Earthquake early warning system in Liaoning, China based on PRESTo

1.   Introduction

Liaoning is located in northeast China and is considered as part of the Northeast Asia Active Block (Zhang, 2003). The region is under the influence from the east by the subduction of the Pacific Plate under the Eurasian Plate and from the southwest by the collision of the Indian Plate with the Eurasian Plate. As shown in Figure 1, the topography of Liaoning Province varies in ESE direction with uplifts on both the east and west border regions and a depression in between. A major NE-trending fault zone, the Tancheng-Lujiang fault zone, also known as Tanlu fault zone (TLFZ), crosses the middle of the province in a northeast direction. The TLFZ has a mainly right-lateral strike-slip motion and is considered to be the eastern margin of the North China basin (Allen et al., 1997; Yin, 2010). It is seismically active and faults in and around the TLFZ are predominantly striking northeast. All major historical earthquakes in eastern China, and in Liaoning Province in particular, are related to the TLFZ. Although not as seismically active as other earthquake-prone regions such as the western Pacific or the Himalayas, there are frequent earthquakes in Liaoning and surrounding areas (Figure 1) with a dozen or so strong historical earthquakes, such as the 1944 M_W6.6 earthquake in Dandong (39.887°N, 124.148°E) near the China-Korea border; the famous 1975 Haicheng earthquake (M_S7.5), which was arguably the first successfully “predicted” major earthquake in human history (Wang et al., 2006); and the disastrous 1976 Tangshan earthquake (M_W7.6) which occurred in the neighboring Hebei Province and killed more than 200,000 people. There are also occasional moderate (M5~6) earthquakes with very shallow depths that can cause extraordinary damages, such as the 2013 M_S5.1 Dengta earthquake (Su et al., 2020). Therefore, Liaoning is a region with a high earthquake hazard potential and, given the high population density and recent economic development, earthquake early warning (EEW) is very important for the mitigation of seismic disasters.

Figure  1.  Map of Liaoning Province and surrounding areas. Background color shows the topography. Black triangles indicate locations of the seismic stations of Liaoning seismic Network (LNNet). The black lines represent major faults (fault data are from https://gmt-china.org/data/) in the region, and the thick red lines depict the Tanlu fault zone. Black dots are epicenters of earthquakes of 2.5 ≤ M_L < 3.0. Red squares and stars mark the epicenters of 25 earthquakes of magnitudes 3.0 < M_L ≤ 3.5 and 3.5 < M_L ≤ 6.0, respectively, during 2009–2019 whose records are used in this study to implement PRESTo for LNNet. Orange and red circles are strong historical earthquakes of magnitudes 6.0 ≤ M < 7.0 and M ≥ 7.0 since 1900, respectively. White circles show major cities in Liaoning Province

DownLoad: Full-Size Img PowerPoint

In the past few decades, with the development in seismic monitoring networks and telecommunications technologies, EEW systems have been developed and put into routine operation in many parts of the world, including Japan (UrEDAS, Kamigaichi et al., 2009), Mexico (SASMEX, Espinosa-Aranda et al., 2009), California (ShakeAlert, Kohler et al., 2018), Italy (PRESTo, Zollo et al., 2009; Satriano et al., 2011), Turkey (SOSEWIN, Fleming et al., 2009) and Taiwan (Hsiao et al., 2009). Since the turn of the century, especially after the disastrous M_W7.9 Wenchuan earthquake in 2008, EEW systems have been established quickly in earthquake prone provinces in Chinese mainland, such as in Sichuan and Yunnan region (Peng et al., 2013; 2015, 2017; 2019; 2020; Peng and Yang, 2019), and in Fujian (Zhang et al., 2016). This study is the first effort in using the recently deployed seismic network in Liaoning Province in northeast China to build an effective EEW system.

