Sie sind auf Seite 1von 17

Bulletin of the SeismologicalSociety of America, Vol. 85, No. 4, pp.

1127-1143, August 1995

A Comparison and Test of Various Site-Response Estimation Techniques, Including Three That Are Not Reference-Site Dependent
by E d w a r d H. Field and Klaus H. Jacob

We compare various site-response estimation techniques using aftershock data of the 1989 Loma Prieta earthquake collected in Oakland, California. Because of recent interest in comparing results from weak and strong motion (to infer any nonlinearity) and between direct S and coda waves, we pay particular attention to the uncertainties. First, sediment to bedrock spectral-ratio estimates between pairs of sites are compared with those obtained from various generalizedinversion approaches where the source and site effects of multiply recorded events are solved for simultaneously. We find that the site amplification factors are very similar among these approaches, but that the uncertainties can be significantly different depending on how the data are weighted. We also examine and test three site-response estimation techniques that do not rely on a reference site to estimate source and path effects. The first involves a parameterized source- and path-effects inversion. Even when the bedrock data are excluded from consideration, this approach is found to reveal the frequency dependence of site response at each of the sediment sites. The second technique involves taking horizontal- to vertical-component spectral ratios (receiver-function-type estimates) of shear-wave aftershock data. These are also found to reveal the frequency dependence of site response at the sediment sites, and the results for the bedrock site are relatively fiat and near unity. The third estimate is formed by taking horizontalto vertical-component ratios of ambient seismic noise, and these are shown to reveal the fundamental resonant frequency of the sediment sites. Unfortunately, discrepancies exist among all of the site-response estimates (and with one-dimensional predictions) with respect to a frequency-independent scaling factor. Nevertheless, the highly frequency-dependent character of site response is well constrained, and the fact that non-reference-site-dependent methods are capable of revealing this is promising for site-specific hazard assessments in regions that lack adequate reference sites.

Abstract

Introduction It has long been understood that earthquake ground motions can be significantly amplified by sedimentary deposits near the Earth's surface (e.g., Milne, 1898). The potentially severe consequences of this phenomenon were recently demonstrated in the damage patterns of the 1985 Michoacan, Mexico, earthquake (e.g., Singh et al., 1988), the 1988 Armenian earthquake (e.g., Borcherdt et al., 1989), the 1989 Loma Prieta earthquake (e.g., Hough et al., 1990; Borcherdt and Glassmoyer, 1992), and the recent Northridge earthquake in Los Angeles, California (EERI Preliminary Reconnaissance Report, 1994). Numerous other studies have also demonstrated the ability of surface geologic conditions to alter seismic motions (e.g., Borcherdt, 1970; King and Tucker, 1984; Aki, 1988; Field et al., 1992). Given the greatly increased level of damage that can be produced, it is of practical importance to develop methods for assessing the nature of, and potential for, sediment amplification, especially when choosing the location and design of critical and essential facilities. The greatest challenge in estimating site response from earthquake data is removing the source and path effects. The most common procedure, introduced by Borcherdt (1970), is to divide the spectrum observed at the site in question by that observed at a nearby reference site (preferably on competent bedrock). If the two sites have similar source and path effects, and if the reference site has a negligible site response, then the resulting spectral ratio constitutes an estimate of site response. Andrews (1986) recast the me,hod of spectral ratios into a generalized-inverse problem by solving the data of multiply recorded events for all source/path ef-

