Validation of The Probabilistic Seismic Hazard Assessment by The Taiwan Earthquake Model Through Comparison With Strong Ground Motion Observations

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Validation of the Probabilistic Seismic

Hazard Assessment by the Taiwan


Earthquake Model through Comparison
with Strong Ground Motion Observations
Jia-Cian Gao1,2, Yi-Hsuan Tseng3, and Chung-Han Chan*1,4

Abstract
To validate the probabilistic hazard assessment proposed by the Taiwan Earthquake
Model (TEM), we compared it with the strong ground motion observations. We
accessed the Taiwan Strong Motion Instrumentation Program (TSMIP) database and
reported the maximum ground shaking of each strong-motion station. Comparing
the TSMIP observations and the TEM hazard model reveals similar spatial patterns.
However, some records indicate significantly higher shaking levels than the model does
due to the occurrence of some large events, for example, the 1999 M w 7.6 Chi-Chi earth-
quake. Such discrepancies cannot be explained by model parameter uncertainties but
by unexpected events in the given short observation period. We have confirmed that,
although each seismogenic structure in Taiwan is unlikely to rupture within a short
period, the summarized earthquake potentials from all of the structures are significant.
In addition, we discuss the impacts of some model parameters, including epistemic Cite this article as Gao, J.-C.,
Y.-H. Tseng, and C.-H. Chan (2022).
uncertainties of source parameters, truncation of standard deviation for ground-motion
Validation of the Probabilistic Seismic
prediction equations, the Gutenberg–Richter law for area source, and the time-depen- Hazard Assessment by the Taiwan
dent seismicity rate model. The outcomes of this study provide not only crucial infor- Earthquake Model through Comparison
with Strong Ground Motion Observations,
mation for urban planning on a city scale and building code legislation on a national Seismol. Res. Lett. XX, 1–15, doi: 10.1785/
scale but also suggestions for the next generation of probabilistic seismic hazard assess- 0220210186.

ment for Taiwan as well as other regions. Supplemental Material

Introduction considering an acceleration threshold at a set of recording sites.


A probabilistic seismic hazard analysis (PSHA) evaluates the The results show that the probabilistic seismic hazard models in
probability of various levels of ground motion during a certain France overestimated the number of sites with exceedance for
period. It provides a reference for minimizing the seismic risk, low-acceleration levels or short return periods. In Taiwan,
selecting the sites for infrastructure, developing the legislation Wang, Lee, et al. (2016) compared the estimated strong ground
for building safety, and calculating the correlated insurance motions in a return period of 23 yr and maximal shaking level
premiums. Thus, the Taiwan Earthquake Model (TEM) has for each Taiwan Strong Motion Instrumentation Program
published two versions of the Taiwan seismic hazard maps (TSMIP; Shin, 1993) station in an observation period of 23 yr
(Wang, Chan, et al., 2016; Chan et al., 2020) by implementing (between 1993 and 2015). The results showed comparable shak-
updated databases and state-of-the-art earthquake models. ing levels between model and observation, except in central
The credibility of proposed hazard maps must be proven. For Taiwan, where the 1999 M w 7.6 Chi-Chi earthquake and sub-
example, Albarello and D’Amico (2008) indicated a significant sequent aftershocks took place (fig. 5 of Wang, Lee, et al., 2016).
mismatch between the PSHA and observations in Italy. One
possible problem is that the seismic hazard computational
model has been parametrized by considering data included in 1. Earthquake-Disaster & Risk Evaluation and Management (E-DREaM) Center,
National Central University, Taoyuan, Taiwan, https://orcid.org/0000-0003-1875-
the control time period. Another reason is the model is time 3652 (C-HC); 2. Graduate Institute of Applied Geology, National Central University,
independent. That assumes time stationarity and ergodicity, Taipei, Taiwan; 3. Department of Civil Engineering, National Taiwan University,
Taipei, Taiwan; 4. Department of Earth Sciences, National Central University, Taipei,
leading to any time interval being equivalent to the other ones. Taiwan
Tasan et al. (2014) compared the distribution of the expected *Corresponding author: hantijun@googlemail.com
number of sites with exceedance with the observed number © Seismological Society of America

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 1

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
The occurrence of the 2018 M w 6.3 Hualien earthquake raised short periods and therefore might not result in significant
the importance of a time-dependent rupture model (Chan et al., seismic loss; we thus eliminate these cases in our subsequent
2019). This event took place on the Milun fault, which ruptured analysis. To discover similar cases, we evaluated the ratio
in 1951. According to the time-independent Poisson distribu- between the two horizontal components through the following
tion model, its rupture probability for 50 yr before the event was equations:
53%, whereas a significantly higher probability of 80% would
   
have been suggested based on the time-dependent Brownian  PGAEW   PGAEW 
Ratio  ; when 
 PGA  ≤ 1; else; 1
passage time (BPT) model (Ellsworth et al., 1999). Although PGANS 
EQ-TARGET;temp:intralink-;df2;320;678

NS
the BPT model has been applied to many seismic hazard assess-
ments, for example, National Seismic Hazard Maps for Japan
   
(Fujiwara et al., 2009), its forecasting ability is seldom compared  PGANS   PGANS 

Ratio   
; when  ≤ 1; 2
with that of other time-dependent models. PGAEW  PGAEW 
EQ-TARGET;temp:intralink-;df2;320;626

In this study, we aim to evaluate the credibility of the TEM


seismic hazard assessment proposed in 2020 (Chan et al., 2020) in which PGAEW represents PGA in the east–west orientation,
according to strong ground motion observations and paleoseis- and PGANS represents PGA in the north–south orientation.
mological observations. We access the strong ground motion If the ground shakings in the two components are close to each
database from the TSMIP seismic network and report the other in the maximum shaking level, the ratio will be close to 1.0,
maximum ground shaking of each station. The observations and vice versa. We found that most of the records reported high
are compared with the TEM PSHA to test credibility and indi- ratios, with 85% of them being higher than 0.6. Because the
cate limitations of the model. For illustrating the temporal evo- remaining 15% have a significantly lower ratio, we defined a
lution of earthquake probability, we implement several time- ratio of 0.6 as a threshold for examining the reliability of the
dependent models, including the BPT model, and examine waveforms. If a waveform obtained a spike in only one compo-
them by comparing them to paleoseismic data. nent (e.g., the record in the north–south component by
TTN034, Fig. S1a), and it is dissimilar in maximum shaking
Ground-Motion Observations level to the observation made by its neighboring station (e.g.,
Starting in 1993, the Central Weather Bureau established more the record by TTN043, Fig. S1b), we implemented the waveform
than 700 seismic stations for the TSMIP network (Shin, 1993). in the component without a spike, for example, 532.613 gals
To obtain its records, we accessed the Geophysical Database recorded by the east–west component of TTN034 (Fig. S1a).
Management System. To unify a fixed observation period, we Based on the analysis mentioned earlier, the maximum PGA
implemented only stations operating during 1995 and 2019, in the observation period for each station is obtained (Fig. 1a). It
consistent with an observation period of 25 yr. Based on this shows significantly higher ground-shaking levels in the regions
criterion, a total of 536 stations were selected for analysis. We close to some significant events, for example, the 1994 M 6.2
have confirmed that the removed stations are not concentrated Nan-Ao earthquake, the 1999 Chi-Chi sequence, the 2003 M 6.4
in any specific area. Chengkung earthquake, the 2006 M 7.0 Pingtung doublet, and
To obtain the maximum strong ground shaking of each sta- the 2018 M 6.3 Hualien earthquake. For the stations in northern
tion during the observation period, we discarded the records Taiwan, by contrast, where there was no large earthquake during
out of function in some components, calculated the geometric the observation period, the maximum ground shaking recorded
mean peak ground acceleration (PGA) of the two horizontal during the observed period was relatively low.
components (east–west and north–south), and reported the
maximum PGA among all of the earthquake records in the The Probabilistic Seismic Hazard Model
observation period. To validate the TEM PSHA proposed in 2020 (TEM
The analysis procedure mentioned earlier, however, might PSHA2020) (Chan et al., 2020), we followed their assessment
sometimes lead to high peaks in hazard levels. For the case of procedure, obtained parameter uncertainties, implemented
the TTN034 station during the 2003 M w 6.8 Chengkung earth- ground-motion prediction equations (GMPEs), and evaluated
quake (location shown in Fig. 1a), for example, the record hazard level to compare with observed ground shakings.
shows high-PGA peaks in the north–south and vertical com-
ponents (Fig. S1a, available in the supplemental material to this Seismogenic sources and corresponding
article), whereas the record in the east–west component has a parameters with uncertainties
similar shaking level to those found in the records of other sta- In the TEM PSHA2020, the seismic sources were divided into
tions, for example, TTN043 (Fig. S1b). The recorded spikes three independent categories: (1) specific seismogenic structure
could be attributed to various factors, for example, source, sources, meaning the earthquakes caused by rupture on a par-
path, or site effects (Albarello et al., 2015) or station construc- ticular fault or structure; (2) shallow-background sources, mean-
tion (Wen et al., 2001). Such spikes, however, last for only ing the earthquakes that cannot be associated with a specific

