Content uploaded by Debi Lyman Kilb
Author content
All content in this area was uploaded by Debi Lyman Kilb on Jul 07, 2014
Content may be subject to copyright.
Rapid Earthquake Characterization Using MEMS Accelerometers
and Volunteer Hosts Following the M7.2 Darfield,
New Zealand, Earthquake
by Jesse F. Lawrence, Elizabeth S. Cochran, Angela Chung, Anna Kaiser,
Carl M. Christensen, Richard Allen, Jack W. Baker, Bill Fry, Thomas
Heaton, Deborah Kilb, Monica D. Kohler, and Michela Taufer
Abstract We test the feasibility of rapidly detecting and characterizing earthquakes
with the Quake-Catcher Network (QCN) that connects low-cost microelectromechan-
ical systems accelerometers to a network of volunteer-owned, Internet-connected com-
puters. Following the 3 September 2010 M7.2 Darfield, New Zealand, earthquake we
installed over 180 QCN sensors in the Christchurch region to record the aftershock se-
quence. The sensors are monitored continuously by the host computer and send trigger
reports to the central server. The central server correlates incoming triggers to detect
when an earthquake has occurred. The location and magnitude are then rapidly esti-
mated from a minimal set of received ground-motion parameters. Full seismic time
series are typically not retrieved for tens of minutes or even hours after an event. We
benchmark the QCN real-time detection performance against the GNS Science GeoNet
earthquake catalog. Under normal network operations, QCN detects and characterizes
earthquakes within 9.1 s of the earthquake rupture and determines the magnitude within
1 magnitude unit of that reported in the GNS catalog for 90% of the detections.
Introduction
Over the past decade, several cyber-social-seismic net-
works have been developed, including the Personal Seismic
Network (Cranswick et al., 1993), NetQuakes (Luetgert et al.,
2009), the Quake-Catcher Network (QCN;Cochran, Law-
rence, Christensen, and Chung, 2009;Cochran, Lawrence,
Christensen, and Jakka, 2009), and the Community Seismic
Network (Clayton et al., 2011). New sensor technology and
computational techniques provide an avenue for creating very
large cyber-social-seismic networks by reducing instrument
costs, minimizing needed infrastructure, and harnessing
public interest. Small low-cost ($30–$3000) microelectrome-
chanical systems (MEMS) triaxial sensors provide ground-ac-
celeration measurements of moderate to large earthquakes
(Cochran, Lawrence, Christensen, and Chung, 2009;Co-
chran, Lawrence, Christensen, and Jakka, 2009;Chung et al.,
2011;Cochran et al., 2011). Data from these low-cost sensors
are transmitted to a central server either through an Internet-
connected computer or via any available wireless connection
(Luetgert et al., 2009;Cochran, Lawrence, Christensen, and
Chung, 2009;Cochran, Lawrence, Christensen, and Jakka,
2009;Clayton et al., 2011). These networks minimize the
costs associated with monitoring the sensors by utilizing
the host’s computing resources, A/C power, Internet, and shel-
ter (Luetgert et al., 2009;Cochran, Lawrence, Christensen,
and Chung, 2009;Cochran, Lawrence, Christensen, and
Jakka, 2009;Clayton et al., 2011).
The QCN represents one type of cyber-social-seismic
network. In the QCN architecture, MEMS sensors are con-
nected directly to Universal Serial Bus (USB) ports on a
host’s computer; the computer monitors the sensor and sends
time series and ground-motion parameters to a central server.
This is a low-cost paradigm compared to traditional sensor
networks and even other cyber-social-seismic networks such
as the NetQuakes program (Luetgert et al., 2009), which pla-
ces self-contained independent systems in homes and offices.
The USB sensors used by QCN range in cost from $30 to $150.
Since QCN’s inception in 2007, the network has grown to over
2000 volunteers worldwide. The stations are distributed glob-
ally (Cochran, Lawrence, Christensen, and Chung, 2009;Co-
chran, Lawrence, Christensen, and Jakka, 2009) as well as in
dense arrays in particular regions of interest, including the
rupture zones of major earthquakes to monitor aftershock se-
quences (Chung et al., 2011;Cochran et al., 2011).
Berkeley Open Infrastructure for Network Computing
(BOINC), a cyber-infrastructure software package (Anderson
and Kubiatowicz, 2002;Anderson, 2004;Christensen et al.,
2005), is used to initiate and manage communication between
a host computer and the central server (Cochran, Lawrence,
BSSA Early Edition / 1
Bulletin of the Seismological Society of America, Vol. 104, No. 1, pp. –, February 2014, doi: 10.1785/0120120196
Christensen, and Chung, 2009;Cochran, Lawrence, Christen-
sen, and Jakka, 2009). BOINC allows hosts to donate any de-
sired portion of their CPU and Internet bandwidth to a variety
of scientific projects (Anderson, 2004). Hosts also provide a
USB port and 15–30 min of their time to install the sensor and
report the sensor’s location through a Google Map online in-
terface. Sensors are typically shipped to volunteers through
the mail or may be transported via luggage on commercial
airlines, particularly for rapid aftershock deployments.
The MEMS sensors used by the QCN network have
steadily improved since the network began in 2007. The noise
floor of the MEMS accelerometers has decreased from
4×10−2m=s2in 2008 to 2×10−3m=s2in 2010, and
<6×10−4m=s2for the sensors currently used (Cochran,
Lawrence, Christensen, and Chung, 2009). MEMS accelerom-
eter amplitude and phase responses are relatively flat from 0 to
10 Hz (Holland, 2003;Farine et al., 2004;Evans et al., 2013).
The digitizers have improved from 8-bit in 2008–2010 (Co-
chran, Lawrence, Christensen, and Chung, 2009), to 14-bit in
2010–2011 (Cochran et al., 2011, and data used in this study),
to 16-bit since 2011. As a result of MEMS accelerometer and
digitizer improvements, the lowest-magnitude earthquake re-
corded with reasonable fidelity through QCN has dropped
from M4.5 in 2008 (Cochran, Lawrence, Christensen, and
Chung, 2009)toM2.6 in 2010.