In establishing the EEW system for the Liaoning region, we employ the open source software PRobabilistic and Evolutionary early warning SysTem (PRESTo). PRESTo was developed by the RISSC (RIcerca in Sesmologia Sperimentale Conputazionale) laboratory of the University Federico II in Naples, Italy. All the PRESTo-related files including software as well as documentations such as installation instruction and user manual can be freely downloaded from its official website (http://www.prestoews.org). The software integrates recent algorithms for real-time, rapid earthquake detection, location, magnitude estimation and damage assessment into an easily configurable and portable package (Satriano et al., 2011). PRESTo has been under active experimentation in southern Italy on the Irpinia Seismic Network (ISNet). It is a readily adaptable and user-friendly platform and has been adopted in the EEW operations in many seismic networks worldwide (e.g. Picozzi et al., 2015; Pitilakis et al., 2016).

2.   Method and seismic data

We first make a brief introduction about the main concept of PRESTo for the benefit of discussion in Section 3 on our implementation to the Liaoning region. Theoretical and technical details on PRESTo can be found on the official website of the package (http://www.prestoews.org) as well as references listed therein.

PRESTo is composed of four core algorithms for event detection, location, magnitude determination and ground motion prediction (Satriano et al., 2008). The first one is FilterPicker (FP) for automatic, real-time phase picking (Lomax et al., 2012; Vassallo et al., 2012). FP is designed on the basis of the classical short-term average/long-term average (STA/LTA) algorithms (Allen, 1982; Baer and Kradolfer, 1987) and can realize real-time phase picking from continuous data streams with high efficiency and accuracy. FP adopts two picking thresholds S₁ and S₂, and the picking is carried out when the value of a characteristic function exceeds S₁ and meanwhile the integral of the characteristic function exceeds S₂. The second core algorithm in PRESTo is real-time evolutionary earthquake location algorithm (RTloc) (Satriano et al., 2008) for real-time evolutionary earthquake location. It starts locating the event as soon as the first station is triggered (i.e. when the P-wave is detected and picked by FP), and stations that are not triggered can also be included to reduce the uncertainty of the location result. The third core algorithm is RTmag (Lancieri and Zollo, 2008), which uses an empirical relationship to estimate the magnitude M of the earthquake from the peak ground displacement (PGD), d_peak and the epicentral distance R. The fourth core algorithm of PRESTo is for ground motion estimation, which calculates the ground motion based on the event magnitude using appropriate ground motion prediction equations (GMPEs). The user interface of PRESTo for Liaoning is shown in Figure 2.

Figure  2.  Screenshot of the user interface of PRESTo in processing the archived records from the LNNet near the origin time of the 10 January 2013 M_L3.9 earthquake (event #2 in Table 1). The left panel shows the z-component of velocity data streams and picking process. Waveforms in the time windows highlighted in yellow are used to estimate the magnitude. The right panels illustrate the location result (top: latitude and longitude; middle: depth) and magnitude estimation process (bottom). The red star denotes the hypocenter with the estimated magnitude also shown. The tetrahedrons are the stations of the LNNet, with the color indicating the peak displacement at the station. The yellow and red circles represent the wave front of P- and S- wave, respectively. The red line in the bottom panel denotes the origin time of the earthquake. The location and magnitude estimation results are given at the bottom.

DownLoad: Full-Size Img PowerPoint

In this study, our goal is to implement the PRESTo platform to realize the EEW system for Liaoning region using the Liaoning seismic network LNNet. The station distribution of LNNet is shown in Figure 1. Currently, the LNNet has 37 stations, 5 of which belong to the national network equipped with JCZ-1, CTS-1E, CTS-1EP, CTS-1E and CTS-1EF instruments, respectively, while the rest are equipped with BBVS-60 seismometers. The implementation involves configuring the region-specific files and finding the optimal parameters in the four core algorithms in PRESTo. For this purpose, we use archived LNNet broadband records of 25 earthquakes of magnitudes 3.0 < M_L ≤ 6.0 in Liaoning from November 2012 to March 2017. The 25 events are listed in Table 1 and their locations are shown in Figure 1. Once PRESTo can run successfully for archived data, the EEW system can be put into practical operation by plugging in real-time data streams from the Liaoning seismic network.