2 Seismological Research Letters www.srl-online.org • Volume XX • Number XX • – 2022

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
(a) (b)

Maximum observed PGA PSHA for mean PGA


in 25 yr period for 50% in 25 yr
25°

Max PGA recorded 50 km


by TSMIP stations
from 1995 to 2019
1994
Chelungpu M6.2
fault

2018
M6.3
24°

1999
M7.6

2003
M6.4

Ground shaking
23°
TTN043 in PGA (in g)
TTN034 0.45

0.40

0.35
VS30 (in m/sec)
0.30

180 760 0.25

0.20

0.15 22°
0.10
2006
M7.0 & 7.0 0.05

0.00

120° 121° 122°

structure, illustrated by area source and smoothing models; and Figure 1. (a) Distribution of Taiwan Strong Motion
(3) subduction sources, meaning the earthquakes that took place Instrumentation Program (TSMIP) stations with maximum
in a subduction zone. Because the seismic sources of the three ground-shaking levels recorded during 1995 and 2019 (colored
triangle), the V S30 map (base map), and epicenters of some
categories are independent of one another, all categories are
significant earthquakes (stars). The distribution of V S30 was
implemented for PSHA without a logic tree. Previous studies interpolated from the Taiwan region from the Engineering
(e.g., Wang, Lee, et al., 2016) concluded that the seismic hazards Geological Database for TSMIP, such that the V S30 at most sites
for most of the Taiwan regions stem mainly from crustal sources were obtained from measurements (Kuo et al., 2012). The
(i.e., either specific seismogenic structure sources or shallow- alignment of the Chelungpu fault that resulted in the 1999 Chi-
Chi earthquake is illustrated via the black line. The locations of
background sources). In addition, there is no significant subduc-
stations TTN034 and TTN043 are denoted. (b) Seismic hazard
tion event during the observation period. We thus incorporated maps in mean value in a return period of 25 yr. The alignments of
only the subduction sources into the hazard assessment without the seismogenic structures are presented in black lines, and the
a discussion on their impact. Because the TEM PSHA2020 did Chelungpu fault is shown in dark red. Geometries of the area
not implement parameter uncertainties, which could provide sources are presented via green polygons. The color version of
a quantitative constraint to compare with the observations, this this figure is available only in the electronic edition.
study followed the procedure of the TEM PSHA2020 but
included source parameter uncertainties. Chan et al. (2012) con-
cluded a trivial difference in seismicity rates for smoothing mod- Seismogenic structure sources. To quantify the uncer-
els using different bandwidth functions within one standard tainty of the recurrence interval for each seismogenic structure,
deviation, and the impact of the parameter uncertainties for we obtained the mean, minimum, and maximum values of slip
the smoothing model might be insignificant. Thus, we obtained rate (shown in Table 1) based on the geomorphological evi-
parameters and corresponding uncertainties for the seismogenic dence, updated after Shyu et al. (2020). The constraint of
structures and area sources, which are detailed subsequently. slip-rate deviations could be associated with an understanding

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 3

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
of each structure. For some well-investigated faults or struc- model) (Youngs and Coppersmith, 1985). In addition, we
tures, for example, the Chelungpu fault (ID 17), Meishan fault proposed another parameter set for the area sources, consid-
(ID 20), and Milun fault (ID 32), the ranges of slip rates are ering variable a and b values for each area (shown in Table 3)
minimal, whereas for some faults or structures with insufficient to compare with the TEM PSHA2020.
geomorphological evidence, for example, the Hukou fault (ID
4), the Tunglo structure (ID 11), and the Sanyi fault (ID 14), GMPEs
the uncertainties are relatively large. For our seismic hazard assessment, ground-motion behaviors
To evaluate recurrence rates for the seismogenic struc- were simulated by GMPEs representing the ground-shaking
tures, we followed the procedure of the TEM PSHA2020. level as a function of magnitude, distance, and other factors.
We first considered the displacement in an earthquake with To confirm the credibility of the TEM PSHA2020, we followed
a characteristic magnitude and acquired its corresponding its procedure and implemented GMPEs by Lin et al. (2011)
recurrence rate with possible deviation, which is to say, dis- and Lin and Lee (2008) for the crustal and subduction sources,
placement divided by slip rate, shown in Table 1. The hazard respectively. Both GMPE sets were obtained based on the
assessment further considered cases of potential multiple- observations of the Taiwan events. For the seismic hazard
structure rupture and partial rupture based on historical assessment, we first followed the procedure of TEM
records or geological relationships among neighboring struc- PSHA2020 to truncate at two standard deviations (2σ) for
tures. To partition the seismic rate into the cases of single- the GMPEs and then discussed the feasibility of various trun-
and multiple-structure ruptures, we followed the procedure cations.
of the TEM PSHA2020 by implementing the Gutenberg–
Richter law (Gutenberg and Richter, 1944) and rupture PSHA result for 50% probability of exceedance in
parameters (including rupture area, slip, and magnitude) 25 yr
to evaluate recurrence rates for each case (Table 2). To quan- A proper PSHA using source parameters with their uncertain-
tify time-dependent rupture probabilities for the seismogenic ties requires the application of a logic tree. For the seismic
structures with rupture histories, the TEM PSHA2020 sources, we considered the mean, minimum, and maximum
adopted the BPT model (Ellsworth et al., 1999), which incor- slip rate and assumed a weighting of 60%, 20%, and 20%,
porated the occurrence time of the previous rupture. By respectively. For shallow-background sources, we first followed
considering multiple-structure rupture and time dependency, the setting of the TEM PSHA2020 model to consider the area
the revised recurrence intervals for corresponding structures source and smoothing models with equal weightings (50% for
were calculated as shown in parentheses in Table 1. This each). In addition, for the area source model, we assumed a
model expects a higher rupture probability for a seismogenic weighting of 60%, 20%, and 20% for the medium, upper, and
structure that has a longer time elapsed and a shorter recur- lower bounds, respectively, of a-value uncertainties. Based on
rence interval, for example, the Milun fault (ID 32), whereas the source parameters and the settings of the logic tree, we
the earthquake probability becomes lower for the structure could assess the probabilistic seismic hazard in Taiwan.
that has just recently ruptured, for example, the Chelungpu We presented hazard levels for 50% exceedance in 25 yr to
fault (ID 17). compare with the same observation period for the TSMIP rec-
ord, that is, 50% of the observations are expected to be higher
Area source models. To obtain parameters for each area than the levels of our model. By considering all of the branches
source, we followed the procedure of the TEM PSHA2020. in the logic tree, we first presented a hazard map at the mean
We acquired the earthquake catalog summarized by Wu et al. level (see Fig. 1b). Because a short return period is assumed, the
(2008) and implemented the complete part of the catalog (i.e., seismic hazard could be contributed from the regions close to
M ≥4.0 from 1973 to 1993 and M ≥3.0 from 1993 to 2016). seismogenic structures with short recurrence intervals, area
Then we removed foreshocks and aftershocks via the Gardner sources with large a-values, and large site amplification (i.e.,
and Knopoff (1974) declustering approach. Based on the low V S30 shown in Fig. 1a), for example, Hualien City with
declustered catalog, we described the seismicity for each area the Milun fault (ID 32) in area S17A; the Chianan Plain with
source based on the Gutenberg–Richter law (Gutenberg and the Muchiliao-Liuchia fault (ID 22) and Chungchou structure
Richter, 1944) by regression of the catalog. We first followed (ID 23) in the western part of S07; and the Ilan plain with the
the procedure of the TEM PSHA2020 to obtain a unified b- Northern Ilan structure (ID 37) and Southern Ilan structure
value of 1.10 for the entire study region. To discuss the impact (ID 38) in area S14A. By contrast, the hazard levels are rela-
of uncertainty, we then determined the a-value with uncer- tively lower in areas S04, S05A, and S08A, where the seismo-
tainty for each area (shown in Table 3). The uncertainty could genic structures (recurrence intervals significantly longer than
be associated with either deviation of magnitude determination 25 yr) and shallow-background sources (low a-values for the
or special seismic activity behaviors that do not follow area sources and low seismicity rate for the smoothing model)
the Gutenberg–Richter law (e.g., the characteristic earthquake are relatively inactive.

