Content uploaded by Debi Lyman Kilb
Author content
All content in this area was uploaded by Debi Lyman Kilb on Nov 06, 2022
Content may be subject to copyright.
Testing the ShakeAlert Earthquake Early
Warning System Using Synthesized
Earthquake Sequences
Maren Böse*1 , Jennifer Andrews2, Colin O’Rourke3, Deborah Kilb4, Angela Lux5,
Julian Bunn2, and Jeffrey McGuire6
Abstract
Cite this article as Böse, M.,
J. Andrews, C. O’Rourke, D. Kilb, A. Lux,
J. Bunn, and J. McGuire (2022). Testing the
ShakeAlert Earthquake Early Warning
System Using Synthesized Earthquake
Sequences, Seismol. Res. Lett. XX,1–17,
doi: 10.1785/0220220088.
Supplemental Material
We test the behavior of the United States (US) West Coast ShakeAlert earthquake early
warning (EEW) system during temporally close earthquake pairs to understand current
performance and limitations. We consider performance metrics based on source param-
eter and ground-motion forecast accuracy, as well as on alerting timeliness. We generate
ground-motion times series for synthesized earthquake sequences from real data by com-
bining the signals from pairs of well-recorded earthquakes (4:4≤M≤7:1) using time
shifts ranging from −60 to +180 s. We examine fore- and aftershock sequences, near-
simultaneous events in different source regions, and simulated out-of-network and off-
shore earthquakes. We find that the operational ShakeAlert algorithms Earthquake
Point-source Integrated Code (EPIC) and Finite-Fault Rupture Detector (FinDer) and the
Propagation of Local Undamped Motion (PLUM) method perform largely as expected:
EPIC provides the best source location estimates and is often fastest but can under-
estimate magnitudes or, in extreme cases, miss large earthquakes; FinDer provides
real-time line-source models and unsaturated magnitude estimates for large earthquakes
but currently cannot process concurrent events and may mislocate offshore earthquakes;
PLUM identifies pockets of strong ground motion, but can overestimate alert areas.
Implications for system performance are: (1) spatially and temporally close events are
difficult to identify separately; (2) challenging scenarios with foreshocks that are close
in space and time can lead to missed alerts for large earthquakes; and (3) in these situa-
tions the algorithms can often estimate ground motion better than source parameters.
To improve EEW, our work suggests revisiting the current algorithm weighting in
ShakeAlert, to continue developments that focus on using ground-motion data to aggre-
gate alerts from multiple algorithms, and to investigate methods to optimally leverage
algorithm ground-motion estimates. For testing and certification of EEW performance in
ShakeAlert and other EEW systems where applicable, we also suggest that 25 of our 73
scenarios become part of the baseline data set.
Introduction
Achieving robust performance in earthquake early warning
(EEW) systems during earthquake sequences is challenging.
The EEW system of the Japan Meteorological Agency (JMA),
for instance, generated a significant number of false alarms dur-
ing the 2011 M9 Tohoku-Oki and 2016 M7Kumamotoearth-
quake sequences due to incorrectly determined source locations
and magnitudes for concurrent earthquakes. In the first 19 days
following the Tohoku-Oki mainshock, JMA issued warnings for
45 earthquakes: about 15 of these warnings were correct as seis-
mic intensities of 5-lower or greater on the JMA intensity scale
were observed. Warnings for seven events were missed. Ground
motions were grossly overestimated in 11 events, in which the
observed intensities did not exceed 2 at any of the observation
stations (Hoshiba et al.,2011). The system performed signifi-
cantly better in the smaller M7 Kumamoto sequence five years
1. Swiss Seismological Service (SED), ETH Zurich, Zurich, Switzerland, https://orcid
.org/0000-0003-4639-719X (MB); 2. California Institute of Technology (Caltech),
Pasadena, California, U.S.A., https://orcid.org/0000-0002-5679-5565 (JA);
https://orcid.org/0000-0002-3798-298X (JB); 3. United States Geological Survey,
University of Washington, Seattle, Washington, U.S.A., https://orcid.org/0000-
0001-5403-4685 (COR); 4. Scripps Institution of Oceanography, University of
California, San Diego, La Jolla, California, U.S.A., https://orcid.org/0000-0002-
7252-4167 (DK); 5. University of California, Berkeley, Berkeley, California, U.S.A.,
https://orcid.org/0000-0002-3767-6018 (AL); 6. United States Geological Survey,
Moffett Field, California, U.S.A., https://orcid.org/0000-0001-9235-2166 (JMG)
*Corresponding author: maren.boese@sed.ethz.ch
© Seismological Society of America
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 1
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
later and did not miss or strongly underpredict strong motion in
any large aftershock. Still, due to the simultaneous occurrence of
small earthquakes within close proximity, the system issued
overpredicted warnings for four events (Kodera et al.,2016).
The United States (US) West Coast ShakeAlert® EEW system
(Given et al., 2018;Kohler et al.,2020) has also experienced peri-
ods of increased seismicity. During the first two weeks of the
2019 M7.1 Ridgecrest earthquake sequence in southern
California, ShakeAlert created alert messages for 75% of the
87 M≥4:0 aftershocks (Chung et al.,2020). No earthquakes
were detected for almost 13 min following the mainshock,
due to challenges in associating phase picks with events in rapid
succession, as well as station telemetry issues (Stubailo et al.,
2020). ShakeAlert currently employs two EEW algorithms:
Earthquake Point-source Integrated Code (EPIC; Chung et al.,
2019) and Finite-Fault Rupture Detector (FinDer; Böse et al.,
2018). EPIC uses a short-term average (STA) to long-term aver-
age (LTA) triggering algorithm (Allen, 1982), which can tempo-
rarily lose its sensitivity when background amplitudes are high,
leading to missed event detections. FinDer is less sensitive to
background noise because it tracks seismic wave amplitudes
rather than phase picks and was able to detect four additional
M4–5 aftershocks in the immediate aftermath of the M6.4 and
7.1 Ridgecrest fore- and mainshocks (Chung et al., 2019).
However, ShakeAlert currently does not issue alerts based on
FinDer alone and requires EPIC to have also made an associated
event detection (Kohler et al.,2020).
In this article, we test the behavior and performance of the
US West Coast ShakeAlert EEW system during synthesized
earthquake sequences created from real data by combining
the signals from pairs of well-recorded earthquakes. We run
systematic tests for fore- and aftershock sequences, simultane-
ous events in different source regions, and out-of-network and
offshore scenarios, which all include a large M7.1 earthquake
because we want to understand ShakeAlert performance in
potentially damaging earthquakes. For each scenario we test
the performance of EPIC, FinDer, and the ShakeAlert
Solution Aggregator (SA; Kohler et al., 2020) and complement
our analysis by adding the Propagation of Local Undamped
Motion (PLUM; Kodera et al., 2018) EEW algorithm, which
is being considered for future integration into ShakeAlert
(Cochran et al., 2019).
In our composite data sets, two earthquakes occur within
three minutes or less of each other. In reality, such scenarios are
rare, but can be significant. For instance, historically the 1906
San Andreas fault earthquake and subsequent fires that
destroyed San Francisco were preceded by a widely felt fore-
shock about 20 s earlier (Bolt, 1968;Lomax, 2005). Analyzing
the more recent ComCat catalog (see Data and Resources), we
find that out of 77,481 M5+ global earthquakes (1975–2021),
1406 pairs (2%) are close in space and time (within 200 km and
three minutes). Reducing these data to those within the
ShakeAlert alerting region, this data set is reduced to 317 events,
of which 16 pairs (5%) are close in space and time (Table S1 and
Fig. S1 in the supplemental material available to this article).
Furthermore, some earthquakes are actually composites
of two or more smaller quakes. The 2010 M7.2 El Mayor
Cucapah earthquake in Baja California, for example, initiated
as a moderate (M∼6) normal event, and 15 s later ruptured
bilaterally along two fault segments with dominantly strike-slip
motion corresponding to M∼7(Wei et al., 2011). Although
usually interpreted as a single complex event, earthquakes like
El Mayor Cucapah can be modeled by a quick succession of
foreshock–mainshock pairs (here M∼6andM∼7).
From California ShakeAlert performance we know that dif-
ferentiating between spatially and temporally close events is a
challenge for source-based methods and current ShakeAlert
system logic. For example, during the 10 August 2020
Bombay Beach sequence FinDer merged an M3.5 foreshock
and the 26 s later M4.6 mainshock into a single event detec-
tion, and simply updated event parameters (such as magni-
tude) without assigning a new event ID. This is expected
behavior because the methodology for FinDer to detect two
near simultaneous events is not yet implemented and may
not be realistic in extreme cases as studied here. EPIC, on
the other hand, correctly created a second event. As a conse-
quence, for these data the SA did not optimally associate, and
weight, the reports from EPIC and FinDer.
The same scenario can happen in a larger (M>6) main-
shock preceded close in time by a foreshock. An instructive
example is the recent 20 December 2021 M6.2 earthquake
sequence near Petrolia, Northern California, with an M5.7
event occurring just 11 s before the main M6.2 energy release.
The smaller M5.7 earthquake, currently interpreted as a fore-
shock to the M6.2 mainshock, was located offshore along the
Mendocino transform fault, whereas the M6.2 was located
∼30 km to the east and onshore, near Petrolia. ShakeAlert
detected the smaller event ∼2 s before the mainshock initiated,
that is, 9 s after the smaller event origin. Although EPIC
located the event offshore (with a peak magnitude of
M5.7), closely tracking the evolution of the smaller foreshock,
FinDer moved its line-source model onshore (with a final
magnitude estimate of M6.3), closely tracking the evolution
of the mainshock. The ShakeAlert SA associated the two
solutions into a single detection, while heavily down weighting
the FinDer contribution. In this example, the largest reported
magnitude estimate was 0.5 magnitude units smaller than
the mainshock (compare ShakeAlert M5.7 and true M6.2).
PLUM issued a detection 11 s after the origin time of the
M5.7 foreshock, and its detection had a duration of 78 s.
