Вы находитесь на странице: 1из 11

Bulletin of the Seismological Society of America, Vol. 106, No. 2, pp. –, April 2016, doi: 10.

1785/0120150214

Uncertainty in V S30-Based Site Response


by Eric M. Thompson and David. J. Wald

Abstract Methods that account for site response range in complexity from simple
linear categorical adjustment factors to sophisticated nonlinear constitutive models.
Seismic-hazard analysis usually relies on ground-motion prediction equations (GMPEs);
within this framework site response is modeled statistically with simplified site param-
eters that include the time-averaged shear-wave velocity to 30 m (V S30 ) and basin depth
parameters. Because V S30 is not known in most locations, it must be interpolated or
inferred through secondary information such as geology or topography. In this article,
we analyze a subset of stations for which V S30 has been measured to address effects of
V S30 proxies on the uncertainty in the ground motions as modeled by GMPEs. The
stations we analyze also include multiple recordings, which allow us to compute the
repeatable site effects (or empirical amplification factors [EAFs]) from the ground mo-
tions. Although all methods exhibit similar bias, the proxy methods only reduce the
ground-motion standard deviations at long periods when compared to GMPEs without
a site term, whereas measured V S30 values reduce the standard deviations at all periods.
The standard deviation of the ground motions are much lower when the EAFs are used,
indicating that future refinements of the site term in GMPEs have the potential to sub-
stantially reduce the overall uncertainty in the prediction of ground motions by GMPEs.

Introduction
Recently, we reported on empirical amplification func- probabilistic seismic-hazard analysis (e.g., Petersen et al.,
tions (EAFs) from the Ancheta et al. (2014) database (referred 2014) and real-time hazard products, such as ShakeMap
to as the Next Generation Attenuation [NGA]-West2 data- (Wald et al., 1999; Worden et al., 2010).
base) for the purpose of mapping EAFs in California (Thomp- Our previous efforts on the accuracy of V S30 proxies
son and Wald, 2014). However, the resulting suite of EAFs (e.g., Thompson et al., 2010, 2014) assessed the accuracy of
also provides the opportunity to reflect on the overall accuracy alternative V S30 estimates in terms of V S30 residuals. However,
of V S30 -based site-response models as well as the proxies that from the seismic-hazard perspective, the accuracy of the
are often used to estimate V S30 , which we present in this ar- resulting ground motions is more important. The pertinent
ticle. Because our original goal was to map EAFs in California, metric of accuracy should focus on how accurately the ground
the geographic region under consideration in this article is also motions are represented. Therefore, in this article, we focus on
California. The goal of this article is to compare the accuracy assessing the accuracy of different V S30 estimates in terms of
of V S30 -based site models, in which V S30 values are estimated ground-motion intensity measures (IMs; i.e., peak accelera-
from a wide range of available methods, including site-spe- tion, peak velocity, or spectral acceleration). Thus, we must
cific measurements and different mapping techniques. In par- employ a V S30 -based model of site response, such as Choi
ticular, we make an attempt to quantify the degree to which the and Stewart (2005), Walling et al. (2008), Seyhan and Stewart
use of V S30 (through direct measurements or proxies) im- (2014), and Kamai et al. (2014). We arbitrarily adopt one par-
proves our ability to model ground motions and the extent to ticular site-response model for our comparison of V S30 ampli-
which it is possible to improve upon current V S30 -based site- fications, because the current generation of V S30 -based site-
response models. response models are based on similar methods and similar
Although more detailed models can include the full data. We expect that this choice will have little effect on
resonant response of S waves for a layered velocity model the results; however, we acknowledge that the small portion
and the strain dependence of the soil properties as amplitudes of our results regarding basin depth parameters will be more
increase (e.g., Hartzell et al., 2004), this study focuses only sensitive to this choice because there is greater variety in the
on V S30 -based models of site response. These simpler site- proposed models and less data available to constrain and as-
response models are still important because they are widely sess those models.
available and used in ground-motion prediction equations Unlike the equations for relating V S30 to amplification,
(GMPEs; e.g., Boore et al., 2014) that are, in turn, used for the different methods of estimating V S30 vary substantially in

BSSA Early Edition / 1


2 E. M. Thompson and D. J. Wald

(a) (b) 1.0 (c) 1.0


0.4

Data: RJB ≤ 80 km
0.8 0.8
0.2

c (ln units)

φ (ln units)
τ (ln units)
0.6 0.6
0.0
0.4 0.4
−0.2
0.2 0.2
−0.4
0.0 0.0

1.0 1.0
Data: RJB ≤ 400 km

0.4
0.8 0.8
0.2
c (ln units)

φ (ln units)
τ (ln units)
0.6 0.6
0.0
0.4 0.4
−0.2
0.2 0.2
−0.4
0.0 0.0
0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10
Period (s) Period (s) Period (s)
Mean BSSA14: M 4.5, RJB = 80 km
95% CI BSSA14: M 4.5, RJB = 400 km

Figure 1. Mixed-effects regression parameters: (a) fixed effect (c), (b) standard deviation of the interevent residuals (τ), and (c) standard
deviation of the intraevent residuals (ϕ). The solid black lines indicate the maximum likelihood estimate, and the shaded gray areas give the
95% confidence intervals (CI). The top row uses only recordings with Joyner–Boore distance RJB < 80 km; the bottom row uses only
recordings with RJB < 400 km. Note that the there is only one curve for the Boore et al. (2014, or BSSA14) model of τ, because it does
not depend on distance.