1127

1128 fects and site effects simultaneously. Others have since applied various forms of this generalized-inversion approach (Boatwright et aL, 1991a; Hartzell, 1992; Seekins and Boatwright, 1994), and the question remains as to whether it provides any advantage over simply averaging spectral ratios for each site separately (the traditional approach introduced by Borcherdt, 1970). The first goal of this article, therefore, is to compare site-response estimates obtained from traditional spectral ratios with those obtained from various generalized-inversion spectral-ratio approaches (also determined relative to a reference site). These techniques for estimating site response depend on the availability of an adequate reference site (with a negligible site response). Since such a site may not always be available, as in the New Madrid seismic zone, it is desirable to develop alternative methods that do not depend on a reference site. Toward that end, we also examine three recently developed non-reference-site-dependent techniques by comparing results with those obtained from sediment-to-bedrock spectral ratios. The first alternative method employs a generalized-inversion scheme introduced by Boatwright et al. (1991a), where shear-wave spectra are represented with a parameterized source- and path-effect model and a frequency-dependent site-response term for each site. The source parameters for each event, the path-effect parameters, and the response for each site are then solved for simultaneously in the least-squares sense. The second non-reference-site-dependent technique involves dividing the horizontal-component shear-wave spectra at each site by the vertical-component spectrum observed simultaneously at that site (Lermo et al., 1993). This procedure, which is analogous to the so-called receiver-function technique applied to studies of the upper mantle and crust from teleseismic records (e.g., Langston, 1979), assumes that the local site conditions are relatively transparent to the motion that appears on the vertical component. The last technique that we examine is that introduced by Nakamura (1989) for analyzing ambient seismic noise. He hypothesized that site response could be estimated by dividing horizontal-component noise spectra by vertical-component noise spectra. Several studies have since shown that Nakamura's procedure can be successful in identifying the fundamental resonant frequency of sedimentary deposits (Omachi et aL, 1991; Lermo et at., 1992; Field and Jacob, 1993b; Field et al., 1994a). We use aftershock data obtained in Oakland, California, following the 1989 Loma Prieta earthquake to test these siteresponse estimation techniques. Details regarding the deployment can be found in Hough et aL (1990) or Field et al. (1992). Figure la shows the location of the five sites occupied in this East Bay study: sites S1, $2, and $5 were located on Bay Mud deposits that cover Quaternary alluvium; site $3 was placed directly on the alluvium layer; and a reference site, $4, was placed on the Franciscan Complex which constitutes bedrock in this region. Each site was equipped with Mark Products L22-D geophones (three-component), which have a corner frequency at 2 Hz. Table 1 gives pertinent

E.H. Field and K. H. Jacob

EMERYVlL TREASURE IS.

~' ...: '~l ;, '?


:"~ll~l~llllJ

~?/~;',~,'.-,

Ill II
~ ' ~ 2 / . , ' ~ 2 ":'..'.::.:';-.':!'.:(:..i::.....'.:.'!::...
, ~ \l \ ~ -]

'- I . j I.I

,zkm
COLLAPSED UNCOLLAPSED EXPLANATION ~ B a y Mud (Holocene) I I I':""~. " . ~ : . 1 0 - 3 0 m thick I ~luvium (quaternary) I t" ~ ~ /10-180m thick I ~]~F~ ~ Complex (Mesozoic)l I I I I I I I III Basement l

III

WEAK MOTION SITES

38"1

122"40'

.~ _.~:.

122"
t

121"22'
I

"~...:.

EARTHQUAKES-e WEAK MOTION SITES-II

37*

%
36* 45' Figure 1. (a) Map of Oakland, California, region showing the surface geology and the locations of the five sites (S1, $2, $3, $4, and $5) occupied during the aftershockdeployment. (b) Regional map showing the epicenter locations (circles) and recording sites (squares).

Various Site-Response Estimation Techniques information on the 18 aftershocks used in this study, and Figure lb shows their locations relative to our recording sites. The magnitudes range between 2.5 and 4.4 and the epicentral distances range between ~75 and - 1 1 5 km. Given the low aftershock magnitudes and large epicentral distances, the site-response estimates obtained here are for weak motion. However, some evidence of nonlinear site response during strong ground motion has recently appeared (e.g., Chin and Aki, 1991; Darragh and Shakal, 1991; Beresnev et al., 1994). The extent to which nonlinear effects are significant remains a controversial issue. Nevertheless, weak-motion site response will remain a logical starting point for determining the response of sediments even if nonlinear effects are influential for stronger motions than those considered in this study. Because a successful identification of nonlinearity will require demonstrating that the strong-motion response is significantly different from the weak-motion response, special care must be taken in determining the natural variability and uncertainty of the estimates. The level of uncertainty will also determine how much effort is warranted in testing and distinguishing between various theoretical models of site response. Finally, the uncertainties also need to be understood before applying the results to assessments of seismic hazard. For these reasons, we devote considerable attention to the uncertainties obtained from the various site-response estimation techniques. In this study we have not examined the coda method of estimating site response, which was introduced by Phillips and Aki (1986) and has been applied in several site-response studies (Mayeda et al., 1991; S u e t al., 1992; Koyanagi et al., 1992). Unfortunately, the event-triggered windows in our East Bay study are not long enough to get an adequate

