# Bulletin of the Seismological Society of America

## Abstract

To mitigate the potential seismic disasters dominated by the Longmen Shan fault zone (LFZ) in the Sichuan region, building up suitable magnitude‐estimation models for earthquake early warning (EEW) systems is important. In this article, the records from the mainshocks and aftershocks of the 2008 Wenchuan (*M*_{w} 7.9) and 2013 Lushan earthquakes (*M*_{w} 6.6), which occurred in the Sichuan region, were used to develop the maximum predominant and characteristic periods ( and *τ*_{c}) and peak ground displacement (*P*_{d}) parameters for estimating earthquake magnitude. The developed correlations were evaluated and compared with previous studies in other regions. The parameter is correlated with magnitudes in the 4–6 and 6–8 ranges, whereas *τ*_{c} parameter scales are correlated with the entire magnitude range without evident saturation. Nevertheless, the linear slope is slightly lower than the previous studies in other regions show. As for *P*_{d} parameter, it is confirmed as a good estimator and performs most similarly with the compared regions with the magnitude of 4–6 range. Indeed for large magnitude, we observe the saturation effect. The longer time‐window length coupled with narrowing the filtering bandwidth can improve the saturation; however, the linear slope decreases. Speculatively, the different performances can perhaps be attributed to the regional characteristics in the LFZ region. Our works offer an insight into the feasibility of the EEW system in Sichuan, China.

## Introduction

On 12 May 2008, an earthquake of *M*_{s} 8.0 (*M*_{w} 7.9) shocked the Wenchuan County, Sichuan province in China. The epicenter of this earthquake is 31.0° N, 103.4° E with the focal depth of 14 km and is located in the mid‐north of the Longmen Shan fault zone (LFZ). The seismic source mechanism inversion shows that the Wenchuan earthquake is a thrust faulting, with small amounts of right‐lateral strike‐slip, and the fault strike is 225° (Zhang *et al.*, 2008) with the direction of long axis rupture toward the northeast. Five years later, on 20 April 2013, another large earthquake shocked Lushan County, Sichuan, again. The surface‐wave magnitude of the Lushan earthquake is 7.0 (*M*_{w} 6.6), its epicenter is 30.3° N, 103.0° E with a focal depth of 13 km and is located in the south of the LFZ, with the fracture direction of long axis moved toward the southwest. The focal mechanism solutions show that the Lushan earthquake is caused by thrust faulting (Zhang, Xu, *et al.*, 2013). The focal mechanism solutions and the epicenter locations of the mainshocks and aftershocks of the two earthquakes happened within a week and are shown in Figure 1a. The sketch of the central fault, mountain front fault, and mountain back fault of the LFZ is shown in Figure 1b.

The two strong earthquakes in the Sichuan region are relatively independent, but both of them connect with the seismogenic structure. Their focal rupture types are very similar, mainly characterized by thrust fault property, and occurred in the LFZ, which is formed by the eastward extrusion from the Qinghai–Tibet plate, whereas the Sichuan basin plays the role of a hard crust block, located from northeast to southwest along the edges of Sichuan basin (Zhang, Deng, *et al.*, 2013). The Longmen Shan fault generally trends from Luding, Tianquan, and goes through Mao County, Wenchuan, Beichuan, Guangyuan to Mianxian County in Shaanxi Province, which is about 500 km long and 30–70 km wide. In recent years, the LFZ has trended to dominate large earthquakes in the Sichuan region. The fault ruptures in the 2008 Wenchuan and 2013 Lushan earthquakes are located in the mid‐north and south of the Longmen Shan fault, respectively. There is a rupture gap in the mid‐south of the fault, as shown in Figure 1a, and the rupture gap is perhaps a potential risk source for causing large earthquake. As a result, strengthening the monitoring of potential seismic risk in the LFZ, especially for developing earthquake early warning (EEW) systems for the Sichuan region, becomes particularly important.