scale and applicability. Thus, it is important to consider the Data


accuracy of different V S30 proxies. As discussed by Thomp-
son et al. (2011), we generally expect that the accuracy of a For record selection, we use the same basic screening cri-
V S30 mapping approach will be inversely related to its spatial teria described by Boore et al. (2014; herein called BSSA14).
coverage. The topographic slope method of Wald and Allen One complicating issue is that, although their equations are
(2007) (updated by Allen and Wald, 2009) and the geomor- applicable up to a maximum Joyner–Boore distance (RJB ; the
phic terrain method of Yong et al. (2012) give global cover- horizontal distance to the surface projection of rupture) of
age, whereas the Wills et al. (2015) and Thompson et al. 400 km, they exclude records beyond RJB  80 km to deter-
(2014) maps only cover California. Thus, in the same manner mine some of the terms for the phase 2 regression. So it is
in which we only consider one site-response model, we com- important for us to investigate the influence of the larger dis-
pare only one method of determining V S30 at each of the fol- tance records on the residuals. Figure 1 plots the bias (c),
lowing spatial scales: global, regional, and site. We use the interevent and intraevent residual standard deviations (τ and
Allen and Wald (2009) method at the global scale and the ϕ; see the Methods section for additional details), and the
Wills et al. (2015) map at the regional scale. In this article, 95% confidence limits of these parameters against period for
we refer to the Allen and Wald (2009) method as the topo- RJB < 80 km (top row) and for RJB < 400 km (bottom row).
graphic method and the Wills et al. (2015) map as the geol- For comparison, we also plot lines to indicate the model
ogy method. τ and ϕ given by BSSA14. One concern is that the bias
For a site-specific investigation, many different methods increases at long periods when data beyond RJB of 80 are
of estimating V S30 are available. V S30 is computed from the included. However, we decided that it is more important to
shear-wave velocity profile, which is estimated from geo- include these data in our analysis than to try to avoid this
physical techniques with varying levels of precision and small yet persistent bias.
depth of penetration (see Boore and Asten, 2008, for a blind With multiple recordings at a site, the EAF can be esti-
comparison of different techniques). We do not wish to con- mated from multiple ground motions at that site (e.g., Joyner
sider the different geophysical techniques as additional var- and Boore, 1993; Tinsley et al., 2004; Lin et al., 2011;
iables in our study, and so we simply treat all the V S30 entries Rodriguez-Marek et al., 2011). Because the EAF is estimated
in the NGA-West2 site database that were derived from a by comparing the recorded IMs with the IMs for a constant
shear-wave velocity profile that extends to a depth of 20 m reference condition, they are not limited by the simplicity of
or more as site-specific V S30 measurements (see Seyhan et al., the V S30 proxy or other common site-response modeling as-
2014, for more details). sumptions, such as a 1D vertical wave propagation. Thus, the

BSSA Early Edition


Uncertainty in V S30 -Based Site Response 3

difference sources (see Ancheta et al., 2014, for details). Fol-


lowing the approach of BSSA14, the site amplification term
FS is expressed as
150

EQ-TARGET;temp:intralink-;df1;313;697 FS  lnFlin   lnFnl   Fδz1 ; 1


Number of stations

in which Flin is the function that represents linear site effects,


100 Fnl is for the nonlinear site effects, and Fδz1 is for the effects
of basin depth. For convenience, Table 1 provides definitions
of selected variables employed in this article. Note that Fδz1
is centered such that it is relative to the value of z1 expected
50
for the V S30 at that site.
The first step in computing the EAF relative to a GMPE
reference condition is to compute the event terms, which are
the average of the residuals for each event. They vary with
period and must be subtracted from the data to isolate the site
0 effects. We refer to the recorded IM as y, which is a function
0.01 0.1 1 10
of oscillator period T. To compute the event terms, we adjust
Period (s) the recorded data to a consistent reference site condition; for
this, we adopt the BSSA14 equations with a V S30 of 760 m=s
Figure 2. The number of stations that fulfill the data-screening
criteria as a function of period.
(which means that FS  0). We refer to the V S30  760-
adjusted recorded values as yr  y= expFS  so the unad-
justed and reference-condition-adjusted residuals are
EAF is able to capture any repeatable site effects, including
velocity structure that is deeper than 30 m and deviations EQ-TARGET;temp:intralink-;df2;313;457 R  lny − ln^y; 2
from 1D behavior such as horizontally propagating surface
waves (Graves, 1993; Baise et al., 2003) and seismic scatter-
ing (Thompson et al., 2009). The EAF, however, is still sub- EQ-TARGET;temp:intralink-;df3;313;412 Rr  lnyr  − ln^y; 3
ject to uncertainty because it is estimated from a limited
number of recorded motions. This uncertainty decreases as respectively, in which y^ is the BSSA14 prediction for
the number of recordings at each station increases. There- V S30  760 m=s. The event terms are then computed by
fore, we only include stations in our analysis for which we separating the total residuals defined in equation (3) into
can compute the EAF from four or more records for which interevent and intraevent residuals using the mixed-effects
the BSSA14 nonlinearity term (Fnl ; see the Methods section) framework:
is within 5% of unity to ensure that linear amplification can
be assumed (the records must also be from earthquakes for EQ-TARGET;temp:intralink-;df4;313;341 Rr i;j  c  ηi  εi;j ; 4
which event terms could be computed using the BSSA14 re-
quirement of four records within 80 km). These selection cri- in which i is the earthquake event index, j is the record index,
teria yield 555 stations. Of those 555 stations, 194 have c is the mean residual (i.e., the overall bias of the residuals
measured V S30 values. This is the subset of data that we ana- relative to the BSSA14 predictions), ηi is the mean residual
lyze in this article. The number of stations that fulfill these for the ith event (i.e., event term, interevent residual, or be-
criteria, shown in Figure 2, varies with period for two rea- tween-event residual), and εi;j is the intraevent (i.e., within
sons: (1) the longest usable period varies from record to rec- event) residual. The standard deviation of ηi is τ, and the
ord, diminishing the number of available records as period standard deviation of εi;j is ϕ. We estimate the residuals
increases; and (2) nonlinear effects are stronger at shorter and their standard deviations using the mixed-effects regres-
periods, diminishing the number of available records as sion package nlme in R (Pinheiro et al., 2013).
period decreases. The pertinent residual for defining the EAF is to subtract
the event term from the unadjusted residuals, giving