Table  1.  Events used for the implementation of PRESTo

Event No.	Event date	Long. (°E)	Lat. (°N)	Depth (km)	M	ΔLong. (°)	ΔLat. (°)	PRESTo Depth (km)	PRESTo Mag.	Use
1	2012-11-01	122.383	40.483	–	3.3	0.003	0.012	7.805	3.4	S
2	2013-01-10	123.500	39.500	10	3.9	0.021	0.056	6.727	3.7	R
3	2013-01-21	122.400	42.900	8	3.9	0.021	0.042	11.039	3.9	R
4	2013-01-23	123.217	41.483	7	5.1	0.021	0.003	2.234	5.1	B
5	2013-03-30	122.390	40.520	6	3.6	0.004	0.024	7.805	3.4	R
6	2013-04-22	122.400	42.900	6	5.3	0.020	0.024	11.039	5.2	R
7	2013-04-25	122.370	42.930	7	3.6	0.008	0.009	10.680	3.6	R
8	2013-05-10	122.340	42.930	7	3.7	0.050	0.070	9.781	3.7	R
9	2014-04-17	122.867	40.650	–	3.2	0.030	0.022	8.344	2.9	S
10	2014-04-28	122.300	40.467	–	3.4	0.024	0.016	6.367	3.3	S
11	2014-06-26	122.317	40.467	–	3.4	0.030	0.023	8.164	3.6	S
12	2014-07-10	121.117	39.317	–	3.1	0.173	0.187	6.100	2.9	S
13	2014-08-19	122.217	40.133	–	3.1	0.044	0.035	4.570	2.8	S
14	2014-08-22	122.317	40.467	6	3.8	0.018	0.044	11.938	3.8	B
15	2015-08-04	122.433	40.483	6	4.3	0.025	0.012	9.242	4.4	B
16	2015-08-19	122.400	40.517	–	3.1	0.053	0.035	2.953	2.8	S
17	2015-08-25	122.917	39.483	–	3.1	0.022	0.223	0.000	3.1	S
18	2015-11-23	122.450	40.817	9	4.0	0.014	0.015	7.805	4.3	B
19	2016-04-24	122.783	39.567	–	3.1	0.004	0.128	0.617	3.0	S
20	2016-05-22	122.100	41.630	6	4.6	0.008	0.015	24.516	4.3	R
21	2016-05-22	122.080	41.620	6	4.3	–	–	–	–	–
22	2016-07-03	122.633	40.700	–	3.4	0.022	0.045	0.000	3.1	S
23	2016-10-29	119.750	41.330	7	3.8	0.080	0.010	19.484	3.7	R
24	2017-01-17	121.367	41.367	–	3.2	0.038	0.029	12.117	2.8	S
25	2017-03-04	122.500	42.067	–	3.4	0.024	0.004	10.680	3.0	S
Note: The Liaoning regional earthquake catalog does not have event depth. Depths in this table for events of magnitude ML ≥ 3.5 are from the catalog of China Earthquake Networks Center (CENC, http://www.cenc.ac.cn/). The two catalogs have the same longitude and latitude for the same event. ΔLong. and ΔLat. are differences in event longitude and latitude, respectively, and are defined as the PRESTo values minus the catalog values. The last column gives how the event is used in PRESTo implementation in this study (S: simulation; R: regression for magnitude estimation equation; B: both). Events #20 and #21 are too close in time, and only event #20 can be processed by PRESTo. The differences between catalog and PRESTo in epicenters, depths and magnitudes are shown in Figure 5

 | Show Table

DownLoad: CSV

3.   Implementation of PRESTo to LNNet

In general, PRESTo is a user-friendly software platform that can easily be implemented in different regions using local network such as the LNNet data. However, a number of parameters used by the core algorithms in PRESTo must be tuned based on data from specific regions. Therefore, our aim in this study is to determine the optimal set of parameters for the LNNet and discuss the performance of the EEW system. For the implementation and offline testing purposes, PRESTo can be run in simulation mode in which it reads the SAC files from the archived records and converts them into data streams to simulate the actual early warning operation using real-time data. The implementation of PRESTo involves four major tasks: configuration of region-specific files according to the network information; building the travel-time table for all seismic stations involved; supplying the equation for magnitude estimation and ground motion prediction; and setting the optimal values of miscellaneous parameters for the algorithms in PRESTo.