4 Seismological Research Letters www.srl-online.org • Volume XX • Number XX • – 2022

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
TABLE 1
Source Parameters for the 44 Seismogenic Structures in the TEM Database

Slip Rate (mm/yr) Recurrence Rate (Per Year)

ID Mw Displacement (m) Min Mean Max Min Mean Max

1 7.01 1.29 1.10 1.66 2.94 0.0008 0.0013 0.0023

2 6.24 0.72 0.06 0.13 0.52 0.0001 0.0002 0.0007

3 6.03 0.60 0.09 0.18 0.72 0.0002 0.0003 0.0012

4 6.77 1.16 0.19 0.46 2.34 0.0002 0.0004 0.0020

5 6.66 0.95 1.80 3.18 10.15 0.0019 0.0033 0.0106

6 6.41 0.83 0.28 0.66 2.91 0.0003 (0.0002) 0.0008 (0.0004) 0.0035 (0.0016)

7 6.91 1.31 0.47 1.12 5.35 0.0004 0.0009 0.0041

8 6.48 0.90 0.62 1.44 6.81 0.0007 (0.0003) 0.0016 (0.0003) 0.0076 (0.0035)

9 6.52 0.80 0.12 0.13 0.14 0.0002 (0.0001) 0.0002 (0.0001) 0.0002 (0.0001)

10 6.84 1.22 0.78 1.84 8.77 0.0006 (0.0003) 0.0015 (0.0007) 0.0072 (0.0034)

11 6.17 0.68 0.19 0.50 2.63 0.0003 0.0007 0.0039

12 6.19 0.69 0.36 0.84 3.89 0.0005 0.0012 0.0056

13 6.61 0.99 0.61 1.38 5.32 0.0006 (0.0000) 0.0014 (0.0000) 0.0054 (0.0001)

14 7.04 1.45 0.29 0.85 4.61 0.0002 0.0006 0.0032

15 6.64 0.94 0.27 0.50 1.70 0.0003 (0.0000) 0.0005 (0.0000) 0.0018 (0.0000)

16 7.57 2.35 0.95 1.87 6.97 0.0004 (0.0006) 0.0008 (0.0011) 0.0030 (0.0041)

17 7.60 2.45 6.94 6.94 6.94 0.0029 (0.0000) 0.0029 (0.0000) 0.0029 (0.0000)

18 6.96 1.38 0.47 1.06 4.88 0.0003 0.0008 0.0035

19 6.95 1.37 1.87 4.66 23.39 0.0014 0.0034 0.0173

20 6.60 0.89 2.50 2.51 2.54 0.0028 (0.0000) 0.0029 (0.0000) 0.0029 (0.0000)

21 7.21 1.71 1.40 3.36 16.12 0.0008 (0.0001) 0.0020 (0.0003) 0.0094 (0.0016)

22 6.85 1.23 4.40 5.75 7.10 0.0036 (0.0065) 0.0048 (0.0085) 0.0059 (0.0105)

23 6.89 1.28 9.02 12.20 18.71 0.0074 0.0100 0.0153

24 6.38 0.69 0.80 2.65 4.50 0.0012 (0.0006) 0.0038 (0.0020) 0.0065 (0.0034)

25 6.07 0.61 6.10 7.07 8.72 0.0096 (0.0000) 0.0111 (0.0000) 0.0137 (0.0000)

26 6.68 0.97 0.72 1.10 1.50 0.0007 0.0011 0.0015

27 6.30 0.75 0.81 1.78 8.04 0.0011 0.0024 0.0108

28 6.66 0.95 0.17 0.32 1.05 0.0002 0.0003 0.0011

29 7.10 1.62 0.57 0.98 3.01 0.0004 (0.0002) 0.0006 (0.0003) 0.0019 (0.0009)

30 6.85 1.20 5.74 6.15 6.62 0.0047 (0.0020) 0.0050 (0.0021) 0.0054 (0.0023)

31 6.31 0.77 1.87 3.22 6.96 0.0024 0.0042 0.0090

32 6.56 0.85 9.92 10.15 10.47 0.0122 (0.0410) 0.0125 (0.0420) 0.0129 (0.0433)

33 7.52 5.60 11.35 17.10 0.0025 (0.0009) 0.0050 (0.0018) 0.0075 (0.0028)

34 7.38 2.00 4.76 7.28 11.16 0.0024 0.0037 0.0057

The recurrence intervals in parentheses were revised according to slip partitioning for multiple-structure ruptures and the time-dependent Brownian passage time (BPT) model.
Note that the mean and deviations for slip rate and recurrence interval of each structure were listed. The alignments of the seismogenic sources are presented in Figure 5. TEM,
Taiwan earthquake model.
(Continued next page.)

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 5

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
TABLE 1 (continued)
Source Parameters for the 44 Seismogenic Structures in the TEM Database

Slip Rate (mm/yr) Recurrence Rate (Per Year)

ID Mw Displacement (m) Min Mean Max Min Mean Max

35 6.24 0.71 3.55 5.28 8.02 0.0052 (0.0001) 0.0077 (0.0001) 0.0117 (0.0001)

36 6.73 1.10 5.74 7.32 9.03 0.0052 0.0067 0.0082

37 6.90 1.14 0.96 3.29 6.27 0.0008 0.0029 0.0054

38 6.43 0.64 4.47 5.48 6.92 0.0068 0.0083 0.0105

39 6.00 0.57 2.03 5.01 9.00 0.0037 0.0091 0.0163

40 6.07 0.48 0.56 0.94 2.56 0.0012 0.0020 0.0053

41 7.24 1.74 0.45 0.92 3.50 0.0003 (0.0001) 0.0005 (0.0001) 0.0020 (0.0004)

42 6.58 0.96 0.85 1.73 6.53 0.0009 0.0018 0.0069

43 6.41 0.83 0.92 1.64 5.46 0.0011 0.0020 0.0065

44 6.81 1.19 0.40 0.92 4.24 0.0003 0.0008 0.0036

The recurrence intervals in parentheses were revised according to slip partitioning for multiple-structure ruptures and the time-dependent Brownian passage time (BPT) model.
Note that the mean and deviations for slip rate and recurrence interval of each structure were listed. The alignments of the seismogenic sources are presented in Figure 5. TEM,
Taiwan earthquake model.