Methods EQ-TARGET;temp:intralink-;df5;313;177 Rec i;j  R − ηi − c; 5


Modern GMPEs typically model site response as a func-
tion of V S30 combined with parameters to approximate basin in which the “ec” subscript indicates that this is the event-
depth, such as the depth at which the shear-wave velocity corrected residual. Up until this point, we have used the term
profile exceeds 1:0 km=s (z1 ). In this article, we use z1 from EAF qualitatively. Quantitatively, we define the EAF as the
the NGA-West2 flatfile, which is compiled from a number of mean Rec i;j at a site:
BSSA Early Edition
4 E. M. Thompson and D. J. Wald

Table 1 from the 3D velocity models for northern and southern


Definitions of Selected Variables California). There are 98 stations where z1 is not available,
Variable Definition
and so Fδz1  0 at these stations for the calculation of these
parameters.
IM Ground-motion IM; for example, peak ground
acceleration and spectral acceleration
• μE and ϕE are the μ and ϕ for which FS is computed as
FS Site amplification based on V S30 ; natural logarithm of EAF; thus, μE is uninteresting and equal to zero by def-
amplification ratio inition. Note that ϕE is the same as single-station σ as
Flin Linear site amplification defined by Atkinson (2006) and termed ϕSS by Al Atik
Fnl Nonlinear site amplification et al. (2010).
Fδz1 Site amplification due to basin depth
y Recorded ground-motion IM
yr  y=FS V S30  760-adjusted recorded ground-motion IM Example Sites
R Logarithmic ground-motion residual relative to the
V S30  760 prediction Comparing the EAF with V S30 -based amplifications at a
Rr Logarithmic V S30  760-adjusted ground-motion few example sites illustrates the different types of uncertain-
residual relative to V S30  760 prediction
c Ground-motion prediction equation (GMPE) bias
ties in modeling the site term. Unless otherwise noted, in this
τ Interevent residual standard deviation section we assume Fδz1  0. Figure 3 plots the EAF along
ϕ Intraevent residual standard deviation with FS for four different stations. Summary information for
Rr i;j Rr for event i and record j each station is given in Table 2. The proxy-based V S30 values
Rec i;j R adjusted for the interevent residual for both geology and topography result in a good match
EAF Mean Rec i;j at a site
μX Bias of the site term; mean of EAF minus FS ; subscript
between the EAF and FS at the Chino station; in this case,
indicates how FS is computed we expect that a more accurately measured V S30 , or even a
ϕX Standard deviation of EAF minus FS ; subscript site-specific analysis, could not provide significant improve-
indicates how FS is computed ment to the proxy-based V S30 site-response model.
ϕE , ϕSS Special case of ϕX in which FS equals the EAF; The EAF at China Lake is not well modeled by the geol-
referred to as single-station σ by Atkinson (2006) and
defined as ϕSS by Al Atik et al. (2010)
ogy- and topography-based V S30 estimates: both result in
δS2SS EAF − FS amplifications greater than one, whereas the EAF indicates
ϕS2S Standard deviation of δS2SS that the site amplification is smaller than that for the refer-
ence V S30 of 760 m=s. Yong et al. (2013) measured the V S
IM, intensity measure; EAF, empirical amplification factor.
profile at China Lake and also reported that the local geology
is granite, whereas the Wills et al. (2015) map classifies this
1X n
location as steep Quaternary alluvium (V S30  351:9 m=s).
EAF  R  ; 6
n j1 ec i;j If we remove the classification error that results from the spa-
EQ-TARGET;temp:intralink-;df6;55;372

