WEBVTT Kind: captions Language: en-US 00:00:01.600 --> 00:00:03.680 Today I will be talking about a fixed-network 00:00:03.680 --> 00:00:06.776 smartphone-based approach to earthquake early warning. 00:00:06.800 --> 00:00:11.336 I will report on an implementation of this in Costa Rica that we call ASTUTI. 00:00:11.360 --> 00:00:15.920 I’d like to acknowledge a large group of collaborators and USAID for funding. 00:00:18.160 --> 00:00:21.016 The motivation for our work is listed in these bullets. 00:00:21.040 --> 00:00:24.080 The earliest projections of earthquake early warning system performance and 00:00:24.080 --> 00:00:27.680 benefits to society have been modified and scaled back as the theoretical, 00:00:27.680 --> 00:00:32.216 empirical, practical, and social limitations of EEW are elucidated. 00:00:32.240 --> 00:00:35.680 Low-cost consumer-grade sensors in smartphones are now sensitive enough 00:00:35.680 --> 00:00:40.616 to be effectively used in EEW systems and even in crowdsourced modes. 00:00:40.640 --> 00:00:45.120 A one-size-fits-all approach to EEW, however, may not be appropriate. 00:00:45.120 --> 00:00:48.640 Achieving drop, cover, and hold on, for instance, has different requirements 00:00:48.640 --> 00:00:52.880 and costs than automatic mitigation measures for expensive infrastructure. 00:00:52.880 --> 00:00:56.696 Here we examine the low-cost end of this EEW spectrum. 00:00:56.720 --> 00:01:00.240 Specifically, we ask, can off-the-shelf smartphones deployed in a fixed 00:01:00.240 --> 00:01:04.560 network provide EEW in a reliable and effective enough manner so that 00:01:04.560 --> 00:01:07.760 significant populations receive warnings in time for drop, cover, 00:01:07.760 --> 00:01:11.096 and hold on protective actions to be undertaken? 00:01:11.120 --> 00:01:14.240 The symbology of these figures will be repeated through this talk. 00:01:14.240 --> 00:01:17.280 The copper-color map shows a 1-kilometer population data set. 00:01:17.280 --> 00:01:20.320 The capital city, San José, is denoted with an SJ and is the 00:01:20.320 --> 00:01:23.840 region of densest Costa Rican population in the center of the country. 00:01:23.840 --> 00:01:26.480 The white triangles are phones installed in the field. 00:01:26.480 --> 00:01:29.600 From September to December of 2019, we constructed the 00:01:29.600 --> 00:01:32.720 ASTUTI network in Costa Rica. Currently the network comprises 00:01:32.720 --> 00:01:35.760 82 stations with plans for a future densification phase. 00:01:35.760 --> 00:01:38.320 The colored circles show earthquake epicenters coded 00:01:38.320 --> 00:01:41.040 according to magnitude. The colored squares show 00:01:41.040 --> 00:01:44.856 Did You Feel It? reported Modified Mercalli Intensity values. 00:01:44.880 --> 00:01:47.920 Costa Rica’s seismic hazard is due primarily to earthquakes generated 00:01:47.920 --> 00:01:50.696 by oblique subduction at the Middle America Trench. 00:01:50.720 --> 00:01:54.400 Since 1853, eight earthquakes greater than magnitude 7 have 00:01:54.400 --> 00:01:58.216 occurred either on the northern Nicoya or southern Osa peninsulas. 00:01:58.240 --> 00:02:03.920 As shown on the right, recently the magnitude 7.6 2012 Nicoya Peninsula 00:02:03.920 --> 00:02:07.760 earthquake caused shaking throughout the entire country with PGA values 00:02:07.760 --> 00:02:13.440 as high as 0.5 to 1.4 g and MMI 5 to 7 reported in San José. 00:02:13.440 --> 00:02:15.840 It is generally accepted that magnitude 6 and above 00:02:15.840 --> 00:02:18.056 earthquakes are felt country-wide. 00:02:18.080 --> 00:02:20.800 Given the size of the tectonic framework, we designed ASTUTI 00:02:20.800 --> 00:02:25.040 with these principles. First, our EEW efforts are focused on warning people, 00:02:25.040 --> 00:02:28.400 and our targeted user response is drop, cover, and hold on. 00:02:28.400 --> 00:02:30.960 Second, we prioritized detecting and alerting for 00:02:30.960 --> 00:02:33.040 Middle America Trench earthquakes. 