Earthquake early warning (EEW) systems aim at providing seconds to tens of seconds to the target area for people to take emergency measures and reduce the loss before the arrival of destructive waves. Up until now, EEW systems have been used in many countries and regions, such as Japan (Nakamura, 1984, 1988; Odaka *et al.*, 2003; Kamigaichi, 2004; Horiuchi *et al.*, 2005), Mexico (Espinosa‐Aranda *et al.*, 2009), Romania (Bose *et al.*, 2007), Turkey (Alcik *et al.*, 2009), Italy (Zollo *et al.*, 2006, 2009), the United States (Allen and Kanamori, 2003; Allen *et al.*, 2009; Bose *et al.*, 2009), Taiwan (Wu and Kanamori, 2005b; Hsiao *et al.*, 2009), and so on. China is no exception, after the 5‐year testing of EEW systems in the Beijing capital region, last year the government launched a 5‐year program (total budget: $284 million U.S.) for establishing the EEW systems based upon a densely seismic network of more than 15,000 seismic observatories throughout the country. Among the above applications, there are two types of EEW systems, regional (network‐based) and onsite (stand‐alone) warning systems. For regional warnings, the seismograph obtained from stations or arrays is used to estimate an earthquake’s scale; for onsite warning, the seismograph used for identifying a *P* wave is from a single station and is used for earthquake‐scale estimation. Generally, the regional warning has a higher accuracy than the onsite one due to the use of more stations and dense sensors arrays. However, different from Japan or California, etc., up to the present moment, the seismic dense network has not been developed in China with the exception of the Beijing capital region. Hence, EEW systems of developing based on a single station, rather than on a network, is likely to be more valuable in the Sichuan region for the moment.

Among EEW systems, the real‐time magnitude estimation is one of the most important and difficult works. At present, many seismologists in the world have developed some real‐time magnitude‐estimation parameters, such as , *τ*_{c}, *τ*_{log}, *P*_{d}, and squared velocity integral (*IV*2) (Allen and Kanamori, 2003; Kanamori, 2005; Wu and Li, 2006; Festa *et al.*, 2008; Zollo *et al.* 2010; Ziv, 2014). Furthermore, some methods or models for magnitude estimation are proposed, such as the threshold‐based method (Zollo *et al.*, 2010), evolutionary approach via inversion of displacement spectra (Caprio *et al.*, 2011), and the application of *τ*_{c}×*P*_{d} (Huang *et al.*, 2015). Until now, magnitude‐estimation algorithms in EEW systems have not been developed in the Sichuan region, one of the largest seismic risk regions in China. As a result, it is worth performing a study on the parameters for rapid magnitude estimation in the Sichuan region.

In this study, we focus on three of the magnitude‐estimation parameters, including , *τ*_{c}, and *P*_{d}. We use the earthquake records from the mainshocks and aftershocks of the 2008 Wenchuan and 2013 Lushan earthquakes, which occurred in the Sichuan region, to build up the , *τ*_{c}, and *P*_{d} regression models for rapidly estimating earthquake magnitudes. Then, the three models were evaluated for their performances in the Sichuan region with different *P*‐wave time windows and filtering bandwidth for *P*_{d}, especially in regard to magnitude underestimation (saturation phenomenon) in large‐scale earthquakes. Finally, the three regression models are compared with those developed based on the dataset of southern California, Japan, Taiwan, Euro‐Mediterranean, and Chile, which were determined in the previous studies, and some explanations for their differences were discussed.

## The Magnitude‐Estimation Parameters in This Study

### Maximum of Predominant Period,

In 1988, Nakamura put forward an algorithm to calculate the seismic predominant period based on real‐time velocity records, which has been applied to the Japanese Shinkansen EEW system, called UrEDAS (Nakamura, 1988). Allen and Kanamori (2003) believed that the maximum of *τ*_{p}(*I*), calculated based on the few seconds following the arrival of the *P* wave, is proportional to the earthquake’s magnitude, and so developed an algorithm called method (Allen and Kanamori, 2003), as shown in equation (1). This method has been applied to California EEW systems (Wurman *et al.*, 2007) and Earthquake Alarm Systems (Wu and Kanamori, 2005b). It should be noted that the correlation between and the strong event magnitude (*M*_{w}>6) is not related solely to the source properties but is derived from a combination of attenuation, site effect, and filtering effects (Wolfe, 2006; Yamada and Ide, 2008). (1)in which and , *x*_{i} is the vertical velocity signal, and *α* is a smoothing constant, which is set as 0.999 in this study.

