of 8
740
S
eismological Research Letters
V
olume 80, Number 5
S
eptember/October 2009
doi: 10.1785/gssrl.80.5.740
INTRODUCTION
The Virtual Seismologist (VS) method is a Bayesian approach
to regional network-based earthquake early warning (EEW)
that estimates earthquake magnitude, location, and the distri-
bution of peak ground motion using observed ground motion
amplitudes, predefined prior information, and appropriate
attenuation relationships (Cua 2005; Cua and Heaton 2007).
The application of Bayes’s theorem in earthquake early warn-
ing (Cua 2005) states that the most probable source estimate
at any given time is a combination of contributions from prior
information (possibilities include network topology or station
health status, regional hazard maps, earthquake forecasts, the
Gutenberg-Richter magnitude-frequency relationship) and
a likelihood function, which takes into account observations
from the ongoing earthquake. Prior information can be con-
sidered relatively static over the timescale of a given earthquake
rupture. The changes in the source estimates and predicted
peak ground motion distribution, which are updated each sec-
ond, are due to changes in the likelihood function as additional
arrival and amplitude data become available. The potential use
of prior information differentiates the VS approach from other
regional, network-based EEW algorithms, such as ElarmS
(Allen and Kanamori 2003).
Implementation of the VS algorithm in California is
an ongoing effort of the Swiss Seismological Service (SED)
at ETH Zurich. We prioritized the development of codes
involved in real-time data processing, which corresponds to the
likelihood function in our Bayesian framework; code devel
-
opment to implement the contribution of prior information
is to follow. The VS algorithm is one of three early warning
algorithms being implemented and tested in real time as part
of the California Integrated Seismic Network (CISN) early
warning project; the other two are the ElarmS algorithm of
Allen and Kanamori (2003) and the onsite algorithm of Wu
and Kanamori (2005). These algorithms send reports to the
Southern California Earthquake Center (SCEC) early warn
-
ing testing Web site, which evaluates performance based on the
accuracy and time of availability of magnitude, location, and
peak ground motion estimates. Real-time testing will allow the
scientific community to establish whether EEW systems can
deliver reliable, timely information that can be used in post-
earthquake, pre-shaking damage mitigation. In this article,
we describe the VS likelihood function, its code architecture,
and processing flow and summarize its real-time performance
in terms of magnitude and location accuracy in southern
California from July 2008 through April 2009.
DESCRIPTION OF THE VS LIKELIHOOD FUNCTION
Conceptually, the VS likelihood function is a set of relation
-
ships used to map available arrival and ground motion envelope
amplitude information from an ongoing earthquake into esti-
mates of earthquake magnitude, location, depth, origin time,
and the distribution of peak ground shaking. All earthquake
source parameter and ground motion estimates are updated
each second, as additional data become available. A short-
term/long-term average (STA/LTA) algorithm based on Allen
(1978) is used for automatic picking. The
Binder
Earthworm
phase associator codes (Dietz 2002) are adopted to estimate
location, depth, and origin time based on available picks.
Magnitude estimation and ground motion prediction require
the following relationships derived by Cua (2005) and Cua and
Heaton (2007): 1) a
P
-
S
discriminant, 2) a single-station mag-
nitude estimate based on ground motion ratios, 3) envelope
attenuation relationships, and 4) a multiple-station magnitude
and location estimate. These relationships are based on ground
motion envelope values, which are defined as the maximum
absolute value on a given channel over a one-second window.
The functional forms for these relationships are given below.
r
e
al-time
p
er
formance of the Virtual
s
e
ismologist
e
a
rthquake
e
a
rly
w
a
rning
a
l
gorithm in
s
o
uthern
c
a
lifornia
Georgia Cua, Michael Fischer, Thomas Heaton, and Stefan Wiemer
Georgia Cua,
1
Michael Fischer,
1
Thomas Heaton,
2
and Stefan Wiemer
1
1.
S
wiss Seismological Service, ETH Zurich, Switzerland
2.
C
alifornia Institute of Technology, Pasadena, California, U.S.A.
Seismological Research Letters
V
olume 80, Number 5
S
eptember/October 2009
7
41
p
-
s
Discriminant
PSZA
ZV
HA
=
+
0 4
0 55
0 46
0 5
10
10
10
. log ( ) . log ( )
. log ( ) . 55
0 1
10
log ( )
.
HV
PSPS
if
then wave, else wave.