This duration is longer than the median PLUM detection
durations, which are typically ∼20 s. We attribute this longer
duration to the amplified ground motions from the combined
two earthquakes. The spatial footprint of the PLUM alert
region encompassed both the foreshock and mainshock
locations.
2Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
ShakeAlert Operation
ShakeAlert is an EEW system implemented by the U.S.
Geological Survey (USGS) along with State and university part-
ners (Given et al.,2018). The current system alerts for earth-
quakes along the US West Coast, using a reporting region
that covers the states of California, Washington, and Oregon.
Alerts can be received as Wireless Emergency Alerts on mobile
devices, as well as through mobile applications (e.g., MyShake,
QuakeAlertUSA, and ShakeReadySD), or through a ShakeAlert-
powered earthquake alert feature that is integrated into the
Android Operating System. As of March 2022, ShakeAlert dis-
tribution partners issue public alerts via mobile devices for
earthquakes M4.5+ inside the ShakeAlert reporting region.
The alert thresholds depend on the means of delivery and cur-
rently range from Modified Mercalli Intensity (MMI) III (weak
or larger shaking) to MMI V (moderate or larger shaking) and
magnitude estimates of 4.5 or 5.0 (see Chung et al.,2020). Other
ShakeAlert partners operate automated systems for specific
applications that are initiated at M4.0 and higher and MMI
III and higher depending on the specific system.
EPIC is a point-source algorithm that determines earth-
quake hypocenters from Pphase picks and magnitudes from
empirical peak displacement versus distance relations (Kuyuk
and Allen, 2013). Because seismically observed displacement
amplitudes saturate in large earthquakes (M>7), EPIC is best
suited for small and moderate-size earthquakes. EPIC is usu-
ally the fastest ShakeAlert algorithm and in areas with dense
instrumentation can often alert within 4–5 s of the event origin
(Kohler et al., 2020).
FinDer uses a template matching approach that compares
the spatial distribution of peak ground acceleration (PGA)
amplitudes observed at seismic stations with theoretical tem-
plate maps that are predicted from ground-motion models for
different line-source parameters (location, length, and strike).
FinDer can process events as small as M2.5, but is best suited
for events with clear finite-source character (M>6). FinDer is
sensitive to spatial ground-motion distributions rather than
absolute amplitudes and, therefore, does not suffer from mag-
nitude saturation in large earthquakes (Böse et al., 2018).
Source parameter estimates from EPIC and FinDer are com-
bined by the ShakeAlert SA, which passes on information for
alert distribution if certain criteria are met (Given et al.,
2018). The SA logic has been updated since last described in
Kohler et al. (2020) to improve performance when testing with
large historic earthquakes. The current SA location and magni-
tude weighting for MSA <6:0 is approximately 83% for EPIC
and 17% for FinDer (in which MSA is the SA magnitude com-
puted from the weighted magnitude contributions from EPIC
and FinDer). This reflects the final weightings once EPIC
reaches its maximum station number for lowest uncertainty,
and typically only a few updates are needed to reach
this ratio (Chung et al.,2020). For MSA ≥6:0, MFinDer is used
if MFinDer >MEPIC, else the earlier ratio is applied. FinDer is
currently not allowed to alert alone but allowing it to do so
is being considered. The ShakeAlert eqInfo2GM module
(Thakoor et al.,2019) converts source parameter estimates from
the SA into ground-motion grid and contour products, which
are used to define alert areas. As a distance metric, eqInfo2GM
uses the epicentral distance for MSA <6:0, and the closest dis-
tance to the FinDer line source (if available) for MSA ≥6:0.
PLUM is not yet adopted by ShakeAlert but is being consid-
ered for future integration (Cochran et al.,2019,2022;Minson
et al., 2020;Kilb et al.,2021). Unlike the other two algorithms,
PLUM does not determine source parameters, and instead
extrapolates the observed shaking intensities to nearby regions
and produces ground-motion predictions. In the version of
PLUM being considered here, alerts are issued when two or more
neighboring stations observe MMI values above configurable
thresholds (see the Data and Playback section). Regions within
a given radius of those stations receive the alert, and as other
neighboring stations observe additional shaking the alert region
cangrowovertime.OneadvantageofPLUMisavoidingtheneed
to calculate source parameters before converting to ground-
motion predictions (Cochran et al.,2019). Key to successful
PLUM detections is tuning parameters to avoid overextending
alert areas for small events and underextending alert areas for
large earthquakes. Potential disadvantages of PLUM are delayed
alerts compared to EPIC and FinDer if the faster P-wave ampli-
tudes are not above the PLUM trigger thresholds (Cochran et al.,
2022). In addition, PLUM lacks a magnitude estimate to use as
an alert decision variable and, depending on parameter settings,
can alert for small earthquakes that produce ground motions
above PLUM’s MMI trigger threshold (Kilb et al., 2021).
Data and Playback
To explore the performance of ShakeAlert and PLUM during
earthquake pairs, we synthesize composite signals by summing
two earthquake ground-motion time series recorded at the
same station using time shifts ranging from −60 to +180 s.
We use six earthquakes as “building blocks”(Table 1, Fig. 1)
and create three data sets. Building blocks RC71, RC64,
RC54, and RC44* correspond to earthquakes in the southern
California Ridgecrest area with magnitudes ranging from
M4.4 to 7.1. Building blocks EM45 and EM60* correspond
to earthquakes in the El Monte (Los Angeles) area with magni-
tudes of M4.5 and 6.0. Events marked with asterisks were scaled
to larger (e.g., EM60* was upscaled from EM45) or smaller (e.g.,
RC44* was downscaled from RC54) magnitude earthquakes
using a frequency-dependent transfer function (Roh, 2021).
The data sets each comprise three-component waveforms at
about 400 stations from the CI, AZ, CE, SN, NN, NP, and BC
networks, with locations across southern California and
northern Baja California, Mexico (∼32°–37° N, ∼114°–121° W;
Fig. 1). Data set number 1 simulates fore- and aftershock
sequences in the southern California Ridgecrest area (building
blocks RC71, RC64, RC54, and RC44*; Fig. 2); data set number
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 3
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
TABLE 1
Building Blocks and Algorithm Performance, Including Reporting Time, Estimated Magnitude (First and Last
Update), Peak Modified Mercalli Intensity (MMI), and Location Error Δ(First and Last Update)
ANSS Catalog EPIC FinDer PLUM
Building
Block
Location (Event ID) Origin Time
(t0) (yyyy/mm/dd), UTC (hh:mm:ss)
(Moment) Magnitude, Peak MMI
(ShakeMap Grid and Station)
Latitude/Longitude
Detection Time After t0,
Magnitude (First–Last),
Peak MMI, Location
Error (First–Last)
Detection Time After t0,
Magnitude (First–Last),
Peak MMI, Location
Error (First–Last)
Detection Time After
t0, Magnitude, Peak
MMI (First–Last),
Location Error
RC71 Ridgecrest (ci38457511)
2019/07/06, 03:19:53 4.3 s 4.8 s 4.0 s
M7.1 M5.6–6.4 M5.5–7.1 –
MMImax
obs 8:7, MMIstn
obs 8:8––MMImax
pred 6:2–8:2
35.770/−117.599 Δ6:4–3:8km Δ6:9–4:4km –
RC64 Ridgecrest (ci38443183)
2019/07/04, 17:33:49 6.1 s 6.2 s 6.0 s
M6.4 M5.8–6.2 M5.5–6.6 –
MMImax
obs 8:4, MMIstn
obs 7:5––MMImax
pred 4:7–7:6
35.70/−117.504 Δ0:4–2:4km Δ7:1–15:4km –
RC54 Ridgecrest (ci38450263)
2019/07/05, 11:07:53 4.1 s 5.2 s 4.0 s
M5.4 M6.0–5.7 M5.0–5.7 –
MMImax
obs 7:1, MMIstn
obs 5:7––MMImax
pred 6:0
35.760/−117.575 Δ7:4–1:2km Δ8:5–4:5km –
RC44* Ridgecrest (ci38450263*)
2019/07/05, 11:07:53 4.2 s 7.1 s (not detected)
M4.4 M4.4–4.2 M4.0–4.7
MMImax
obs 4:6, MMIstn
obs 3:6––
35.760/−117.575 Δ6:8–0:8km Δ4:5–2:7km
EM45 El Monte (ci38695658)
2020/09/19, 06:38:46 3.6 s 4.8 s 3.0 s
M4.5 M4.5–4.7 M4.6–5.1 –
MMImax
obs 6:1, MMIstn
obs 5:2––MMImax
pred 4:5–5:4
34.038/−118.080 Δ3:2–3:1km Δ2:1–11:7km –
EM60* El Monte (ci38695658*)
2020/09/19, 06:38:46 2.8 s 6.3 s 5.0 s
M6.0 M5.3–6.5 M5.5–6.3 –
MMImax
obs 9:1, MMIstn
obs 8:0––MMImax
pred 6:1–7:8
34.038/−118.080 Δ6:0–1:8km Δ2:1–1:2km –
We provide both the maximum MMI observed at any station (MMIstn
obs) and the maximum MMI reported by ShakeMap (MMImax
obs ). ANSS, Advanced National Seismic System; EPIC,
Earthquake Point-source Integrated Code; FinDer, Finite-Fault Rupture Detector; and PLUM, Propagation of Local Undamped Motion.
*Events upscaled or downscaled from other events.
4Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
2 simulates near-simultaneous earthquakes in Ridgecrest and
Los Angeles (building blocks RC71, EM45, and EM60*); data
set number 3 simulates offshore and out-of-network scenarios
based on two Ridgecrest events (building blocks RC71 and
RC54). For data set number 3, we iteratively remove near-
source stations around the Ridgecrest mainshock to mimic off-
shore events that are located between 25 and 175 km from our
artificial shoreline. The maximum distance was chosen to align
with the offshore extent of the ShakeAlert alerting region
(Kohler et al., 2020). We chose to simulate these out-of-net-
work events, rather than use real events, due to the lack of suit-
able systematic observations (see the Discussion section).