### Characteristic Period, *τ*_{c}

In 2005, Kanamori (2005) improved the method and put forward the *τ*_{c} parameter by adopting the fixed‐interval integration instead of the original step‐by‐step integration. The value of *τ*_{c} indicates the average of the periods within *t* seconds after the arrival of *P* waves in the vertical component, as shown in equation (2). Currently, this model has been applied to Virtual Seismologist systems (Cua *et al.*, 2009). *τ*_{c} has similar physical meanings with , but they have different results for the same seismic records, which may be related to the noise level before the *P*‐phase arrival (Shieh *et al.*, 2008). (2)in which *u*(*t*) is the vertical displacement and *t* is the time‐window length, which is mostly set at 3 s.

### Peak Ground Displacement, *P*_{d}

In 2005, Wu and Kanamori (2005b) found that the peak initial displacement amplitude (*P*_{d}) correlates well with the peak ground velocity (PGV) and earthquake magnitude after high‐pass filtering for seismic data within 3 s after the arrival of the *P* wave. Wu and Li (2006) put forward the empirical relationship between *P*_{d} and the hypocentral distance *R* and the magnitude *M* based on the earthquake records in southern California, and used this relationship to estimate magnitude, shown as (3)in which *P*_{d} is the peak initial displacement amplitude after applying a 0.075‐Hz high‐pass recursive Butterworth filter to remove the low‐frequency drift after the numerical integration; *M* is the magnitude; *R* is the hypocentral distance; and *A*, *B*, and *C* are constants determined from the regression analysis.

Unlike Wu and Zhao, Zollo *et al.* (2006) investigate low‐pass‐filtered peak displacements based on strong‐motion data, and compute the *P*_{d} relationship for both *P* and *S* phases with equation (4) after distance is corrected. Lancieri and Zollo (2008) did further studies and believed that 2 or 4 s seismic data after the *P*‐wave arrival can also be used to estimate earthquake magnitude based on the seismic records within an epicenter’s distance of 50 km. This method has been applied on RTMag systems (Kanamori, 2005) and PRESTo systems (Lancieri and Zollo, 2008; Iannaccone *et al.*, 2009; Zollo, *et al.* 2009). Hereinafter, *P*_{d} will designate the parameter computed following Zollo *et al.* (2006): (4)

## Seismic Record Selections (Dataset)

A total of 937 earthquake records were used to develop the EEW parameter models for estimating magnitude in this study and are provided by the China Strong‐Motion Network Centre (CSMNC), China Earthquake Administration, in which the mainshocks and aftershocks of the *M*_{w} 7.9 Wenchuan (498 records) and *M*_{w} 6.6 Lushan earthquakes (439 records) are included. The catalog magnitudes are local magnitude *M*_{L} for events with *M*_{L}<6 and moment magnitude *M*_{w} for larger events (*M*_{w}≥6), which are re‐assigned using the Global Centroid Moment Tensor moment magnitude catalog and denotes both types of magnitudes simply by *M* in this study. Both of the event locations and magnitudes have been revised. We set up three criteria for seismic record selections: (1) the magnitude of the event is greater than 4; (2) to ensure good station coverage for each event and to avoid the near‐source effect (Yamada and Mori, 2009) and the complexity of path effects for *P* waves at longer distances, the hypocentral distances are in 20–100 km range; and (3) only the vertical component of the record is used. Figure 2 shows the distributions of the analyzed strong‐motion records as a function of magnitude and hypocentral distance.

The *P*‐phase arrivals from the vertical components and the first *S*‐phase arrivals from the horizontal components of all the selected strong‐motion records have been identified and manually picked, those whose signal‐to‐noise ratios (SNR) are high enough to trace the *P*‐ and *S*‐phase arrivals clearly. The *S*‐phase selection is used to determine the longer *P*‐phase time window to be used. The *P*‐phase time windows which are shorter than 2 s are not commonly used because the parameter measures are affected by trigger uncertainties, location error, and amplitude bumps introduced by the filter. Hence, starting from the manually picked *P*‐phase arrivals, we considered three different time windows (2, 3, and 4 s wide) on the 0.075‐Hz high‐pass‐filtered records to measure the *τ*_{p} and *τ*_{c}. To measure the peak ground displacement *P*_{d}, we considered three different time windows (2, 3, and 4 s) and three filtering bandwidths (0.075 Hz high‐pass filtered, 0.075–3 Hz band‐pass filtered, and 0.25–3 Hz band‐pass filtered).