TABLE 2
Parameters for the Potential Multiple-Structure Ruptures

Recurrence Rate (Per Year)

ID Seismogenic Structure Name Type Mw Min Mean Max

6, 8 Hsinchu fault, Hsinchu frontal structure R, R 6.65 0.00029 0.00067 0.00313

9, 10 Touhuanping structure, Miaoli frontal structure SS, R 6.92 0.00027 0.00061 0.00284

13, 15 Shihtan fault, Tuntzuchiao fault R, SS 6.83 0.00001 (0.00000) 0.00001 (0.00000) 0.00005 (0.00000)

20, 21 Meishan fault, Chiayi frontal structure SS, R 7.24 0.00023 (0.00045) 0.00034 (0.00066) 0.00106 (0.00203)

21, 41 Chiayi frontal structure, Tainan frontal structure R, R 7.43 0.00032 0.00075 0.00342

25, 41 Houchiali fault, Tainan frontal structure R, R 7.25 0.00022 0.00030 0.00065

29, 30 Chaochou fault, Hengchun fault SS/R, SS/R 7.20 0.00095 0.00090 0.00164

35, 33c Luyeh fault, Longitudinal Valley fault (south) R, R/SS 6.97 0.00065 (0.00056) 0.00112 (0.00096) 0.00169 (0.00145)

33a Longitudinal Valley fault (north) R/SS 6.99 0.00024 0.00048 0.00072

33b Longitudinal Valley fault (central) R/SS 7.26 0.00047 0.00095 0.00143

33c Longitudinal Valley fault (south) R/SS 6.95 0.00021 0.00043 0.00065

33ab Longitudinal Valley fault (north, central) R/SS 7.42 0.00070 0.00143 0.00215

33bc Longitudinal Valley fault (central, south) R/SS 7.40 0.00067 0.00136 0.00205

The recurrence intervals in parentheses were revised according to the time-dependent BPT model. Note that the mean and deviations for the recurrence interval of each case are
listed. The alignments of the seismogenic sources are presented in Figure 5.

Comparison between the Observations 25 yr. We compared the general pattern in the entire
and the Models Taiwan region and discussed the credibility of our model,
To understand the relationship between the observations and followed by a detailed discussion on the case of the 1999
the models under a uniform condition, we analyzed the Chi-Chi earthquake associated with the Chelungpu fault
TSMIP records during 1995 and 2019 and assessed probabi- and the 2018 Hualien earthquake associated with the
listic seismic hazard for 50% probability of exceedance in Milun fault.

6 Seismological Research Letters www.srl-online.org • Volume XX • Number XX • – 2022

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
TABLE 3
Parameters for the 28 Area Sources

Area a-Value a-Value Uncertainty b-Value a-Value a-Value Uncertainty b-Value

S01 4.090 0.092 1.100 3.670 0.009 1.000

S02 4.480 0.122 4.010 0.014 0.990

S03 4.010 0.151 3.240 0.017 0.920

S04 3.450 0.236 2.250 0.052 0.810

S05A 3.960 0.094 4.560 0.008 1.250

S05B 4.750 0.082 5.000 0.008 1.160

S06 5.470 0.281 4.550 0.022 0.920

S07 5.260 0.157 4.740 0.015 0.990

S08A 4.710 0.259 3.810 0.038 0.910

S08B 4.050 0.208 3.040 0.022 0.870

S09 4.940 0.172 4.290 0.014 0.960

S10 5.230 0.291 4.160 0.023 0.880

S11 5.120 0.142 5.260 0.022 1.130

S12 5.230 0.086 5.180 0.009 1.090

S13 4.950 0.144 4.910 0.031 1.090

S14A 4.830 0.244 3.740 0.025 0.860

S14B 5.420 0.109 5.570 0.012 1.130

S14C 5.240 0.132 5.100 0.022 1.070

S15 5.730 0.142 5.280 0.011 1.010

S16 5.940 0.235 5.070 0.021 0.920

S17A 5.630 0.112 5.580 0.014 1.090

S17B 4.720 0.175 4.720 0.047 1.100

S18A 5.420 0.094 5.420 0.01 1.100

S18B 4.890 0.115 4.400 0.009 0.990

S19A 5.500 0.244 4.550 0.006 0.910

S19B 4.880 0.137 4.470 0.018 1.010

S20 4.910 0.137 4.910 0.025 1.100

S21 5.380 0.299 4.280 0.018 0.880

The geometries of the area sources are presented in Figure 1b.

General patterns as the 1994 M 6.2 Nan-Ao event, the 2003 M 6.4 Chengkung
Comparing the observed ground shakings (see Fig. 1a) and event, and the 2006 M 7.0 Pingtung doublet. Although high
probabilistic seismic hazard model (see Fig. 1b) shows similar ground shakings were recorded in Hualien City (during the
spatial patterns in both, that is, high hazard levels in the 2018 M 6.3 Hualien earthquake), the modeled shaking levels
Chianan Plain and in Hualien City and low hazard levels in are even higher due to the hazard contribution from both
northern (areas S04 and S05A) and southwestern (area S08A) the Milun fault (ID 32) and the area source (S17A). Although
Taiwan. However, there are some discrepancies between the the Milun fault had just ruptured in 2018, considering its short
observations and the models. For example, some stations in cen- recurrence interval, its probability of rupture in the near future
tral Taiwan recorded large ground-shaking levels during the remains high (see Table 1). In addition, the ratio of the stations
Chi-Chi mainshock and subsequent aftershocks; some signifi- with observed ground-shaking levels higher than the modeled
cantly higher shakings were observed during large events, such ones is 41.9% and therefore lower than the median (because