In 2010 QCN initiated a Rapid Aftershock Deployment
Program (RAMP) following the M8.8 Maule, Chile, earth-
quake (Chung et al., 2011). A set of earthquake detection,
location, and characterization algorithms were developed
in retrospective tests using the continuous aftershock dataset
(Chung et al., 2011). The results of that study suggested that
moderate to large earthquakes could be detected, located, and
their magnitudes estimated using a limited set of ground-mo-
tion parameters transmitted to the central server from the
QCN stations. For events reported by both QCN and the
U.S. Geological Survey National Earthquake Information
Center (USGS NEIC), the retrospective tests resulted in sim-
ilar estimated locations, magnitudes, and shaking intensities
(Chung et al., 2011).
Here, we describe the data collected during a RAMP ini-
tiated following the M7.2 2010 Darfield, New Zealand,
earthquake. Specifically, we report the results of real-time
rapid earthquake detections using only parametric data pro-
vided by QCN stations. The emphasis here is on testing the
performance of the rapid earthquake detection algorithms
and the feasibility of using social-cyber-seismic network data
for earthquake characterization.
Data
On 3 September 2010, an M7.2 earthquake occurred in
Darfield, New Zealand, located just northwest of the urban
center of Christchurch (Gledhill et al., 2010). QCN and
GNS Science collaborated to initiate a RAMP to augment
the existing GNS strong-motion network with USB accelerom-
eters (Cochran et al., 2011). Volunteers were recruited to host
QCN sensors through local media outlets. Within 11 days of
the mainshock, over 180 QCN sensors were deployed in
homes in and around Christchurch (Fig. 1). Note that the
placement of sensors was limited by availability of hosts.
Therefore, the stations tend to be located in population
centers rather than distributed evenly around the aftershock
sequence. The majority of sensors were aligned, leveled,
and mounted to the ground floors or basements of residences.
Data were recorded at 50 samples per second. See Cochran,
Lawrence, Christensen, and Chung, 2009,Cochran, Law-
rence, Christensen, and Jakka, 2009, and Chung et al. (2011)
for additional details of the network architecture. Cochran
et al. (2011) showed that closely spaced QCN and GeoNet
stations provide similar peak acceleration, velocity, and dis-
placement time series. And, although the QCN sensors are
lower resolution and recorded fewer events, they found that
the GeoNet and QCN stations provide comparable peak
ground accelerations (PGAs) and peak ground velocities
(PGVs) as well as root mean square scatter in the observations.
QCN stations can record in either continuous or triggered
mode. The New Zealand RAMP stations were set up to record
in continuous mode, with time series uploaded to the central
server in 10-minute-long packets. Even in continuous mode,
the host computer monitors the three component data for sig-
nificant ground motion using a variant of a short-term average
over long-term average (STA/LTA; e.g., Earle and Shearer,
Earthquake depths
Magnitude distribution
6
12
18
24
30
3.2 4.0 4.8 5.6 6.4
Magnitude
Frequency (%)
612182430
0
10
20
30
Depth (km)
Frequency (%)
0 57
(b) (c)
172.5 173.0
172.0
M7.2
0
Christchurch
QCN Sensor
Earthquake
Major Earthquake
0 5 10 km
M6.3
M6.3
M6.3
(a)
Figure 1. Seismic sensors and earthquakes detected by the low-
cost sensor network between 25 September 2010 and 1 April 2011.
(a) Map of sensors (blue triangles), detected earthquakes (red stars),
and major earthquakes (green stars); inset map shows the study area
(red star). (b) Magnitude distribution and (c) depth distribution of
detected earthquakes.
2J. F. Lawrence et al.
BSSA Early Edition
1994) to generate triggers (Chung et al., 2011). QCN
replaces the LTA with a long-term standard deviation (σLT).
The ratio STA=σLT is the statistical definition of signal-to-
noise ratio (SNR) and is the inverse to the coefficient of varia-
tion (a measure of dispersion; McDonough and Whalen,
1995). This formulation of SNR is only applicable for non-
negative numbers, so we apply the statistic to the vector mag-
nitude of demeaned acceleration. QCN uses a 0.06 s short-term
window and a 60 s long-term window. We require that the SNR
exceed the 99% confidence interval (∼3) for a trigger to be
issued. To reduce multiple triggers associated with one strong
new signal, we require that all triggers have STA values and
SNR ratios 10% greater than any STA and SNR values deter-
mined in the previous second.
Following a trigger, the maximum acceleration at the
time of the trigger is determined. The PGAs defined here dif-
fers from traditional measurements. The traditional PGA is
not ideal for rapid event detection because it requires waiting
until the entire wavetrain from an event has passed by a sta-
tion before issuing a measurement. Therefore, we introduce a
variant of PGA at a given time lag τ:PGAτ. This time lag is
relative to a trigger, for example, the detection time of a
strong new motion. The peak acceleration at the time of the
trigger, τ0, is given by PGA0. The peak acceleration up to
4 s following the trigger is denoted by PGA4.
