- © 2006 by the Seismological Society of America
ShakeMap is a computer program to map earthquake ground-motion parameters in near real time. It originally was developed and implemented using a dense recording network in California (Wald et al. 1999); typical interstation spacing in southern California is about 5 to 50 km, though some stations in the Los Angeles metropolitan area are less than 2 km from each other, as seen at http://www.data.scec.org/stations/CI_station_map.php). For our implementation of ShakeMap in southern Ontario, we have developed an automated earthquake detection program, “EventDetector,” to continuously detect events and trigger ShakeMap in a sparse network where typical interstation spacing is about 100 km. The EventDetector is keyed to ShakeMap objectives in that it detects events based on the exceedance of specified ground-motion amplitudes at five or more stations. The specified threshold is set as peak ground velocity (PGV) = 0.003 mm/s. For our network configuration, this criterion should result in ShakeMap triggering on all events of M > 2.8 within the southern Ontario network.
When the program detects an event, EventDetector automatically triggers the ShakeMap program and continues to process the incoming data. The main challenge in event detection is to screen out false triggers caused by teleseismic events or other spurious signals. We distinguish between local/regional and teleseismic events by using the standard deviation of the peak ground velocity, normalized by the mean (denoted NSD) as a discriminant. Events that have NSD < 0.7 are teleseismic events (with a rate of 10% false triggers). To locate the regional earthquake once an event is confirmed by the above criteria, we use the ground-motion centroid concept, originally developed by Kanamori (1993). The centroid provides an estimate of the earthquake moment magnitude (M) and location from the point of view of ground motion, and it may not coincide with the actual earthquake epicenter and magnitude. The estimation is based on fitting the observed amplitude data to ground-motion attenuation relations. An additional check on whether the event is a significant earthquake is based on the determined centroid magnitude; events of M < 2.8 are not considered further.
We have compared the accuracy of estimated ground motions obtained using the epicenter location with those obtained using the ground-motion centroid for more than 30 local/regional earthquakes of magnitude 2.5 to 5.0. Overall the centroid performance in terms of estimating recorded ground motions appears to be satisfactory. In fact, the Nuttli magnitude (MN) 5.4 6 March 2005 Rivière-du-Loup earthquake demonstrates that estimated PGV based on ShakeMap agrees quite well with the recorded PGV (Kaka and Atkinson 2005b). Use of the epicenter might improve estimates of ground motion near the epicenter in cases where no nearby stations are available. However, the centroid is available in near real time while the epicenter is not.
The ShakeMap concept originally was conceived by Wald et al. (1999) and implemented in California using a dense strong-motion network (see figure 1 of Wald et al. 1999; typical inter-spacing 5 to 50 km). ShakeMaps are Web-based near-real-time maps that show the spatial distribution of recorded and predicted peak ground motions and estimate the corresponding felt intensity. ShakeMap was designed as a rapid response tool to portray the extent and variation of ground shaking at locations throughout the region, for purposes of providing rapid public, planning, and emergency response and post-earthquake information immediately after local and regional earthquakes.
A modified version of ShakeMap has been developed for earthquakes in southern and central Ontario; solutions are posted automatically to www.shakemap.carleton.ca. Ontario ShakeMap is based on seismographic data from the Portable Observatories for Lithospheric Analysis and Research Investigating Seismicity (POLARIS) network (www.polarisnet.ca) and aims to provide rapid ground-shaking information on small to moderate regional earthquakes. This requires continuous access to near-real-time data and automated procedures to detect and locate earthquakes as they occur. Thus we have developed an automated earthquake detection program, “EventDetector,” to detect events. An estimate of the earthquake location and magnitude is needed in assigning ground-motion values to areas where there are no seismographic recording stations. We have modified Kanamori's ground-motion centroid program (Kanamori 1993) to locate earthquakes and determine their magnitude for ShakeMap in Ontario. The focus in these applications is on the prediction of ground-motion amplitudes rather than the actual earthquake magnitude and location. A particular challenge in ShakeMap development in Ontario and other areas of eastern North America is developing algorithms that are effective in a sparse network. The interstation spacing in southern Ontario is typically about 100 km.