tial precision of the digitized geology map by using the mean


V S30 for crystalline rock of 710:1 m=s reported by Wills et al.
in which n is the number of recordings at that site. Thus, the
(2015), we see that the agreement between FS and the EAF is
median amplification factor is expEAF. We have computed
improved (this corresponds to the FS curve labeled “True”
EAF from the analysis of the interevent residuals from a
geology in Fig. 3 for China Lake). However, this value of
mixed-effects regression. Stafford (2014) has pointed out
V S30 still does not reproduce the deamplification (relative to
that that this is not optimal because the site terms will affect
V S30  760 m=s) that is observed at China Lake. If we use
the event terms, and so a more formal analysis would
the site-specific measured V S30 of 1463 m=s reported by
compute the site and the event terms simultaneously.
Yong et al. (2013), then we see that the resulting FS
To quantify the accuracy of the linear components of FS ,
adequately reproduces the EAF.
we compute the bias (μ) and standard deviation (ϕ) of the
Figure 3 also shows that the proxy-based V S30 amplifi-
site-response residual Rec i;j − FS  using different assump-
cations do not accurately match the EAF for the Saticoy site.
tions for FS :
At long periods, the measured V S30 value gives better results
• μX and ϕX are the μ and ϕ for which FS  0 and Fδz1  0. than the proxy-based V S30 values but still underpredicts the
In other words, no site term has been applied. EAF. The median EAF at Saticoy reach a maximum value of
• μT and ϕT are the μ and ϕ for which FS is computed from about 5–6 at long periods (around 2–10 s) and indicates that
the Allen and Wald (2009) topographic slope-based V S30 this is a deep-soil site. Adding the Fδz1 term to the FS
and Fδz1  0. calculation improves the fit further by increasing the long-
• μG and ϕG are the μ and ϕ for which FS is computed from period amplification to about a factor of 4. The formulation
the Wills et al. (2015) geology-based V S30 and Fδz1  0. of the Fδz1 term in BSSA14 only affects periods greater than
• μM and ϕM are the μ and ϕ for which FS is computed from 0.65 s, and so the shorter periods remain unaffected. The z1
measured V S30 and Fδz1  0. parameter helps because it accounts for the fact that the
• μM;Z and ϕM;Z are the μ and ϕ for which FS is computed soil at this site is deeper than the average site with the same
from measured V S30 and Fδz1 is computed from z1 (taken V S30 value.
BSSA Early Edition
Uncertainty in V S30 -Based Site Response 5

10
Chino China Lake

Amplification: exp(FS), exp(EAF)


1

0.1

10
Saticoy Barrett
Amplification: exp(FS), exp(EAF)

0.1
0.01 0.1 1 10 0.01 0.1 1 10
Period (s) Period (s)

EAF FS
exp(EAF) Geology VS30 Measured VS30
95% CI Topography VS30 Measured VS30, with Fδz1
"True" Geology VS30

Figure 3. Four example sites comparing the empirical amplification factor (EAF) to V S30 -based amplifications (FS ) for different estimates
of V S30 . In all cases Fnl  0 and Fδz1  0 unless otherwise noted. The shaded gray areas indicate the 95% CI for the EAF. Geology V S30 is
from Wills et al. (2015), and topography V S30 is from Allen and Wald (2009). The color version of this figure is available only in the
electronic edition.

The EAF at Barrett shows amplifications factors of about sured V S30 value here is 511 m=s, indicating that it is a stiff-
2–3 at short periods (0.01–0.3 s) and slight deamplification at soil–soft-rock site. Although the FS does not match the EAF,
mid-to-long periods (> 0:6 s). This is a site where no V S30 the shape of the EAF is consistent with this interpretation of
value will be able to model the shape of the observed am- the site conditions associated with the value of V S30 in the
plification, simply because FS does not produce a peak at sense that we expect stiff sites to resonate at shorter periods.
these shorter periods for any V S30 –z1 combination. The mea- The z1 taken from the velocity profile at this site (from Yong

Table 2
Site Parameters for Stations in Figure 2
V S30 m=s
Station Name Longitude (°) Latitude (°) Measured Geology Topography z1 (km)

Chino −117.68 33.999 292 293.5 319 0.320


China Lake −117.597 35.81574 1464 351.9 538 0.003
Saticoy −119.1646 34.293 249 351.9 433 0.740
Barrett −116.6722 32.68005 511 710.1 539 —

BSSA Early Edition


6 E. M. Thompson and D. J. Wald

4
FS = 0 Topography−based VS30 Geology−based VS30

(Rec)i,j − FS 0

−2

μX = 0.433 φX = 0.754 μT = −0.023 φT = 0.720 μG = −0.008 φG = 0.711


−4

4
Measured VS30 Measured VS30 with Fδz1 FS = FEAF

2
(Rec)i,j − FS

−2

μM = −0.028 φM = 0.591 μM,Z = −0.032 φM,Z = 0.596 μE = 0.000 φE = 0.448


−4
100 1000 100 1000 100 1000
VS30 (m/s) VS30 (m/s) VS30 (m/s)

Figure 4. Plots of the site-response residual Rec i;j − FS with V S30 for T  1 s and for different assumptions for the site term (FS ). In
each plot, the abscissa is the measured V S30 value, even in panels for which FS is computed without the use of measured V S30 . The black dots
and vertical bars give the binned means and 1 standard deviations.