For each scenario, we create a reference ShakeMap to assess
the performance of the algorithms with regards to the pre-
dicted (peak ground motion) alert areas. For data sets number
1 and number 3, we assume that the USGS ShakeMap (see Data
and Resources) for the largest earthquake in each scenario is a
good proxy for the ground truth, because the secondary
smaller quake creates ground motions that are much smaller
than the mainshock. For the RC71 + EM45 scenarios in data
set number 2, we combine the USGS ShakeMaps for the M7.1
Ridgecrest and M4.5 El Monte earthquakes by taking the
maximum MMI value at each grid point. We do the same
for the RC71 + EM60* scenarios, but upscale the ShakeMap
for El Monte with an MMI ratio value that we determined
empirically by comparing the MMI values computed from
the scaled (EM60*) and unscaled (EM45) waveforms.
We run full-waveform playbacks of each scenario using the
Earthworm Tankplayer software (see Data and Resources). The
EPIC and FinDer codes and configurations in our playbacks are
near identical to those currently used in the ShakeAlert produc-
tion system (as of March 2022). The only difference in our play-
backs is that FinDer is configured to allow processing of very
large earthquakes (up to M9), which has no impact on the
results shown here. For PLUM we adoptMMI trigger thresholds
of MMI 4 and 3 at the first and second station, respectively (Kilb
et al.,2021), which is more conservative than what was used in
earlier studies and reduces the unwanted detection of smaller
magnitude earthquakes (Cochran et al.,2019;Minson et al.,
2020). To determine PLUM alert regions, the MMI at the obser-
vation station is extrapolated to 30 km (Cochran et al., 2022).
Results
We start with a playback of the six building blocks to establish
a performance benchmark for each algorithm. The algorithms
achieve a very good performance (Table 1): they detect all
earthquakes within 4–6 s after event origin (data latencies
neglected) and with location errors typically much smaller
than 10 km. The peak magnitude errors are <0.7 magnitude
units, with the largest error observed for RC71 (M7.1
Ridgecrest mainshock), for which EPIC estimates the final
magnitude as M6.4. This underestimation is expected because
of amplitude saturation. FinDer is typically 0.1–3.5 s slower for
initial detections and has slightly larger location errors (1–
15 km) than EPIC but provides very good magnitude esti-
mates, including M7.1 for RC71. PLUM does not provide
source parameters, only MMI estimates, and peak MMI values,
MMImax
pred, agree well with the observed target values as taken
either from the reference ShakeMap grid, MMImax
obs , or from the
waveforms converted to MMI using the relations of Worden
et al. (2012), MMIstn
obs. As expected, PLUM does not detect the
smaller RC44* event, because the initial PLUM MMI 4 trigger-
ing threshold is not exceeded.
Source parameters
The playback results for the three composite data sets are sum-
marized in Tables 2–4. Overall, the algorithms perform largely
as expected: EPIC detects the largest number of earthquakes
and provides the best epicenter location estimates. However,
it can miss large earthquakes if events occur close in time,
or if events are contained within the wavetrain of an earlier
earthquake in the same or different source regions (here
Figure 1. The study region in southern California. Locations of
earthquakes in Ridgecrest (RC, earthquake magnitudes also
listed) and El Monte (EM) used as building blocks in our synthesized
earthquake pairs (Table 1). RC and EM are ∼200 km apart,
implying that the seismic waves from one location will reach the
other location and waveforms from (near-) simultaneous events will
begin to overlap after 30 s (Pwave) and 50 s (Swave). The triangles
show the approximately 400 seismic stations used in this study. The
black solid lines show the artificial “shorelines”used for offshore
simulations of RC71 at 25, 100, and 175 km distance.
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 5
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
200 km apart). FinDer provides the best magnitude estimates
for large earthquakes (M>6); for onshore events its line-
source model is typically in excellent agreement with the fault
rupture, and the growth of these line sources matches well with
the evolving ground motions. However, the current FinDer
code cannot process multiple events at the same time, and
its location estimates are degraded in offshore and out-of-net-
work events (>75 km). PLUM responds accurately to the
ground-motion continuum of near-simultaneous earthquakes
and identifies pockets of strong ground motion, such as in sedi-
mentary basins. However, PLUM does not solve for source
parameters (which does not integrate well with the current
ShakeAlert design) and makes no effort to differentiate
between multiple temporally and spatially close earthquakes
(which probably does not matter if the goal is to send alerts
to those experiencing shaking above a certain threshold).
Ground-motion
parameters
To explore the effect of the
source parameter outputs on
the predicted and observed
ground motions, we adopt the
ShakeAlert eqInfo2gm module
(Thakoor et al.,2019)andpre-
dict MMI maps using the source
parameters estimated by the
algorithms: for EPIC we use
the epicenter and magnitude,
for FinDer its line-source model
and derived magnitude, and for
the SA the weighted magnitude
estimate and either line-source
(if MSA ≥6:0) or epicenter.
For PLUM we use its direct
ground-motion output. Figure 3
shows resulting MMI maps in
comparison to the reference
ShakeMaps for four of the sce-
narios, selected to demonstrate
key performance issues. Sample
videos of the ground-motion
evolution are shown in the sup-
plemental videos S1–S4.
Rapid foreshocks (data set
number 1). In our source
parameter analysis, we found
that EPIC can miss large earth-
quakes if multiple events occur
close in time, for example, if
we have a 10–60 s earlier
M4.4–5.4 foreshock as in the
scenarios analyzed here. As
shown in Table 2this behavior is not systematic. EPIC’s missed
event detections mostly result from the STA/LTA detection
scheme. An elevated LTA value caused by the first earthquake
can hamper the detection of the second event. In this case, the
combined SA magnitude may not exceed 6.0 because EPIC
missed the mainshock and its foreshock detection is incorrectly
associated with the FinDer mainshock report. As a conse-
quence, the FinDer solution is incorrectly down weighted
because of the current algorithm weighting adopted by the
SA. This scenario will cause a strong underestimation of the
ground-motion alert areas computed from SA source param-
eters, with a maximum magnitude underestimation of 1.5
magnitude units seen across the tested scenarios (Table 2
and Fig. 3a). The PLUM alert area, by contrast, is significantly
overestimated. However, this area depends on how PLUM’s
alert regional extents are defined. There is a tradeoff between
Foreshock, me shi= –60 s
Foreshock, me shi= –30 s
Aershock, me shi= +30 s
Aershock, me shi= +60 s
Aershock, me shi= +120 s
1 min
Figure 2. Example composite waveforms that combine building blocks RC71 and RC54 (Table 1)at
station CI.PASC.00 with time shifts of −60, −30, +30, +60, and +120 s (from top to bottom).
6Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
TABLE 2
Algorithm Performance for Data Set Number 1 (Building Blocks RC71, RC64, RC54, and RC44* in Table 1;27
Scenarios), Including Final Magnitudes, Epicenter Location Errors (Δ), and Final Peak MMI
Scenario
Time
Shift Algorithm M 6.4 (RC64) M 5.4 (RC54) M 4.4 (RC44*)
Foreshock +
M7.1 (RC71)
−60 s EPIC M6.3, Δ1:0km
†M5.8, Δ0:8km
†M4.3, 6.6, Δ7:1=3km
†
FinDer M7.0, Δ15 km M7.0, Δ6:4km M7.0, Δ6:4km
SA M 6.9,Δ1:3 km M 5.9,Δ1:0 km M 4.8, 6.6,Δ7:0=3:0km
PLUM MMImax
pred 8:2 MMImax
pred 8:2 MMImax
pred 8:2
−30 s EPIC M6.3, Δ1:5km M5.8, 7.0, = 1.5/2.6 km†M4.4, 6.8, Δ2:2=3km
†
FinDer M7.0, Δ15 km M7.0, Δ6:4km M7.0, Δ6:4km
SA M 6.9,Δ1:8 km M 6.0, 7.0,Δ1:3=2:6 km M 4.7, 6.8,Δ2:4=3:0km
PLUM MMImax
pred 8:2 MMImax
pred 8:2 MMImax
pred 8:2
−10 s EPIC M6.3, Δ0:2km M5.8, Δ0:1km
†M5.2, Δ1:9km
†
FinDer M7.0, Δ15 km M7.0, Δ6:4km M7.0, Δ6:4km
SA M 6.9, Δ0:7 km M 5.9, Δ0:3 km M 5.5, Δ2:0km
PLUM MMImax
pred 8:1 MMImax
pred 8:0 MMImax
pred 8:2
−5s EPIC M6.3, Δ1:4km M5.8, Δ1:3km M5.8, Δ1:3km
FinDer M7.0, Δ15 km M7.0, Δ6:4km M7.0, Δ6:4km
SA M 6.9, Δ1:7 km M 7.0, Δ1:3 km M 6.9, Δ1:5km
PLUM: MMImax
pred 8:1 MMImax
pred 8:0 MMImax
pred 8:2
M7.1 (RC71) +
aftershock
+15 s EPIC M6.4, Δ1:0km M6.4, Δ3:3km
†M6.4, Δ1:1km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ1:2 km M 6.9, Δ3:3 km M 6.9, Δ1:2km
PLUM MMImax
pred 8:3 MMImax
pred 8:2 MMImax
pred 8:2
+30 s EPIC M6.4, Δ1:0km M6.4, Δ1:1km M6.4, Δ1:1km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ1:2km M 6.9, Δ1:3 km M 6.9, Δ1:2km
PLUM MMImax
pred 8:0 MMImax
pred 8:2 MMImax
pred 8:2
+60 s EPIC M6.4, Δ1:9km M6.4, Δ1:5km M6.4, Δ1:6km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ2:0 km M 6.9, Δ1:5 km M 6.9, Δ1:7km
PLUM MMImax
pred 8:2 MMImax
pred 8:2 MMImax
pred 8:2
+120 s EPIC M6.5, 6.8, Δ1:6=0:9km M6.5, Δ1:0km M6.4, Δ1:0km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, 6.8, Δ1:7=0:9 km M 6.9, Δ1:2 km M 6.9, Δ1:2km
Separate values are given (and indicated with “/”) when multiple events were detected within the considered 3-minute-long time window. Bold entries highlight performance of
SA (which combines EPIC and FinDer and thus gives the total ShakeAlert performance).
