligson H. 2006. Value of injuries in the Northridge earthquake. Earthq. Spectra 22:555–
63
Ross ZE, Meier MA, Hauksson E, Heaton TH. 2018. Generalized seismic phase detection with deep learning.
Bull. Seismol. Soc. Am. 108:2894–901
Ruhl CJ, Melgar D, Chung AI, Grapenthin R, Allen RM. 2019a. Quantifying the value of real-time geodetic
constraints on earthquake early warning using a global seismic and geodetic dataset. arXiv:1901.11124
[physics.geo-ph]
Ruhl CJ, Melgar D, Geng J, Goldberg DE, Crowell BW, et al. 2019b. A global database of strong-motion
displacement GNSS recordings and an example application to PGD scaling. Seismol. Res. Lett. 90(1):271–
79
Ruhl CJ, Melgar D, Grapenthin R, Allen RM. 2017. The value of real-time GNSS to earthquake early warning.
Geophys. Res. Lett. 44(16):8311–19
Rydelek P, Horiuchi S. 2006. Earth science: Is earthquake rupture deterministic? Nature 442(7100):E5
Satriano C, Elia L, Martino C, Lancieri M, Zollo A, Iannaccone G. 2010. PRESTo, the earthquake early
warning system for southern Italy: concepts, capabilities and future perspectives. Soil Dyn. Earthq. Eng.
31:137–53
Saunders JK, Goldberg DE, Haase JS, Bock Y, Ofeld DG, et al. 2016. Seismogeodesy using GPS and low-
cost MEMS accelerometers: perspectives for earthquake early warning and rapid response. Bull. Seismol.
Soc. Am. 106:2469–89
Seki T, Okada T, Ikeda M, Sugano T. 2008. Early warning “area mail.” NTT Tech. Rev. 6(12):1–6
Sheen D-H, Park J-H, Chi H-C, Hwang E-H, Lim I-S, et al. 2017. The rst stage of an earthquake early
warning system in South Korea. Seismol. Res. Lett. 88(6):1491–98
Shoaf KI, Nguyen LH, Sareen HR, Bourque LB. 1998. Injuries as a result of California earthquakes in the
past decade. Disasters 22:218–35
Strauss JA, Allen RM. 2016. Benets and costs of earthquake early warning, Seismol. Res. Lett. 87(3):765–
72
Tréhu AM, Wilcock WS, Hilmo R, Bodin P, Connolly J, et al. 2018. The role of the Ocean Observatories
Initiative in monitoring the offshore earthquake activity of the Cascadia subduction zone. Oceanography
31(1):104–13
Vallée M, Ampuero JP, Juhel K, Bernard P, Montagner JP, Barsuglia M. 2017. Observations and modeling of
the elastogravity signals preceding direct seismic waves. Science 358(6367):1164–68
Webb SC. 1998. Broadband seismology and noise under the ocean. Rev. Geophys. 36(1):105–42
www.annualreviews.org •Earthquake Early Warning
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47CH15_Allen ARjats.cls May 13, 2019 14:21
Wood MM, Mileti DS, Kano M, Kelley MM, Regan R, Bourque LB. 2012. Communicating actionable risk
for terrorism and other hazards. Risk Anal. 32(4):601–15
Worden CB, Wald DJ, Allen TI, Lin K, Garcia D, Cua G. 2010. A revised ground-motion and intensity
interpolation scheme for ShakeMap. Bull. Seismol. Soc. Am. 100(6):3083–96
Wright TJ, Houlié N, Hildyard M, Iwabuchi T. 2012. Real-time, reliable magnitudes for large earthquakes
from 1 Hz GPS precise point positioning: the 2011 Tohoku-Oki ( Japan) earthquake. Geophys. Res. Lett.
39(12):L12302
Wu Y-M, Hsiao N-C, Chin T-L, Chen D-Y, Chan Y-T, Wang K-S. 2014. Earthquake early warning system in
Taiwan. In Encyclopedia of Earthquake Engineering, ed. M Beer, IA Kougioumtzoglou, E Patelli, S-K Au.
Berlin: Springer. https://doi.org/10.1007/978-3-642-36197-5_99-1
Wu Y-M, Mittal H, Huang T-C, Yang BM, Jan J-C, Chen SK. 2018. Performance of a low-cost earthquake
early warning system (P-alert) and shake map production during the 2018 Mw6.4 Hualien, Taiwan,
earthquake. Seismol. Res. Lett. 90(1):19–29
Yin L,Andrews J, Heaton T. 2018. Rapid earthquake discrimination for earthquake early warning: a Bayesian
probabilistic approach using three-component single-station waveforms and seismicity forecast. Bull.