>−
(
1)
In Equation 1,
ZA,
Z
V,
H
A,
and
HV
denote the vertical
acceleration, vertical velocity, and root mean square (rms) hori-
zontal acceleration and rms horizontal velocity envelope val-
ues, respectively. This relationship quantifies the concept that
P
waves will have larger amplitudes on the vertical channel, while
S
waves will have larger amplitudes on the horizontal channels.
Single-station Magnitude Estimate
ZADZA
ZD
P
M
ZAD
=
=−
0 36
0 93
1
10
10
. log ( ) . log ( )
if wave
..
. ,
.
.
63
8 94
0 45
1 46
⋅ +
=
=− ⋅
ZAD
S
M
ZAD
M
ZAD
ZAD
σ
if wave
++
=
8 05
0 41. ,
.
σ
M
ZAD
(
2)
In Equation 2,
ZAD
is the ground motion ratio between
the vertical acceleration (
ZA
) and vertical displacement (
ZD
)
envelope values that best correlates with magnitude (Cua 2005;
Cua and Heaton 2007).
ZAD
is inversely proportional to the
size of the event. It is relatively larger for small, point-source-
type events, which are richer in high frequency energy, and
smaller for events that require finite rupture characterization,
which are richer in lower frequency energy. The single-station
magnitude estimate
M
ZAD
can be calculated as soon as two
seconds of
P
-wave amplitude data are available following the
STA/LTA pick.
Envelope Attenuation Relationships
log ( , )
( ) log ( )
( )
( )
(
10
10
YMRaMbRMdRMe
RMRCM
CM
= + +
+
= +
))
arctan(
) exp
= ⋅
− ⋅ ⋅ −
( )
(
)
cMcM
1
5
2
5
(
3)
In Equation 3,
M
denotes magnitude, while
R
is epicen-
tral distance in km for
M
<
5
e
vents, and closest distance
to the fault or Joyner-Boore distance (Boore and Atkinson
2008), when available for
M
>
5
events. There are 24 sepa
-
rate sets of coefficients (
a,
b
,
c
1
,
c
2
,
d
,
e
)
for maximum
P
- and
S
-wave amplitudes for horizontal and vertical channels of
acceleration, velocity, and displacement on rock and soil sites.
These envelope attenuation relationships are valid up to 200
km away from events in the magnitude range 2
<
M
<
8
. They
are used in the multiple-station magnitude and location esti-
mation step (next section), as well as for predicting the geo
-
graphical distribution of peak ground acceleration and veloc-
ity given a magnitude and location estimate. Derivation of
these extended magnitude relationships is described in Cua
and Heaton (2009).
Multiple-station Magnitude and Location Estimate
L M
L M
L
ij
j
P S
i
( , , )
(
)
,
lat lon
,lat,lon
stations
=
==
∑∑
11
((
)
(
( ))
(log
,
M
ZAD
Z M
Y
ij
j
ZAD
obs i
i
,lat,lon
=
+
2
2
10
2
σ
jjk
ijk
ijk
k
Y
M
=
log
(
)
10
2
2
1
4
2
,lat,lon
σ
(
4)
i
n
n
=
1, ,
where is the number of stations
wwith detections
wave, wave
(for
P
j P
S
k
=
=
1 4, ,
ZZV, HA, HV, HDchannels)
Z M
M
iP
( )
.
. ,
=− +
0 62 5 50
σ
ZZ M
iS
Z M
P
S
Z M
M
( )
( )
.
( )
.
. ,
.
=
=− +
=
0 28
0 69 5 52
0 25
σ
In Equation 4,
ZAD
is as described in Equation 2, log
10
Y
ijk
is as
described in Equation 3, and log
10
Y
obs,ijk
are available observed
envelope amplitudes on
ZV,
H
A,
H
V,
and
H
D
channels. The
equations for
Z
(
M
)
iP
and
Z
(
M
)
iS
are the inverses of the single-
station magnitude relationships
M
ZAD
in Equation 2.
The magnitude and location coordinates that minimize
Equation 4 correspond to the most probable magnitude (
Mvs
)
and location estimates given the available observed envelope
values. The location estimate corresponds to a strong motion
centroid, which is the location of a point source event that
best fits the distribution of ground motion amplitudes given a
magnitude estimate and a particular attenuation relationship
(Kanamori 1993). This centroid location estimate is extremely
robust when constrained by a large number of observations
but not very stable or precise when using data from only a few
stations. Instead, we use the
Binder
location estimate (Dietz
2002), which then reduces the determination of
Mvs
to a
1-dimensional search over magnitude space in Equation 4.