et al., 2013) is 25 m, and Figure 3 shows that this gives a standard deviation for each case is labeled in the figure. Be-
slight improvement in fit between FS and the EAF at long cause we are plotting Rec i;j − FS in Figure 4, it contains
periods but is unable to meaningfully improve the site-re- two components of variability, as discussed by Al Atik et al.
sponse model at this location. For a stiff site such as this, (2010): the deviation of the EAF (i.e., mean Rec ) at each site
the z1 parameter would ideally increase FS at short periods from FS (δS2SS ; i.e., EAF − FS ), and the remaining record-
and decrease FS at longer periods to reflect the shift in res- to-record residual Rec i;j − EAF. Although the variability
onance frequency to shorter periods as the ground becomes of Rec i;j is useful, because it is analogous to ϕ reported in
stiffer. GMPEs, δS2SS (given in Fig. 5) is useful because it does not
These examples illustrate that the EAF may be accu- include the record-to-record component of variability at each
rately modeled with simple V S30 proxies or that it can be lim- site. Another benefit of δS2SS is that it gives equal weight to
ited by the uncertainty of the proxy. The Saticoy and Barrett each site, whereas Rec i;j places more weight on sites with
sites further demonstrate that, even when the V S30 is known more recordings. The standard deviation of δS2SS is defined
with precision and even when the Fδz1 term is included and as ϕS2S by Al Atik et al. (2010), and we report ϕS2S for differ-
z1 is known accurately, the EAF may be impossible to rep- ent estimates of V S30 in Figure 5. There are only five panels
licate with existing V S30 models. The four example sites in in Figure 5 because we did not include the plot of δS2SS for
Figure 3, however, cannot give an indication of the preva- FS  EAF, which is zero by definition.
lence of these different types of errors; in other words, it The outlier in Figure 5 at low V S30 values is the Salton
is important to know whether sites like Barrett or Chino Sea Wildlife Refuge (SSWR) station. The negative residual
are more common. We address this issue in the next section, indicates that the EAF is overpredicted by FS for T  1 s
where we compute summary statistics for all of the sites spectral acceleration, which is based on a measured V S30
within the NGA-West2 database for which we can estimate of 191 m=s. Although it is beyond the scope of this article
the EAF and for which the V S30 has been measured. to analyze this individual site in detail, we note that this sta-
tion is located on Imperial Valley alluvium and that there is
Uncertainty in Site-Response Models no reason to suspect that the V S profile or V S30 is inaccurate.
However, the station is located about 300 m southeast of
To visualize the data across all sites, we plot the site-re- Rock Hill, which rises about 25 m above the alluvium. Rock
sponse residuals Rec i;j − FS  for each of the different FS Hill is one of a series of rhyolite domes that form a nearly
estimates against V S30 for T  1 s in Figure 4; the bias and linear northeast–southwest-trending transect in the Salton
BSSA Early Edition
Uncertainty in V S30 -Based Site Response 7

4
FS = 0 Topography−based VS30 Geology−based VS30

δS2Ss 0

−2

φS2S = 0.625 φS2S = 0.593 φS2S = 0.568


−4
100 1000 100 1000 100 1000
VS30 (m/s) VS30 (m/s) VS30 (m/s)
4
Measured VS30 Measured VS30 with Fδz1

2
δS2Ss

−2
SSWR

φS2S = 0.419 φS2S = 0.423


−4
100 1000 100 1000
VS30 (m/s) VS30 (m/s)

Figure 5. Plots of δS2SS with V S30 for T  1 s and for different assumptions for the site term (FS ). In each plot, the abscissa is the
measured V S30 value, even in panels for which FS is computed without the use of measured V S30 . The black dots and vertical bars give the
binned means and 1 standard deviations. The Salton Sea Wildlife Refuge station (SSWR) is labeled because it is an outlier and is discussed
in the text.

Sea geothermal field (Robinson et al., 1976). Thus, we sus- to amplify the ground motion; the GMPE without the site-
pect that velocity structure in the vicinity of this station ex- term results in underprediction (positive bias). When the site
hibits complex 3D variability, which is the source of the term is added, all the different methods decrease the bias dra-
anomalous linear site-response behavior. matically. The standard deviations give a different trend: the
All of the V S30 estimates substantially reduce the bias FS from proxy-based V S30 values do not reduce the standard
and standard deviation for T  1 s (Figs. 4 and 5). To see deviations at short periods (T < ∼1 s) but do provide minor
the trends across a broad range of periods, we plot the bias reductions at longer periods. The measured V S30 reduces the
and standard deviations of the site-response residuals for standard deviations further. We do not see much improve-
periods from 0.01 to 10 s in Figure 6; the analogous figure ment with the addition of the z1 parameter, but this is likely
for δS2SS is given in Figure 7. We include both bias and stan- because this parameter is unavailable at more than half of the
dard deviations in these plots to provide a complete descrip- sites in our analysis.
tion of accuracy. That said, the standard deviation is much
more important, because the bias can be small even in the Discussion
presence of large scatter in the residuals. In Figure 6, we also
The reduction in the bias, relative to FS  0, is arbitrar-
include the Rodriguez-Marek et al. (2013) model of ϕSS
ily determined by the assumed reference V S30 . For example,
(which is the same as ϕE ). Their model is a function of mag- the bias of the “no site” curve in Figures 6 and 7 will be
nitude and rupture distance (RRUP ); and, for the comparison, reduced if we assume a reference site condition of
we select RRUP  84 km (the median RRUP from our data V S30  373 m=s (the median V S30 for the sites in our analy-
set) and M 5 (the median in our data is 4.2, but their equa- sis). Therefore, it is important not to attribute all of this re-
tions are constant for M < 5). The comparison between the duction in bias to the use of V S30 . The bias plots in Figures 6
Rodriguez-Marek et al. (2013) model of ϕSS to our estimates and 7, however, do demonstrate that, on average, measured
of ϕE shows very good agreement at long periods, but we V S30 values do not reduce the bias relative to the use of
obtained consistently larger values than their model as period proxy-based V S30 estimates.
decreases. To further illustrate the reduction in ϕ and ϕS2S for the
The large positive bias seen in Figures 6 and 7 for the different site-term methods, we compute the percentage
curves that assume FS  0 is because linear site effects tend change in standard deviation relative to the standard devia-
BSSA Early Edition
8 E. M. Thompson and D. J. Wald

(a) 0.8 (b) 1.0

μX μM Rodriguez−Marek et al. (2013)


μT μM,Z φSS(M=5, RRUP=84 km)
0.6 μG μE 0.8

Standard deviation (ln units)


Bias (ln units)
0.4 0.6

0.2 0.4

0.0 0.2 φX φM
φT φM,Z
φG φE

−0.2 0.0
0.01 0.1 1 10 0.01 0.1 1 10
Period (s) Period (s)

Figure 6. (a) Bias and (b) standard deviation of Rec i;j as a function of period for different assumptions for the site term (FS ). The
Rodriguez-Marek et al. (2013) model of ϕSS is included to compare with our estimate (labeled ϕE ).