RECEIVING NEAR-REAL-TIME DATA
The Seismic Network Acquisition Systems (SNAQS) program developed by the Geological Survey of Canada (GSC) is used to stream near-real-time data from Ontario POLARIS stations into the ShakeMap program. The locations of POLARIS stations that we currently monitor in near real time are shown in figure 1 along with MN ≥ 2.7 events reported by the GSC during 2003–2005. The data come in packets of six seconds in duration and are saved in circular “ringbuffers.” The SNAQS distribution contains a program called “ringsniffer” that looks at the ringbuffer for one component of one station and organizes information about data packets to ensure that the data is read in time-contiguous order (as opposed to the raw unordered signal that comes over the network connection) (T. Cote, 2003, personal communication).
EventDetector continuously reads data, detects events, and triggers ShakeMap. It is designed to detect events based on the exceedance of specified ground-motion amplitudes at five or more stations. We use the vertical component of peak ground velocity as the amplitude parameter to define an event. Velocity is chosen because the instruments measure velocity. The vertical component is chosen because the site amplification of the vertical component is small enough to be neglected in most cases. (Note: The horizontal to vertical component ratio technique for estimating site response is based on this assumption; see Beresnev and Atkinson 1997; Siddiqqi and Atkinson 2002). When the vertical component amplitude of PGV exceeds 0.008 mm/s at any POLARIS Ontario station, after removal of long period noise and any linear trends, EventDetector examines the incoming data at all stations for a 2-minute window. The length of the time window was chosen by determining the length of time required for the strongest portion of the seismic signal (typically the S waves or Lg waves) to reach the majority of the stations within the network, assuming that the first triggering station was nearest to the event. The amplitude of 0.008 mm/s is near the felt threshold (figure 2). ShakeMap is configured such that if five or more stations (including the trigger station) record PGV ≥ 0.003 mm/s, or eight or more stations record PGV ≥ 0.002 mm/s within the 2-minute window, then the EventDetector will initiate the centroid program to calculate the location and magnitude. These initial triggering criteria are based on correlations between PGV, magnitude, and distance of past recorded events, as shown in figure 2. Figure 2(A) shows the recorded vertical component PGV as a function of distance and magnitude for 24 recorded events of MN ≥ 2.3 that occurred in the region during the time period 28 May 2002 to 31 March 2005. Figure 2(B) shows plots of the recorded vertical component PGV for five typical felt events in the region as a function of distance. For our network configuration (figure 1), we can deduce by inspection of figure 2 that all felt events should be detected if we achieve a PGV threshold of 0.003 mm/s at five or more stations. This assessment presumes that the five felt events plotted on figure 2(B) (highlighted in figure 1) represent a typical sample, albeit limited. In figure 2(B), note that the 0.003 mm/s criterion is exceeded as follows (all events of MN 2.7 to 2.8): 11 stations for 13 January 2005, six stations for 26 February 2005, five stations for 2 May 2003, five stations for 20 August 2003, and eight stations for 17 March 2004. We conclude that with our [5 stations ≥ 0.003 mm/s] trigger criteria, ShakeMap should be triggered if the centroid location is inside the network (42°–48°N and 76°–82°W) and the centroid magnitude is above 2.8. If the centroid determined by ShakeMap falls outside the network, then we have configured the program to post a ShakeMap only if the centroid magnitude is greater than 4.5; the generated ShakeMap will not be accurate in this case, but it will at least provide some initial information as to the levels of motion within the network. Figure 3 depicts a flowchart of the automated procedures to produce ShakeMap in near real time.
An important challenge in event detection is to screen out false triggers such as teleseismic events or spurious signals. The normalized standard deviation of the PGV (normalized to the mean), denoted NSD, can be used to discriminate between local/regional and teleseismic events in most cases: where NSD is the normalized standard deviation and is the mean PGV.
The NSD is a simple way to discriminate teleseismic events based on the presumption that such events should produce similar ground motions across the network. By contrast, regional events will produce large motions at nearby stations and small motions at a distance, resulting in a large NSD. Based on a selection of study events as listed in tables 1(A) and (B), we determined that events that have NSD ≤ 0.7 can be considered teleseismic events (with a rate of 10% false triggers), as illustrated in figure 4. Note that the earthquakes considered in the development of this discriminant include local events within the network as well as regional events for which ShakeMaps would not be posted because they fall outside the network (and have M < 4.5); we included both types of events in the development of the NSD tool to ensure that we did not falsely screen large regional events outside the network as being “teleseismic.” The NSD discriminant was added to EventDetector before the initiation of ShakeMap.