00:02:33.040 --> 00:02:36.560 Third, we attempt to issue warnings at the earliest occurrence of detected 00:02:36.560 --> 00:02:40.616 events of potential concern rather than waiting until they exceed some size. 00:02:40.640 --> 00:02:43.040 Fourth, because of local ground motion variability 00:02:43.040 --> 00:02:46.320 and because it is unlikely that earthquake ruptures are deterministic, 00:02:46.320 --> 00:02:49.920 our detection alerting is entirely non-parametric. 00:02:49.920 --> 00:02:54.000 Fifth, we evaluate the scenario where every alert will be issued 00:02:54.000 --> 00:02:56.960 for the entire country. And finally, sixth, because of 00:02:56.960 --> 00:02:59.760 the previous design priorities, there will likely be a number of 00:02:59.760 --> 00:03:03.120 false alerts issued for smaller events when users will feel no shaking 00:03:03.120 --> 00:03:05.120 although an event was correctly detected. 00:03:05.120 --> 00:03:08.720 This will require rapid post-event messaging to remind users that 00:03:08.720 --> 00:03:11.920 the system performed correctly even if they did not feel shaking. 00:03:11.920 --> 00:03:15.360 We stress, however, that no public alerts have yet been issued, aside from those 00:03:15.360 --> 00:03:17.760 sent to our small group of beta testers. 00:03:17.760 --> 00:03:22.499 We leave thorough investigation of this topic to a future presentation. 00:03:23.680 --> 00:03:26.560 Phones with Android operating systems were installed in the ground floor 00:03:26.560 --> 00:03:29.280 of buildings to protect the boxes. Criteria for phone choice was 00:03:29.280 --> 00:03:32.376 based on a combination of in-country availability and cost. 00:03:32.400 --> 00:03:35.200 Our control and sending app on the phones is called QED 00:03:35.200 --> 00:03:38.640 for Quick Event Detection. We also use a mobile management 00:03:38.640 --> 00:03:41.896 platform to permit remote root access to the phones. 00:03:41.920 --> 00:03:46.080 As shown in the diagram on the right, the key flow of information is from 00:03:46.080 --> 00:03:51.040 the phones to a UDP receiver and an MQTT broker that then permits 00:03:51.040 --> 00:03:55.200 distribution of the data to any number of processing and archiving clients. 00:03:55.200 --> 00:03:59.416 This makes it easily scalable and trivial to add new networks. 00:03:59.440 --> 00:04:03.360 From December 2019 to late August 2020, phones streamed 00:04:03.360 --> 00:04:05.520 accelerometer data at 10 hertz. 00:04:05.520 --> 00:04:08.800 Subsequently, we increased streaming rate to 100 hertz. 00:04:08.800 --> 00:04:15.288 At 100 hertz, each site records roughly 70 to 100 megabytes per day. 00:04:15.840 --> 00:04:19.520 Any EEW processing algorithm that operates on accelerometer data can 00:04:19.520 --> 00:04:23.280 subscribe to the MQTT broker. Here we report on a new algorithm, 00:04:23.280 --> 00:04:27.760 PGAN, for Peak Ground Acceleration with N vertices, that is an adaption 00:04:27.760 --> 00:04:31.280 of the PLUM – Propagation of Local Undamped Motion – method 00:04:31.280 --> 00:04:35.120 for a network of smartphones. PGAN is similar to PLUM in that it 00:04:35.120 --> 00:04:39.280 is a non-parametric ground motion- based algorithm which does not attempt 00:04:39.280 --> 00:04:42.936 to estimate anything about a detected earthquake source. 