(a) 0.8 (b) 1.0

0.6 0.8
Standard deviation (ln units)
Bias (ln units)

0.4 0.6

0.2 0.4

0.0 0.2

−0.2 0.0
0.01 0.1 1 10 0.01 0.1 1 10
Period (s) Period (s)

No site Topography Geology Measured Measured, z1

Figure 7. (a) Bias and (b) standard deviations of δS2SS (i.e., ϕS2S ) as a function of period for different assumptions for the site term (FS ).

tions obtained with no site term in Figure 8. Surprisingly, the for the percentage change in ϕ for the EAF. It is also possible
proxy values provide no reduction in ϕ and ϕS2S at short peri- that, if the GMPE developers had regressed their models
ods (T < 1 s). The proxy values reduce ϕ by 5%–12% and against a particular proxy-based V S30 , then that method
ϕS2S by 8%–22% at longer periods (T > 1 s). Measured V S30 might have performed slightly better in our comparisons.
(with or without z1 ) reduces ϕ by 3%–9% and ϕS2S by That said, we expect this effect to be minimal, particularly
5%–18% for shorter periods (<∼0:2 s) and reduces ϕ by since the developers used a mixture of inferred and measured
15%–24% and ϕS2S by 26%–40% at periods longer than V S30 values in their regression.
0.5 s. It is important to remember that the reduction of ϕ It has long been recognized that, besides V S30 , basin
is always less than that of ϕS2S , because the latter does depth is an important factor (e.g., Campbell, 1997; Field,
not include the record-to-record variability that remains after 2000). For the site term in BSSA14, basin depth is modeled
adjusting for the repeatable site effects (i.e., the standard with the z1 parameter. At present, although this parameter
deviation of Rec i;j − EAF, or ϕE ). As a percentage of can provide improvement over V S30 alone for individual sites
the total variability, ϕE is represented in Figure 8 by the curve (such as Saticoy in Fig. 3), our results indicate that on

BSSA Early Edition


Uncertainty in V S30 -Based Site Response 9

(a) 10 Percentage change in φ (b) Percentage change in φ S2S


0

−10
Percentage change
−20

−30

−40

−50

−60
0.01 0.1 1 10 0.01 0.1 1 10
Period (s) Period (s)

Topography Geology Measured Measured, z1 EAF

Figure 8. (a) Percentage change of ϕ and (b) ϕS2S relative to using no site term.

average it does not provide a significant benefit across all values and compared the standard deviations for different
sites. However, this is primarily because z1 is not known formulations of the site term. The focus of the analysis, how-
for more than half of the sites. It is important to obtain es- ever, was to assess the value of using V S30 as the site-term
timates of z1 at more sites, and this is particularly true for versus site-class categories (e.g., soil versus rock). The re-
stiff-soil sites where the current basin depth terms are not sults for T  2 s yielded (base 10) logarithmic standard de-
able to reflect the increased amplifications at high frequen- viations of 0.25 for no site class, 0.25 for a binary soil–rock
cies (such as Barrett in Fig. 3). Fundamental frequency (f 0 ) site classification, 0.21 for National Earthquake Hazards Re-
has also been proposed as a viable parameter to augment duction Program site classes, and 0.20 for V S30 as a continu-
V S30 (e.g., Cadet et al., 2012; Di Alessandro et al., 2012), ous random variable; the corresponding percentage change is
which could fulfill a similar role as z1 . An important benefit −20%, which is similar to the value that we report at T  2 s
to augmenting V S30 with f 0 is that it is relatively easy to ob- in Figure 8 (∼23%). This difference is not surprising, given
tain estimates of f0 in locations where a deep site investiga- the differences in the data and GMPE in the analysis.
tion is not feasible and a 3D velocity model is unavailable. Differences between our analysis and that of Boore (2004)
Regardless of which parameter is used, it is important that include the following:
the site term be able to reflect the shift in site period to short
• We focus on the performance of alternative V S30 proxies
periods for stiff sites, as seen at the Barrett station (Fig. 3).
rather than categorical versus continuous site terms. The
Our analysis has some similarities to the analyses pre-
use of V S30 in GMPEs is much more pervasive (although
sented by both Rodriguez-Marek et al. (2013) and Boore
still not without controversy) than at the time Boore (2004)
(2004; section 4.1.2). The reduction in the interevent stan-
was published.
dard deviations that we observe in Figure 8 (from ϕM to ϕE ) • We show the period dependence of the effects on standard
is analogous to the change from the use of ergodic interevent deviation.
standard deviations to partially nonergodic interevent stan- • The use of the EAF provides an estimate of the lower bound
dard deviations (i.e., single-station ϕ) as discussed in Rodri- on standard deviation that we can hope to attain with future
guez-Marek et al. (2013). Although Rodriguez-Marek et al. refinements to the site term in GMPEs.
(2013) developed models of single-station ϕ (i.e., ϕSS ) for
use in partially nonergodic probabilistic seismic-hazard
analysis (PSHA), our purpose in computing ϕE is to assess Conclusions
the relative accuracy of the site terms in GMPEs. Within this
context, ϕE represents an unobtainable minimum standard In this article, we compute and analyze EAFs from the
deviation for the site term. For PSHA applications, it is better NGA-West2 database to quantify the difference in the
to directly use models of single-station ϕ than to compute it ground-motion standard deviations for different methods
from models of ergodic ϕ paired with modification terms of computing the site-response term. We consider different
such as those in Figure 8. estimates of V S30 , as well as the use of z1 , as a proxy for
Boore (2004) plotted T  2 s response spectral resid- basin depth. We find almost no difference in bias for all
uals as a function of V S30 for a subset of sites with measured of these methods of computing the site term. In terms of