Further, we use the maximum vector magnitude of the
acceleration, jaj, rather than assessing the horizontal or ver-
tical magnitudes separately. We demean each component
prior to taking the magnitude of the vector. The peak vector
magnitude of acceleration PGAjajis typically dominated by
the peak horizontal acceleration, PGAH. Note that PGAjaj
removes the partitioning of energy between vertical and hori-
zontal, which may vary due to site conditions, source mecha-
nism, and earth structure. In some cases within this paper, we
refer to PGAjajand in other cases we refer to PGAjajτ. The
peak vector magnitude accelerations at the time of the trigger
and up to 4 s following the trigger are given by PGAjaj0and
PGAjaj4. We note that near the source, PGA and PGAjaj4are
typically very similar.
Trigger information and ground-motion parameters de-
rived from the time series are transferred to the servers in real
time. However, there is a finite time delay between the time
of an initial detection on the volunteer computer and when
that trigger is registered in the QCN database. This latency
accrues as the trigger information and associated data are
saved to the volunteer computer, the trigger information is
uploaded to the central server, and the server parses and
registers the trigger information. Figure 2illustrates the prob-
ability and cumulative distribution functions for the trigger
latencies of sensors in California and worldwide. The mean
latencies are ∼3:4and 4.2 s for California and worldwide,
respectively, and over 90% of all triggers are registered
within 6 and 7 s for California and worldwide, respectively.
There are several factors that may contribute to lower laten-
cies within California, including proximity to the QCN server
located in California, faster average Internet connections,
and/or computers with greater processing capabilities.
For less than 0.09% of triggers, latencies are greater than
10 s. In some instances, Internet dropouts can cause triggers
to be delayed by up to several days. In such cases, the trigger
data are clearly not useful for rapid earthquake detection and
characterization. Following a major earthquake it is possible
for power and Internet to fail, causing the network to lose
sensitivity. Power and Internet outages following the
Mw6.3 21 February 2011 Christchurch earthquake reduced
the number of QCN sensors reporting aftershocks for
∼48 hrs. This is a critical limitation of QCN; however, tradi-
tional networks may also lose contact with stations during
major events (e.g., Kilb et al., 2007).
Rapid Earthquake Detection
On 25 September 2010, QCN implemented preliminary
real-time earthquake detection and characterization using
only trigger information from the dense QCN network in
Christchurch. The rapid earthquake detection algorithm mon-
itors the triggers registered by the real-time database, includ-
ing the trigger time, peak vector amplitude of acceleration at
0 2 4 6 8 10
0
1
2
3
4
(a) (b)
Trigger latency (s)
Frequency (%)
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
Trigger latency (s)
Cumulative Distribution
California
World
Trigger upload time probability distribution Trigger upload time cumulative distribution
Figure 2. Latencies (s) for triggers detected on volunteers’computers to be transferred to the QCN central server located in California.
(a) Histogram and (b) cumulative distribution of trigger latencies. Blue indicate latencies from all QCN stations distributed globally and red
are latencies for stations in California. Dashed black line indicates the cumulative distribution of 0.9.
Rapid Earthquake Characterization Using MEMS Accelerometers Following the M7.2 Darfield Earthquake 3
BSSA Early Edition
the time of the trigger (referred to here as PGAjaj0), and station
information. The triggers transfer rapidly from the volunteer
computer to the server (2–5 s; Fig. 2) through BOINC’s trickle
protocol. The initial transfer of compact data descriptions,
rather than full waveforms, greatly reduces communication
latencies between the host computer and central server.
Triggers are temporally and spatially correlated at the
QCN server to evaluate if they are associated with a regional
earthquake or due to isolated, single-station noise. We evalu-
ate triggers received at 0.2 s intervals, comparing each trigger
with all other triggers that occurred within the previous
100 s. If triggers occur within 100 s and the stations are lo-
cated within 200 km they are considered roughly correlated
(Chung et al., 2011). We apply an SNR detection algorithm
(as above) to the number of roughly correlated triggers with a
minimum of 99% confidence and a minimum requirement of
5 correlated signals. Here, the short-term window is 100 s
and the long-term window is 5 days.
Once the number of roughly correlated triggers exceeds
the SNR (typically 6–7) or the minimum requirement (5), we
attempt to locate the earthquake hypocenter using a 3D grid
search (Chung et al., 2011). Further details of the earthquake
location algorithm are available in Chung et al. (2011).
Figure 3shows the correlation between the estimated travel
times and the observed travel times for the 21 February 2011
M6.3 earthquake. The QCN hypocenters are within ∼7km
of the GNS reported locations, on average, which have re-
ported errors on the order of 1 km in latitude and longitude.
Differences in the estimated hypocenter locations may result
from the higher noise level of the QCN sensors as well as
different station distributions, seismic-velocity models, and
location algorithms. The excellent correlation (e.g., Fig. 3,
in which R2>0:94) between observed and theoretical travel
times for the best-fit source locations suggests the events are
well located, given the a priori seismic-velocity model.
Online accessible figures are automatically generated for
each iteration based on the best-fit location determined by a
3D grid search. This correlation between observed and esti-
mated travel times are used in the rapid earthquake location
scheme. Only best-fit locations with correlations better than
R2>0:5are considered probable earthquakes.