00:04:42.960 --> 00:04:46.640 PGAN requires multiple neighboring stations to experience anomalous 00:04:46.640 --> 00:04:51.440 accelerations in order to trigger an alert. The network is divided into a polygonal 00:04:51.440 --> 00:04:56.136 mesh with a configurable number of vertices greater than or equal to 3. 00:04:56.160 --> 00:05:02.296 Here we report primarily on quadrilateral PGAN-4 configurations. 00:05:02.320 --> 00:05:05.680 Once-per-second PGA values are measured at each station, and the 00:05:05.680 --> 00:05:09.736 values are compared with the primary threshold, currently 0.6% g. 00:05:09.760 --> 00:05:13.200 If any station is above the threshold, the polygon it is part of is marked as 00:05:13.200 --> 00:05:16.240 potentially triggered. The PGA values at each of the 00:05:16.240 --> 00:05:19.200 remaining stations in the polygon it is part of are compared with a 00:05:19.200 --> 00:05:24.400 secondary threshold, currently 0.55% g. If all stations in a potentially triggered 00:05:24.400 --> 00:05:27.440 polygon are above the thresholds, an alert is issued. 00:05:27.440 --> 00:05:30.880 An example of this is the green triangle of polygons from the 00:05:30.880 --> 00:05:35.200 earthquake – the red star in the figure. Once a polygon is triggered, an alert 00:05:35.200 --> 00:05:40.229 is sent out using the Amazon Web Services Simple Notification system. 00:05:41.520 --> 00:05:44.800 Here is an example of some of the raw data from the accelerometers. 00:05:44.800 --> 00:05:47.680 We combined the raw acceleration data from the three orthogonal 00:05:47.680 --> 00:05:51.235 MEMS accelerometers – V, H2, and H1 in the figure – 00:05:51.267 --> 00:05:54.640 into a PGA data type sampled and sent at 1 hertz. 00:05:54.640 --> 00:05:58.880 This is called the P, for PGA, message. The P message is the vector norm 00:05:58.880 --> 00:06:02.640 of the individually de-meaned 100 hertz acceleration values and 00:06:02.640 --> 00:06:06.400 has units of meters per second squared. An example of an earthquake and 00:06:06.400 --> 00:06:09.896 two small aftershocks are shown on the left figure. 00:06:09.920 --> 00:06:13.200 To mitigate the transient noise spikes common to many smartphones, 00:06:13.200 --> 00:06:16.880 while permitting real seismic accelerations to pass, the P message 00:06:16.880 --> 00:06:22.296 implements a simple filter that takes the 30% highest in a 1-second window. 00:06:22.320 --> 00:06:26.338 An example of the noise mitigation is on the right. 00:06:29.520 --> 00:06:33.120 We characterize site and network data quality by examining histograms of 00:06:33.120 --> 00:06:37.200 PGA data in the figure on the left. Generally, the entire network exhibits 00:06:37.200 --> 00:06:42.720 mean P values of 0.25% g – the dark gray histogram in the figure on the left. 00:06:42.720 --> 00:06:45.200 Over periods of time varying from minutes to days, 00:06:45.200 --> 00:06:49.073 individual stations may experience elevated deviations from this background 00:06:49.073 --> 00:06:53.120 behavior, which would be the red-line histogram in the figure on the left. 00:06:53.120 --> 00:06:56.720 The causes of these transient elevated periods of background noise are varied 00:06:56.720 --> 00:07:00.240 and likely related to temporally changing site characteristics, 00:07:00.240 --> 00:07:04.160 including anthropogenic noise. On the right, from six months of 00:07:04.160 --> 00:07:06.960 continuously streamed data, we find that data latency varies 00:07:06.960 --> 00:07:11.496 from 0.35 to 0.45 seconds, depending on the time of day. 00:07:11.520 --> 00:07:14.960 For comparison, published data latencies for the Italian, Chinese, 00:07:14.960 --> 00:07:16.800 and ShakeAlert earthquake early warning networks 00:07:16.800 --> 00:07:21.120 are 0.9, 2, and 1 to 3 seconds, respectively. 