BSSA Early Edition


10 E. M. Thompson and D. J. Wald

the standard deviation of the site-response residual, we find Campbell, K. W. (1997). Empirical near-source attenuation relationships for
the following: horizontal and vertical components of peak ground acceleration, peak
ground velocity, and pseudo-absolute acceleration response spectra,
• Geology- and topography-based V S30 values do not reduce Seismol. Res. Lett. 68, 154–179.
Choi, Y., and J. P. Stewart (2005). Nonlinear site amplification as function of
the standard deviation at short periods but do reduce the
30 m shear wave velocity, Earthq. Spectra 21, 1–30.
standard deviation by 8%–22% for periods longer than 1 s. Di Alessandro, C., L. F. Bonilla, D. M. Boore, A. Rovelli, and O. Scotti
• Measured V S30 (with or without z1 ) reduce the standard (2012). Predominant-period site classification for response spectra
deviation by 3%–9% for periods less than 0.2 s and by prediction equations in Italy, Bull. Seismol. Soc. Am. 102, 680–695.
26%–40% for periods longer than 0.5 s. Field, E. H. (2000). A modified ground-motion attenuation relationship for
• Applying the EAFs reduces the intraevent standard devia- southern California that accounts for detailed site classification and a
basin-depth effect, Bull. Seismol. Soc. Am. 90, S209–S221.
tion by about 25% for periods shorter than 0.1 s and by Graves, R. W. (1993). Modeling three-dimensional site response effects in
more than 40% for periods longer than 1 s. the Marina District basin, San Francisco, California, Bull. Seismol.
Soc. Am. 83, 1042–1063.
Hartzell, S., L. F. Bonilla, and R. A. Williams (2004). Prediction of nonlinear
Data and Resources soil effects, Bull. Seismol. Soc. Am. 94, 1609–1629.
Joyner, W. B., and D. M. Boore (1993). Methods for regression analysis of
The ground-motion data used in this article were com-
strong-motion data, Bull. Seismol. Soc. Am. 83, 469–487.
piled, processed, and combined with metadata as part of the Kamai, R., N. A. Abrahamson, and W. J. Silva (2014). Nonlinear horizontal
Pacific Earthquake Engineering Research (PEER) Next Gen- site amplification for constraining the NGA-West2 GMPEs, Earthq.
eration Attenuation (NGA)-West2 project (http://peer. Spectra 30, 1223–1240.
berkeley.edu/ngawest2/; last accessed July 2015). Detailed Lin, P.-S., B. Chiou, N. Abrahamson, M. Walling, C.-T. Lee, and C.-T.
Cheng (2011). Repeatable source, site, and path effects on the standard
site characterizations for selected sites were provided by Alan
deviation for ground-motion prediction, Bull. Seismol. Soc. Am. 101,
Yong; the detailed reports are available at http://pubs.usgs. 2281–2295.
gov/of/2013/1102/ (last accessed July 2015). We used the nlme Petersen, M. D., M. P. Moschetti, P. M. Powers, C. S. Mueller, K. M. Haller,
package (Pinheiro et al., 2013) in R to compute the mixed- A. D. Frankel, Z. Yuehua, S. Rezaeian, S. C. Harmsen, O. S. Boyd,
effects regression. et al. (2014). Documentation for the 2014 update of the United States
national seismic hazard maps, U.S. Geol. Surv. Open-File Rept. 2014–
1091, 243 pp., doi: 10.3133/ofr20141091.
Acknowledgments Pinheiro, J., D. Bates, S. DebRoy, D. Sarkar, and Development Core Team
(2013). nlme: Linear and nonlinear mixed effects models, R package
We thank David Boore, Ronnie Kamai, Dino Bindi, and an anonymous v.3., 1–113.
reviewer for providing detailed reviews this article. Discussions with Albert Robinson, P. T., W. A. Elders, and L. J. P. Muffler (1976). Quaternary vol-
Kottke and Kenneth Campbell have also improved this article. We thank canism in the Salton Sea geothermal field, Imperial Valley, California,
Chris Wills for providing the preprint and data files for his latest V S30 map Geol. Soc. Am. Bull. 87, 347–360.
of California and Alan Yong for providing detailed profile information. Rodriguez-Marek, A., F. Cotton, N. A. Abrahamson, S. Akkar, L. Al Atik,
B. Edwards, G. A. Montalva, and H. M. Dawood (2013). A model for
References single-station standard deviation using data from various tectonic re-
gions, Bull. Seismol. Soc. Am. 103, 3149–3163.
Al Atik, L., N. Abrahamson, J. J. Bommer, F. Scherbaum, F. Cotton, and N. Rodriguez-Marek, A., G. A. Montalva, F. Cotton, and F. Bonilla (2011).
Kuehn (2010). The variability of ground-motion prediction models and Analysis of single-station standard deviation using the KiK-net data,
its components, Seismol. Res. Lett. 81, 794–801. Bull. Seismol. Soc. Am. 101, 1242–1258.
Allen, T. I., and D. J. Wald (2009). On the use of high-resolution topographic Seyhan, E., and J. P. Stewart (2014). Semi-empirical nonlinear site ampli-
data as a proxy for seismic site conditions (V S30 ), Bull. Seismol. Soc. fication from NGA-West 2 data and simulations, Earthq. Spectra 30,
Am. 99, 935–943. 1241–1256.
Ancheta, T. D., R. B. Darragh, J. P. Stewart, E. Seyhan, W. J. Silva, B. S. J. Seyhan, E., J. P. Stewart, T. D. Ancheta, R. B. Darragh, and R. W. Graves
Chiou, K. E. Wooddell, R. W. Graves, A. R. Kottke, D. M. Boore, et al. (2014). NGA-West 2 site database, Earthq. Spectra 30, 1007–1024.
(2014). NGA-West2 database, Earthq. Spectra 30, 989–1005. Stafford, P. J. (2014). Crossed and nested mixed-effects approaches for en-
Atkinson, G. M. (2006). Single-station sigma, Bull. Seismol. Soc. Am. 96, hanced model development and removal of the ergodic assumption in
446–455. empirical ground-motion models, Bull. Seismol. Soc. Am. 104, 702–719.
Baise, L. G., S. D. Glaser, and D. Dreger (2003). Site response at Treasure Tinsley, J, S. E. Hough, A. Yong, H. Kanamori, E. Yu, V. Appel, and C. Wills
and Yerba Buena Islands, California, J. Geotech. Geoenvir. Eng. 129, (2004). Geotechnical characterization of TriNet sites: A status report,
415–426. Seismol. Res. Lett. 75, 505–514.
Boore, D. M. (2004). Can site response be predicted? J. Earthq. Eng. 8, Thompson, E. M., and D. J. Wald (2014). Site response mapping with one
1–41. less proxy: Collaborative research with SDSU and the USGS, Final
Boore, D. M., and M. W. Asten (2008). Comparisons of shear-wave slow- Technical Rept. USGS Award Number G13AP00070.
ness in the Santa Clara Valley, California, using blind interpretations of Thompson, E. M., L. G. Baise, R. E. Kayen, and B. B. Guzina (2009). Imped-
data from invasive and non-invasive methods, Bull. Seismol. Soc. Am. iments to predicting site response: Seismic property estimation and mod-
98, 1983–2003. eling simplifications, Bull. Seismol. Soc. Am. 99, 2927–2949.
Boore, D. M., J. P. Stewart, E. Seyhan, and G. M. Atkinson (2014). NGA- Thompson, E. M., L. G. Baise, R. E. Kayen, E. C. Morgan, and J. Kakla-
West 2 equations for predicting PGA, PGV, and 5%-damped PSA for manos (2011). Multiscale site-response mapping: A case study of
shallow crustal earthquakes, Earthq. Spectra 30, 1057–1085. Parkfield, California, Bull. Seismol. Soc. Am. 101, 1081–1100.
Cadet, H., P.-Y. Bard, A.-M. Duval, and E. Bertand (2012). Site effect as- Thompson, E. M., L. G. Baise, R. E. Kayen, Y. Tanaka, and H. Tanaka
sessment using KiK-net data: Part 2–site amplification prediction (2010). A geostatistical approach to mapping site response spectral
equation based on f0 and Vsz, Bull. Earthq. Eng. 10, 451–489. amplifications, Eng. Geol. 114, 330–342.