Next, we estimate the earthquake magnitude by fitting
the PGA as a function of distance (e.g., Gutenberg and
Richter, 1956;Donovan, 1973). The scatter in PGA is often
quite high due to a regionally heterogeneous geologic
material and the site conditions at each sensor (e.g., Cochran
et al., 2011). However, PGA measured at multiple stations
can provide a stable magnitude estimate. For the QCN
New Zealand data, we derived the empirical relationship be-
tween magnitude, distance, and PGAjaj:
M0:03Δ1:09 lnPGAjaj4:28;
in which Δis distance in km and PGAjajis the vector sum of
the three-component PGA in m=s2. This relationship is
similar to those previously reported for New Zealand crustal
earthquakes using ground-motion prediction equations
(GMPEs) of Zhao et al. (1997),McVerry et al. (2006), and
Campbell and Bozorgina (2008), suggesting previously
determined PGA–distance–magnitude relationships can be
used (Fig. 4). Figure 4shows the more typical PGA computed
from the full waveform, rather than PGAτ. The relationship
0.96
Estimated time (s)
Observed time (s)
Iteration Time Correlation
71
0.94114
2
4
6
8
10
0
6810402
21 February 2011: M6.3
Figure 3. Correlation of modeled travel times to observed travel
times for the M6.3 earthquake on 21 February 2011, with values
shown for detection iteration 1 (red) and 4 (blue).
Campbell and Bozorgnia, 2008
McVerry et al., 2006
PGA (m s–2)
10
1
10
0
10
–1
101
100102
This Study
Source distance (km)
Zhao et al., 1997
M 6.0 M 4.0
10
–2
Figure 4. Comparison between the PGA relationship obtained
here and the GMPEsofZhao et al. (1997),McVerry et al. (2006),
and Campbell and Bozorgnia (2008). Note the relationships deviate
at greater distances; however, there are fewer QCN PGA measure-
ments available to constrain the relationship at these distances (gray
region).
4J. F. Lawrence et al.
BSSA Early Edition
used here is within the range of previously reported GMPEs.
Note that here we use vector PGA rather than the geometric
mean of the horizontals more typically applied to GMPEs;
however, the vertical ground motions only contribute mini-
mally to the vector PGA. There are some differences in the
GMPEs, which may be due to limited data quantities and scat-
ter in measured accelerations due to source and site effects.
At close distance ranges all of the models, including the
model derived from QCN data, converge to similar PGAs.
The majority of the ground-acceleration measurements avail-
able in the dataset used here are at source-to-station distances
of 5–30 km.
QCN host computers transmit updated PGAjajmeasure-
ments at τ1, 2, and 4 s after the initial trigger (PGAjaj1–4),
which are used to iteratively improve magnitude estimates.
New data between each iteration may include new triggers
(PGAjaj0) reported by additional sensors or from new wave
arrivals at the same sensor and updated PGAjaj1–4from sta-
tions that had previously triggered. The number of iterations
for an earthquake depends on the earthquake location and
magnitude relative to the distribution of seismic stations.
Between 25 September 2011 and 1 April 2011 the QCN
sensors detected 116 earthquakes ranging in size from M3.6
to M6.3 (Figs. 1and 5). The minimum and median detection
times were 3.1 and 9.8 s after the estimated earthquake origin
time, respectively (Fig. 5a). This time includes: (1) source-to-
sensor propagation (variable), (2) trigger detection (<1s),
(3) transfer of trigger information to the QCN server (2–5 s),
(4) trigger association and event declaration (requires at least
five associated triggers, variable), (5) event detection
(<0:1s), earthquake location (<0:02 s), magnitude estima-
tion (<0:01 s), (6) AlertMap generation (∼0:4s), and (7) post-
ing to the web (∼0:4s concurrent with AlertMap generation).
Approximately half of the delay between the earthquake ori-
gin and detection is accounted for in wave propagation from
the hypocenter to seven or more sensors (∼4:3s on average).
The median magnitude estimate increases by 0.5 magnitude
units during the first four iterations, likely due to the fact that
smaller earthquakes often result in only one or two iterations,
whereas larger events typically result in a greater number of
iterations because more sensors trigger with equal or greater
PGAjaj0–4. However, some the observed increase in magni-
tude with number of iterations may be due to updated
PGAjaj0–4estimates which can increase as more of the wave-
field is observed at a station, including the Swave and sub-
sequent coda.
To determine the accuracy of the real-time detection and
characterization we compare the QCN earthquake catalog
16
16 24
Time difference (s)
4
8
12
0 6 12 18 24 30
Detection time (s)
6
12
18
24
0 8 32 40
Distance difference (km)
6
12
18
24
30
01020304050 ––
––
Depth difference (km)
6
12
18
048
(a)
(d) (e) (f)
0
0
0
0
3
Latency (s)
06912
(b)
0
Frequency (%)
Magnitude difference
6
12
18
24
012
(c)
0
6
12
18
24
30
1
3
5
7
9
11
13
Iteration
Frequency (%)
Frequency (%)
Frequency (%)
Frequency (%) Frequency (%)
Figure 5. Rapid earthquake detection statistics for 116 earthquakes between 25 September 2010 and 1 April 2011. (a) Time between
estimated earthquake origin to initial earthquake detection. (b) Latency from host signal detection to trigger registry on the QCN server.
Histograms comparing QCN rapid earthquake characterization and GNS catalog of: (c) magnitude; (d) epicentral location; (e) depth;
and (f) origin time (s). Histograms are shown for the first iteration; lines indicate median values for each iteration. The magnitudes are
within 0.5 and 1.0 magnitude units for 63% and 90% of all earthquakes, respectively.
Rapid Earthquake Characterization Using MEMS Accelerometers Following the M7.2 Darfield Earthquake 5
BSSA Early Edition
with the published GNS Science catalog. The accuracy/pre-
cision of the real-time estimates varied over time because of
attrition of QCN volunteers, improvements to the monitoring
process, and power and communications failures following
large quakes. Although we conducted retrospective testing of
the earthquake detection algorithms following the M8.8
Maule, Chile, earthquake (Chung et al., 2011), the algo-
rithms had not been tested in real time. We implemented a
preliminary version of the real-time detection algorithms on
25 September 2011 and, over the following months, added
features, improved algorithms, and expanded the real-time
output features.