00:07:22.480 --> 00:07:26.240 We test the full operation of the ASTUTI system using the smartphone’s 00:07:26.240 --> 00:07:30.880 programmable vibration feature. We used the M7.6 2012 Nicoya 00:07:30.880 --> 00:07:34.160 earthquake as a scenario event and programmed S wave arrival time 00:07:34.160 --> 00:07:37.760 for each site. Over a period of three days, we vibrated the phones in the 00:07:37.760 --> 00:07:42.800 Nicoya Peninsula three times an hour, resulting in a total of 216 simulations. 00:07:42.800 --> 00:07:45.280 An example of one simulation is on the left. 00:07:45.280 --> 00:07:48.880 The magenta title shows detection time, and the magenta circle shows the 00:07:48.880 --> 00:07:51.360 S wave position at the time of detection. 00:07:51.360 --> 00:07:54.160 The green triangles show the triggered stations. 00:07:54.160 --> 00:07:58.000 On the right, we show that triggered detection latency for the PGAN-4 00:07:58.000 --> 00:08:02.856 method has a mean value of 12 to 13 seconds for all of the simulations. 00:08:02.880 --> 00:08:06.400 We measured the alert latency by sending text message alerts to a set of 00:08:06.400 --> 00:08:10.320 15 people with cell phones in and around the greater San José region. 00:08:10.320 --> 00:08:14.160 We find that the mean value for alerting latency is on the order of 4 seconds. 00:08:14.160 --> 00:08:16.960 Because this is from a small number of phones that were relatively close 00:08:16.960 --> 00:08:20.400 together, it is not clear how representative this metric is of latencies 00:08:20.400 --> 00:08:22.880 that would result from sending many thousands of alerts across 00:08:22.880 --> 00:08:25.600 a wider geographic region. To the best of our knowledge, 00:08:25.600 --> 00:08:28.800 none of the other smartphone EEW projects, such as MyShake, 00:08:28.800 --> 00:08:33.893 or the Earthquake Network Project, have reported alerting latency data. 00:08:34.800 --> 00:08:37.680 Because of COVID-19-related travel restricting the development of our 00:08:37.680 --> 00:08:40.560 alerting messaging algorithm, the entire system was not operational 00:08:40.560 --> 00:08:43.520 until August of 2020. Accordingly, here we evaluate 00:08:43.520 --> 00:08:48.320 performance based on offline playback from December 2019 to June 2020. 00:08:48.320 --> 00:08:51.520 During the period of assessment, ASTUTI, using the PGAN-4 algorithm, 00:08:51.520 --> 00:08:55.200 detected 5 of the 13 events that were accompanied by Did You Feel It? 00:08:55.200 --> 00:08:57.440 reports of shaking somewhere in Costa Rica. 00:08:57.440 --> 00:09:00.880 The majority of these events – 9 of 13 – occurred outside 00:09:00.880 --> 00:09:04.936 of the network, and they ranged in magnitude from 5.3 to 5.9. 00:09:04.960 --> 00:09:08.720 Only one of the detected events, from March 7th of 2020, had an 00:09:08.720 --> 00:09:13.016 epicenter entirely within the network. That is the example shown in the figure. 00:09:13.040 --> 00:09:15.840 In addition to the associated Did You Feel It? data, we plot the 00:09:15.840 --> 00:09:19.520 estimated position of the S wave front at the time that the alert was issued 00:09:19.520 --> 00:09:23.280 with the solid magenta circle. In the plots, we add to the detection 00:09:23.280 --> 00:09:27.040 time 5 seconds – the median value of the alert time – and 15 seconds 00:09:27.040 --> 00:09:30.376 for total estimate of drop, cover, hold on available time. 00:09:30.400 --> 00:09:34.320 That’s shown as the dashed circle. In this example, detection was 00:09:34.320 --> 00:09:39.120 at 11 seconds after origin time. At least 20 seconds passed until 00:09:39.120 --> 00:09:42.320 that S wave would reach the San José dense population, 00:09:42.320 --> 00:09:45.736 suggesting that drop, cover, and hold on could be achievable. 