BSSA Early Edition


Uncertainty in V S30 -Based Site Response 11

Thompson, E. M., D. J. Wald, and C. B. Worden (2014). A V S30 map for Yong, A., S. E. Hough, J. Iwahashi, and A. Braverman (2012). A terrain-
California with geologic and topographic constraints, Bull. Seismol. based site-conditions map of California with implications for the con-
Soc. Am. 104, 2313–2321. tiguous United States, Bull. Seismol. Soc. Am. 102, 114–128.
Wald, D. J., and T. I. Allen (2007). Topographic slope as a proxy for seismic Yong, A., A. Martin, K. Stokoe, and J. Diehl (2013). ARRA-funded V S30
site conditions and amplification, Bull. Seismol. Soc. Am. 97, no. 5, measurements using multi-technique approach at strong-motion sta-
1379–1395. tions in California and central-eastern United States, U.S. Geol. Surv.
Wald, D. J., V. Quitoriano, T. H. Heaton, H. Kanamori, C. W. Scrivner, and Open-File Rept. 2013–1102, 60 pp.
C. B. Worden (1999). “Trinet ShakeMaps”: Rapid generation of instru-
mental ground motion and intensity maps for earthquakes in southern
California, Earthq. Spectra 15, 537–556.
Walling, M., W. J. Silva, and N. A. Abrahamson (2008). Non-linear site U.S. Geological Survey
amplification factors for constraining the NGA models, Earthq. Spec- Denver Federal Center
tra 24, 243–255. P.O. Box 25046, Mail Stop 966
Wills, C. J., C. I. Gutierrez, F. G. Perez, and D. M. Branum (2015). A next- Lakewood, Colorado 80225
generation V S30 map for California based on geology and topography, emthompson@usgs.gov
Bull. Seismol. Soc. Am. 105, 3083–3091, doi: 10.1785/0120150105.
Worden, C. B., D. J. Wald, T. I. Allen, K. Lin, D. Garcia, and G. Cua (2010).
A revised ground-motion and intensity interpolation scheme for Manuscript received 15 January 2016;
ShakeMap, Bull. Seismol. Soc. Am. 100, no. 6, 3083–3096. Published Online 01 March 2016

BSSA Early Edition

Вам также может понравиться