Seismol. Soc. Am. 108:2054–67
Allen •Melgar
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47_FrontMatter ARI 26 April 2019 11:24
Annual Review
of Earth and
Planetary Sciences
Volume 47, 2019 Contents
Big Time
Paul F. Hoffman ppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp1
Unanticipated Uses of the Global Positioning System
Kristine M. Larson pppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp19
Dynamics in the Uppermost Lower Mantle: Insights into the Deep
Mantle Water Cycle Based on the Numerical Modeling of
Subducted Slabs and Global-Scale Mantle Dynamics
Takashi Nakagawa and Tomoeki Nakakuki ppppppppppppppppppppppppppppppppppppppppppppppppp41
Atmospheric Escape and the Evolution of Close-In Exoplanets
James E. Owen pppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp67
The Sedimentary Cycle on Early Mars
Scott M. McLennan, John P. Grotzinger, Joel A. Hurowitz, and Nicholas J. Tosca pppppp91
New Horizons Observations of the Atmosphere of Pluto
G. Randall Gladstone and Leslie A. Young pppppppppppppppppppppppppppppppppppppppppppppppp119
The Compositional Diversity of Low-Mass Exoplanets
Daniel Jontof-Hutter pppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp141
Destruction of the North China Craton in the Mesozoic
Fu-Yuan Wu, Jin-Hui Yang, Yi-Gang Xu, Simon A. Wilde,
and Richard J. Walker ppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp173
Seawater Chemistry Through Phanerozoic Time
Alexandra V. Turchyn and Donald J. DePaolo pppppppppppppppppppppppppppppppppppppppppppp197
Global Patterns of Carbon Dioxide Variability from Satellite
Observations
Xun Jiang and Yuk L. Yung ppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp225
Permeability of Clays and Shales
C.E. Neuzil pppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp247
Flood Basalts and Mass Extinctions
Matthew E. Clapham and Paul R. Renne ppppppppppppppppppppppppppppppppppppppppppppppppp275
Repeating Earthquakes
Naoki Uchida and Roland B¨urgmann pppppppppppppppppppppppppppppppppppppppppppppppppppppp305
viii
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
EA47_FrontMatter ARI 26 April 2019 11:24
Soil Functions: Connecting Earth’s Critical Zone
Steven A. Banwart, Nikolaos P. Nikolaidis, Yong-Guan Zhu, Caroline L. Peacock,
and Donald L. Sparks pppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp333
Earthquake Early Warning: Advances, Scientific Challenges,
and Societal Needs
Richard M. Allen and Diego Melgar ppppppppppppppppppppppppppppppppppppppppppppppppppppppp361
Noble Gases: A Record of Earth’s Evolution and Mantle Dynamics
Sujoy Mukhopadhyay and Rita Parai pppppppppppppppppppppppppppppppppppppppppppppppppppppp389
Supraglacial Streams and Rivers
Lincoln H Pitcher and Laurence C. Smith ppppppppppppppppppppppppppppppppppppppppppppppppp421
Isotopes in the Water Cycle: Regional- to Global-Scale Patterns and
Applications
Gabriel J. Bowen, Zhongyin Cai, Richard P. Fiorella, and Annie L. Putman pppppppppp453
Marsh Processes and Their Response to Climate Change
and Sea-Level Rise
Duncan M. FitzGerald and Zoe Hughes ppppppppppppppppppppppppppppppppppppppppppppppppppp481
The Mesozoic Biogeographic History of Gondwanan Terrestrial
Vertebrates: Insights from Madagascar’s Fossil Record
David W. Krause, Joseph J.W. Sertich, Patrick M. O’Connor,
Kristina Curry Rogers, and Raymond R. Rogers pppppppppppppppppppppppppppppppppppppppp519
Droughts, Wildfires, and Forest Carbon Cycling: A Pantropical
Synthesis
Paulo M. Brando, Lucas Paolucci, Caroline C. Ummenhofer, Elsa M. Ordway,
Henrik Hartmann, Megan E. Cattau, Ludmila Rattis, Vincent Medjibe,
Michael T. Coe, and Jennifer Balch ppppppppppppppppppppppppppppppppppppppppppppppppppppp555
Exoplanet Clouds
Christiane Helling ppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppppp583
Errata
An online log of corrections to Annual Review of Earth and Planetary Sciences articles
may be found at http://www.annualreviews.org/errata/earth
Contents ix
Annu. Rev. Earth Planet. Sci. 2019.47:361-388. Downloaded from www.annualreviews.org
Access provided by University of California - Berkeley on 06/03/19. For personal use only.