00:09:45.760 --> 00:09:50.240 Generally, our detection times for PGAN-4 ranged from 11 to 30 seconds. 00:09:50.240 --> 00:09:54.560 This range for subduction zone events compares well with systems such as 00:09:54.560 --> 00:09:58.560 SASMEX and ShakeAlert. We classify alerting and felt shaking 00:09:58.560 --> 00:10:02.216 outcomes for users as true positive shaking – TP-S; 00:10:02.240 --> 00:10:07.120 TP-ns – true positive no shaking; NA – no alert – the system correctly 00:10:07.120 --> 00:10:10.616 detects and event, but a user does not receive an alert prior to shaking; 00:10:10.640 --> 00:10:14.560 FA – false alert – the system incorrectly detects an event when none occurred; 00:10:14.560 --> 00:10:16.800 and MA – missed alert – a felt event occurs, 00:10:16.800 --> 00:10:19.870 but the system does not issue an alert. 00:10:22.080 --> 00:10:26.000 For each event, we plot P message records sections showing the expected 00:10:26.000 --> 00:10:30.480 range of P and S wave velocities of 6 to 3 kilometers per second, 00:10:30.480 --> 00:10:34.320 respectively, shown as the pink envelope. The figure on the right 00:10:34.320 --> 00:10:39.040 is a zoom of the record section on the left showing only the triggering stations, 00:10:39.040 --> 00:10:41.816 which are indicated in bold in the figure on the left. 00:10:41.840 --> 00:10:45.336 The triggering time is shown as the vertical magenta line. 00:10:45.360 --> 00:10:49.840 Clearly, the trigger occurs with the earliest PGA elevated values, 00:10:49.840 --> 00:10:53.336 which correspond with predicted P wave arrival times. 00:10:53.360 --> 00:10:56.560 We find that generally two of the five ASTUTI-detected 00:10:56.560 --> 00:10:59.840 events triggered on the P wave. 00:11:02.080 --> 00:11:05.520 Generally, the ASTUTI-detected events were accompanied by stronger shaking 00:11:05.520 --> 00:11:08.800 that was felt by much larger percentages of the population. 00:11:08.800 --> 00:11:13.440 The detected events had median and max MMI levels of 4.3 and 6, 00:11:13.440 --> 00:11:17.840 with roughly 17 to 73% of the population experiencing felt shaking, 00:11:17.840 --> 00:11:21.920 while the non-detected events had median and max MMI levels of 2.9 00:11:21.920 --> 00:11:26.960 and 3.8, with zero to 19% of the population experiencing felt shaking. 00:11:26.960 --> 00:11:30.800 The figure on the right shows all the non-detected events and the paucity of 00:11:30.800 --> 00:11:36.087 Did You Feel It? reports indicates the low levels of shaking from these events. 00:11:37.360 --> 00:11:40.960 For comparison, here is an offshore event that ASTUTI detected, but rather 00:11:40.960 --> 00:11:44.720 than P, it triggered on the S wave. See the record sections on the right 00:11:44.720 --> 00:11:49.680 for clear S triggering. The 29-second detection time for this event was more 00:11:49.680 --> 00:11:52.960 than 2 times slower than other events for two related reasons. 00:11:52.960 --> 00:11:56.800 The first is the event was offshore. And the second is it triggered 00:11:56.800 --> 00:12:00.880 on the S wave. Despite this, more than 60% of the Costa Rican 00:12:00.880 --> 00:12:04.080 population would have gotten the alert prior to S wave arrival. 00:12:04.080 --> 00:12:06.320 But that would have only permitted a couple of seconds 00:12:06.320 --> 00:12:09.820 for mitigating actions in San José. 00:12:11.120 --> 00:12:14.640 The alerting outcomes for the five detected events are shown in the table. 