1129 window of coda. We do, however, discuss some implications that our results have with respect to the coda method. Representation of Ground Motion Suppose we have a network of I sites over which J events have been recorded (each site does not need to have recorded all J events). Then the amplitude spectrum of the jth event recorded at the ith site, Oij(f), can be written in the frequency domain as a product of a source term, Ej(f), a path term, Pu(f), and a site-effect term, Si(f):

Oij(f) "~ Hi(f) Pij(f) Si( ~"


Taking the natural logarithm, equation (1) becomes In Oo(f) = In Ej(f) + In Po(f) + In S,(f).

(1)

(2)

This linear expression often forms the basis of attempts to separate the source, path, and site effects. Spectral-Ratio Estimates In this section, we compare traditional spectral-ratio site-response estimates, introduced by Borcherdt (1970), with those obtained from various forms of the generalized inversion approach introduced by Andrews (1986). Motivated by recent interest in comparing site response between weak and strong motion, and between direct S and coda waves (e.g., Margheriti et aL, 1994; Kato et al., 1995), a rather detailed development of the uncertainty estimates is given. We find that the site amplification factors obtained from the generalized inversion approach are very similar to

Table 1 Event Information (-- Means the Event Was Not Recorded)
Julian Day 294 298 297 298 297 295 294 297 299 295 294 294 295 298 300 295 298 294 Hour: Minute 00:49 01:27 04:48 05:38 07:02 08:17 08:32 08:56 09:01 09:44 10:57 12:54 12:55 13:00 13:29 14:24 22:01 22:14 Latitude (N) 37.038 37.068 37.046 37.074 37.171 37.171 37.112 37.178 37.039 36.916 37.158 37.137 37.057 36.889 36.952 36.984 36.987 37.055 Longitude (W) 238.141 238.178 238.199 238.175 238.045 237.929 238.006 238.041 238.118 238.328 238.017 238.065 238.210 238.363 238.275 238.190 238.191 238.121 Depth (kin) 11.6 9.4 4.2 9.7 5.0 13.2 13.0 4.9 10.8 6.0 7.2 5.3 6.8 3.7 13.9 15.0 14.7 12.7 Mag. (m0) 4.1 4.2 3.2 2.8 2.8 2.6 2.7 2.5 3.6 3.3 2.6 3.1 2.2 3.7 3.0 3.7 3.7 4.4 Site S1 * * * * ----* * -* -* * * * * Site $2 * * * * * * * -* * * * * * * * * * Site $3 * * * -* * * * * * * * -* -* * * Site $4 * * * * * * * * * * * * * * * * * * Site $5 -* * * * * -* -* --* * -* -*

1130 the traditional spectral ratios. The uncertainty estimates, on the other hand, can be significantly different depending on the data weighting scheme used. If all data are weighted equally, then the uncertainties obtained from the generalized inversion approach can be up to a factor of --0.7 less than those of traditional spectral ratios. The data weighting scheme introduced by Andrews (1986), on the other hand, is shown to produce somewhat arbitrary uncertainties and we advocate an alternative approach. Finally, the expressions provided in this section can be used for deployment planning purposes; they indicate how much data need to be collected in order to obtain estimates of a given level of precision. Readers who are not interested in the details of the spectral-ratio technique can easily skip this section, to the discussion of the non-reference-site-dependent techniques, without a loss of continuity. Traditional Spectral Ratios Suppose we have a reference site (i = R) that we assume to have a negligible site response (In SR = 0). If the interstation spacing is small compared to the epicentral distances, so that Po ~- PR; then the site response at each site can be estimated from

E.H. Field and K. H. Jacob causes random differences between the true input and the reference-site observation), then one would want to minimize this source of uncertainty by using the standard deviation of the mean (equation 5). On the other hand, if the scatter in spectral ratios is produced by an intrinsic variability in the site response itself (by virtue of different incidence angles, backazimuths, and perhaps wave types from the different sources) then one would want to account for this by applying the standard deviation (equation 4) to the uncertainty estimates. Furthermore, to convincingly identify a nonlinear site response from spectral ratios, it will be necessary to show that the response for the event in question is outside the range of scatter represented by the standard deviation. The above analysis implicitly assumes that the level of noise in each observation, say Cro0,is the same for each event (aou = ao,). If we further assume that the level of noise is the same at both site i and the reference site (ao, = aoR = ao), and that the two are uncorrelated, then we have
std ~ R Si

= E~, + ~,]o.~

f i ~o,

and equation (5) can be rewritten as

In SfR =

.= In

ORj

~-"

-~ j=l ~ (In O o -- In OR1), (3)

ag~ = ~ a o .