An additional component of the event detector is the development of an automatic e-mail alert capability. Currently there are two optional e-mail alert systems. A rapid PGV alert is issued by e-mail to interested parties, such as Ontario Power Generation (OPG), when the recorded PGV exceeds 1 mm/s (about MMI ≅ 3) at any specified stations of concern; the e-mail provides the PGV value and the time it was recorded at each of the stations. A “shake-mail” alert is sent automatically whenever ShakeMap is triggered. More detailed information is contained in the shake-mail message, including the estimated magnitude, location, and PGV values along with an expected Modified Mercalli Intensity (MMI) value for all stations. An alternative to this shake-mail capability is to log on to the Web site (www.shakemap.carleton.ca) and get full details on the ShakeMap events as they are posted. Both systems can be configured according to the needs of the recipient.
The centroid program was originally developed by Kanamori (1993) to locate earthquakes using ground-motion amplitude data. The program aims to determine the centroid of the ground motions, which is the point from which the ground motions appear to radiate. This may not coincide with the actual epicenter, particularly in the case of an extended source. Kanamori (1993) used peak ground acceleration (PGA) as the amplitude parameter while we use the vertical component of peak ground velocity (PGV). PGV was chosen because it is the simplest and most rapidly available ground-motion parameter from our network stations. An important advantage in using the centroid program is near-real-time estimation of the ground-motion centroid (rather than waiting until an epicenter becomes available), thus speeding the generation of ShakeMaps. Our use of the centroid approach is a departure from practice in California and other regions of the United States, where the epicenter is used. To our knowledge, the centroid concept has not been implemented in any other region, possibly because more rapid and reliable epicenters may be available elsewhere, where stations are more densely distributed.
Centroid Location and Magnitude
When an earthquake is detected, the centroid program uses the vertical component peak amplitude data at all stations within the 2-minute window to find the location and magnitude of the event. The centroid is a geographic location near the largest recorded PGV from which the ground motion appears to radiate, based on the pattern of observed amplitudes. The centroid magnitude is the earthquake magnitude that best explains the observed ground-motion amplitudes given the centroid location and regional ground-motion relations that give PGV as a function of magnitude and distance.
To determine the centroid location and magnitude, we fit the vertical component PGV recorded values for the event at all stations to the empirical ground-motion relations developed by Kaka and Atkinson (2005a): (1) where Y is the vertical component PGV in mm/s, R is hypocentral distance in km, and M is moment magnitude. A grid search technique is used to find the magnitude and location that minimizes the misfit to equation 1.
We then create a rectangular grid of phantom stations, spaced at 0.3° apart. The PGV values are estimated for all phantom stations using equation 1, based on the centroid magnitude and location. The PGV estimation at phantom stations is necessary to constrain ShakeMap interpolations in areas where no actual records are available. However, phantom stations are ignored if there is an actual record available within 10 km of a phantom station. ShakeMap then uses a combination of the estimated PGV at phantom stations and the actual measured PGV at all seismographic recording stations to create a contour map of PGV.
Before creating a contour map, the vertical component PGV values are converted to the equivalent horizontal component (in order to generate a horizontal component PGV map). We use the average horizontal-to-vertical ratio (H/V ratio) of PGV at each POLARIS station to convert from vertical to horizontal. We assume [PGV(horizontal) = PGV(vertical) × 1.21] as a default ratio for rock stations with unknown H/V-ratio values (Siddiqqi and Atkinson 2002). Soil sites are assumed to be further amplified by a soil response factor of 2.38 for PGV, applicable to generic stiff soil/soft rock sites (National Earthquake Hazards Reduction Program [NEHRP] C), as suggested by Adams and Halchuk (2003). This factor is based on amplification through a typical seismic velocity gradient, according to the quarter-wavelength approach (Boore and Joyner 1997).
The horizontal component PGV map is also translated into a map of felt intensity, using the relation of Kaka and Atkinson (2004): (2) where MMI is Modified Mercalli Intensity, Y is the horizontal component PGV in mm/s, and R is hypocentral distance in km. At this stage of development, it is assumed when predicting intensity that all sites across Ontario are NEHRP C (firm ground) (except the POLARIS stations, which are classified by their H/V ratio). This can be refined as surficial geology information becomes available to develop a grid of NEHRP site classification and amplifications across the province.