*RC44 is downscaled from RC54.
†Nine scenarios, which include cases of both good and poor performance (in which the mainshock is missed and strongly underestimated), are proposed to become part of the
ShakeAlert baseline test suite.
(Continued next page.)
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 7
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
how much warning time can be given and the spatial extent of
the alerting region, which is controlled by the radial distance to
which ground motions around the observing station are
extrapolated (here set to 30 km; Cochran et al., 2022).
Overlapping waveforms from near-simultaneous
events (data set number 2). EPIC can miss a second event
ifitiscontainedwithinthewavetrainofanearlierearthquake
in another source region (here 200 km apart; Fig. 1). In this
case, the SA alert areas may still be estimated reasonably well
(Fig. 3b), because the large event is detected by the amplitude-
based FinDer algorithm and its line-source model improves
ground-motion estimates; however, the maximum magnitude
reported by the SA and used to compute ground motions may
be underestimated by as much as 1.9 magnitude units
(Table 3) due to SA association behaviors described for data
set number 1. If the larger event occurs first, FinDer will not
report parameters for the second (smaller) event, because
FinDer multievent processing is currently not fully imple-
mented. If two events occur close in time, FinDer will assign
the same core parameters (in particular the epicenter) to both
earthquakes, whereas a line-source model is computed for
the larger event. Nevertheless, in this scenario the alert areas
for SA source parameters may also be estimated well (Fig. 3c,
here correctly combining EPIC’s point-source estimate for
EM60* and FinDer’s line-source estimate for RC71) if EPIC
issues independent alerts for the two events. However, if the
smallereventislessthan60safterthelargerevent,evenifitis
of significant magnitude (M6) it will not currently be issued
an alert (missed by EPIC and FinDer), and ground motions in
the epicentral area would be underestimated by the system
(Table 3). PLUM successfully detects both events, either as
separate detections or by merging the later event into the
ongoing first alert.
Offshore and out-of-network earthquakes (data set
number 3). For the offshore scenarios, the EPIC location
and magnitude solutions are typically well estimated (average
12 km location error, Table 4), though EPIC sometimes cre-
ates poor location estimates (150–300 km location error) for
the second event due to slightly late triggers on a few stations
caused by the STA/LTA algorithm not being able to trigger
sooner as the LTA remains high after the first event. The
FinDer location estimates, by contrast, are degraded in off-
shore earthquakes >75 km from shoreline because the events
tend to be mapped close to the “coast”and magnitudes are
decreased accordingly. On the other hand, peak MMI predic-
tions resulting from the EPIC, FinDer, and SA source esti-
mates (Table S2) suggest that the closer and smaller
FinDer line-source estimate often achieves better agreement
with observed values for the offshore events than the larger
more distant EPIC solution. The secondary higher MMI esti-
mate in EPIC in Table S2 results from events with incorrect
“onshore”location estimates (Table 4). Aside from scenarios
with MSA ≥6:0, in which the FinDer line source is used to
define distance metrics for ground-motion prediction, the
SA solution for these offshore scenarios is mostly controlled
by EPIC, and the SA alert areas are again estimated well
(Fig. 3d). A new feature of FinDer to use fault-specific tem-
plates, such as for Cascadia, is currently being developed and
is expected to improve FinDer’s performance in offshore
earthquakes in the future. PLUM can only detect motions
where stations are located (currently only onshore) and
the PLUM alert area is again somewhat overestimated due
to the chosen radial alerting distance.
Overall, we find that most scenarios that appear critical in
the source parameter space are less critical in terms of ground-
motion alert areas, including the occurrence of simultaneous
events with overlapping waveforms (data set number 2) and
TABLE 2 (continued)
Algorithm Performance for Data Set Number 1 (Building Blocks RC71, RC64, RC54, and RC44* in Table 1;27
Scenarios), Including Final Magnitudes, Epicenter Location Errors (Δ), and Final Peak MMI
Scenario
Time
Shift Algorithm M 6.4 (RC64) M 5.4 (RC54) M 4.4 (RC44*)
PLUM MMImax
pred 8:2 MMImax
pred 8:2 MMImax
pred 8:0
+180 s EPIC M6.5, Δ3:0km
†M6.4, Δ1:1km M6.4, Δ1:9km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ3:0 km M 6.9, Δ1:2 km M 6.9, Δ2:0km
PLUM MMImax
pred 8:2 MMImax
pred 8:2 MMImax
pred 8:2
Separate values are given (and indicated with “/”) when multiple events were detected within the considered 3-minute-long time window. Bold entries highlight performance of
SA (which combines EPIC and FinDer and thus gives the total ShakeAlert performance).
*RC44 is downscaled from RC54.
†Nine scenarios, which include cases of both good and poor performance (in which the mainshock is missed and strongly underestimated), are proposed to become part of the
ShakeAlert baseline test suite.
8Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
TABLE 3
Algorithm Performance for Data Set Number 2 (Building Blocks RC71, EM60, and EM45; Table 1)
Scenario Time Shift Alg. M 4.5 (EM45) M 6.0 (EM60*)
El Monte before Ridgecrest −60 s EPIC M4.8, 6.7, Δ3:0km=3:0km†M6.5, Δ3:6km†
FinDer M7.0, Δ3:9km M7.0, Δ1:4km
SA M 5.1, 6.6, Δ2:9=3:0 km M 6.9, Δ3:5km
PLUM MMImax
pred 5:4, 8.1 MMImax
pred 8:2
−30 s EPIC M4.8, Δ3:3km†M6.6, Δ3:4km†
FinDer M7.0, Δ3:9km M7.0, Δ3:9km
SA M 5.2, Δ3:2 km M 6.9, Δ3:3km
PLUM MMImax
pred 5:4, 8.2 MMImax
pred 8:2
−10 s EPIC M4.9, 6.6, Δ4:1=0:6km M6.6, 6.9, Δ3:9=2:3km
FinDer M7.0, Δ3:9km M7.0, Δ3:9km
SA M 5.1, 6.6, Δ4:0=0:6 km M 7.0, 6.9, Δ3:7=2:3km
PLUM MMImax
pred 5:4, 8.2 MMImax
pred 8:2, 3.9
−5s EPIC M4.9, 6.5, Δ3:4=2:7km M6.5, 6.5, Δ1:9=2:5km
FinDer M7.0, Δ3:9km M7.0, Δ3:9km
SA M 5.2, 6.4, Δ3:3=2:7 km M 7.0, 6.4, Δ1:9=2:5km
PLUM MMImax
pred 5:4, 8.2 MMImax
pred 8:2, 4.0
Ridgecrest before El Monte +15 s EPIC M6.3, Δ1:1km M6.8, Δ1:1km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ1:2 km M 7.0, Δ1:2km
PLUM MMImax
pred 8:2, 4.2 MMImax
pred 8:2, 4.2
+30 s EPIC M6.4, Δ1:6km M7.0, Δ1:1km†
FinDer M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ1:7 km M 7.0, Δ1:2km
PLUM MMImax
pred 8:0, 4.2 MMImax
pred 8:2, 4.2
+60 s EPIC M6.4, Δ1:1km M6.7, 6.7, Δ0:7=0:3km†
FinDer M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ1:2 km M 6.9, 6.7, Δ0:8=0:3km
PLUM MMImax
pred 8:2, 4.2 MMImax
pred 8:2, 4.2
+120 s EPIC M6.5, Δ1:5km M6.5, 6.8, Δ1:0=1:4km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.5, Δ1:5 km M 6.9, 6.7, Δ1:2=1:4km
PLUM MMImax
pred 8:2, 3.8 MMImax
pred 8:0, 4.2
+180 s EPIC M6.4, Δ1:1km M6.4, 6.8, Δ1:6=2:0km
FinDer M7.0, Δ5:1km M7.0, Δ5:1km
SA M 6.9, Δ1:2 km M 6.9, 6.7, Δ1:7=2:0km
PLUM MMImax
pred 8:2, 4.2 MMImax
pred 8:2, 4.2
This includes 18 scenarios, of which 6 are recommended for the test suite. Follows Table 2.
*EM60 is upscaled from EM45.