The reader is referred to Cua (2005), Cua and Heaton
(2007), and Cua and Heaton (2009) for details on the deriva
-
tion of these various relationships.
The offline performance of the full Bayesian VS approach
(including contributions from both the likelihood function
and the prior term) on waveform datasets recorded by the
SCSN from the 1999
M
7.3 Hector Mine, 2002
M
4.8 Yorba
Linda, 2003
M
6.5 San Simeon, and 2004
M
6.0 Parkfield
earthquakes is discussed in detail in Cua (2005) and Cua and
Heaton (2007).
SYSTEM ARCHITECTURE
Hauksson
et
a
l.
(
2006) developed a real-time processing envi-
ronment that provides an interface between the real-time net-
work data streams and the EEW algorithms participating in
the CISN project. What we refer to collectively as the VS codes
are the three subsystems (or collections of modules) shown in
742
S
eismological Research Letters
V
olume 80, Number 5
S
eptember/October 2009
Figure 1. Each subsystem has a particular task. The
Waveform
C
ollector
subsystem performs basic waveform processing such
as picking, gain correction, baseline removal, filtering, and
down-sampling. The ground motion envelope amplitudes of
acceleration, velocity, and filtered displacement (AVDs in
Figure 1) required as inputs to the VS likelihood function (Cua
and Heaton 2007) are calculated here. The
Messenger
subsys-
tem sends information from the
Waveform
C
ollector
subsystem
to the
EEWVS
(which stands for Earthquake Early Warning
Virtual Seismologist) subsystem. The
EEWVS
subsystem 1)
filters and weighs incoming picks, 2) estimates location and
origin time based on acceptable picks using the Earthworm
Binder
phase associator, 3) estimates magnitude given the
Binder
location estimate and the available envelope amplitudes
using the VS likelihood function relationships, 4) evaluates the
reliability of the magnitude and location estimate, and 5) logs
the estimated magnitude, location, and predicted peak ground
shaking to an event summary file. All event summary files are
stored locally for subsequent performance analysis. Event sum-
mary files within initial magnitude estimates larger than
M
2.7 are automatically sent to the SCEC early warning testing
Web site. In an operational early warning system, step 5 would
include pushing the EEW information to users.
DEALING WITH NOISE IN REAL TIME
To maximize the available warning time, real-time early warn
-
ing algorithms must decide whether the available data is from
an earthquake or noise as early as possible. The following auxil
-
iary algorithms were developed or adopted to assist in this task:
1) pick and envelope filtering, 2) the
Binder
Earthworm phase
associator codes (Dietz 2002), and 3) event filtering.
Pick and Envelope Filtering
The VS installation at the SCSN sees real-time data from 170
broadband and strong motion stations; automatic STA/LTA
pickers (Allen 1978) report an average of 50,000 picks each
day. Ninety percent of these picks are unrelated to earthquake
activity. The
Binder
associator (Dietz 2002) can determine
whether a given pick can be associated with an earthquake
hypocenter or is due to noise, but only when a large number of
picks are available. In the interest of minimizing the number
of false alarms, we assign quality factors to the incoming picks
based on the signal-to-noise ratio. This section describes some
empirically determined criteria we have developed for assigning
pick quality to picks coming from a standard STA/LTA (Allen
1978) picker algorithm. In the VS codes, an event is declared
once four valid (with sufficient quality) picks are available.
The first set of requirements for a pick to be valid are: 1)
the maximum velocity on the vertical channel exceeds 0.0001
cm/s (which is the maximum vertical
P
-wave velocity expected
from an
M
3.0 event 50 km away using the Cua and Heaton
(2009) envelope attenuation relationships), and 2) the average
velocity envelope amplitude three seconds after the pick time
exceeds the average velocity envelope amplitude before the pick
time. About 80% of raw picks are rejected by these criteria and
not used in location estimation; the rejected picks play a later
role in distinguishing between correct and incorrect event dec-
larations.