A question that arises is: Why use the vertical PGV for ShakeMap when horizontal component data are available and conversion to the horizontal component is required to produce the final products? It is advantageous to base ShakeMap on the vertical component in order to minimize the effects of site response when the centroid location and magnitude are being determined. The vertical component also will give a more robust interpolation of the base ground motions across the network. Use of the vertical component allows us to decouple the site effects from the other aspects of the program. This results in better overall performance in a sparse network. As additional information on site response in various regions is obtained, it can be readily incorporated into ShakeMap by changing the grid of amplification factors.
ACCURACY OF PGV ESTIMATION
For ShakeMap, we are interested in the prediction of ground motions rather than the determination of the earthquake epicenter and magnitude. To examine how well the centroid program estimates PGV values, we analyzed 24 local/regional earthquakes of MN 2.3 to 5.4, as listed in table 2, and compared estimated PGV values obtained using the centroid location to those obtained using the epicentral location. In figure 5, we plot the PGV residuals (calculated by dividing each observed PGV by the predicted PGV) for the centroid and the epicenter as a function of distance for all stations. In figure 6, PGV residuals are plotted as a function of moment magnitude (M). M values for most of the events are taken from Atkinson (2004). However, for some small events, M values were estimated from the empirical relationship given by Sonley and Atkinson (2005).
Figure 5 shows an apparent trend of decreasing residuals with increasing distance for both the epicentral and centroid representations. There is a significant underprediction of PGV near the earthquake source for the centroid solutions; this is partly due to the inaccuracy of the centroid in representing the actual location of the earthquake. For the epicenter case, the near-source residuals are more satisfactory, although the overall trend with distance is of course the same, resulting in overprediction of amplitudes at distances beyond 200 km. This suggests that for cases where the centroid is close to the epicenter, the bias in estimated ground motion at distances less than 200 km will be small.
Similar residual trends with distance have been inferred by studying Fourier amplitude data from POLARIS stations in southern Ontario, in comparison to values expected based on regional ground-motion models derived for all of eastern North America (Snider 2005). Snider (2005) did not find any such trend when examining just those stations from eastern Ontario, which suggests that the residual-distance trend may be attributable to attenuation in southern Ontario that is slightly different from the average attenuation for all of eastern North America.
To interpret this unexpected residual-distance trend, we prepared a number of plots of the data trends of PGV with distance in various magnitude bins. Figure 7 shows typical plots (other magnitude bins indicate similar trends), in which we compare the attenuation of PGV for the data just in southern Ontario (used to evaluate the centroid algorithm) with the attenuation of data from the larger eastern North America dataset used in deriving empirical ground-motion relations developed by Kaka and Atkinson (2005a). The equation used to predict PGV from distance and magnitude for the upper and lower magnitude values of the data plotted is also shown (equation as per Kaka and Atkinson 2005a).
Figure 7 suggests that two factors cause the residual trends noted in southern Ontario. The first is an apparent “Moho bounce” effect (e.g., Atkinson 2004) that makes the actual attenuation more complicated than the simple functional form used; this is manifested by a flattening in the observed decay of PGV near 100 km, as waves reflected from the Moho join the direct waves at sites in this distance range. The equation smooths through this detail, leading to inaccuracies in the prediction. The inaccuracy caused by neglecting the Moho bounce effect is likely responsible for the larger ground-motion amplitudes seen at distances closer to the epicenter. This trend implies an underestimate of intensity (by up to 1 MMI unit) at close distances (< 30 km). A second factor is that it appears that attenuation in southern Ontario is somewhat faster with distance than for eastern North America as a whole, particularly in the near-source region. The near-source data are still sparse, and this observation needs to be verified with additional data collected over time.
Figure 7 has highlighted two specific issues that were not addressed in the development of empirical ground-motion relations (which were performed in an earlier phase of the study): (1) “Moho bounce” complications in the shape of the attenuation function; and (2) the apparently faster regional decay of ground-motion amplitudes in southern Ontario. Furthermore, the database used in deriving the empirical relations on which ShakeMap was founded does not include 22 recent (2004–2005) significant regional events (see table 3). This highlights the need to continue to refine the models of attenuation that are included in ShakeMap as additional data are gathered by the program. The relations currently used in ShakeMap are the “first-generation” and can be significantly improved over the next year or two as data of this type make the definition of regional attenuation trends more robust. In the meantime, it should be kept in mind that limitations in the current relations imply that ShakeMap may underestimate MMI in the near-source region by up to 1 unit, and it may overestimate at large distances (hundreds of km) by a similar amount.