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 9
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
TABLE 4
Algorithm Performance for Data Set Number 3 (Building Blocks RC71 and RC54; Table 1) for Different “Offshore
Distances”
M 5.4 Foreshock (RC54) M 5.4 Aftershock (RC54)
Distance Algorithm −60 s −30 s −10 s +120 s
25 km MMIstn 8.9 8.9 8.9 8.9
EPIC M5.8, 6.7, Δ2=3:6km M5.8, 5.7, Δ2=257 km M5.8, 5.7, Δ3:1=141 km M6.4, Δ1:7km
FinDer M7.2, Δ33 km M7.1, Δ33 km M7.1, Δ33 km M7.1, Δ29 km
SA M 7.2, 6.6, Δ3=3:6 km M 7.1, 5.7, Δ3=257 km M 7.1, 5.7, Δ3:1=141 km M 7.1, Δ2:6km
PLUM MMImax
pred 4:4, 8.2 MMImax
pred 8:2 MMImax
pred 8:2 MMImax
pred 8:2
50 km MMIstn 6.6†6.6†6.6 6.6†
EPIC M5.8, 6.9, Δ7:4=16 km M5.8, Δ10 km M5.9, 6.0, Δ3:1=141 km M6.4, Δ3:5km
FinDer M6.2, Δ51:4km M6.2, Δ51:4km M6.1, Δ46 km M6.4, Δ55 km
SA M 5.8, 6.8, Δ7:7=16 km M 5.8, Δ10:6 km M 6.0, 6.0, Δ4:4=141 km M 6.5, Δ5:2km
PLUM MMImax
pred 3:9, 6.2 MMImax
pred 3:9, 5.9 MMImax
pred 6:1 MMImax
pred 6:1
75 km MMIstn 6.6†6.6†6.6 6.6
EPIC M5.8, Δ6:1km M5.8, 5.6, Δ6=257 km M5.9, 6.0, Δ4:6=141 km M6.4, Δ4:0km
FinDer M6.6, Δ100 km M6.5, Δ100 km M6.6, Δ71 km M6.6, Δ69 km
SA M 5.9, Δ7:4 km M 5.9, 5.6, Δ6=257 km M 6.6, 6.0, Δ5:7=141 km M 6.6, Δ6:0km
PLUM: MMImax
pred 5:9 MMImax
pred 5:9 MMImax
pred 6:1 MMImax
pred 6:1
100 km MMIstn 6.6†6.6†6.6 6.6
EPIC M5.8, Δ7:4km M5.8, Δ22 km M5.9, 6.0, Δ16=141 km M6.4Δ4:1km
FinDer M6.6, Δ100 km M6.5, Δ100 km M6.5, Δ100 km M6.5, Δ101 km
SA M 5.9, Δ5:3 km M 6.0, Δ18 km M 6.6, 6.9, Δ12=141 km M 6.5, Δ7:1km
PLUM: MMImax
pred 5:9 MMImax
pred 5:9 MMImax
pred 5:9 MMImax
pred 6:1
125 km MMIstn 6.6†6.6†6.6 6.6†
EPIC M5.8, 7.2, Δ6:1=29 km M5.8, 5.6, Δ13=179 km M6.0, 6.0, Δ0:7=141 km M6.4, Δ2:2km
FinDer M6.4, Δ135 km M6.5, Δ116 km M6.5, Δ116 km M6.3, Δ125 km
SA M 5.9, 7.2, Δ9=29 km M 5.9, 5.5, Δ14=179 km M 6.4, 6.0, Δ3:5=141 km M 6.4, Δ6:3km
PLUM MMImax
pred 6:1 MMImax
pred 6:1 MMImax
pred 5:9 MMImax
pred 6:1
150 km MMIstn 6.6 6.6 6.6 6.6
EPIC M5.8, Δ6:8km M5.8, Δ18:3km M6.0, Δ1:6km M6.5, Δ14:4km
FinDer M5.9, Δ137 km M6.0, Δ142 km M6.4, Δ146 km M5.9, Δ138 km
SA M 5.8, Δ6:4 km M5.8, Δ13:6 km M 6.4, Δ4:2 km M 6.4, Δ17:9km
PLUM MMImax
pred 5:9 MMImax
pred 5:9 MMImax
pred 5:9 MMImax
pred 6:1
175 km MMIstn 6.0 6.0 6.0 6.0
EPIC M5.4, Δ76 km M5.7, Δ34 km M6.0, Δ49 km –
FinDer M4.6, Δ164 km M6.4, Δ173 km M6.1, Δ173 km M6.1, Δ173 km
This includes with 28 scenarios, of which 10 are recommended for the test suite. Follows Table 2. MMI values refer to “onshore”locations. See Table S2 in the supplemental
material for peak MMI predictions derived from the EPIC, FinDer, and SA source estimates.
(Continued next page.)
10 Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
offshore earthquakes (data set number 3). Problematic, how-
ever, are large M6 + events in which the SA magnitude does
not exceed the current M6.0 threshold (e.g., because of a small
foreshock) and the larger magnitude FinDer solution is incor-
rectly down weighted (data set number 1).
Alert timeliness
To better understand these results in terms of alert timeliness, we
compare in Figure 4the temporal evolution of observed and pre-
dicted ground-motion intensities, MMIstn
obs and MMIstn
pred,ateach
station. We employ the procedure and performance metrics
detailed in Chung et al. (2020) and Meier et al. (2020) and test
three alerting thresholds: MMIalert 2:5, 3.5, and 4.5. The target
ground-motion warning threshold MMItw ,whichweuseto
determine timeliness, is set to 4.5 for all tests. That is, we measure
the time interval between when an algorithm first predicts shak-
ing above MMIalert and when MMItw is exceeded. We assume
that a site will get alerted as soon as MMIpred t≥MMIalert,in
other words, that there is no alert delivery latency. We count cor-
rectly alerted sites with MMIstn
obs ≥MMItw as true positives (TP);
if the alert is created in time, that is, before MMIstn
obst≥MMItw ,
we assign TP(timely) and TP(untimely) otherwise; if MMIalert ≤
MMIstn
obs <MMItw and the alert is in time, we assign TP(timely,
<MMItw ). Sites with MMIstn
obs <MMIalert that get alerted count as
false alerts (false positive, FP). Failing to alert a site with MMIstn
obs ≥
MMIalert counts as a missed alert (false negative, FN). Stations
with low observed and predicted MMI values are counted as true
negatives (TN).
We also compute the empirical cumulative distribution
functions (CDFs) of warning time provided by each algorithm
(Fig. S2). These CDFs account for the entire alerting evolution
in each scenario, encompassing both events in the sequence. If
an algorithm successfully detects a foreshock, it might be able
to provide longer warning times than another algorithm that
misses it, provided that the predicted ground motions exceed
MMIalert. The warning time calculations presented here do not
account for delays caused by data latencies and alert delivery
time, which typically are on the order of 1–2 s, and up to
several seconds, respectively (e.g., Trugman et al., 2019;
Allen et al., 2020).
Figure 4shows these warning time metrics applied to the
scenarios in Figure 3, using stations in southern California with
a distance cutoff of 314, 551, and 768 km from the Ridgecrest
mainshock corresponding to the theoretical distance to sites
with median ground motion of MMI 4.5, 3.5, and 2.5, respec-
tively (Boore and Atkinson, 2008;Atkinson and Boore, 2011;
Worden et al., 2012). This means that for the two lower
MMIalert levels (2.5 and 3.5) we use almost all stations shown
in Figure 1. A larger distance cutoff will generally increase the
percentage of TN because more stations with small ground
motions are included. The lack of TN and FP stations in
the MMIalert 2:5 panels (Fig. 4, left), however, is due to
the spatial footprint of stations used in this study not extending
far enough to include locations with observed MMI below 2.5.
Comparing the EPIC and FinDer results for the scenario
with a 10 s earlier foreshock (data set number 1, Fig. 4a)
we find that using the more accurate magnitude estimate from
FinDer, rather than the weighted SA magnitude, would result
in 5% more TP alerts for MMIalert 2:5, 22% more TP for
MMIalert 3:5, and 1% more TP alerts for MMIalert 4:5.
The use of PLUM would have resulted in 20% more timely
TP alerts at MMIalert 4:5 compared to the SA, at the expense
of 30% FP alerts (overalerting). The occurrence of a foreshock
can be beneficial to providing long warning times, as the fore-
shock often alerts the area many seconds before the mainshock
occurs (although the spatial extent of the alert region generated
by the smaller foreshock would be reduced compared to the
spatial extent generated by the larger mainshock). For instance,
for MMIalert 2:5 the algorithms provide >30 s (PLUM) and
>70 s (EPIC, FinDer, SA) of warning to 50% of sites experi-
encing 4.5 ≤MMI ≤5.5 (Fig. S2a, green lines). For MMIalert
3:5 these numbers drop to ∼0 s (EPIC, SA), ∼30 s (PLUM),
and ∼45 s (FinDer), respectively. PLUM continues providing
positive median warning time (∼8 s) for MMIalert 4:5.
For both scenarios with events occurring close in time in
Ridgecrest and El Monte (data set number 2; Fig. 4b,c), the
threshold of MMIalert 2:5 results in successful TP alerting
for all sites and all algorithms, because a large area is alerted
regardless of the details of source parameter estimation. At
the higher MMIalert 3:5 and 4.5, differences in algorithm
TABLE 4 (continued)
Algorithm Performance for Data Set Number 3 (Building Blocks RC71 and RC54; Table 1) for Different “Offshore
Distances”
M 5.4 Foreshock (RC54) M 5.4 Aftershock (RC54)
Distance Algorithm −60 s −30 s −10 s +120 s
SA M 5.1, Δ94 km M 5.8, Δ38 km M 6.0, Δ53 km M 6.0, Δ178 km
PLUM: MMImax
pred 4:6 MMImax
pred 4:8 MMImax
pred 4:8 MMImax
pred 4:8
This includes with 28 scenarios, of which 10 are recommended for the test suite. Follows Table 2. MMI values refer to “onshore”locations. See Table S2 in the supplemental
material for peak MMI predictions derived from the EPIC, FinDer, and SA source estimates.
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 11
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
performance become more apparent. For the scenario of the
El Monte event occurring first (Fig. 4b), EPIC’sfailureto
detect the Ridgecrest event leads to higher numbers of FN
alerts, whereas FinDer’s delayed detection of the Ridgecrest
source leads to higher numbers of untimely TP alerts. For
the scenario of the El Monte event occurring second
(Fig. 4c), EPIC’s detection of both events results in very suc-
cessful warning time performance. FinDer’s failure to detect
the smaller, later El Monte event (caused by FinDer’s
current inability to issue two alerts at the same time) results
in a high number of FN alerts due to ground-motion under-
estimation in the LA area. Once again, at the higher
MMIalert 4:5, PLUM’s design leads to high numbers of
timely TP alerts at the expense of high numbers of FP alerts.
PLUM calculates MMI differently from the alerting threshold
being used here. Specifically, PLUM incorporates PGA as well
as peak ground velocity (PGV) data and uses the vector maxi-
mum of the three components of data (north–south, east–
west, and up–down). By contrast, for this warning time analy-
sis we compute MMI from the data following the ShakeMap
implementation, which uses the maximum of the horizontal
components of PGV in Worden et al. (2012).These
Figure 3. Modified Mercalli Intensity (MMI) maps predicted by
(from left to right) Earthquake Point-source Integrated Code
(EPIC), Finite-Fault Rupture Detector (FinDer), Propagation of
Local Undamped Motion (PLUM), and Solution Aggregator (SA)
(EPIC and FinDer) compared to the reference ShakeMap (right)
for four scenarios. (a) M7.1 Ridgecrest earthquake with 10 s
earlier M5.4 foreshock; (b) M6.0 earthquake in El Monte (Los
Angeles) followed 60 s later by M7.1 Ridgecrest earthquake; (c)
M7.1 Ridgecrest earthquake followed 60 s later by M6.0
earthquake in El Monte; and (d) M7.1 Ridgecrest earthquake
with 60 s earlier M5.4 foreshock simulated at an offshore
distance of 125 km. MMI 3.5 and 4.5 contour lines marking the
IV and V alert areas are shown by dashed and solid black lines,
respectively, and other MMI contours are shown as solid colored
lines. The spatial extent of these maps is the same as in Figure 1.