Because of the low-resolution nature of the QCN sensors,
many (∼51) earthquakes with magnitude greater than M3.5
located within 15 km of the array were not automatically de-
tected. Continuous waveform records for most of the stations
were uploaded to the central server, so data from the events not
automatically detected were not lost. During the first 32 days,
11 false detections occurred, usually with a low number (5–6)
of triggering stations. After several algorithm improvements
implemented in October 2010, no false detections and char-
acterizations were reported through April 2011. The most im-
portant algorithm change to eliminate false detections was to
ignore triggers from a subset of the stations that repeatedly
triggered due to chronic sensor or computer malfunction
and/or extremely frequent local noise sources.
The network size reduced gradually throughout the de-
ployment period (11 September 2010–1 April 2011) as hosts
were initially recruited for an ∼6week period to record the
early aftershock period following the 3 September 2011
M7.2 Darfield earthquake. Immediately prior to the 21 Feb-
ruary 2011 M6.3 Christchurch earthquake, only 40 10
sensors were operating in the region. GNS Science rede-
ployed some sensors following the February earthquake,
bringing the network to 50 10 sensors. As expected, the
network capabilities diminish as the number of sensors de-
creased. Nevertheless, dozens of earthquakes were detected
and recorded prior to the M6.3 earthquake even with fewer
sensors reporting.
Immediately following the 21 February 2011 M6.3
Christchurch earthquake the capabilities of the real-time
detection were further reduced due to the following power
outages and a rapid succession of aftershocks. Vigorous
aftershock sequences can challenge even very robust net-
works (e.g., Kilb et al., 2007). QCN failed to detect 54 after-
shocks with magnitude greater than M3.5 in the 48 hrs
following the M6.3 Christchurch earthquake. With ongoing
and planned algorithm changes to discriminate between
event updates and new events, we expect fewer complica-
tions for future vigorous aftershock sequences; however,
widespread power failures would still reduce or eliminate
the ability to detect events.
We compare the origin time, hypocenter, and magnitude
for the QCN detected events to the GeoNet catalog (Fig. 5).
On average, we find that the median error in earthquake lo-
cation is ∼7km, the QCN earthquake origin times are biased
1.5 s early, and the magnitude estimates are ∼0:5magnitude
units too high. These differences can be attributed primarily
to different station distributions, velocity models, and attenu-
ation relationships. Because of the limited distribution of
available hosts, QCN sensors are clustered in a small region
(10 ×10 km2) with the majority of the earthquakes occur-
ring outside of the network (Fig. 1). This tends to increase
the location error, biasing earthquake locations farther from
the network, and leading to overestimation of magnitude and
earlier origin times (Fig. 5). However, given these limita-
tions, the QCN rapid earthquake detections compare favor-
ably to the GNS Science catalog.
QCN automatically generates an AlertMap of observed,
interpolated, and predicted peak shaking intensities (e.g.,
Allen and Kanamori, 2003) for each detected earthquake.
Here, intensity is estimated as instrumental intensity (Wald,
Quitoriano, Heaton, and Kanamori, 1999;Wald, Quitoriano,
Heaton, Kanamori, et al., 1999), using PGAjaj0–4instead of
PGA. For most of the events evaluated here, PGAjaj0–4is sim-
ilar to PGA because the earthquakes are small and recorded at
short source–receiver distances; thus, the maximum acceler-
ations occur within the first 4 s of the record. For large mag-
nitude events and/or greater source–receiver distances,
PGAjaj0–4may underestimate the true instrumental intensity.
AlertMaps are generated when the earthquake is first detected
and improved with each iteration as additional data are re-
ported. Because the median event detection time is approxi-
mately 10 s, the AlertMap often provides a prediction of
expected peak shaking intensities for locations away from
the epicenter. Figure 6shows an AlertMap of estimated peak
ground-shaking intensities and the estimated time until
shaking arrives at each location. The individual PGAjaj0–4ob-
servations suggest fine-scale variability in ground-shaking
intensities. The smooth background contours provide an in-
terpolated estimate of the ground-shaking intensities between
observations and, for locations farther from the hypocenter,
the predicted shaking.
Discussion
The results of the real-time earthquake detection and
characterization tests using only ground-motion parameters
derived from the temporary array of QCN stations are prom-
ising. The rate at which stations were installed (over 180
stations in 11 days) is ideal for recording and detecting after-
shocks shortly after a mainshock; however, the station dis-
tribution is somewhat limited to urban areas based on
available volunteers. The resources required to transport a
large number of QCN sensors into the aftershock area were
minimal compared to traditional RAMP experiments due to
the lightweight and compact size of the QCN sensors.
The QCN network and real-time algorithms were able to
determine the magnitude and location of events with median
detection times of 9.1 s after the origin time. The primary de-
lay is propagation of the Pwaves from the source to 5–7
stations as well as latencies to transfer the ground-motion
6J. F. Lawrence et al.
BSSA Early Edition
parameters from the station to the central server over the In-
ternet. Errors in the location (∼7km) and magnitude (1
magnitude unit) are similar to errors from previously reported
earthquake early warning algorithms (4–13 km and 0.5–0.75
magnitude units; e.g., Allen, 2006;Allen et al., 2009).
It is unclear how QCN’s real-time algorithms will per-
form during large or great earthquakes (>M 6.5) measured
at close source–receiver distances (<40 km). PGA and other
waveform measurements can saturate, resulting in an under-
estimate of magnitude (e.g. Kanamori, 2005). So, although
use of PGAjaj0–4was largely successful in this experiment,
other measurements, such as the maximum period of dis-
placement used in earthquake early warning (EEW) algo-
rithms, may be more reliable (e.g., Allen and Kanamori,
2003;Cua and Heaton, 2007;Allen et al., 2009). Never-
theless, retrospective application of these algorithms to the
Maule, Chile, aftershock sequence provided accurate magni-
tude estimates for events up to M6.9 (Chung et al., 2011).