00:12:14.640 --> 00:12:17.680 For the PGAN-4 algorithm, the zero percent false alarm rate 00:12:17.680 --> 00:12:21.040 is lower than all other EEW systems reported false alarm rates, 00:12:21.040 --> 00:12:25.096 aside from retrospective testing of PLUM with West Coast U.S. data. 00:12:25.120 --> 00:12:29.120 Detection times range from 11 to 30 seconds, depending on 00:12:29.120 --> 00:12:33.440 offshore or onshore origins. The detection times offshore 00:12:33.440 --> 00:12:36.656 are notably longer than from events closer to shore. 00:12:36.680 --> 00:12:45.016 TP-S percentages vary from 6 to 63%. TP-ns percentages vary from 27 to 71%. 00:12:45.040 --> 00:12:48.536 And the NA percentages are less than 10%. 00:12:48.560 --> 00:12:52.480 As expected, a low detection threshold criteria, combined with a country-wide 00:12:52.480 --> 00:12:57.016 alerting region, also leads to significant TP-ns outcomes. 00:12:57.040 --> 00:13:00.720 It’s not clear what penalty there may be in terms of user engagement if 00:13:00.720 --> 00:13:04.480 an EEW system provides alerts without felt shaking, especially if 00:13:04.480 --> 00:13:07.280 users were to receive a rapid follow-on message stating that 00:13:07.280 --> 00:13:11.176 an event had been correctly detected, although they did not feel shaking. 00:13:11.200 --> 00:13:14.720 In general, recent social science research suggests that the boy who cried 00:13:14.720 --> 00:13:19.360 wolf assumption may be overstated. From the little social science work done 00:13:19.360 --> 00:13:25.200 in EEW, it appears that EEW users may tend to be tolerant of TP-ns outcomes. 00:13:25.200 --> 00:13:30.960 For instance, Nakayachi et al. found that 75% of EEW users rated the Japanese 00:13:30.960 --> 00:13:35.760 EEW system positively, even though a significant population 00:13:35.760 --> 00:13:39.976 received TP-ns, what they term false alarm, outcomes. 00:13:40.000 --> 00:13:42.720 Furthermore, users reported that these outcomes were not the 00:13:42.720 --> 00:13:45.816 main reasons for negative perceptions of EEW. 00:13:45.840 --> 00:13:49.920 Clearly, a high priority must be EEW pre- and post-event education 00:13:49.920 --> 00:13:53.626 and messaging and research in this – on this topic. 00:13:54.640 --> 00:13:58.080 So the take-home messages are operating in Costa Rica over 00:13:58.080 --> 00:14:01.120 six months, the ASTUTI network detected five events that caused 00:14:01.120 --> 00:14:04.160 widespread felt shaking and had zero false alarms. 00:14:04.160 --> 00:14:06.560 ASTUTI detected the events that mattered. 00:14:06.560 --> 00:14:09.520 Large percentages of the population experienced shaking for detected 00:14:09.520 --> 00:14:12.080 events, while very small percentages of the population 00:14:12.080 --> 00:14:14.216 experienced shaking for missed events. 00:14:14.240 --> 00:14:16.960 The detections and alerts would have provided time for drop, cover, 00:14:16.960 --> 00:14:20.000 and hold on protective actions to be taken before S wave arrival 00:14:20.000 --> 00:14:23.096 for large percentages of the Costa Rican population. 00:14:23.120 --> 00:14:26.560 Two of the five events were triggered by P waves on the phones, suggesting 00:14:26.560 --> 00:14:30.696 that smartphone-based EEW could be more effective than previously thought. 00:14:30.720 --> 00:14:34.160 And finally, fixed-network smartphone-based EEW can provide 00:14:34.160 --> 00:14:37.040 performance on par with scientific-grade EEW, 00:14:37.040 --> 00:14:41.280 but capital and annual costs for ASTUTI – around $20,000 each – 00:14:41.280 --> 00:14:44.857 are two orders of magnitude greater than for ShakeAlert.