Aside from the first scenario with a small foreshock, the SA alert
areas (fourth column) are well estimated. PLUM tends to over-
estimate the alert areas (here for MMI V). This behavior, however,
depends on how the PLUM alert regional extents are defined:
there is a trade-off between detection speed and the spatial
extent of the alerting region, which is controlled by the radial
distance that ground motions are extrapolated (here set to
30 km) from the observing station.
12 Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
differences in MMI sometimes result in lower predicted MMI
values and higher apparent FP rates. The number of FP alerts
for PLUM trades off with the warning times through the alert
radius (here 30 km). For larger values of this radius, the warn-
ing times would be longer (Cochran et al., 2022), but the FP
would also increase.
For the scenario of the El Monte event occurring first, the
algorithms provide on average between ∼30 s (PLUM) and
>100 s (EPIC, FinDer, SA) of warning to sites experiencing
4.5 ≤MMI ≤5.5 when using an MMIalert 2:5 threshold
(Fig. S2b, green lines). For MMIalert 3:5 these numbers drop
to ∼25 s (PLUM), ∼45 s (FinDer), and ∼80 s (EPIC, SA),
respectively. For the scenario of the El Monte event occurring
second, the algorithms provide on average between <30 s
(PLUM) and ∼60 s (EPIC, FinDer, SA) of warning to sites
experiencing 4.5 ≤MMI ≤5.5 when using an MMIalert 2:5
threshold (Fig. S2c). For MMIalert 3:5 these numbers drop to
∼20 s (PLUM), ∼35 s (EPIC), and ∼45 s (FinDer, SA), respec-
tively. PLUM continues providing a positive median warning
time (∼10 s) for MMIalert 4:5.
Observations for the 125 km offshore scenario (data set
number 3; Fig. 4d) are similar for EPIC and FinDer at
MMIalert 2:5, but FinDer and PLUM show increased
TP(untimely) alerts at MMIalert 3:5 and 4.5 compared to
EPIC. PLUM has more TP(timely) alerts than the other algo-
rithms at the MMIalert 4:5 threshold, at the cost of an
increased percentage of FP. For MMIalert 2:5 the algorithms
provide on average between ∼25 s (PLUM) and >100 s (EPIC,
FinDer, SA) of warning to sites experiencing 4.5 ≤MMI ≤5.5
(Fig. S2d). For MMIalert 3:5 these numbers drop to ∼25 s
(PLUM), ∼30 s (EPIC), and ∼45 s (FinDer, SA), respectively.
PLUM continues providing a positive but very small median
warning time (∼3 s) for MMIalert 4:5.
The alerting performance of EPIC, FinDer, and SA is generally
similar. For MMIalert 2:5 they achieve a successful alerting per-
formance with 30%–50% TP (timely) and 50%–70% TP (timely,
<MMItw ) and warning times >20 s. Differences in algorithm
source parameter estimates are somewhat muted by the
Figure 4. Algorithm alerting performance broken down by
warning time at each station for the four scenarios in Figure 3for
ground-motion warning threshold MMItw 4:5 and alerting
thresholds of MMIalert 2:5 (left panels), MMIalert 3:5 (middle
panels), and MMIalert 4:5 (right panels). (a) M7.1 Ridgecrest
earthquake with 10 s earlier M5.4 foreshock; (b) M6.0
earthquake in El Monte (Los Angeles) followed 60 s later by
M7.1 Ridgecrest earthquake; (c) M7.1 Ridgecrest earthquake
followed 60 s later by M6.0 earthquake in El Monte; and (d)
M7.1 Ridgecrest earthquake with 60 s earlier M5.4 foreshock
simulated at an offshore distance of 125 km. X-axis gives the
fraction of cases: FN, false negative (missed alert); FP, false
positive (false alert); TN, true negative; and TP, true positive.
Different rows for a given algorithm correspond to minimum
warning times (labeled on right). The total number of stations
per scenario is given in the brackets. Only southern California
stations are included. The lowest MMI observed is 2.8, slightly
higher than the MMIalert level of 2.5, with the consequence that
the MMIalert 2:5 plots do not have FP or TN.
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 13
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
simplifications of alert settings (i.e., MMI thresholding) for
MMIalert 2:5 but become more apparent for the higher
MMIalert 3:5. By contrast, due to its fundamentally different
design and configuration (groundmotionsareassumedtobe
undamped to a user-configured distance) warning times in
PLUM are usually shorter (Fig. S2). For MMIalert 3:5and
4.5, EPIC, FinDer, and SA correctly alert (TP(timely)) up to
20% and 40% of sites, respectively, mostly with warning times
<5 s; PLUM achieves up to 40% TP(timely) alerts, but usually
with short warning times. Because of overestimated ground
motions (mostly by PLUM) for MMIalert 4:5, the algorithms
create up to 35% false alerts (FP) and miss 5%–45% alerts
(FN) due to reduced warning times resulting from the higher
MMIalert level.
Further general differences between PLUM’s alerting per-
formance and that of the other algorithms are that ground
motions are likely to be overestimated rather than underesti-
mated (because the estimates are undamped), resulting in a
higher FP percentage (30%–35%) and lower FN percentage
(2%–10%). PLUM may overestimate MMI by up to 4 units,
which is much larger than is observed for EPIC or FinDer
(Fig. S3). Beneficially this also results in more TP warnings
(up to 40% timely and up to 10% untimely) for the higher
MMIalert 4:5. Furthermore, for MMIalert 2:5, there is the
potential for a higher FN percentage and TP(untimely) com-
pared to the other algorithms, because PLUM’s MMI alert
thresholds (4 and 3 at the first and second station, respectively)
are higher than 2.5.
Discussion
In all scenarios studied here, the warning times are for large
M7.1 earthquake cases where moderate shaking is widespread
far from the epicenter. Warning times for smaller mainshocks
will be shorter (e.g., Chung et al., 2020) and TP(timely) alerts
will make up a much smaller fraction of the results, particularly
for high MMIalert. We purposely selected the larger M7.1
earthquake as a focus of this study to gain a better understand-
ing of how ShakeAlert might behave in earthquake scenarios
with high damage potential affecting large areas.
Systematic testing of ShakeAlert is a critical component of
its ongoing evolution and improvement (Cochran et al., 2017).
The growing suite of data used within the ShakeAlert testing
framework seeks to investigate scenarios not yet seen by the
system, alongside well-recorded historic events of interest.
Here, we created data sets of synthesized earthquake pairs
designed to systematically test scenarios that are expected to
expose limitations in algorithm design. Our data sets are
restricted to shallow, crustal, strike-slip earthquake signals that
occurred in well-instrumented (data sets number 1 and num-
ber 2) and not well-instrumented (data set number 3) areas of
the ShakeAlert alerting region. Although the events used as
building blocks all occurred in southern California, the perfor-
mance is expected to be similar for crustal earthquakes in
northern California, Oregon, and Washington and in regions
of similar geologic composition. Our data sets do not include
signals for interface or intraslab earthquakes (a major source of
seismic hazard for Oregon and Washington from the Cascadia
subduction zone) because of a dearth of such events with good
records from the US west coast. In addition, we expect any
targeted improvements to algorithms based on our data sets
(which refer to event detection and association of estimates
from multiple algorithms in earthquake sequences) to also
apply to other faulting styles, though future study into those
scenarios may be warranted.
In real aftershock sequences following a large earthquake we
will have to reckon with several hundreds to thousands of events
within short time periods. Associating phase picks to a particular
event under such extreme conditions is highly challenging.
Several solutions have been proposed: based on a Bayesian for-
mulation, Liu and Yamada (2014) proposed a likelihood function
suitable for classifying multiple concurrent earthquakes using
amplitude information. Tamaribuchi et al. (2014) constructed
a likelihood function using the source parameters and solved this
optimization problem using the particle filter method. Wu et al.
(2014) developed a two-step algorithm with a Bayesian model
class selection scheme to estimate the number of concurrent
events and the Rao–Blackwellized Importance Sampling method
to estimate the earthquake parameters. Applied to the first
two months of the M9 Tohoku-Oki sequence, this algorithm
produced far fewer incorrect warnings (over 90% of incorrect
warnings removed) compared to what the JMA EEW system
achieved. Yamada et al. (2021) proposed the Extended
Integrated Particle Filter method to deal with continuous wave-
forms and merge all Japanese real-time seismic networks into a
single framework. Roh (2021) proposed an envelope-based two-
part search algorithm that matches templates to the incoming
observed ground-motion envelopes to find the optimal estimates
of the earthquake source parameters. Ross et al. (2018) demon-
strated that convolutional neural networks are extremely sensi-
tive and robust in detecting seismic phases even in the presence
of high background noise. Applying their deep learning
PhaseLink approach to data sets in California and Japan, Ross
et al. (2019) showed that phases can be precisely associated to
events that occur only ∼12 s apart in their origin time.
Algorithms like PLUM and FinDer can have advantages in
intense earthquake sequences because they do not rely on phase
picks. Retrospective tests for the M9 Tohoku-Oki and M7.1
Kumamoto earthquake sequences have shown that PLUM
may perform well during intense seismicity (Kodera et al.,
2016,2018). PLUM was, therefore, introduced into the opera-
tional JMA EEW system in 2018. In our results, we find that
FinDer and PLUM indeed respond accurately to the ground-
motion continuum of near-simultaneous earthquakes, and
PLUM successfully identifies pockets of strong ground motion.