3.1   Region-specific information

When implementing PRESTo to a new seismic network, several text files containing information on the seismic network and stations such as network and station names and coordinates, need to be configured. In particular, a list of target sites can be supplied by a text file, to which EEW alerts can be sent.

Once the target region for the EEW coverage is determined based on the station distribution, a map file for the target region needs to be provided for the user interface. For the EEW using LNNet records, we choose the target region with longitude and latitude ranges of 118.4°E–127.0°E and 38.3°N–43.6°N (only partly shown in the user interface in Figure 2), respectively. The depth range of the target region is from the surface down to depth of 40.0 km.

3.2   Travel-time table calculation

The speed of operation is one of the key requirements for the EEW system. A standard approach used by seismic networks is to establish a pre-calculated travel-time table that can be looked up during real-time operations to speed up the process. In order to cover all possible locations of earthquakes, a network composed of grid points is set up in the entire target region and the travel-time table contains the travel times of P and S waves between all the grid points to all the stations. The grid size is determined based on the density of the stations and the expected performance of the EEW system. Given the station and target site locations, the grid and the crustal velocity model, the travel-time table can be generated by NLLoc (Lomax et al., 2000). We tested a 7-layer model of Haicheng (HC) for the Haicheng region (Figure 1) used by Zhao and Chen (2017) to study the ground motion of the 1975 Haicheng earthquake.

We tested the location algorithm using the two velocity models in Table 2 with different grid sizes for generating travel time tables and compared the event locations result to find the most appropriate velocity model and the optimal grid size. Table 3 provides the locations for the 23 January 2013 M_L5.1 earthquake (event #4 in Table 1). The results show clearly that the grid size of 4 km is too coarse, leading to relatively large uncertainties in location and origin times, and the size of 1 km does not improve the performance further while increasing the calculation cost. Therefore, we use the model HC with a grid size of 2 km in PRESTo for LNNet.

Table  2.  Crustal velocity model HC (Zhao and Chen, 2017) tested in this study

Layer	Depth to layer top (km)	vP (km/s)	vS (km/s)
1	0.0	2.5	1.07
2	1.0	6.10	3.53
3	15.0	6.20	3.59
4	20.0	6.10	3.53
5	22.0	6.50	3.76
6	26.0	7.10	4.12
7	32.0	8.02	4.46

 | Show Table

DownLoad: CSV

Table  3.  Location results for the 23 January 2013 earthquake

Model	Grid size (km)	Origin time	RMS (s)	Long. (°E)	εx (km)	Lat. (°N)	εy (km)	Depth (km)	εz (km)
HC	4	04:18:16.73	1.207	123.205	4.5	41.5172	3.3	22.375	0.0
HC	2	04:18:15.39	0.486	123.196	1.4	41.4863	0.9	2.234	1.2
HC	1	04:18:15.45	0.478	123.191	1.9	41.4848	1.7	4.543	1.7
Note: εx, εy and εz are PRESTo’s location uncertainties in longitude, latitude and depth, respectively

 | Show Table

DownLoad: CSV

3.3   Magnitude estimation equation

As the same of many other EEW platforms, PRESTo estimates the magnitude of a detected event using the peak ground displacement (PGD) in the first few seconds (usually 2–4 s) after the P-wave arrival. Once the value of d_peak is determined at a station following the detection of an event, the magnitude M can be calculated by the region-specific empirical equation (Wu and Zhao, 2006; Zollo et al., 2006; Lancieri and Zollo, 2008):

where R is the hypocentral distance in kilometer. a, b and c are region-specific constants to be determined by regression using the magnitudes and d_peak measurements for earthquakes in existing catalog.