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 7

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
we presented hazard levels for 50% probability of exceedance, levels for the cases of short recurrence intervals (in the cases
the expected percentage of stations with higher observations for the probabilities of exceedance higher than 20%, corre-
is 50%). Several studies also obtained overpredicted shaking lev- sponding to a recurrence interval of 112 yr). The result infers
els relative to the observations, for example, the cases of France that this model might not be acceptable for short recurrence
(Tasan et al., 2014), Italy (Stein et al., 2017), Japan (Brooks et al., intervals, although the deviation could be covered by epistemic
2016), and California (Salditch et al., 2020). We explain the uncertainty of seismic sources (Fig. 2).
overpredictions in our case in the following.
Sigma truncation for GMPEs
Epistemic uncertainty in PSHA The analysis mentioned earlier were based on truncation at 2σ
Our model, presented earlier, is based on the hazard at the mean for GMPEs, based on the setting of the TEM PSHA2020.
hazard of a logic-tree approach (Fig. 1b). Here, we will confirm Because the outcomes overpredict the observations for short
whether a PSHA has a wider coverage of the possible ground- recurrence intervals (the lower ratios for high probabilities
shaking levels by implementing uncertainty from parameter of exceedance for 2σ in Table 4), we further assessed the shak-
deviations. The uncertainties from the slip rate (see Table 1) ing levels considering truncation at 1.5 and 1:0σ, respectively,
and a-value (see Table 3) for each seismogenic structure source and reported their performance (Table 4). The outcomes
and area source were implemented, respectively. Through imple- suggest that the overpredictions for high probabilities of
menting a logic tree, different percentiles of hazard levels could exceedance were minimized and the performance for low
be assessed, that is, we quantified the deviation of modeled shak- probabilities of exceedance were good. The performances of
ing levels attributed from epistemic uncertainties of the seismic 1:5σ are always better than those of 2:0σ, regardless of recur-
sources. Here, we present ground-shaking levels for the 50th, rence interval.
15th, and 85th percentiles, shown in Figure 2c, respectively. Traditional GMPEs are usually ergodic (Anderson and Brune,
The higher the percentiles are assumed to be, the larger the 1999), that is, ground-motion variability from different earth-
ground-shaking levels are expected to be, and vice versa. The quake sources recorded by various sites are included in the analy-
pattern of the shaking map for the 50th percentile (see sis. Because an application of PSHA focuses on one source for
Fig. 2a) is similar to that of the mean level (see Fig. 1b), which one site, only the aleatory variability or the single-path variability
infers no extreme seismic hazard levels, considering the param- should be in the integration. However, application of global or
eter deviations. The percentage of stations with observations regional GMPEs obtained by ground-motion data recorded at
higher than those of the 50th percentile shaking levels is different sites in different areas results in a large sigma and over-
42.6%. In addition, 50.5% and 31.8% of observations obtain estimation in PSHAs. One possible solution is implementing
higher levels than the 15th and 85th percentile models, respec- nonergodic GMPEs based on different approaches for specific
tively. This range covers our model setting (50% exceedance) and regions (Dawood and Rodriguez-Marek, 2013; Landwehr et al.,
suggests that our assessment is accepted by the observations 2016; Lanzano et al., 2016; Ameri et al., 2017; Kuehn et al., 2019;
when epistemic uncertainty is incorporated. Kuehn and Scherbaum, 2016; Kuehn and Abrahamson, 2020;
Gao et al., 2021). A site-dependent GMPE set obtains a signifi-
PSHA for various probabilities of exceedance cantly smaller scattering of residuals compared with the existing
We have validated our model in the form of 50% exceedance in regional GMPEs.
25 yr (i.e., about 36 yr recurrence interval). To evaluate the For engineering application, safety evaluation of infrastruc-
model for various recurrence intervals, we assessed seismic tures is expected through a conservative PSHA, that is, trun-
hazard at each station for several probabilities of exceedance cation at a large standard deviation (e.g., 3 or even 5σ) for
in 25 yr (Table 4). For each probability of exceedance, we GMPEs for a long recurrence interval (e.g., 2475 yr). The out-
reported the percentage of stations with higher ground-shak- comes of PSHA are influenced by the number of sigma trun-
ing observation than that of the model. That is, if the model is cations and sigma in GMPE. A suitable truncation is important
perfect, the percentage would be identical to the probability of when we incorporate GMPEs with large ground-motion vari-
exceedance; a lower percentage suggests that the model over- ability. Our analyses suggest overpredictions if GMPEs were
predicts the observations, whereas a higher percentage suggests truncated at a large standard deviation, providing insights into
that the model generally underpredicts the observations. the application of PSHAs.
When we truncated at two standard deviations (2σ, corre-
sponding to a confidence interval of 95%) for the GMPEs for Impacts of the Gutenberg–Richter law for area
our PSHA, the performance for long recurrence intervals was source
good, that is, the reported percentages for the models for 1% The TEM PSHA2020 obtained a unified b-value for the entire
(recurrence interval of 2487 yr) and 2% (recurrence interval of study region and then determined the a-value for each area by
1237 yr) of exceedance are 0.6% and 1.7%, respectively. regression of the catalog. Such a procedure could obtain a stable
However, the models significantly overpredicted the shaking b-value, especially for a region with a small number of events for

8 Seismological Research Letters www.srl-online.org • Volume XX • Number XX • – 2022

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
50th percentile 15th percentile 85th percentile

(a) (b) (c)


observation (station)

Seismic
Model (map) and

hazard
(in g)
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00

(d) (e) (f)


model and observation
Residual between

Residual (in g)
(model-obs)
+0.10

+0.05

0.00

−0.05

−0.10

regression (Fujiwara, 2014; Wang, Chan, et al., 2016). However, Figure 2. Comparisons between the seismic hazard maps for the
these area source parameters with a unified b-value might not (a) 50th, (b) 15th, and (c) 85th percentiles and the TSMIP
be able to represent the details of the seismic behavior of observations (colored triangles in a–c) and the differences among
them (colored triangles in d–f). The alignment of the seismogenic
each individual area. Thus, this study proposed another area
structures is presented via black lines. The color version of this
source model, considering variable a- and b-values for each area figure is available only in the electronic edition.
(shown in Table 3). In this model, lower b-values (lower than
1.1, obtained for the entire study region) were obtained for many
of the areas, suggesting higher seismic rates for large magni-
tudes. Based on this model, a higher hazard level was expected

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 9

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
TABLE 4
Percentages of Stations with Higher Ground-Shaking Observation than the Modeled Shaking Level for Several
Probabilities of Exceedance in 25 Yr for Various Truncation Thresholds for GMPEs

Probability of
Exceedance 1.0% 2.0% 5.0% 10.0% 20.0% 30.0% 40.0% 50.0% 60.0% 70.0% 80.0% 90.0%

2.0 sigma 0.6% 1.7% 3.4% 8.2% 15.9% 23.7% 33.6% 41.9% 48.8% 55.0% 62.4% 71.8%

1.5 sigma 1.3% 2.1% 5.6% 10.5% 18.3% 28.8% 37.8% 45.4% 52.3% 59.6% 68.2% 76.8%

1.0 sigma 2.1% 3.7% 7.1% 13.8% 23.0% 33.8% 42.8% 50.8% 57.4% 66.5% 72.9% 81.5%

TABLE 5
Percentages of Stations with Higher Ground-Shaking Observations than the Modeled Shaking Levels for Several
Probabilities of Exceedance in 25 Yr for the Models Considering the Variable b-Value for Each Area and a Unified
b-Value for the Entire Study Region

Exceed Percentage for Average


Model 10% 20% 30% 40% 50% 60% 70% 80% 90% Deviation

Fixed b-value (obs > model) 8.20% 15.90% 23.70% 33.60% 42.00% 48.90% 55.00% 62.50% 71.80% 9.8%

Variable b-value (obs > 4.85% 8.58% 15.11% 22.76% 33.02% 45.90% 59.33% 68.66% 81.16% 12.3%
model)

The average deviations from each probability of exceedance for the two models are reported. Note that a smaller average deviation suggests a better fit with observations.