A central design strategy of ShakeAlert is to combine multi-
ple algorithms (currently EPIC and FinDer) to leverage their
14 Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
complementary strengths and to compensate for their weak-
nesses. This study has revealed two weaknesses in the current
ShakeAlert implementation. First, we found that only in a few
scenarios are the algorithms capable of differentiating between
temporally and spatially close earthquakes. This can lead
to incorrectly associated algorithm reports from EPIC and
FinDer, mainly because the current FinDer code cannot proc-
ess multiple earthquakes at the same time. FinDer will always
start reporting parameters for the first event. If a second earth-
quake initiates within 2 or 3 min afterward, FinDer will merge
both events into a single detection and assign the same event
ID: the epicenter and origin time will refer to the first event, the
line-source model (and magnitude) to the largest event.
Because the SA currently associates estimates from EPIC and
FinDer based on epicenter locations, this can mean that in
some (extreme) scenarios the reports are not, or are incor-
rectly, associated, which may cause a significant over- or
underestimation of SA magnitudes and alert areas.
An alternative association method has recently been devel-
oped, which uses the time and location information of the
ground-motion observation data for each algorithm alert to
determine if different algorithms are likely observing the same
earthquake. This will hopefully obviate problems of missing
or incorrect alert association from multiple algorithms in the
future. The described inconsistency between FinDer epicenter
and line-source estimates was corrected after this study was com-
pleted, and the new code will be pushed to the ShakeAlert
production system in the next software update.
As a second weakness of the current ShakeAlert implementa-
tion, we found that in some scenarios the SA magnitude does not
reach the current MSA 6.0 magnitude-weighting threshold, and so
the FinDer solution is incorrectly downweighted (meaning that
the true magnitude is greater than 6, but the estimated is smaller
than 6). This happens in scenarios in which EPIC misses the
mainshock, either because of a smaller immediate foreshock (data
set number 1) or if the waveforms from earthquakes in different
source regions overlap (data set number 2). In other scenarios,
EPICmissedeventsandFinDerisnotallowedtoalertalone.
In these scenarios ShakeAlert would not issue a warning at all.
To address the issue of differentiating events close in space
and time, changes could be made to individual algorithms
and to the method of combining algorithms. First, the FinDer
code could be modified to handle concurrent earthquakes,
and logic added for separating peak amplitudes into multiple
events. In addition, when events are very close in time and cannot
be separated, FinDer alerts could be modified so that line-source
and epicenter estimates are consistent (the latter was imple-
mented after this study was completed). These changes should
improve the likelihood of correct association with EPIC solutions.
Second, continuing the SA development of using ground-motion
data rather than source parameters to aggregate alerts would both
allow integration of PLUM into ShakeAlert and reduce the
dependence on consistent source parameter estimation.
To reduce the impact of incorrect magnitude estimation due
to misassociation of algorithm solutions, we suggest several ave-
nues for revisiting algorithm weighting and alert-alone strategies.
First, focus on allowing areas with high estimated ground motion
to be alerted, even if source parameters from single algorithms
cannot be reconciled. This also includes exploring how PLUM
could help to confirm large ground motions. Second, pursue
additional ground-motion metrics and visualizations (e.g., sum-
mary values that emphasize the accuracy and timeliness for high
hazard sites) for assessing system performance and directing
system development.
Finally, we suggest including a subset of 25 representative
scenarios from our 73 composites as part of the baseline test suite
for future testing and certification in ShakeAlert (Cochran et al.,
2017) and possibly other EEW systems (see Data and Resources).
These events (marked in Tables 2–4) are effective stress tests for
EEW algorithms with known challenges and measurable metrics
for success. Several scenarios are known to cause algorithms to
fail, which can be used to test future improvements, and we show
that other scenarios result in successful alerts, which can serve as
a backstop against accidental regression in future system updates.
The identification of the source parameters of individual
fore- and aftershocks is generally difficult to correctly obtain
and in these complex situations the EEW algorithms can often
estimate ground motion better than source parameters. Using
EPIC, FinDer, and PLUM in concert will likely provide more
robust ground-motion estimates and optimized warning times.
This is also in line with the study of Meier et al. (2020), which
compared performance of the three algorithms in large crustal
earthquakes in Japan and concluded that they may have com-
plementary strengths.
Even with research-based improvements as described here,
EEW systems may be unable to successfully detect individual
aftershocksforseveralminutesorevenlongerafteramajor
earthquake. Postalert messaging such as proposed by McBride
et al. (2020) is, therefore, all the more important: “… given
the number of aftershocks, you may not receive an alert for
all earthquakes. If you feel shaking, take protective action like
Drop, Cover, and Hold On.”
Conclusions
We have tested the performance of the EPIC, FinDer, and
PLUM EEW algorithms and ShakeAlert SA (EPIC + FinDer)
using ground-motion records from synthesized earthquake
pairs, which we created from well-recorded earthquakes. We
examined three different scenario data sets: fore- and aftershock
sequences (data set number 1, 27 scenarios), near-simultaneous
events in different source regions (data set number 2, 18 scenar-
ios), and simulated out-of-network and offshore earthquakes
(data set number 3, 28 scenarios). Some of our findings were
as expected, for example, EPIC typically detected the events
fastest and FinDer provided the best magnitude estimates
and line-source models for the larger events (M>6). Our study
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 15
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
also revealed that earthquakes in quick succession have the
potential to contribute to the underestimation of the EPIC earth-
quake magnitudes, suggesting different weighting schemes in
ShakeAlert could be explored to address underestimation.
PLUM performed well when events were in different source
regions. Although the PLUM algorithm was not able to identify
offshore ground motions, it successfully identified large onshore
motions and has the potential to warn of high MMI shaking in
low station-density areas. This suggests that PLUM’sdetection
and alert information can potentially be leveraged to confirm or
improve the estimates from other source-based algorithms.
In terms of ground-motion MMI threshold alerting, we find
that at the lower MMIalert 2:5 threshold the performance of
EPIC, FinDer, and SA is generally similar and differences in algo-
rithm source parameter estimates are somewhat muted, but these
differences become more apparent with the higher MMIalert val-
ues. By contrast, due to its fundamentally different design and
configuration, PLUM warning times for offshore earthquakes,
or earthquakes in regions of sparse station coverage, are generally
shorter (typically <5 s) than those provided by the other algo-
rithms. However, for earthquakes in regions of dense station cov-
erage, PLUM’swarningtimesareonparwiththeother
algorithms. PLUM can provide the highest probability for timely
TP at close epicentral distance in these sequence scenarios, as well
as for offshore scenarios with MMIalert >4:5. However, this is at
the expense of a significantly increased likelihood of overalerting
sites at larger distances from the epicenter. In general, we find that
temporally and spatially close earthquakes are difficult to indi-
vidually identify and that for scenarios studied in this work
the algorithms can often estimate ground motion better than they
estimate source parameters. This also motivates the development
of a ground-motion aggregator for future inclusion in ShakeAlert.
Data and Resources
The seismic waveform data for our building blocks (Table 1)wasdown-
loaded from the Southern California Earthquake Data Center (SCEDC,
https://scedc.caltech.edu, last accessed March 2022). The ground-motion
waveformsforthe25scenariosthatwe recommend becoming part of the
baseline data set for testing and certification in ShakeAlert and poten-
tially other earthquake early warning (EEW) systems can be downloaded
as miniSEED files from the Southern California Earthquake Data Center
at Caltech: https://scedc.caltech.edu/data/eewtesting.html#sim_seq (last
accessed July 2022). We constructed our reference ShakeMaps in
Figure 3from U.S. Geological Survey (USGS) ShakeMaps downloaded
from https://earthquake.usgs.gov/data/shakemap/ (last accessed
November 2021). Additional information about the Earthworm
Tankplayer software can be found at http://www.earthwormcentral.
org/ (last accessed July 2022). The supplemental material includes pre-
dicted peak Modified Mercalli Intensity (MMI for data set number 3, and
additionalwarningtimeandground-motionplots,aswellasvideosillus-
trating the alerting performance of Earthquake Point-source Integrated
Code (EPIC), Finite-fault Rupture Detector (FinDer), Solution
Aggregator (SA), and Propagation of Local Undamped Motion
(PLUM) for the scenarios in Figure 3. The material also includes a table
and map showing M 5+ earthquake pairs extracted from the Advanced
National Seismic System (ANSS) ComCat (Guy et al., 2015;https://
earthquake.usgs.gov/data/comcat/, last accessed January 2022) catalog
[1975–2021]).
Declaration of Competing Interests
The authors acknowledge that there are no conflicts of interest
recorded.
Acknowledgments
This material is based upon work supported by the U.S. Geological
Survey (USGS) under Grant/Cooperative Agreement Numbers
G19AC00252 and G21AC10532 to ETH Zurich, G19AC00296,
G21AC10561, and G19AC00125-04 to Caltech, an Intergovernmental
Personnel Act (IPA) to UC San Diego, and G21AC10525 to UC
Berkeley. Any use of trade, firm, or product names is for descriptive
purposes only and does not imply endorsement by the U.S.
Government. This work benefited from helpful discussions with mem-
bers of the ShakeAlert team. The authors would like to thank Minh
Huynh, Deborah Smith, and Andrew Good for developing the codes
to generate the warning time figures. The authors would also like to
thank Grace A. Parker and Geneva Chong (USGS), two anonymous
reviewers, and Editor-In-Chief Allison Bent for their thorough reviews.
References
Allen, R. (1982). Automatic phase pickers: Their present use and
future prospects, Bull. Seismol. Soc. Am. 72, no. 6B, S225–S242,
doi: 10.1785/BSSA07206B0225.
Allen, R. M., S. Allen, Q. Kong, S. Patel, R.F. Mejia, S. Pothan, J. A.
Strauss, and S. Thompson (2020). MyShake: Lessons from the first
year of public earthquake early warning delivery in California,
American Geophysical Union (AGU), Fall Meeting 2020,1–17
December 2020, abstract #S044-01.