The rapid updates to the location, magnitude, and Alert-
Maps for each detected event suggest that, as triggers from
more distant stations are included, earthquake characteriza-
tion improves regardless of PGA saturation. Dense networks
such as the RAMP described here provide a unique opportu-
nity to evaluate earthquakes in real time. Even as we test a
variety of EEW parameters (e.g., Allen and Kanamori, 2003;
Cua and Heaton, 2007;Allen et al., 2009), real-time knowl-
edge of PGAjaj0–4will likely provide useful information re-
garding the spatial distribution of ground-shaking intensities,
which can deviate from the idealized distance-dependent
models. Rapid estimates of ground-shaking intensities and
the distribution of strong ground motions can be used to infer
the true size of an event.
Because initial locations, magnitudes, and AlertMaps
are generated within seconds of the earthquake origin, this
information could be potentially used to provide an EEW
alert to individuals or organizations of incoming strong
ground motion (Hoshiba et al., 2008). The initial test of this
notification system was conducted using a small pool of in-
house participants and no notifications were provided to the
general public. Beginning 5 March 2011, QCN implemented
a tool to automatically disseminate event information to a
select group of users. The notification (sent via e-mail) in-
cluded the earthquake metadata (origin time, location, time
of detection, and magnitude) and an AlertMap. The notifica-
tions can be filtered based on input participant-defined pref-
erences for epicentral distance and magnitude cutoff. These
notifications were typically received by the in-house partic-
ipants within 1 s of event detection. For this cursory test, we
used Simple Mail Transfer Protocol (SMTP)(Hughes, 1998)
and Internet Message Access Protocol (IMAP)(Heinlein and
Hartleben, 2008). SMTP and IMAP are not ideal for imme-
diate mass data dissemination, but the rudimentary test illus-
trated that earthquake information detected by the low-cost
QCN network could be distributed rapidly.
With an average detection time of 9.1 s, and an additional
∼1s for issuing a notification, participants in New Zealand
Depth = 6.7 Oct 18 2010 22:32:16
–20
–10
0
QCN: M4.8 Lon = 172.59 Lat = –43.57
Detected: Oct 18 2010 22:32:40
Christchurch
Depth = 4.5 Oct 18 2010 22:32:16
0
10
QCN: M4.6 Lon = 172.56 Lat = –43.56
20
Detected: Oct 18 2010 22:32:21
Christchurch
Instrumental
Intensity
Peak Vel (cm/s)
Peak Acc (%g)
Potential
Damage
Perceived
Shaking
IIII V VI VII VIII IX X+
<0.12 >115
<0.17 >114
none none none Very Light Light Moderage Moderate/Heavy Heavy Very Heavy
Not Felt Weak Light Moderate Strong Very Strong Severe Violent Extreme
(a) (b)
172° 173° 172°
–44°
–43.5°
–43°
173°
Figure 6. AlertMaps generated (a) 5 s and (b) 24 s after the earthquake origin time by QCN’s rapid earthquake detection system for an
M4.8 Christchurch earthquake. The gray circles illustrate isochrons for S-wave arrival time relative to the detection time (positive times
indicate expected future arrivals, negative times indicate waves have arrived). Initial AlertMaps are very similar to later iterations. The
intensity scale is from (Wald, Quitoriano, Heaton, and Kanamori, 1999;Wald, Quitoriano, Heaton, Kanamori, et al., 1999).
Rapid Earthquake Characterization Using MEMS Accelerometers Following the M7.2 Darfield Earthquake 7
BSSA Early Edition
could have received alerts if such notifications had been is-
sued to individuals external to QCN’s personnel. For ∼83%
of the aftershocks one or more participants around Christ-
church could have received an alert before the strong surface
waves hit (assuming an ∼3km=s surface-wave velocity). This
statistic is skewed because three participants hosted sensors >
50 km from Christchurch. A 3km=s surface wave expands to
30.3 km in 10.1 s (the mean time to a theoretical alert). If the
network had been deployed with more sensors at the perimeter
of Christchurch, many of the detection times would have been
reduced, and the radius of first alert would have decreased.
A dense QCN network similar to the one deployed in
Christchurch, New Zealand is not possible or useful every-
where. QCN requires pre-existing power, Internet, and volun-
teer computers, so it is limited to populated regions with
sufficient resources to afford a base-level cyber-infrastruc-
ture. Anecdotally, QCN has found that retaining volunteers
may depend on several factors: (1) feedback to the volunteer
community to maintain interest, (2) low cost to volunteers of
transmitted data per byte, (3) cultural/language differences,
or (4) regional groups supporting the effort. For comparison,
a year after the M8.8 Maule, Chile, RAMP (Chung et al.,
2011) only two participants remained in the region, whereas
there were more than 40 RAMP volunteers participating more
than 1 yr after the M7.2 Darfield and M6.3 Christchurch,
New Zealand, earthquakes.
Perhaps the greatest benefit of a network like QCN is in
countries that cannot (or will not) afford a real-time seismic
network with traditional research-grade equipment. In such
countries, grass roots and academic efforts through cyber-
social-seismic networks like QCN may enable real-time
earthquake monitoring with >100 sensors for less than
$10,000 in material costs. Because QCN’s servers monitor
all regions, no regional server, service, or administration
is required. In addition, all data collected are available online
to access and download. In developing countries like Haiti
and Mongolia, Internet usage is as high as 8% with broad-
band subscriptions exceeding 1%, which means >100;000
potential cyber-social-seismic network volunteers per country.