Picks meeting the above requirements are valid picks and
are given further quality assignments based on
qv
, the maxi
-
mum velocity within three seconds of the pick time divided
by the average background velocity (essentially the signal-to-
noise ratio), and
rp
, the hourly rate of raw picks reported by
the STA/LTA picker at the given station. Picks accepted by the
trigger discriminant function in Figure 2A are considered trig-
gering picks. At least one triggering pick and three valid picks
are required for the initial event declaration. Once an event is
declared, location, magnitude, and peak ground motion esti-
mation is initiated by the
EEWVS
subsystem.
The average background velocity is calculated on the verti-
cal velocity channel. We limit the maximum allowable value
Control & Information Flow
Information Flow
Binder
Earthworm
Filter
Quake
Filter
Pick
Filter
Envelope
Stations
Envelopes
Quakes
Picks
Seismologist
Virtual
Filter Picker Sampler
Waveforms
Messenger
EEWVS
AVDs
Picks
Location
Estimates
Picks
Calibration
Waveform Collector
Report
F
igure 1.
System architecture of the VS early warning algorithm. Rectangles represent processing modules, while drums represent
dynamic data areas.
Seismological Research Letters
V
olume 80, Number 5
S
eptember/October 2009
7
43
of average background velocity to 0.0005 cm/s; without this
empirically determined limit, the signal-to-noise ratio is a poor
indicator for pick quality during the aftershock sequences of
large events. Broadband channels reporting velocities larger
than 0.8 cm/s (approximately 80% of the clip level for an STS2)
are considered clipped, and subsequent amplitude data from
the broadband channels of that station are not used. At sta-
tions with both broadband and strong motion instruments, the
system switches to using the strong-motion waveforms when
the broadband instrument clips.
b
inder
Earthworm Phase Associator
The
Binder
module is the phase associator used by the
Earthworm network processing system (
http://www.isti2.com/
ew/).
Given a set of
P
-wave arrival times, station locations, and
a 1-D velocity model, it determines the smallest set of hypo-
centers consistent with the available picks (Dietz 2002). We
configured the
Binder
module to run as soon as at least one
triggering pick and three valid picks are available (see section
Pick and Envelope Filtering). The location (latitude, longitude,
depth, and origin time) estimated by
Binder
’s Simple Event
Locator (Dietz 2002) is passed to the Virtual Seismologist
module (Figure 1). The Virtual Seismologist module uses the
Binder
location and the available envelope amplitudes to esti-
mate magnitude and the distribution of peak ground accelera-
tion and velocity (currently without site conditions).
Event Filtering
The reliability of the
Binder
location estimate increases with
the number of available picks. The initial location estimates can
have potentially large errors, since it is possible that one or more
of the input picks is from a non-earthquake source. The event
filtering module checks whether the magnitude and location
estimates are consistent with the available envelope amplitudes
and picks. The goal of event filtering is to determine whether a
VS event declaration corresponds to a real earthquake (which
will eventually have a corresponding entry in the network
earthquake catalog) or constitutes a “phantom” event due to
random noise-related picks meeting the triggering criteria
(which will not match any local earthquake in the catalog but
may have an origin time that coincides with a teleseimic event
recorded by the network). An EEW algorithm should attempt
to maximize its number of real event detections while minimiz
-
ing the number of phantom event declarations. (Naturally, an
EEW algorithm should also attempt to maximize the available
warning time by minimizing the time between the earthquake
origin time and when the EEW information is available.)
The single-station magnitude estimates
M
ZAD
described
in Equation 2 are useful in distinguishing between real and
phantom events. We calculate the linear average of the most
current single-station magnitude estimates
M
ZAD,ave
available
from stations with valid picks. At stations close to the epicen-
ter, the most current single-station magnitude estimate may be
based on the
S
-wave envelope amplitudes; at farther stations,
these are based on
P
-wave envelope amplitudes. As discussed
in the previous section,
M
VS
is the most probable magnitude
estimate given the
Binder
location and the available envelope
amplitudes. We empirically determined that phantom events
typically have
MM
ZADVS
,ave
− >
1 5.
; this criterion can help
distinguish between local and teleseismic events.
For real earthquake sources (as opposed to phantom
events), we expect that picks will be available at most stations
with epicentral distances less than the farthest picked station.
Transmission delay or data latency due to different combina-
tions of the type of datalogger and communication link (
i.e.