In this study, we chose 13 earthquakes with magnitudes 3.5 < M_L ≤ 6.0 (see Table 1, red squares and stars in Figure 1). We first use FilterPicker to automatically pick the P-wave arrivals from z-component records. Then the records are removed of means and the signal-to-noise ratios (SNRs) are calculated for the records in the 2-s and 4-s windows following the picked P arrivals. The SNR of a record is defined as the ratio of the maximum amplitude (norm of the vector sum of three components) in the 2-s or 4-s P-wave window to the root mean square (RMS) of the record in the 10-s window before the picked P-wave arrival. We only use records with SNR ≥ 5. Next, the three-component records are converted to ground velocities by removing the magnification factor and bandpass filtered to 0.075–25 Hz. After integration to displacement, the d_peak is obtained by finding the maximum displacement in the 2-s or 4-s time window after P-arrival. We obtain a total of 303 d_Peak values for the 2-s window and 321 for the 4-s window. These d_peak values are used to determine the constants a, b and c that are specific to the Liaoning region based on the M_L reported in the Liaoning regional earthquake catalog. These constants are listed in Table 4 together with their uncertainties.

Table  4.  Constants in magnitude prediction for Italy and Liaoning

Window Length	Region	Component	a	εa	b	εb	c	εc
2 s	Italy	Z,N,E	−7.69	0.06	1.00	0.00	−1.89	0.03
	Liaoning	Z,N,E	−4.05	0.70	0.81	0.11	−1.32	0.25
4 s	Italy	Z,N,E	−7.69	0.06	1.00	0.00	−1.89	0.03
	Liaoning	Z,N,E	−4.53	0.62	0.90	0.09	−1.21	0.23
Note: εa, εb and εc are uncertainties of a, b and c, respectively.

 | Show Table

DownLoad: CSV

In Figure 3, we display the d_peak-determined magnitudes of the 13 earthquakes of magnitudes 3.5 < M_L ≤ 6.0 by equation (1) using Liaoning-specific constants. The results demonstrate the effectiveness of the constants in Table 4 in determining the magnitudes of local earthquake in Liaoning from the d_peak values (average of 2 s and 4 s) using equation (1). It may be expected that the region-specific constants can be further improved to reduce the uncertainties in the constants and magnitudes by using d_peak data from more earthquakes.

Figure  3.  Performances of the magnitude prediction equation for the 13 earthquakes of magnitudes 3.5 < M_L ≤ 6.0 in Liaoning using different region-specific constants in Table 4. Horizontal and vertical axes are the catalog magnitude M_L and magnitude M_PGD calculated by PRESTo, respectively. Black circles with error bars are obtained using Liaoning-specific constants. The diagonal solid line indicates M_PGD = M_L, and the two dashed lines show the range of 0.3 from the solid line

DownLoad: Full-Size Img PowerPoint

3.4   Ground motion prediction equations

Following the detection and location of an event, and the magnitude estimation, PRESTo uses the ground motion prediction equations (GMPEs) to calculate the peak ground acceleration (PGA) and peak ground velocity (PGV) distributions in the EEW coverage region. If the estimated magnitude M_PGD ≤ 4.0, the PGV and PGA are calculated using the USGS ShakeMap Small Regression (Wald et al., 1999) equation:

with an uncertainty ${\sigma }_{1}$. In equation (2), M and R are the event magnitude and epicentral distance in kilometer, respectively, and x can be either PGV (in cm/s) or PGA (cm/s²), with corresponding constants b₁, b₂, b₅, H and ${\sigma }_{1}$ determined by regression using PGV or PGA data from historical earthquakes. Table 5 lists the constants for Italy.

Table  5.  Values of the regression constants in Equation (2) and associated uncertainties for Italy

x	b1	b2	b3	H	σ1
PGV	2.223	0.740	−1.386	6.0	0.3268
PGA	4.037	0.572	−1.757	6.0	0.3667

 | Show Table

DownLoad: CSV

For events of magnitude M_PGD > 4.0, the GMPE used in PRESTo is (Akkar and Bommer, 2007):

with the uncertainty:

Again, x in equation (3) can either be PGV (in cm/s) or PGA (cm/s²) with corresponding constants. The default values in PRESTo for the constants in equations (3) and (4) are given in Table 6.