Atkinson, G. M., and D. M. Boore (2011). Modifications to existing
ground-motion prediction equations in light of new data, Bull.
Seismol. Soc. Am. 101, no. 3, 1121–1135.
Bolt, B. A. (1968). The focus of the 1906 California earthquake, Bull.
Seismol. Soc. Am. 58, no. 1, 457–471.
Boore, D. M., and G. M. Atkinson (2008). Ground-motion prediction
equations for the average horizontal component of PGA, PGV,
and 5%-damped PSA at spectral periods between 0.01 s and
10.0 s, Earthq. Spectra 24, no. 1, 99–138.
Böse, M., D. E. Smith, C. Felizardo, M.-A. Meier, T. H. Heaton, and J.
F. Clinton (2018). FinDer v.2: Improved real-time ground-motion
predictions for M2-M9 with seismic finite-source characterization,
Geophys. J. Int. 212, 725–742, doi: 10.1093/gji/ggx430.
Chung, A., M. Meier, J. Andrews, M. Böse, B. Crowell, J. McGuire, and
D. Smith (2020). ShakeAlert earthquake early warning system per-
formance during the 2019 Ridgecrest earthquake sequence, Bull.
Seismol. Soc. Am. 110, no. 4, 1904–1923, doi: 10.1785/0120200032.
Chung, A. I., I. Henson, and R. M. Allen (2019). Optimizing earth-
quake early warning performance: ElarmS-3, Seismol. Res. Lett. 90,
no. 2A, 727–743, doi: 10.1785/0220180192.
Cochran, E. S., J. Bunn, S. Minson, A. S. Baltay, D. L. Kilb, Y. Kodera,
and M. Hoshiba (2019). Event detection performance of the
PLUM earthquake early warning algorithm in southern California,
Bull. Seismol. Soc. Am. 109, no. 4, 1524–1541.
16 Seismological Research Letters www.srl-online.org •Volume XX •Number XX •–2022
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
Cochran, E. S., M. D. Kohler, D. D. Given, S. Guiwits, J. Andrews, M. -
A. Meier, M. Ahmad, I. Henson, R. Hartog, and D. Smith (2017).
Earthquake early warning ShakeAlert system: Testing and certif-
ication platform, Seismol. Res. Lett. 89, no. 1, 108–117, doi:
10.1785/0220170138.
Cochran, E. S., J. K. Saunders, S. E. Minson, J. Bunn, A. Baltay, D. Kilb,
C. O’Rourke, M. Hoshiba, and Y. Kodera (2022). Alert optimization
of the PLUM earthquake early warning algorithm for the western
United States, Bull. Seismol. Soc. Am. doi: 10.1785/0120210259.
Given, D., R. M. Allen, A. S. Baltay, P. Bodin, E. S. Cochran, K.
Creager, R. M. de Groot, L. S. Gee, E. Hauksson, T. H. Heaton,
et al. (2018). Implementation plan for the ShakeAlert system—
An earthquake early warning system for the West Coast of the
United States, U.S. Geol. Surv. Open-File Rept. 2018-1155.
Guy, M., J. Patton, J. M. Fee, M. Hearne, E. M. Martinez, D. Ketchum,
C. B. Worden, V. Quitoriano, E. J. Hunter, G. M. Smoczyk, et al.
(2015). National Earthquake Information Center systems overview
and integration, U.S. Geol. Surv. Open-File Rept. 2015–1120, 25 pp.
Hoshiba, M., K. Iwakiri, N. Hayashimoto, T. Shimoyama, K. Hirano,
Y. Yamada, Y. Ishigaki, and H. Kikuta (2011). Outline of the 2011
off the Pacific coast of Tohoku earthquake (Mw 9.0)—Earthquake
early warning and observed seismic intensity, Earth Planets Space
63, no. 7, doi: 10.5047/eps.2011.05.031.
Kilb, D., J. J. Bunn, J. K. Saunders, E. S. Cochran, S. E. Minson, A.
Baltay, C. T. O’Rourke, M. Hoshiba, and Y. Kodera (2021). The
PLUM earthquake early warning algorithm: A retrospective case
study of West Coast, USA, Data, J. Geophys. Res. 126, no. 7,
e2020JB021053.
Kodera, Y., J. Saitou, N. Hayashimoto, S. Adachi, M. Morimoto, Y.
Nishimae, and M. Hoshiba (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, doi: 10.1186/
s40623-016-0567-1.
Kodera, Y., Y. Yamada, K. Hirano, K. Tamaribuchi, S. Adachi, N.
Hayashimoto, M. Morimoto, M. Nakamura, and M. Hoshiba
(2018). The propagation of local undamped motion (PLUM)
method: A simple and robust seismic wavefield estimation
approach for earthquake early warning, Bull. Seismol. Soc. Am.
108, no. 2, 983–1003.
Kohler, M., D. Smith, J. Andrews, A. Chung, R. Hartog, I. Henson, D.
Given, R. de Groot, and S. Guiwits (2020). Earthquake early
warning ShakeAlert 2.0, public rollout, Seismol. Res. Lett. 91,
doi: 10.1785/0220190245.
Kuyuk, H. S., and R. M. Allen (2013). A global approach to provide
magnitude estimates for earthquake early warning alerts, Geophys.
Res. Lett. 40, 6329–6333, doi: 10.1002/2013GL058580.
Liu, A., and M. Yamada (2014). Bayesian approach for identification
of multiple events in an early warning system, Bull. Seismol. Soc.
Am. 104, 1111–1121, doi: 10.1785/0120130208.
Lomax, A. (2005). A reanalysis of the hypocentral location and related
observations for the great 1906 California earthquake, Bull.
Seismol. Soc. Am. 95, no. 3, 861–877.
McBride, S. K., A. Bostrom, J. Sutton, R. M. de Groot, A. S. Baltay, B.
Terbush, P. Bodin, M. Dixon, E. Holland, R. Arba, et al. (2020).
Developing post-alert messaging for ShakeAlert, the earthquake
early warning system for the West Coast of the United States
of America, Int. J. Disaster Risk Reduct. 50, doi: 10.1016/
j.ijdrr.2020.101713.
Meier, M. A., Y. Kodera, M. Böse, A. Chung, M. Hoshiba, E. Cochran,
S. Minson, E. Hauksson, and T. Heaton (2020). How often can
earthquake early warning systems alert sites with high-intensity
ground motion? J. Geophys. Res. 125, no. 2, e2019JB017718.
Minson, S., J. Saunders, J. Bunn, E. Cochran, A. Baltay, D. Kilb, M.
Hoshiba, and Y. Kodera (2020). Real-time performance of the
PLUM earthquake early warning method during the 2019 M6.4
and 7.1 Ridgecrest, California, earthquakes, Bull. Seismol. Soc.
Am. 110, doi: 10.1785/0120200021.
Roh, B. (2021). Matching waveform envelopes for earthquake early
warning, Dissertation (Ph.D.), California Institute of Technology,
doi: 10.7907/hw8k-zx98.
Ross, Z., M.-A. Meier, E. Hauksson, and T. Heaton (2018). Generalized
seismic phase detection with deep learning, Bull.Seismol.Soc.Am.
108, doi: 10.1785/0120180080.
Ross, Z. E., Y. Yue, M. A. Meier, E. Hauksson, and T. H. Heaton
(2019). PhaseLink: A deep learning approach to seismic phase
association, J. Geophys. Res. Solid Earth 124, 856–869, doi:
10.1029/2018JB016674.
Stubailo, I., M. Alvarez, G. Biasi, R. Bhadha, and E. Hauksson (2020).
Latency of waveform data delivery from the southernCalifornia seis-
mic network during the 2019 Ridgecrest earthquake sequence and its
effect on ShakeAlert, Seismol. Res. Lett. doi: 10.1785/0220200211.
Tamaribuchi, K., M. Yamada, and S. Wu (2014). A new approach to
identify multiple concurrent events for improvement of earth-
quake early warning, Zisin 67, no. 2, 41–55 (in Japanese with
English abstract).
Thakoor, K., J. Andrews, E. Hauksson, and T. Heaton (2019). From
earthquake source parameters to ground-motion warnings near
you: The ShakeAlert earthquake information to ground-motion
(eqInfo2GM) method, Seismol. Res. Lett. 90, no. 3, 1243–1257,
doi: 10.1785/0220180245.
Trugman, D. T., M. T. Page, S. E. Minson, and E. S. Cochran (2019).
Peak ground displacement saturates exactly when expected:
Implications for earthquake early warning, J. Geophys. Res. 124,
no. 5, 4642–4653.
Wei, S., E. Fielding, S. Leprince, A. Sladen, J.-P. Avouac, D.
Helmberger, E. Hauksson, R. Chu, M. Simons, K. Hudnut, et al.
(2011). Superficial simplicity of the 2010 El Mayor-Cucapah earth-
quake of Baja California in Mexico, Nature Geosci. 4, 615–618, doi:
10.1038/ngeo1213.
Worden, C. B., M. C. Gerstenberger, D. A. Rhoades, and D. J. Wald
(2012). Probabilistic relationships between ground-motion param-
eters and modified Mercalli intensity in California. Bull. Seismol.
Soc. Am. 102, no. 1, 204–221, doi: 10.1785/0120110156.
Wu, S., M. Yamada, K. Tamaribuchi, and J. Beck (2014). Multi-events
earthquake early warning algorithm using a Bayesian approach,
Geophys. J. Int. 200, no. 2, 791–808.
Yamada, M., K. Tamaribuchi, and S. Wu (2021). The extended inte-
grated particle filter method (IPFx) as a high-performance earth-
quake early warning system, Bull. Seismol. Soc. Am. 111, no. 3,
1263–1272, doi: 10.1785/0120210008.
Manuscript received 21 March 2022
Published online 22 September 2022
Volume XX •Number XX •–2022 •www.srl-online.org Seismological Research Letters 17
Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220220088/5705171/srl-2022088.1.pdf
by UC San Diego Library, dkilb
on 22 September 2022