In Figure 3a, we show the acceleration waveform for three events that span the large magnitude (*M*_{w}>6) with hypocentral distances nearly 70 km: the Wenchuan main event (*M*_{w} 7.9), the Lushan main event (*M*_{w} 6.6), and the Wenchuan main aftershock (*M*_{w} 6.1). The 2‐ and 4‐s *P*‐phase windows are highlighted in gray before the first *S*‐phase arrival. Figure 3b gives a simplified drawing for calculating , *τ*_{c}, and *P*_{d} from one of the *M*_{w} 7.9 Wenchuan earthquake records which is displayed in Figure 3a at the same station.

## Establishment and Evaluation of versus Magnitude Correlation

The first investigated parameter is the maximum of the predominant period . Using the recursive definition given in equation (1), we compute for the seismic records in our dataset and correlate it with the magnitude based on equation (5) in time‐window lengths of 2, 3, and 4 s after *P*‐phase arrival. The data have been high‐pass filtered at 0.075 Hz using the two‐pole Butterworth filter. The computed result is shown in Figure 4, whereas the detail of the correlated result is reported in Table 1. (5)in which is the maximum of *τ*_{p}(*i*), *M* denotes the catalog magnitude, and *a* and *b* represent constants determined by regression best fitting.

The best‐fitting regression equations for each time‐window length can be obtained based on regression analysis. Figure 4a–c shows the results for fitting the data into different time‐window lengths (2, 3, and 4 s). The parameter shows a large dispersion throughout the magnitude range. Because there are different numbers of seismic records for each earthquake bin, the mean value (triangles in Fig. 4) is used to get the fitting line, which was also used by Allen and Kanamori (2003) and Kanamori (2005) for getting a smaller deviation. The parameters are correlated with magnitudes in the 4–6 and 6–8 ranges, however, with the large overall standard error, especially in the large magnitude. It clearly scales with the final magnitude in the 4–6 range, optimally in 3 s. Nevertheless, for magnitude higher than 6, it shows a weak dependence on the magnitude and evident underestimation with the lower slope. Because of the large weighted standard error and statistical insignificance, the regression in *M* 6–8 range is not considered. saturates for *M*_{w} 6 in our dataset. The similar phenomenon appears in the values retrieved for Chile data (Lancieri *et al.*, 2011) with the similar slope. With the increasing of the time‐window length (longer than 3 s), saturation improves slightly, but the parameters in the 4–6 range disperse in 4 s. The scatter might attribute to the accumulation of a small amount of low‐frequency noise, with the accumulated *X* and *D* terms in equation (1). Based on the comparison between the correlation coefficients, the best‐fitting regression equation for the in this study is (6)in the time‐window length of 3 s and with correlation coefficient of 0.97.

## Establishment and Evaluation of *τ*_{c} versus Magnitude Correlation

Based on the recursive definition given in equation (2), we compute *τ*_{c} and correlate it with magnitude based on equation (7) in several time‐window lengths (2, 3, and 4 s after *P*‐phase arrival). The record has been high‐pass filtered at 0.075 Hz using the two‐pole Butterworth filter. The computed result is shown in Figure 5, whereas the detail of the correlated result is reported in Table 2: (7)in which *τ*_{c} is the computed parameter, *M* denotes the catalog magnitude, and *a* and *b* represent constants determined by regression best fitting.

Figure 5a, 5b, and 5c shows the results for fitting the data in time‐window lengths of 2, 3, and 4 s, respectively. Unlike parameter, *τ*_{c} clearly scales with the entire magnitude range (4–8 range) with the larger overall standard error in 2 and 3 s. It should be noted that the evident saturation does not occur. The discreteness becomes smaller as time‐window length increases to 4 s. Nevertheless, the slope of 0.161±0.012 is slightly lower than the values retrieved for the Japan, California, Taiwan (Kanamori, 2005), and Chile datasets (Lancieri *et al.*, 2011). The lower slope implies less sensitivity of *τ*_{c} for the change of magnitude, meaning that it is less predictive for the larger magnitude. We will discuss this in the Comparison of Regressions for Sichuan Region and Other Regions section. Based on comparing the correlation coefficients, the best‐fitting *τ*_{c} regression equation for our dataset is (8)in the time‐window length of 4 s and with a correlation coefficient of 0.87.