(6)

where J is the number of events recorded at both site i and the reference site (which may vary between sites), and we have dropped the frequency dependence for notational simplicity. Equation (3) constitutes the geometric-average spectral ratio. Using this estimate assumes that the observations are consistent with a lognormal distribution, which we explicitly test later in this article. If the reference site has a non-negligible site response, then the spectral ratios become relative site-response estimates. The standard deviation of In SsR, which represents the scatter of individual spectral ratios, is estimated from asR = Si [(In 0 o - In O~j) - In SsR]2 (4)

std

The standard deviation of the mean (or standard error), which represents the uncertainty with respect to the true mean, is given as
1

a Si SR = -

,/J

]-~ std a ~ Si "

(5)

Note that staas, R and asf are for In Si rather than for Si itself. The distinction between the standard deviation and the. standard deviation of the mean is important for hazard analyses. If the scatter observed among the individual spectral ratios results from some kind of noise in our observations (e.g., due to scattering from lateral heterogeneities, which

This gives the uncertainty of traditional spectral ratios in terms of the noise in the data and the number of ratios averaged. We will later compare this with the uncertainty of generalized inversion spectral ratios. Plots of traditional spectral ratios for the transverse component at two of our East Bay sites (S1 and $3 relative to $4) are shown in Figure 2a. These observations, as well as those for the radial component and the other two sites ($2 and $5), have been shown and discussed previously (Field et al., 1992, 1994b). These estimates, and all others shown in this article, were computed from 20-sec windows beginning at the shear-wave arrival. Only negligible differences were obtained by using 10-sec windows. The spectral estimates were computed using the high-resolution method of Thomson (1982) and Park et al. (1987), which produces an equivalent smoothing width of 0.4 Hz. The dotted lines in Figure 2a, which represent the ___ 2~tdasR limits, define the range where an additional spectral ratio would have 95% likelihood of residing. These reveal a large degree of variability among the individual spectral ratios. The dashed lines, which represent the 95% confidence limits of the mean (___2asR), demonstrate that the site response can be fairly well constrained when several recordings are available. The site response at $3 is dominated by three prominent peaks, which represent the first three resonant modes of the alluvium layer. The response at S 1 is characterized by four prominent peaks, which represent coupled resonances between the mud and alluvium layers at this site. Both of these

Various Site-Response Estimation Techniques

1131

a) traditional spectral ratios

4o41 :. o ,

t
q':' ~i,~ '! ,,!.~

".

." /

;:::

.',

" /"""

~...... :::~ :.i-/:':. ~.:, i~' ,'. : :Y',::: "'..~:,::::'.~i'~


'
',:::',,," ' , ,~,~

-"<L-

-- .

,'

2.0
1.0

-" .
" ...... ".'"'

-...

:.:.:.
":: .......

, ,;,.,
- :. ~ "~.!~,!

' ~

"':" ;":

0.4 0.2

~','~: ,9 i. "l '"~" '/

b) Andrews' weight scheme estimates


"~ ;~ 4020 10 4.0 2.0 1.0 0.4 0.2

S1

$3

40-1

20 ~

c) unit weight scheme estimates " ": S1 -

i:~ (::i'}

$3 ~,
"..
" .

~o--[
4.0 - J ~ - . 2.0 -'[ 0.4 "J 0.2 I
~

,~-. ~ : ~ i
" .

~ :
..";~.'..'.:.~.. ...... :,

"...