,
radio link, internet, or satellite) between the station and the
central processing site ranges between 3 to 13 seconds (with
a mean of 6.5 seconds) at the SCSN (Allen 2008). To make
allowances for stations with low quality picks and the variation
in telemetry delays across different stations, we set the criterion
that at least half the stations with epicentral distance less than
−121
−120
−119
−118
−117
−116
−115
−114
32
33
34
35
36
37
38
CI.SBB2
CI.EDW2
CI.MPI
CI.ARV
CI.LJR
CI.SPF
CI.CHF
CI.MLS
CI.EML
qv
rp, raw picks per hour
triggering picks
non-triggering picks
(A)
(B)
longitude
latitude
threshold
d
F
igure 2.
(A) The trigger discriminant function as a function
of signal-to-noise ratio (qv) and the rate of raw picks. Only trig-
gering picks are allowed to contribute to the initial event decla-
ration. (B) Early warning algorithms must be able to distinguish
between local and teleseismic events. The event filter correctly
rejects the
Binder
location (star) for the
M
7.7 Sea of Okhotsk
since there are more unpicked stations (open circles) than
picked stations (filled circles) within
d
threshold
.
744
S
eismological Research Letters
V
olume 80, Number 5
S
eptember/October 2009
d
threshold
must report picks (regardless of pick quality) for the
event to be valid. In Equation 5,
d
RR
threshold
=
+
max
2
,
(
5)
R
max
is the epicentral distance of the farthest picked station,
and is the average epicentral distance of remaining picked sta-
tions.
Illustrating Event Filtering with M 7.7 Sea of Okhotsk
(Teleseismic) Event
An
M
7.7 earthquake in the Sea of Okhotsk that occurred on
2008/07/05 02:12:03 UTC produced triggers on multiple
SCSN stations. There were enough picks of sufficient quality
to initiate the calculation of a VS trial solution (magnitude,
location, and origin time). Figure 2B shows picked (filled cir-
cles) and unpicked (open circles) SCSN stations and the
Binder
epicentral location estimate (star) based on the available picks.
The large circle encloses the region with epicentral distance less
than
d
threshold
. Since less than half of the stations within this
region reported picks, the candidate event is correctly flagged
as a phantom event.
The average of the most current single-station magnitude
estimates,
M
ZAD,
ave
, is 5.15; the most probable magnitude
given the
Binder
location estimate and the available envelope
amplitudes,
M
VS
, is 3.49. Since
M
ZAD,ave
M
VS
=
1.66, the
e
vent is declared invalid.
Binder
allows for multiple candidate
hypocenters once enough picks are available. For this teleseis-
mic event, the
Binder
module proposed a total of seven candi-
date hypocenters in southern California (all based on only a
few picks), all of which had and were thus flagged as phantom
events and rejected.
Binder
would eventually withdraw these
erroneous hypocenters itself, once enough picks were available.
However, the
MM
ZADVS
,
.
ave
− >
1 5
criteria rejects these
trial locations much earlier.
If the trial source parameters and the available observa-
tions contradict each other, the event estimates are declared
invalid—the event is most likely a phantom event. If the trial
source parameters and the observations are consistent with
each other, the estimates are considered valid and are updated
every second until computation times out when no more new
picks are reported within a 10-second window.
RESULTS AND DISCUSSION
The VS codes described above have been running in real time
at the SCSN since 13 July 2008. Figure 3A shows the real-time
SCSN stations, real-time VS event detections (with magni
-
tudes
M
> 1.0), and
M
3.0 events missed by the VS codes
between 13 July 2008 and 9 April 2009. In generating the
performance statistics discussed in this section, we use events
listed in the SCSN earthquake catalog within the rectangu
-
lar region bounded by 31.5 and 37.5 latitude and –121.25 and
–114 longitude, since the codes are only installed in southern
California. Of the 1,220 reports generated by the VS algorithm
over this time period, 1,201 events (98.4%) had estimated ori
-
gin times within
±
30 seconds of an origin time listed in the
SCSN catalog; only 19 (1.6%) were “phantom events” that did
not match any local event in the catalog. The VS codes missed
60 out of 107
M
3.0 SCSN events over this time period; the
majority of these events are either on the outskirts or outside
the network, offshore, or in areas of low station density. Events
that occurred while the codes were offline (such as the 28 July
2008
M
5.4 Chino Hills earthquake) are also included in the
missed event count. The number of valid missed
M
3.0 events
(or events that the codes should have detected but did not, for
reasons yet unresolved) is on the order of 10 events. The evolu-
tion of VS magnitude and location estimates as a function of
time for the
M
5.4 28 July 2008 Chino Hills event and the
M
5.1 5 December 2008 event near Barstow, California (which
were the two largest events that occurred during the period
covered by this study) are shown in Figures 3B through 3E.