for the regions with lower b-values, including area S14A (Fig. 3). whereas our model does not properly cover some observa-
In this region, the TEM PSHA2020 predicted higher hazard tions with either extremely low or extremely high ground-
levels than the observations at most stations (Fig. 2), whereas shaking levels. The overpredictions (red triangles in Fig. 2f)
the model with variable b-values further deviated from the might be associated with no significant events in the vicinity
observations (Fig. 3). of the seismogenic structures with a short recurrence rate,
To further validate the two models quantitatively, we com- for example, the Ilan Plain. This includes the Northern
pared the observations with different probabilities of exceedance Ilan structure (ID 37) and Southern Ilan structure (ID 38).
for the two models (see Table 5). Both models overpredict Considering their lower bound of recurrence rates (0.0008
the observations, regardless of probabilities of exceedance (as and 0.0068 per year, respectively) and a Poisson distribution
discussed in the Sigma truncation for GMPEs section). In model, the probability that at least one structure ruptures in a
addition, the model with variable b-values has obtained an 25 yr period is 17.3%. Because no large earthquake took place
even lower ratio for observations than predictions, especially during the observation period, our model presented signifi-
for long return periods (i.e., low probabilities of exceedance). cantly higher levels. The underpredictions (blue triangles
This result could be associated with the lower b-values for in Fig. 2f) could be attributed to some events, for example,
several areas, resulting in higher modeled ground shakings the 1994 Nan-Ao, 1999 Chi-Chi, 2003 Chengkung, and
for long return periods (because higher rates are expected for 2006 Pingtung events (stars in Fig. 1a). We discuss each of
large events) but lower modeled ground shakings for shorter the cases in the following section.
return periods (because lower rates are expected for small
events). The outcome of this validation not only confirmed The 1999 Chi-Chi earthquake and the Chelungpu
the credibility of the TEM PSHA2020 procedure for the area fault. The Chi-Chi earthquake is attributed to the rupture
source but also inferred that proposing a single b-value for a along the Chelungpu fault (ID 17). Considering the inverse
larger region could more accurately illustrate the seismic activity of its recurrence rate (shown in Table 1), its expected recurrence
behavior. period is 345 yr. According to a Poisson distribution model, its
rupture probability in 25 yr is ∼7.0% (1-EXP(–25/345)).
Impacts of significant earthquakes We further implemented the BPT model by considering
We further quantified the differences in shaking levels between that the time elapse of the Chelungpu fault before 1999 was
the observations and the model (Fig. 4). Most stations have an 415 yr (Chan et al., 2017). The density function (DF) of the
insignificant difference between the observation and the model, BPT model is expressed as follows:

10 Seismological Research Letters www.srl-online.org • Volume XX • Number XX • – 2022

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
1.0

Shaking level difference Observation > model


(model-observation) 42.0%
25° 0.8
el
od
m

Observed ground shaking (in g)


=
S14A
n
tio
rva
b se
0.6 O

0.4

24°
0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0
Modeled ground shaking for 50% in 25 yr (in g)

Figure 4. Corresponding maximum observed ground shaking in


23° 25 yr and modeled ground shaking for 50% in 25 yr for each
station. Of the stations, 42% have higher observed ground
shaking than the predicted levels.

the 25 yr rupture probability of the Chelungpu fault before the


Chi-Chi event is 13.2%.
Although the BPT model was widely implemented in vari-
22°
ous PSHA studies, its credibility has seldom been compared
with that of other time-dependent models. Thus, we input sev-
120° 121° 122° eral statistical distributions into our model and compared it
with the observed earthquake recurrence intervals. We intro-
Difference in hazard level for 50% in 25 yr (in g)
duced three additional commonly used statistical distributions,
(Fixed b) - (variable b) −0.010 0.000 0.010 that is, the log-normal, gamma, and Poisson distribution mod-
els, described subsequently.
Model - observation −0.10 0.00 0.10
Log-normal model: The 1988 Working Group on California
Earthquake Probabilities implemented this model as a density
Figure 3. (Map view) The difference between the models con- function (DF) that is expressed as follows:
sidering the variable b-value for each area and a unified b-value
for the entire study region and (colored triangles) the differences  
p log t−2
between the model with the variable b-value and the TSMIP −1
DF   2π σt exp − ;
2σ 2
EQ-TARGET;temp:intralink-;;308;236

observations. Note that the map view and the triangles are not
on the same scale. The color version of this figure is available only
in the electronic edition. in which μ is the mean recurrence interval, t is the time since
last earthquake, and σ is the standard deviation.
Gamma model: With this model, Hakimhashemi (2009)
 1=2  
μ t − μ2 calculated the occurrence rate evolution along the Dead Sea
DF  exp − ; fault zone. Based on the model, DF is expressed as follows:
2πα2 t 3 2α2 μt
EQ-TARGET;temp:intralink-;;41;142

in which μ is the mean recurrence interval, t is the time elapsed 1


t α−1 e−β ;
t
DF 
βα Γα
EQ-TARGET;temp:intralink-;;308;119

since last earthquake, and α is the aperiodicity, which is usually


in between 0.3 and 0.7. We followed the procedure of Chan in which Γα is a Gamma function and α and β are constants
et al. (2019) and assumed a fixed α of 0.5. Based on this model, that are obtained by trial and error.

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 11

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
that the BPT distribution had the best fit with the observations,
100 5
confirming its credibility and raising the importance of a time-
Recurrence interval of the Chelunpu
fault observation and model
dependent model for PSHA. The outcome of this test also
4
explains the mismatch between PSHA and observations in
80

Number of the paleoseismic event


Observation and deviation the Italy case by Albarello and D’Amico (2008).
BPT function
Cumulative probability

Log-normal
3
The 2018 Hualien earthquake and the Milun fault. The
60
Gamma 2018 M 6.3 Hualien earthquake initiated offshore and ruptured
Poisson
along the Milun fault (ID 32), resulting in large ground-shak-
40 2
ing levels (Fig. 1a). In comparing the observations with our
model, however, we expect even higher ground shakings
rms
(see Fig. 2f). This could be attributed to the high recurrence
BPT function 0.108
20 1 rate of the Milun fault (highest among the 44 seismogenic
Log-normal 0.193

Gamma 0.159
structures, shown in Table 1) and high seismic rate in the area
Poisson 0.164
source (S17A) (third highest among the 28 area sources, shown
0 0 in Table 3). Because the recurrence rate of the Milun fault is
0 200 400 600 800 high, even considering the time-dependent BPT model and the
Recurrence interval (yr) last rupture in 2018, the expected rupture probability in the
near future remains high (fig. 4a of Chan et al., 2019).
Figure 5. Comparison of the observed recurrence intervals for the
Chelungpu rupture with modeled probability distributions based Earthquakes on unidentified seismogenic struc-
on the Brownian passage time (BPT), log-normal, gamma, and tures. During the observation period from 1995 to 2019, sev-
Poisson distributions. The mean and standard deviation of the
eral events occurred that cannot be attributed to any identified
recurrence interval (obtained by Fig. S2) were implemented in the
BPT model, and the best-fit parameters obtained by trial and seismogenic structures, for example, the 1994 M 6.2 Nan-Ao
error were adopted in the other three distributions. The root earthquake, the 2003 M 6.4 Chengkung earthquake, and the
mean square (rms) in comparison with observations for each 2006 M 7.0 Pingtung doublet (Fig. 1a). Their impact could
distribution is shown. be regarded from shallow-background sources in the forms
of area source and smoothing models, both of which forecast
the general spatial patterns of seismicity well (figs. 2 and 3 of
Poisson distribution model: The Poisson distribution Chan et al., 2019). In the case of the Nan-Ao earthquake, the
expresses DF as follows: earthquake took place on the boundary between areas S14A
and S15. Considering their a- and b- values (see Table 3),
EQ-TARGET;temp:intralink-;;53;340 DF  1 − e−ν×t ; the expected number of M 6.3 events in 25 yr is 1.58 and
0.20, respectively. Although an event of this magnitude is
in which ν is a constant that is obtained by trial and error. expected during the observation period in the two areas, it
DF presents the earthquake probability as a function of the is rather difficult to predict the specific locations of future
time elapsed since the last event t. The corresponding cumulative events. This ambiguity might explain not only the underpre-
probability is a function of t, which is expressed as follows: diction of ground shakings in the vicinity of the epicenter but
R tΔt also their overprediction in the Ilan Plain.
DFdt The Chengkung and Pingtung events occurred in areas
R∞ ;
t DFdt S18A and S08B, respectively. The expected number of events
EQ-TARGET;temp:intralink-;;53;249

of M 6.4 and M 7.0 in 25 yr are 0.13 and 0.01, respectively.