## Establishment and Evaluation of *P*_{d} versus Magnitude Correlation

The last investigated parameter is the peak ground displacement *P*_{d}, focused on the different filtering bandwidths and their impact on the correlations. We compute *P*_{d} by high‐pass filtered records at 0.075 Hz (used by Wu and Li, 2006) and by band‐pass filtered records at 0.075–3 and 0.25–3 Hz (used by Lancieri *et al.*, 2011) using the two‐pole Butterworth filter in various time‐window lengths (2, 3, and 4 s after *P*‐phase arrival), respectively.

To retrieve the magnitude dependence of peak amplitudes, we used equation (3) to correct the *P*_{d} parameter for hypocentral distance effect (assume the *C* coefficient is equal to 1) and to normalize them to a common distance of 1 km (used by Lancieri *et al.*, 2011). A linear regression curve of the form (9)has been determined using the average values of in each magnitude bin weighted by the inverse of the standard deviation. The detail of the correlated result is summarized in Table 3, along with the measured–weighted standard error. Figure 6a, 6b, and 6c shows the single measurements and average estimates of parameters after records band‐pass filtered at 0.075–3 Hz along with error bars and best‐fit regression (including the ±1 weighted standard error [WSE]) in time‐window lengths of 2, 3, and 4 s, respectively. This filter window was fixed in agreement with previous studies (Wu and Kanamori, 2005a; Yamada *et al.*, 2007; Zollo *et al.*, 2007).

The *P*_{d} parameter clearly scales with the final magnitude in the 4–6 range. Nevertheless, for a magnitude higher than 6 (or 6.5), it shows a weak dependence on the magnitude even when there is no evident correlation with magnitude in the short time‐window length. The same saturation problem of *P*_{d} amplitude was also observed for large earthquakes in Taiwan (Wu *et al.*, 2006) and the Hector Mine and Northridge earthquakes (Wu and Li, 2006), with the possible existence of saturation effect of the early *P*_{d} amplitude versus magnitude relationship for **M**>6.5. With the increasing time‐window length, the saturation improves slightly, but the underestimation is still evident in 4 s (Fig. 6c). This observation led us to do further investigation of the different filters and their impact on the correlations.

Figure 7a, 7b, and 7c shows the results for fitting the data in time‐window lengths of 4 s after *P*‐phase arrival using 0.075‐Hz high‐pass filtering, 0.075–3 and 0.25–3‐Hz band‐pass filtering, respectively. The coefficients are reported in Table 3. From Figure 7a–c, narrowing the filtering bandwidth improves saturation, with the decrease of the linear slope. Looking at the slopes evaluated in the 4–6 magnitude range, the value obtained using 0.075‐Hz high‐pass‐filtered data is 0.479, whereas the value obtained using 0.075–3‐Hz band‐pass‐filtered data is 0.375 and the value obtained using 0.25–3‐Hz band‐pass‐filtered data is 0.311. Considering the *R*‐squared correlation coefficient values, although the 0.25–3 Hz band‐pass filter limits the saturation effect, the *R*‐squared correlation coefficient value is quite small (only 0.65) whereas the values for the other two filters are 0.71 and 0.82. The 0.075–3‐Hz band‐pass‐filtered data have the highest goodness of fit when the 0.25–3‐Hz band‐pass‐filtered data have the lowest one. The improvement of saturation might just be due to the small linear slope and matches *P*_{d} in large magnitude, as well as the lower correlation, which is still worth discussing. In addition, the large‐scale earthquakes (especially *M*≥7) give full play to the role of energetic contribution of frequencies smaller than 0.25 Hz, so the 0.25–3 Hz band‐pass filter will invisibly lower the proportion of the large events, seeming to limit the saturation effect.

## Comparison of Regressions for the Sichuan Region and Other Regions

To display the characteristics of the EEW system parameters (, *τ*_{c}, and *P*_{d}) versus earthquake magnitudes in the Sichuan region, China, in this study, the regression equations for those parameters are compared with those of the other region in the previous studies.