Table  6.  Values of the regression constants in Equations (3) and (4)

x	${b}_{1}$	${b}_{2}$	${b}_{3}$	${b}_{4}$	${b}_{5}$	${b}_{6}$	${s}_{1}$	${s}_{1m}$	${s}_{2}$	${s}_{2m}$
PGV	−1.36	1.06	−0.079	−2.95	0.31	5.55	0.85	−0.096	0.31	−0.040
PGA	1.65	0.77	−0.074	−3.16	0.32	7.68	0.56	−0.049	0.19	−0.017

 | Show Table

DownLoad: CSV

It should be noted that to fully establish the GPMEs for Liaoning region it is necessary to determine the regression constants in equations (2–4) using sufficiently large number of strong-motion records of various magnitudes from a range of epicentral distances, especially for events with relatively large magnitudes. More observational efforts are needed to accumulate records over a longer period of time to obtain reliable GMPEs for Liaoning region. It would be a significant undertaking and is beyond the scope of this study. For the moment, we implement the GMPEs using the values in Tables 5 and 6 for Italy, which can be easily replaced once the regression constants are available for Liaoning.

3.5   Miscellaneous parameters

PRESTo also has a number of miscellaneous parameters to control the detection, phase picking and event location processes. We run PRESTo in simulation mode using archived records for 15 earthquakes in Table 1 to find the optimal values of the parameters for LNNet. The values for the parameters are determined by trial-and-error process. The optimal values of the main parameters are given in Table 7.

Table  7.  Optimal values of miscellaneous parameters in PRESTo for ISNet and LNNet

(a) Parameters for picking
Parameter name	Description	ISNet value	LNNet value
picker_filterWindow	filter window length, setting filter frequency band	0.5 s	0.5 s
picker_longTermWindow	long-term window length for calculating average signal amplitude	5.0 s	7.0 s
picker_threshold1	threshold of characteristic function to trigger picking	10.0	8.0
picker_threshold2	threshold of integral of characteristic function to trigger picking	10.0	8.0
picker_tUpEvent	time after reaching threshold1 to check if threshold2 is reached	0.5 s	0.5 s
(b) Parmaters for event binding and location
Parameter name	Description	ISNet value	LNNet value
binder_stations_for_coincidence	minimum number of triggered stations in coincidence time window to declare an event	6	4
binder_secs_for_coincidence	length of coincidence window	3 s	20 s
binder_secs_for_association	length of window for all picks to belong to same event	10 s	50 s
binder_quakes_life	length of time for earthquake parameters be refined	15 s	60 s
binder_quakes_separation	minimum time interval of first picks of two earthquakes	30 s	200 s
binder_apparent_vel_stations_spacing	average distance between stations	30 km	70 km
binder_apparent_vel_max_distance	maximum distance between stations	120 km	700 km
locate_use_non_triggering_stations	whether to use non-triggering stations to locate earthquakes	1 (true)	1

 | Show Table

DownLoad: CSV

Owing to the large region of Liaoning, the picking thresholds of LNNet is smaller than those of ISNet. Also, due to the lower station density of LNNet, the time windows for binding are larger than those of ISNet.

4.   Discussion

In the process of determining the optimal values of the parameters in Table 7, we find that they have different influence on the detection and location of events. By comparing the resulting RMS values of the predicted arrival times and uncertainties in event locations, we can determine the optimal values of the parameters. Tables 8 and 9 present an example for determining the parameters picker_filterWindow and picker_longTermWindow for the 23 January 2013 event. The results show clearly that when the values of these two parameters are too large, the uncertainties in event location increase. Thus, we choose the values of 0.5 s for picker_filterWindow and 7 s for picker_longTermWindow.