in which Δt is the exposure period. Considering the deviation of the a-value, the expected number
Confirming the credibility of the models requires comparing of events similar to the Chengkung earthquake and the
them with the rupture histories. We used the paleoseismological Pingtung earthquake in a 25 yr period is 0.74 and 0.01, respec-
data of the Chelungpu fault (Chen et al., 2007) and found that, tively. The occurrence of these earthquakes raises awareness of
in the past 2000 yr, the recurrence interval of the Chelungpu unexpected events in a region with relatively low hazard.
rupture had been 371 ± 126 yr (see Fig. S2). With the paleoseis-
mological records thus gained, we could validate our models. In Conclusions
the case of the BPT model, we incorporated the mean and stan- The annual recurrence rate could be further converted into
dard deviation of the recurrence interval; for the other three rupture probability in a period of interest considering proper
models, we considered the best-fit parameters obtained by trial temporal models (e.g., the Poisson distribution and the BPT
and error (see Fig. 5). The root mean square (rms) values suggest model). Here, we presented the rupture probability of each

12 Seismological Research Letters www.srl-online.org • Volume XX • Number XX • – 2022

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
seismogenic structure and multiple-structure rupture in the Figure 6. The rupture probability of each seismogenic structure and
next 25 yr (see Fig. 6). Several seismogenic structures in eastern multiple-structure rupture in the next 25 yr. The summarized rup-
Taiwan have obtained higher rupture probabilities, for exam- ture probabilities from all of these structures are reported, con-
sidering magnitude thresholds of 6.5, 6.7, and 7.0, respectively. The
ple, the Milun fault (ID 32), Southern Ilan structure (ID 38),
color version of this figure is available only in the electronic edition.
Taimali coastline structure (ID 36), Longitudinal Valley fault

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 13

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
(ID 33), and Central Range structure (ID 34), implying Acknowledgments
relatively high seismic hazard in their vicinity. Except for these This study was supported by the Ministry of Science and Technology in
active structures, most of the structures or faults obtain rupture Taiwan under the Grant Numbers MOST 109-2116-M-008 -029 -MY3,
probabilities lower than 3%, inferring each might be unlikely to MOST 110-2124-M-002 -008, and MOST 110-2634-F-008-008. This
rupture in the near future. However, when we summarized the work is financially supported by the Earthquake Disaster & Risk
Evaluation and Management Center (E-DREaM) from the Featured
rupture probabilities considering magnitude thresholds, signifi-
Areas Research Center Program within the framework of the Higher
cantly high potentials were reported (86%, 64%, and 22% for
Education Sprout Project by the Ministry of Education in Taiwan.
M ≥6.5, 6.7, and 7.0, respectively). This suggests the occurrence
The authors thank Allison Bent, Chris Rollins, and one anonymous
of a significant event such as the 1999 Chi-Chi earthquake reviewer for their constructive comments.
within a short period in Taiwan, resulting in severe ground
shaking in its vicinity (see Fig. 1a). The rupture probability References
map might be beneficial to urban planning on a city scale Albarello, D., and V. D’Amico (2008). Testing probabilistic seismic
and building code legislation on a national scale. hazard estimates by comparison with observations: An example
During the observation period, several earthquakes took in Italy, Geophys. J. Int. 175, no. 3, 1088–1094.
place on unidentified seismogenic structures, resulting in severe Albarello, D., L. Peruzza, and V. D’Amico (2015). A scoring test on
ground shakings in their vicinity (see Fig. 1a). According to the probabilistic seismic hazard estimates in Italy, Nat. Hazards Earth
TEM PSHA2020, these events were attributed to shallow-back- Syst. Sci. 15, no. 1, 171–186.
ground sources, and their impacts would be smoothed, that is, Ameri, G., F. Hollender, V. Perron, and C. Martin (2017). Site-specific
the model underestimated hazard levels in the vicinity (blue partially nonergodic PSHA for a hard-rock critical site in southern
France: Adjustment of ground motion prediction equations and
triangles near these epicenters shown in Fig. 2f) and slightly
sensitivity analysis, Bull. Earthq. Eng. 15, no. 10, 4089–4111.
overestimated ones in the far field. This shortcoming could
Anderson, J. G., and J. N. Brune (1999). Probabilistic seismic hazard
be improved by identifying seismogenic structures, such as
analysis without the ergodic assumption, Seismol. Res. Lett. 70,
blind faults or décollement (e.g., Hsin-Hua Huang, personal no. 1, 19–28, doi: 10.1785/gssrl.70.1.19.
comm., 2021) and offshore active faults or structures (e.g., Brooks, E. M., S. Stein, and B. D. Spencer (2016). Comparing the per-
Shyu et al., 2005). formance of Japan’s earthquake hazard maps to uniform and ran-
This study validated the TEM PSHA2020 by comparing the domized maps, Seismol. Res. Lett. 87, no. 1, 90–102, doi: 10.1785/
strong-motion observations from 1995 to 2019 and discussing 0220150100.
the impacts of some significant events. Our study could provide Chan, C. H., K. F. Ma, Y. T. Lee, and Y. J. Wang (2019). Rethinking
suggestions for the next generation of probabilistic seismic haz- seismic source model of probabilistic hazard assessment in Taiwan
ard models for Taiwan, as well as other regions. In addition, the after the 2018 Hualien, Taiwan, earthquake sequence, Seismol. Res.
regions with low observation levels might infer a potential haz- Lett. 90, no. 1, 88–96.
Chan, C. H., K. F. Ma, J. B. H. Shyu, Y. T. Lee, Y. J. Wang, J. C. Gao, Y.
ard in the future, raising the need to strengthen architecture and
T. Yen, and R. J. Rau (2020). Probabilistic seismic hazard assess-
infrastructure for hazard mitigation. We expect a subsequent
ment for Taiwan: TEM PSHA2020, Earthq. Spectra 36, 137–159.
study could be conducted when a longer observation period
Chan, C. H., Y. Wang, Y. J. Wang, and Y. T. Lee (2017). Seismic-haz-
is available, providing observed data as contributing factors ard assessment over time: Modeling earthquakes in Taiwan, Bull.
for the seismic hazard model to promote the model’s ability. Seismol. Soc. Am. 107, no. 5, 2342–2352, doi: 10.1785/0120160278.
Chan, C. H., Y. M. Wu, and J. P. Wang (2012). Earthquake forecasting
Data and Resources using the rate-and-state friction model and a smoothing Kernel:
The Taiwan Strong Motion Instrumentation Program (TSMIP) records Application to Taiwan, Nat. Hazards Earth Syst. Sci. 12, no. 10,
are from the Geophysical Database Management System (https:// 3045–3057.
gdmsn.cwb.gov.tw, last accessed July 2020). The source parameters Chen, W. S., C. C. Yang, I. C. Yen, L. S. Lee, K. J. Lee, H. C. Yang, H. C.
of our seismic hazard model are from the Taiwan Earthquake Chang, Y. Ota, C. W. Lin, W. H. Lin, et al. (2007). Late Holocene
Model (TEM) (last accessed June 2021). Our seismic hazard analysis paleoseismicity of the southern part of the Chelunpu Fault in
is calculated using the OpenQuake Engine 3.10 (https://github.com/ Central Taiwan: Evidence from the Chushan excavation site.
gem/oq-engine, last accessed May 2021). The distribution of V S30 is Bull. Seismol. Soc. Am. 97, no. 1B, 1–13.
from the Taiwan region from the Engineering Geological Database Dawood, H. M., and A. Rodriguez-Marek (2013). A method for
for TSMIP (http://egdt.ncree.org.tw/, last accessed February 2019). including path effects in ground-motion prediction equations:
The supplemental material for this article includes waveforms and An example using the Mw 9.0 Tohoku earthquake aftershocks.
timeline of the paleoseismic events along the Chelungpu fault. Bull. Seismol. Soc. Am. 103, no. 2B, 1360–1372.
Ellsworth, W. L., M. V. Matthews, R. M. Nadeau, S. P. Nishenko, P. A.
Reasenberg, and R. W. Simpson (1999). A physically based earth-
Declaration of Competing Interests quake recurrence model for estimation of long-term earthquake
The authors acknowledge that there are no conflicts of interest probabilities, U.S. Geol. Surv. Open-File Rept. 99-522, doi:
recorded. 10.3133/ofr99522.