If only 0.1% of those potential volunteers are interested in par-
ticipating, they could host >100 sensors in each country.
In the near future, it is likely that cyber-social-seismic
networks like QCN will not operate in isolation. By integrat-
ing traditional seismic networks with volunteer networks, the
combined result will likely exceed both in isolation. Volun-
teer networks will benefit from the added stability of low-
noise traditional sensors in quiet sites, whereas traditional
networks will benefit from large quantities of spatially dis-
tributed data. Integrating multiple data types such as maxi-
mum period of displacement, PGV, and PGA will provide
additional constraints.
Conclusion
The Quake-Catcher Network is designed for low-
overhead installation of dense strong-motion seismic networks
capable of providing real-time ground-motion parameters as
well as continuous waveform data. The dense RAMP array
installed following the 2010 M7.2 Darfield, New Zealand,
earthquake illustrates the potential for low-cost seismic net-
works to rapidly provide information on earthquake location
and magnitude, as well as peak shaking intensities. The expan-
sion of scientific data products resulting from low-cost cyber-
social-seismic networks is likely to continue and could pro-
vide an important new component of seismic networks to aid
in the generation of real-time earthquake products, including
ShakeMap (Wald, Quitoriano, Heaton, Kanamori, et al., 1999)
and EEW, for moderate to large earthquakes.
Data and Resources
QCN trigger and waveform data used in this study are
available from http://qcn.stanford.edu/sensor/trdl.php (free
registration required, last accessed February 2013). Maps of
current QCN participants are available from http://qcn
.stanford.edu (last accessed February 2013). List and maps
of earthquakes detected by QCN are available at http://qcn
.stanford.edu/earthquakes (last accessed February 2013).
GNS Science earthquake catalog used for comparison to
QCN detected events is available from http://magma.geonet
.org.nz/quakesearch/ (last accessed December 2013). Some
plots were made using Generic Mapping Tools (www.soest
.hawaii.edu/gmt, last accessed January 2013; Wessel and
Smith, 1991,1998).
Acknowledgments
We thank the hundreds of QCN volunteer hosts and the field crew, with-
out whom this study would never have occurred. Many thanks to Dave An-
derson whose support and augmentation of BOINC has allowed QCN to grow
quickly. We thank Hiroo Kanamori and Dan McNamara for thorough feed-
back during the revisions of this manuscript. This research was supported in
part by NSF EAR 1027802 and the New Zealand Natural Hazards Research
Platform.
References
Allen, R. M. (2006). Probabilistic warning times for earthquake ground shak-
ing in the San Francisco Bay area, Seismol. Res. Lett. 77, 371–376.
Allen, R. M., and H. Kanamori (2003). The potential for earthquake early
warning in Southern California, Science 300, 786–789.
Allen, R. M., H. Brown, M. Hellweg, O. Khainovski, P. Lombard, and
D. Neuhauser (2009). Real-time earthquake detection and hazard
assessment by ElarmS across California, Geophys. Res. Lett. 36, doi:
10.1029/2008GL036766.
Anderson, D. P. (2004). BOINC: A system for public-resource computing
and storage, in Proc. of the Fifth IEEE/ACM International Workshop
on Grid Computing, Pittsburg, Pennsylvania, 8 November 2004, 4–10.
Anderson, D. P., and J. Kubiatowicz (2002). The world-wide computer, Sci.
Am. 286, 40–47.
Campbell, K. W., and Y. Bozorgnia (2008). NGA ground motion model for
the geometric mean horizontal component of PGA, PGV, PGD and 5%
damped linear elastic response spectra for periods ranging from 0.01 to
10 s, Earthq. Spectra 24, 139–171.
Christensen, C., T. Aina, and D. Stainforth (2005). The challenge of volun-
teer computing with lengthy climate model simulations, in First IEEE
International Conference on e-Science and Grid Computing, Institute
8J. F. Lawrence et al.
BSSA Early Edition
of Electrical and Electronics Engineers, Melbourne, Victoria, Aus-
tralia, 8–5 December 2005, 8–15.
Chung, A. I., C. Neighbors, A. Belmonte, M. Miller, H. H. Sepulveda, C.
Christensen, R. S. Jakka, E. S. Cochran, and J. F. Lawrence (2011).
The Quake-Catcher Network rapid aftershock mobilization program
following the 2010 M8.8 Maule, Chile earthquake, Seismol. Res. Lett.
82, 526–532.
Clayton, R. W., T. Heaton, M. Chandy, A. K. Krause, M. Kohler, J. Bunn, R.
Guy, M. Olson, M. Faulkner, M. Cheng, L. Strand, R. Chandy, D.
Obenshain, A. Liu, and M. Aivazis (2011). Community Seismic Net-
work, Ann. Geophys. 54, 738–747.
Cochran, E., J. Lawrence, C. Christensen, and A. Chung (2009). A novel
strong-motion seismic network for community participation in earth-
quake monitoring, IEEE Instrum. Meas. Mag. 12, 8–15.
Cochran, E. S., J. F. Lawrence, C. Christensen, and R. S. Jakka (2009). The
Quake-Catcher Network; citizen science expanding seismic horizons,
Seismol. Res. Lett. 80, 26–30.
Cochran, E. S., J. F. Lawrence, A. Kaiser, B. Fry, A. I. Chung, and
C. Christensen (2011). Comparison between low-cost and traditional
MEMS accelerometers: A case study from the M7.1 Darfield,
New Zealand, aftershock deployment, Ann. Geophys. 54, 728–737.