Table  8.  Influence of picker_filterWindow on location uncertainty for the 23 January 2013 event

picker_filterWindow values (s)	RMS (s)	Long. (°E)	εx (km)	Lat. (°N)	εy (km)	Depth (km)	εz (km)
0.001	0.523	123.185	1.6	41.4935	1.5	0.617	2.2
0.01	0.523	123.185	1.6	41.4935	1.5	0.617	2.2
0.05	0.474	123.196	1.6	41.4863	1.1	2.234	1.3
0.1	0.473	123.196	1.6	41.4863	1.1	2.234	1.4
0.2	0.491	123.196	1.5	41.4863	1.1	2.234	1.5
0.3	0.491	123.196	1.2	41.4863	0.8	2.234	1.7
0.4	0.486	123.196	1.5	41.4863	1	2.234	1.4
0.5	0.486	123.196	1.5	41.4863	1	2.234	1.4
1.0	0.486	123.196	1.5	41.4863	1	2.234	1.4
10.0	0.486	123.196	1.5	41.4863	1	2.234	1.4

 | Show Table

DownLoad: CSV

Table  9.  Influence of picker_ longTermWindow on location uncertainty for the 23 January 2013 event

picker_longTermWindow values (s)	RMS (s)	Long. (°E)	εx (km)	Lat. (°N)	εy (km)	Depth (km)	εz (km)
1	0.492	123.196	1.4	41.4863	1	2.953	2
2	0.484	123.196	1.6	41.4863	1.3	2.234	1.7
3	0.484	123.196	1.6	41.4863	1.3	2.234	1.6
4	0.484	123.196	1.5	41.4863	1	2.234	1.4
5	0.484	123.196	1.5	41.4863	1.2	2.234	1.6
7	0.486	123.196	1.5	41.4863	1	2.234	1.4
10	0.486	123.196	1.2	41.4863	0.8	2.234	2

 | Show Table

DownLoad: CSV

From the simulation results, we find that the parameter picker_threshold2 plays an important role in phase picking and location. Empirically, the values of picker_threshold1 and picker_threshold2 can be simply set equal, which we denote as S. Figure 4 illustrates the phase picking and event location results with different S values in the range of 6–10 s. Combining the results of mean RMS, mean location uncertainty and the mean number of phase picks, we choose S=8 as the optimal picking thresholds for LNNet.

Figure  4.  The influence of the value of S (picker_threshold1 and picker_threshold2) on phase picking and location. (a) Variation of RMS of arrival picks with S. (b) Variation of location uncertainty ε_r (maximum of ε_x, ε_y and ε_z) with S. (c) Variation of the number of phase picks with S. Black, red and blue lines represent the average of all 15 events, events of lower magnitudes (M_L < 3.5) and events of higher magnitudes (M_L ≥ 4.0), respectively

DownLoad: Full-Size Img PowerPoint

The parameter binder_stations_for_coincidence also needs to be considered carefully. If it is too small, the rate of false alarm is high, whereas if it is too large, the lead time (the time interval between first alert and S-wave arrival at a target site, where the S-wave arrival is directly read from the travel-time table) is too short. Thus, its preferable value is between 3 and 5. The length of the location time windows are closely related to the average inter-station distance of the seismic network and the regional average P-wave speed. Therefore, the related parameters binder_secs_for_coincidence, binder_secs_for_association, binder_quakes_life and binder_quakes_separation are determined to be 20 s, 50 s, 60 s and 100 s, respectively.

When the value of S (the value of parameters picker_threshold1 and picker_threshold2) is low, more phases with relatively low SNRs can be picked, which reduces the rate of missed picks, but at the same time increases the rate of false picks. Higher values of S have the opposite effect. On the other hand, for small earthquakes with a large value for the parameter binder_apparent_vel_stations_spacing, if the picking threshold S is large, the results become worse when non-triggering stations are used in location (i.e. locate_use_non_triggering_stations is set as 1). The reason for this is that P waves at some stations that should have been picked are not picked. However, while under the optimal value of S, using non-triggering stations gives better results, as shown in Table 10.