We observed no significant trend of residuals with magnitude (see figure 6). Overall the centroid performance in terms of predicting recorded ground motions appears to be satisfactory, subject to the caveats above. In fact, the MN 5.4 6 March 2005 Rivière-du-Loup earthquake demonstrates that predicted PGV agrees quite well with the recorded PGV (Kaka and Atkinson 2005b). Use of the epicenter would not improve the residuals on average, but it may better predict motions near the epicenter in cases where no nearby stations are available. However, the centroid is available in near real time while the epicenter is not.
PERFORMANCE OF SHAKEMAP
Since the beginning of June 2004, prototype near-real-time ShakeMaps have been produced for earthquakes of M ≥ 2.8 and are posted at our Web site (http://www.shakemap.carleton.ca) within minutes of occurrence. Because our ShakeMap grid region is not as seismically active as other areas of Canada, we use a relatively low magnitude threshold (M 2.8); this typically results in one to three events per month. Information on moderate events is important in this region due to its large population and concentration of critical facilities. Because the events generally are small, ShakeMap is primarily an information service rather than an emergency response tool.
In table 3 and figure 8, we compare automatically posted near-real-time ShakeMap centroid locations to those of the epicentral solutions/earthquake locations produced by the Canadian National Earthquake Database (NEDB) (http://www.seismo.nrcan.gc.ca/nedb/eq_db_e.php). Note that ΔM in the table was calculated by converting NEDB Nuttli magnitude (MN) to M using the empirical relationship given by Sonley and Atkinson (2005):
In figure 9 we plot both solutions for two selected local and regional events listed in table 3 along with the contoured PGV values from ShakeMap. All other ShakeMaps can be viewed at http://www.shakemap.carleton.ca. It is apparent from figures 7 and 8 that the centroid provides a good location estimate in most cases, even with a limited number of recording stations. Reliable ShakeMaps can be produced for events of M ≥ 2.8 located within the network. If the epicenter falls outside the network, the centroid location generally will not coincide with the actual earthquake location; in this case ShakeMap will not provide a reliable estimate of the event location. However, ShakeMap still provides reasonable estimates of PGV and intensity values across the network.
For generating reliable near-real-time ShakeMaps for earthquakes in Ontario, we conclude that:
events of M ≥ 2.8 within the network (42°–48°N and 76°–82°W) can be detected based on the exceedance of peak ground-motion amplitude of 0.003 mm/s at five or more stations with NSD > 0.7 (where NSD is the normalized standard deviation of the PGV normalized by the mean); and
the ground motion centroid provides a suitable estimate of magnitude and location to estimate peak ground-motion values for sparsely instrumented areas within the network.
Implementation of ShakeMaps in Ontario is an ongoing project and we are working to improve the following aspects:
refining of site response factors throughout the region;
evaluation of near-real-time ShakeMaps for other ground-motion parameters including peak ground acceleration (PGA) and response spectra (PSA(f)); and
improvement of ground-motion relations to predict PGV as a function of magnitude and distance.
The ShakeMaps can be viewed as they are posted in near real time at http://www.shakemap.carleton.ca.
ShakeMap is being transferred now to the Geological Survey of Canada (GSC) where it will operate more effectively in the long term. This transfer broadens and improves ShakeMap coverage because it will incorporate the Canadian National Seismographic Network (CNSN) stations. The existing ShakeMap version 2.4B is being upgraded to the latest version 3.1. Transferring ShakeMap to the GSC is a key step in providing near-real-time seismic information on all felt events to a broader audience. In the future, ShakeMap will be expanded to southwestern British Columbia in conjunction with the Pacific Geosciences Centre for ongoing improvements and more widespread application across western Canada.
This work could not have been accomplished without the collaborative efforts of Tim Cote and John Adams of the Geological Survey of Canada and Bill Jack of Carleton University. We thank Siew Ling Soh and Jianling Hua for computer programming and Dariush Motazedian and Eleanor Sonley for their feedback throughout the work. This work was supported financially by Ontario Power Generation and the Ontario Research and Development Challenge Fund.
Department of Earth Sciences, Carleton University