14 Seismological Research Letters www.srl-online.org • Volume XX • Number XX • – 2022

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin
Fujiwara, H. (2014). Seismic hazard maps for Japan, in Encyclopedia of Lin, P. S., C. T. Lee, C. T. Cheng, and C. H. Sung (2011). Response
Complexity and Systems Science, 1–28. spectral attenuation relations for shallow crustal earthquakes in
Fujiwara, H., S. Kawai, S. Aoi, N. Morikawa, S. Senna, N. Kudo, M. Taiwan, Eng. Geol. 121, nos. 3/4, 150–164.
Ooi, K. X. Hao, K. Wakamatsu, Y. Ishikawa, and T. Okumura Salditch, L., M. M. Gallahue, M. C. Lucas, J. S. Neely, S. E. Hough, and S.
(2009). Technical reports on national seismic hazard maps for Stein (2020). California Historical Intensity Mapping Project
Japan, Technical Note of the National Research Institute for (CHIMP): A consistently reinterpreted dataset of seismic intensities
Earth Science and Disaster Prevention, 336. for the past 162 yr and implications for seismic hazard maps,
Gao, J. C., C. H. Chan, and C.T. Lee (2021). Site-dependent ground- Seismol. Res. Lett. 91, no. 5, 2631–2650.
motion prediction equations and uniform hazard response spectra, Shin, T. C. (1993). Progress summary of the Taiwan strong motion
Eng. Geol. 292, 106241, doi: 10.1016/j.enggeo.2021.106241. instrumentation program, Proc. Symposium on Taiwan Strong
Gardner, J. K., and L. Knopoff (1974). Is the sequence of earthquakes Motion Instrumentation.
in Southern California, with aftershocks removed, Poissonian? Shyu, J. B. H., K. Sieh, Y. G. Chen, and C. S. Liu (2005). Neotectonic
Bull. Seismol. Soc. Am. 64, no. 5, 1363–1367. architecture of Taiwan and its implications for future large earth-
Gutenberg, B., and C. F. Richter (1944). Frequency of earthquakes in quakes, J. Geophys. Res. 110, no. B8, doi: 10.1029/2004JB003251.
California, Bull. Seismol. Soc. Am. 34, no. 4, 185–188. Shyu, J. B. H., Y. H. Yin, C. H. Chen, Y. R. Chuang, and S. C. Liu (2020).
Hakimhashemi, A. H. (2009). Time-dependent occurrence rates of Updates to the on-land seismogenic structure source database by the
large earthquakes in the Dead Sea fault zone and applications Taiwan Earthquake Model (TEM) project for seismic hazard analy-
to probabilistic seismic hazard assessments, Ph.D. Dissertation, sis of Taiwan, Terr. Atmos. Ocean. Sci. 31, no. 4, 469–478.
University of Potsdam, Berlin, Germany. Stein, S., M. Liu, T. Camelbeeck, M. Merino, A. Landgraf, E. Hintersberger,
Kuehn, N. M., and N. A. Abrahamson (2020). Spatial correlations of and S. Kübler (2017). Challenges in assessing seismic hazard in intra-
ground motion for non-ergodic seismic hazard analysis, Earthq. plate Europe, Geol. Soc. Lond. Spec. Publ. 432, no. 1, 13–28.
Eng. Struct. Dynam. 49, no. 1, 4–23. Tasan, H., C. Beauval, A. Helmstetter, A. Sandikkaya, and P. Guéguen
Kuehn, N. M., and F. Scherbaum (2016). A partially non-ergodic (2014). Testing probabilistic seismic hazard estimates against accel-
ground-motion prediction equation for Europe and the Middle erometric data in two countries: France and Turkey, Geophys. J. Int.
East, Bull. Earthq. Eng. 14, no. 10, 2629–2642. 198, no. 3, 1554–1571.
Kuehn, N. M., N. A. Abrahamson, and M. A. Walling (2019). Wang, Y. J., C. H. Chan, Y. T. Lee, K. F. Ma, J. B. H. Shyu, and R. J. Rau
Incorporating nonergodic path effects into the NGA-West2 (2016). Probabilistic seismic hazard assessments for Taiwan, Terr.
ground-motion prediction equations, Bull. Seismol. Soc. Am. Atmos. Ocean. Sci. 27, no. 3, 325–340.
109, no. 2, 575–585. Wang, Y. J., Y. T. Lee, C. H. Chan, and K. F. Ma (2016). An investi-
Kuo, C. H., K. L. Wen, H. H. Hsieh, C. M. Lin, T. M. Chang, and K. W. gation of the reliability of the Taiwan Earthquake Model PSHA2015,
Kuo (2012). Site classification and Vs30 estimation of free-field Seismol. Res. Lett. 88, no. 2A, doi: 10.1785/0220160085.
TSMIP stations using the logging data of EGDT, Eng. Geol. Wen, K. L., H. Y. Peng, Y. B. Tsai, and K. C. Chen (2001). Why 1 G
129, 68–75, doi: 10.1016/j.enggeo.2012.01.013. was recorded at TCU129 site during the 1999 1Chi-Chi, Taiwan,
Landwehr, N., N. M. Kuehn, T. Scheffer, and N. A. Abrahamson (2016). Earthquake, Bull. Seismol. Soc. Am. 91, no. 5, 1255–1266.
A nonergodic ground-motion model for California with Wu, Y. M., C. H. Chang, L. Zhao, T. L. Teng, and M. Nakamura
spatially varying coefficients, Bull. Seismol. Soc. Am. 106, no. 6, (2008). A comprehensive relocation of earthquakes in Taiwan
2574–2583. from 1991 to 2005, Bull. Seismol. Soc. Am. 98, no. 3, 1471–1481.
Lanzano, G., M. D’Amico, C. Felicetta, R. Puglia, L. Luzi, F. Pacor, and Youngs, R. R., and K. J. Coppersmith (1985). Implications of fault slip
D. Bindi (2016). Ground-motion prediction equations for region- rates and earthquake recurrence models to probabilistic seismic
specific probabilistic seismic-hazard analysis, Bull. Seismol. Soc. hazard estimates, Bull. Seismol. Soc. Am. 75, no. 4, 939–964.
Am. 106, no. 1, 73–92.
Lin, P. S., and C. T. Lee (2008). Ground-motion attenuation relation-
ships for subduction-zone earthquakes in northeastern Taiwan, Manuscript received 9 July 2021
Bull. Seismol. Soc. Am. 98, no. 1, 220–240, doi: 10.1785/0120060002. Published online 20 April 2022

Volume XX • Number XX • – 2022 • www.srl-online.org Seismological Research Letters 15

Downloaded from http://pubs.geoscienceworld.org/ssa/srl/article-pdf/doi/10.1785/0220210186/5589463/srl-2021186.1.pdf


by National Central Univ Library Serials Office, Yen-Yu Lin

You might also like