Table  10.  Location uncertainties of several earthquakes with optimal value of S = 8

Event date	Use non-triggering stations				Only use triggering stations
	εx (km)	εy (km)	εz (km)		εx (km)	εy (km)	εz (km)
2012-11-01	3.4	2.7	1.8		3.4	3.2	3
2013-01-10	2.4	3.3	2.7		2.4	3.5	2.9
2014-04-17	3.7	2.4	3.8		3.9	2.5	3.8
2015-08-19	4.2	6.6	2.8		5.1	7.1	4.0
2016-04-24	2.1	3.6	1.5		6.3	7.1	4.5
2017-01-17	3.4	3.1	2.9		3.3	3	3.6

 | Show Table

DownLoad: CSV

In Table 1, we also compare the PRESTo location results for the 24 earthquakes of magnitudes M > 3.0 with those in the catalogs (See Figure 5 for their differences). For most events, the discrepancies in latitudes and longitudes between catalog and PRESTo locations are less than 0.05º. Events with discrepancies larger than 0.05º are located near the edge of the network (event #23 in Table 1) with low magnitudes (event #12). This demonstrates the reliability of the event locations by PRESTo.

Figure  5.  (a) Differences between catalog and PRESTo epicenters (blue dots) and depths (red dots). The event numbers are same as in Table 1. (b) Difference between catalog and PRESTo magnitudes

DownLoad: Full-Size Img PowerPoint

For events near the edge of the LNNet, stations far from the epicenter may not be triggered, which will result in few triggered stations and large location errors. One way to solve this problem is to enlarge the target region by including stations in neighboring provinces.

Another important performance factor of an EEW system is the time efficiency, which is reflected by the length of the lead time mentioned earlier. The balance between time efficiency and warning accuracy is controlled by the parameter binder_stations_for_coincidence (see Table 7). Figure 6 shows the variation of lead time (binder_stations_for_coincidence is set to 4) with epicentral distance. We can see that the lead time at 100-km epicentral distance is ~15 s. More stations are needed to improve the time efficiency of the EEW system in LNNet.

Figure  6.  The variation of lead time with epicentral distance. The crosses are the simulation results of earthquakes with different epicentral distance to Shenyang. The dashed line is the linear fit with a slope of 0.2033 s/km and an intercept of –5.2001 s

DownLoad: Full-Size Img PowerPoint

It is worth noting that because of the simple and stable crustal structure in Liaoning region, the amplitude of SmS wave is larger than that of S wave in the distance range around 100 km (generally SmS/S < 2). This affects the peak ground motion as described in Mori and Helmberger (1996), which should be considered in establishing a reliable GMPE.

5.   Conclusions

In this study, we implement the EEW software PRESTo to Liaoning seismic network. Using archived seismic records of past earthquakes, we determin the optimal values for LNNet using PRESTo. The velocity model for Haicheng region with a grid size of 2 km has been used to generate the travel-time table. A magnitude estimation equation is established by regression which can be used to calculate the magnitudes of earthquakes in Liaoning region from the peak amplitudes in the initial 2-s or 4-s of the P-wave signals. Other parameters used by PRESTo are also determined by trial-and-error approach.

We carefully investigat the influence of the picking threshold S. The value for the picking threshold is slightly smaller for LNNet (S = 8) than for ISNet (S = 10), due to LNNet’s larger station spacing than ISNet. The discrepancies between the location results by PRESTo and those in the catalogs are mostly below 0.05º (~5 km). Finally, the lead time of PRESTo-based EEW system for Liaoning is approximately a linear function of the epicentral distance.

With the optimal parameters determined in this study, the PRESTo system works well for Liaoning region, and can be easily put into routine earthquake early warning operation by linking it with the real-time data stream from LNNet. The system can be improved by further parameter tuning using records of future earthquakes.

Acknowledgments

The PRESTo software package can be downloaded at its official website http://www.prestoews.org/. The program NLLoc for generating the travel-time table can be downloaded at http://alomax.free.fr/nlloc/. We thank the developers of the PRESTo platform for making their products available to the public. Data used in this study are provided by Liaoning Earthquake Agency. All figures except for Figure 2 are generated by the Generic Mapping Tools (Wessel and Smith, 1998). The authors acknowledge the course English Presentation for Geophysical Research of Peking University in improving the manuscript.