Figure 4A shows a histogram of when the initial VS esti
-
mates are available relative to the earthquake origin time.
Eighty-two percent and 96.5% of events detected by the VS
codes have initial estimate times within 25 and 30 seconds of
the origin time, respectively. The mean time to the initial VS
estimate is 21.9 seconds. These initial estimate times are domi-
nated by the time for four acceptable picks to be available; the
effects of telemetry delay—SCSN stations have a median telem
-
etry delay of 6.5 seconds (Allen 2008)—and processing time (~3
seconds) are also included. Figure 4B shows contours of approx
-
imate initial VS estimate time (time for
P
waves to propagate to
four stations + 6.5 seconds average telemetry delay + 3 seconds
processing time) in southern California given the locations of
current SCSN stations. More aggressive use of prior informa
-
tion can potentially decrease the initial estimate time and thus
increase the available warning time for potential users. In par-
ticular, the method of Voronoi cells and not-yet-arrived data can
potentially provide location estimates as early as the first
P
wave
is detected (Cua 2005; Cua and Heaton 2007; Horiuchi
et
a
l.
2005; Satriano
et
a
l.
2007). However, such approaches need
significant modification to work with networks such as SCSN
with non-uniform station telemetry delays.
The
Binder
module gives stable location estimates if noise-
related picks are properly filtered out by pick filtering and not
included in the initial event declaration. The median error of
the initial epicentral location estimate for the 1,201 detected
events is 2.6 km. Magnitude estimation performance is shown
in Figure 5. The CISN early warning project aims to evaluate
early warning algorithm performance for
M
3
.0
events. There
have not been many
M
3
.0
events since the codes started
operating; 1,094 out of 1,201 (91%) of the VS detected events
are in the microearthquake range with magnitudes
M
<
3
.0.
The VS algorithm is based on envelope attenuation relation-
ships derived from a dataset spanning the magnitude range
2.0
<
M
7
.5 (Cua and Heaton 2008). It can thus operate at
M
<
3
.0 level, with some systematic overestimation expected
due to the vicinity of the operating range to the lower mag-
nitude limit of the attenuation relationships (Bommer
et
a
l.
2007). Given a reasonable location estimate (94% of events have
initial location estimates within 20 km of the actual epicenter),
Seismological Research Letters
V
olume 80, Number 5
S
eptember/October 2009
7
45
(a)
−121
−12
0
−11
9
−11
8
−117
−11
6
−11
5
−114
31
32
33
34
35
36
37
38
100 km
SCSN EEW stations
real−time VS detection M < 3
real−time VS detection M
3
missed M
3
M5.4 7/29/2008 Chino Hills
M5.1 12/05/2008
(A)
0
10
20
30
40
50
60
0
5
10
15
20
25
30
35
40
45
50
VS estimate time (sec after origin time)
Epicentral location error (km
)
VS location error
M5.4 7/28/2008 Chino Hills (offline)
0
10
20
30
40
50
60
3
4
5
6
7
8
VS estimate time (sec after origin time)
Magnitude
M
VS
0
10
20
30
40
50
60
0
5
10
15
20
25
30
35
40
45
50
VS estimate time (sec after origin time)
Epicentral location error (km
)
VS location error
0
10
20
30
40
50
60
3
4
5
6
7
8
VS estimate time (sec after origin time)
Magnitude
M
VS
M5.1 12/05/2008 (real-time)
M5.4 7/28/2008 Chino Hills (offline)
M5.1 12/05/2008 (real-time)
(B)
(C)
(D)
(E)
F
igure 3.
(A) VS real-time performance in southern California during 13 July 2008 –9 April 2009. The polygon encloses the SCSN area
of responsibility (AOR). (B and C) Evolution of VS epicentral location and magnitude error, respectively, as a function of time for the
M
5.4 28 July 2008 Chino Hills earthquake. The VS codes were of fline at the time of the event due to a scheduled code upgrade. Results
shown were generated using a waveform dataset for the Chino Hills event downloaded from the Southern California Earthquake Data
Center (SCECDC). (D and E) Evolution of VS epicentral location and magnitude error, respectively, as a function of time for an
M
5.1 5
December 2008 earthquake near Barstow, California. Results shown were generated in real time.