Cranswick, E., B. Gardner, R. Banfill, and S. Hammond (1993). Recording
ground motions where people live, Eos Trans. AGU 74, 243–244.
Cua, G., and T. H. Heaton (2007). The virtual seismologist (VS) method: A
Bayesian approach to earthquake early warning, in Earthquake Early
Warning Systems, P. Gasparini, G. Manfredi, and J. Zschau (Editors),
Springer, Berlin, Heidelberg.
Donovan, N. C. (1973). Statistical evaluation of strong motion data, includ-
ing the February 8, 1971, San Fernando earthquake, in Proc. 5th
World Conf. Earthquake Engineering, Rome, Italy, 25–29 June
1973, 1252–1261.
Earle, P. S., and P. M. Shearer (1994). Characterization of global seismo-
grams using an automatic picking algorithm, Bull. Seismol. Soc.
Am. 84, 366–376.
Evans, J. R., R. Allen, A. Chung, E. Cochran, R. Guy, M. Hellweg, and J.
Lawrence (2014). Performance of several low-cost accelerometers,
Seismol. Res. Lett. 85, doi: 10.1785/0220130091.
Farine, M., N. Thorburn, and D. Mougenot (2004). General application of MEMS
sensors for land seismic acquisition—is it time? TLE 23,246–250.
Gledhill, K., J. Ristau, M. Reyners, B. Fry, C. Holden, and G. Science
(2010). The Darfield (Canterbury, New Zealand) Mw7.1 earthquake
of September 2010: A preliminary seismological report, Seismol. Res.
Lett. 82, 378–386.
Gutenberg, B., and C. F. Richter (1956). Earthquake magnitude, intensity,
energy and acceleration, Bull. Seismol. Soc. Am. 46, 105–145.
Heinlein, P., and P. Hartleben (2008). The Book of IMAP: Building a Mail
Server with Courier and Cyrus, No Starch Press, San Francisco, Cal-
ifornia, 352 pp.
Holland, J. (2003). Earthquake data recorded by the MEMS accelerometer,
Seismol. Res. Lett. 74, 20–26.
Hoshiba, M., O. Kamigaichi, M. Saito, S. Tsukada, and N. Hamada (2008).
Earthquake early warning starts nationwide in Japan, Eos Trans. AGU
89, 73–80.
Hughes, L. (1998). Internet E-mail Protocols, Standards and Implementa-
tion, Artech House, Boston, Massachusettes, 457 pp.
Kanamori, H. (2005). Real-time seismology and earthquake damage miti-
gation, Ann. Rev. Earth Planet. Sci. 33, 195–214.
Kilb, D., V. G. Martynov, and F. L. Vernon (2007). Aftershock detection
thresholds as a function of time: Results from the ANZA seismic
network following the 31 October 2001 Ml5.1 Anza, California,
earthquake, Bull. Seismol. Soc. Am. 97,280–792.
Luetgert, J. H., J. R. Evans, J. Hamilton, C. R. Hutt, E. G. Jensen, and
D. H. Oppenheimer (2009). NetQuakes—A new approach to urban
strong-motion seismology (abstract S11B-1707), Am. Geophys. Union
Fall Meeting, San Francisco, California, 14–18 December 2009.
McDonough, R. N., and A. D. Whalen (1995). Detection of Signals in Noise,
Academic Press, San Diego, California, 495 pp.
McVerry, G. H., J. X. Zhao, N. A. Abrahamson, and P. G. Somerville (2006).
New Zealand acceleration response spectrum attenuation relations for
crustal and subduction zone earthquakes, Bull. New Zeal. Soc. Earthq.
Eng. 39, 1–58.
Wald, D. J., J. V. Quitoriano, T. H. Heaton, and H. Kanamori (1999). Rela-
tionship between peak ground acceleration, peak ground velocity, and
modified Mercalli intensity for earthquakes in California, Earthq.
Spectra 15, 557–564.
Wald, D. J., V. Quitoriano, T. H. Heaton, H. Kanamori, C. W. Scrivner, and
C. B. Worden (1999). TriNet “ShakeMaps”: Rapid generation of peak
ground-motion and intensity maps for earthquakes in Southern
California, Earthq. Spectra 15, 537–556.
Wessel, P., and W. H. F. Smith (1991). Free software helps map, and display
data, Eos Trans. AGU 72, 441.
Wessel, P., and W. H. F. Smith (1998). New, improved version of generic
mapping tools released, Eos Trans. AGU 79, 579.
Zhao, J. X., D. J. Dowrick, and G. H. McVerry (1997). Attenuation of peak
ground accelerations in New Zealand earthquakes, Bull. New Zeal.
Soc. Earthq. Eng. 3, 133–158.
Stanford University
397 Panama Mall
Mitchell Building 360
Stanford, California 94305
(J.F.L., A.C., C.M.C., J.W.B.)
United States Geological Survey
525 S. Wilson Avenue
Pasadena, California 91106
ecochran@usgs.gov
(E.S.C.)
GNS Science
P.O. Box 30-368
Lower Hutt 5040, New Zealand
(A.K., B.F.)
University of California
Berkeley, 307 McCone Hall
Berkeley, California 94720
(R.A.)
California Institute of Technology
1200 East California Boulevard MC 104-44
Pasadena, California 91125
(T.H., M.D.K.)
Scripps Institution of Oceanography
University of California
San Diego, 9500 Gilman Dr.
La Jolla, California 92093
(D.K.)
University of Delaware
101 Smith Hall
Newark, Delaware 19716
(M.T.)
Manuscript received 8 June 2012;
Published Online 7 January 2014
Rapid Earthquake Characterization Using MEMS Accelerometers Following the M7.2 Darfield Earthquake 9
BSSA Early Edition