In Figure 8a, the model developed for the Sichuan region is compared with the one developed by Allen and Kanamori (2003) based on the seismic dataset from southern California and the one developed by Lancieri *et al.* (2011) based on the Chile dataset. Allen and Kanamori (2003) used 2‐s *P* phase and 10-Hz low-pass filtering for magnitude in the 3–5 range, 4‐s *P* phase and 3‐Hz low‐pass filtering for magnitude>4.5, whereas Lancieri *et al.* (2011) used 4‐s *P* phase and 0.075‐Hz high‐pass filtering that we use in this study. The figure shows that the slope of linear regression developed in this study is quite similar with that in Lancieri *et al.* (2011), differing from that in Allen and Kanamori (2003). That is foreseeable because of the same time‐window length and filter of this study and Lancieri *et al.* (2011). Nevertheless, the intercepts of the equations are different. The Sichuan’s equation is under Chile’s in the 4–6 magnitude range and becomes higher in the 6–8 range.

Similarly, the *τ*_{c} model developed in this study is compared with the one developed by Kanamori (2005) based on a mixture dataset (including southern California, Japan, and Taiwan) and the one developed by Lancieri *et al.* (2011) based on the Chile dataset, as shown in Figure 8b. Kanamori (2005) used 3‐s *P* phase and 0.075‐Hz high‐pass filtering, whereas Lancieri *et al.* (2011) used 4‐s *P* phase and 0.075‐Hz high‐pass filtering that we use in this study. Unlike , the *τ*_{c} equations perform disparity in different regions despite similar time‐window lengths and filters. In this study, even though the evident saturation does not occur in the *τ*_{c} parameter, the linear slope is small. The lower slope implies less sensitivity of *τ*_{c} for the change of magnitude, meaning that it is less predictive for the larger magnitude.

Finally, in Figure 8c, the *P*_{d} model developed in this study is compared with the one developed by Zollo *et al.* (2006) based on the Euro‐Mediterranean dataset and the one developed by Lancieri *et al.* (2011) based on the Chile dataset. Zollo *et al.* (2006) used 2‐s *P* phase and 0.25–25‐Hz band‐pass filtering and 3‐Hz low‐pass filtering, whereas Lancieri *et al.* (2011) used 2‐s *P* phase and 0.25–3 Hz band‐pass filtering that we use in this study. Among the three parameters investigated in this study, *P*_{d} performs most similarly in different regions under the same time‐window length and filter in the 4–6 magnitude range. Nevertheless, the evident saturations occur in *M*_{w} 6.6 and 7.9, corresponding to the mainshocks of the Lushan and Wenchuan earthquakes.

In general, the above results indicate that the characteristics of the three parameters for the Sichuan region are more or less different from those in other regions. The difference can perhaps be attributed to the causes with different geological and topographic conditions in LFZ, which triggered the Wenchuan and Lushan earthquakes. Additionally, it might involve two speculative reasons, one being that there are clay‐rich sediments that exist in the LFZ region, which may have a damping effect on propagation waves (Verberne *et al.*, 2010), and the other being that the mountain front fault and mountain back fault, which are nearly parallel to the Longmen Shan center fault (Fig. 1b), play a barrier role for wave propagation in some areas, and so the amplitude of some seismic waves is reduced (Liu, 2011). In addition, by comparing the response spectra from the 2008 Wenchuan earthquake with those from other large earthquakes, Lu *et al.* (2010) found that the response spectra from the 2008 Wenchuan earthquake at high frequency are lower than those from other large earthquakes. By the way, the seismic‐characteristic difference between various earthquake types (e.g., subduction event and thrust event) should not be neglected. The above reasons perhaps partly explain why these parameters determined by the Wenchuan and Lushan earthquakes in the Sichuan region perform differently from other regions. Verification of this speculative hypothesis for our dataset will be the subject of future work.

## Conclusion and Discussions

In this study, we focused on three parameters proposed for the real‐time magnitude estimation for the EEW system application in China: maximum of predominant period (), characteristic period (*τ*_{c}), and peak ground displacement (*P*_{d}). Their correlation with magnitude has been previously studied on several datasets. In this work, we investigated the performance of those parameters on the sequence of the mainshocks and aftershocks in the 2008 Wenchuan (*M*_{w} 7.9) and 2013 Lushan (*M*_{w} 6.6) events in LFZ, Sichuan, China.

First, after a brief introduction of the parameters, we constructed the dataset with several criteria from CSMNC, China Earthquake Administration, in which the mainshocks and aftershocks of the Wenchuan earthquake (498 records) and Lushan earthquake (439 records) are included. We manually determined their *P*‐wave arrivals and the first *S*‐wave arrivals with high SNR.

Featured in the second part of this study was the investigation of the correlation between the , *τ*_{c}, and *P*_{d} parameters and the catalog magnitude in different short time‐window lengths, especially different filtering bandwidth for *P*_{d}. As a result, the parameter is correlated with magnitudes in the 4–6 and 6‐8 ranges, with a large overall standard error and smaller linear slope, especially in the large magnitude. Unlike , *τ*_{c} parameter scales with the entire magnitude range (4–8 range) without evident saturation. Nevertheless, the linear slope is slightly lower than the values retrieved for the Japan, California, Taiwan (Kanamori, 2005), and Chile datasets (Lancieri *et al.*, 2011). As for the *P*_{d} parameter, it is confirmed as a good estimator of the 4–6 magnitude range. Similar to the observations of Wu *et al.* (2006) and Wu and Li (2006), *P*_{d} saturates at higher magnitudes. Even though increasing time‐window length and narrowing the filtering bandwidth can improve the saturation, the price is the decrease of the linear slope, meaning that it is less predictive for the larger magnitude. The *P*_{d} correlation and the related saturation have been investigated and argued by many seismologists. Nielsen (2007) and Murphy and Nielsen (2009) used dynamic and kinematic arguments. Differently, Crowell *et al.* (2013) considered that the seismogeodetic data provide important source information (lower frequency information) that cannot be obtained with seismic data alone, which can overcome the saturation. Meier *et al.*, (2016) argued that saturation is a fundamental limitation that has to do with the earthquake source process and the incapacity to predict future rupture development.

Finally, the three parameter regression models are compared with those developed based on the dataset of other regions, which were determined in previous studies (Allen and Kanamori, 2003; Kanamori, 2005; Zollo *et al.*, 2006; Lancieri *et al.*, 2011). For regression, the linear slope is comparable to that of Lancieri *et al.* (2011), but the intercepts of the equations are different. Moreover, even though the evident saturation does not occur in the *τ*_{c} parameter, the linear slope is smaller than the other two. Among the three parameters investigated in this study, *P*_{d} performs most similarly in the compared regions in magnitude of 4–6 range; however, the evident saturations occur in *M*_{w} 6.6 and 7.9. Speculatively, the different performances of the three parameters for the Sichuan region can perhaps be attributed to the particular geological and topographic conditions in the LFZ region. Verification of this speculative hypothesis for our dataset will be the subject of a future work.

The results presented in this study open a new perspective on the feasibility of the EEW system in Sichuan, one of the largest seismic risk regions in China. After the 5‐years testing of the EEW system in the Beijing capital region, the development of the EEW system in the Sichuan region is in progress, and it is imperative. In present circumstances, the source distance cannot be quickly and accurately obtained using real‐time location procedures such as, for instance, the method proposed by Horiuchi *et al.* (2005) in the Sichuan region without the seismic dense network. Hence, we recommended the distance‐independent parameters to estimate the earthquake magnitude in the EEW system practical application for the moment. As a matter of course, the saturation needs further investigation in future works. Moreover, to develop the EEW system in a specific region, it is necessary to consider the regional characteristics and propose the suitable regression model (including the optimal time‐window length and filter) for the given region. The ultimate ideal is to prefigure the real‐time alert system operating in the seismic risk region or vital region (such as the Longmen Shan region and the Beijing capital region), based on the continuous time evolution (Colombelli and Zollo, 2016).

## Data and Resources

The strong‐motion records used in this study were kindly provided by the China Strong‐Motion Network Centre of China Earthquake Administration through the website http://www.csmnc.net/ (last accessed June 2016), but they are not publicly available.

## Acknowledgments

We acknowledge the China Strong‐Motion Network Centre of China Earthquake Administration for making the strong‐motion data accessible. A special thanks to Guanghao Zhai and Daoxing Huang for the sort collate of dataset in Sichuan region. The authors appreciate editors and reviewers for their positive and constructive comments and suggestions on this article. This work was financially supported by Science and Technology Department of Sichuan Province (Program Number 2015SZ0068), in the framework of high‐speed railway earthquake monitoring, early warning, and emergency processing technology project.

- Manuscript received 17 December 2016.