." t / ~ t/

".....'t~".:'"" i ""

,':'":~ :':..:~":U ] " .:- t -;;2

~EN.,~:.i:',]:"!~:,,]
~

~ ' : ' ~ 57~;! , ' " " :,

;-.,:

d) preferred weight scheme estimates

4.0

0.4 0.2 0.2

. 0.4 1.0 2.0 4.0

- ,,...,,, , 10 20 0.2 0.4 1.0 2.0 4.0 , 10 I 20

frequency (Hz)

frequency (Hz)

Figure 2. Spectral-ratio estimates of site response for the transverse components at S 1 and $3, relative to the reference site $4. The dotted and dashed lines represent 2 standard deviation and standard deviation of the mean limits, respectively. (a) Traditional spectral-ratio estimates. (b) Generalized-inversion spectral ratios using Andrews' data weighting scheme. (this does not have 2 standard deviation limits because the data uncertainties are specified explicitly rather than estimated from the data) (c) Generalized-inversion spectral ratios where all data have been given unit weight. (d) Generalized-inversion spectral ratios obtained using our preferred weighting scheme.

1132 observed response estimates were found to agree well with the frequency-dependent character of one-dimensional model predictions based on independent geotechnical data (Field et aL, 1992). However, the spectral-ratio amplitudes are greater than those predicted by a frequency-independent scaling factor of about 2. Scattering of energy from lateral boundaries of the sediments may explain this discrepancy. However, there is no obvious evidence of such wave propagation in the time series observed at the sediment sites (Field, 1994). That is, except for the factor of 2 discrepancy, one-dimensional predictions reproduce the observed waveforms and do not lack any later-arriving energy that might be attributed to lateral propagation within the sediments. Generalized-Inversion Spectral Ratios The method of spectral ratios was recast by Andrews (1986) into a generalized-inverse (GI) problem by solving equation (2) for all the site-effect and source terms simultaneously. A path effect representing geometric attenuation was specified as

E.H. Field and K. H. Jacob determined degree of freedom, which means that all the siteresponse terms can be multiplied by some arbitrary function of frequency, provided all the source terms are divided by the same function. This trade-off will not influence the value of Z2 in equation (8). Andrews (1986) constrained this unresolved degree of freedom by determining each site-response term relative to the network average. Bonamassa and Mueller (1989), and others since (e.g., Boatwright et al., 1991a; Hartzell, 1992), removed the degree of freedom by constraining the amplification at a reference site to be zero (In SR = 0). This latter approach, which we shall adopt here, causes the resulting site-response estimates to be relative to that at the reference site. Since this is analogous to the traditional spectral-ratio estimates, we call these generalizedinversion (GI) spectral ratios. The model-parameter uncertainties are obtained as the square root of the diagonal elements in the model covariance matrix. If all the observations have an equal level of noise (aok = ao), then the model covariance matrix is given as [cov m] = [GrG] - ' a g, (9)

Pij = In

= - l n (ru),

where rij is the known hypocentral distance between the flh event and the ith site. With this applied as a correction to the data, the system of equations can be rewritten as lnEj + ln& = lnOk + lnrk, where k represents the kth observation, and i and j are now implicit functions of k. Following the notation of Menke (1989), this can be written in matrix form as

where [GrG]-~ is often referred to as the unit covariance matrix (Menke, 1989). This equation shows that the modelparameter uncertainties are directly related to the uncertainties in the data. One might wonder whether the generalized-inversion approach produces less uncertainty than the traditional spectral-ratio estimates. For simplicity, assume that all the observations have the same level of noise (aok = ao). Then, if all J events are recorded at all I sites, the source- and siteeffect uncertainties are

o'GI Si ~ ~

0
-1 a'

(lO)

Gm = d,

(7)

where d is a data vector containing the geometric-attenuation-corrected observations, m is a model vector containing the J + I unknown source and site terms, and G is the data kernel matrix, which has J + I columns and K rows (K can be up to J . I to the extent that all events are recorded at all sites). G is a sparse matrix, having only two nonzero elements per row or column. The weighted least-squares solution is obtained as that which minimizes the difference between the observed and predicted data

a~l =

~I+I
JI

(] 1)

Z2 = ~7~ ['(ln Ok + ln rk) -- (ln Ej + ln Si)] 2,


k= 1 O'o/

(8)

where aok is the noise or uncertainty of the kth observation

(In O~).
To the extent that K > J + I, the system of equations appears overdetermined. However, there is actually one un-

respectively, determined by examining the unit model covariance matrix for a wide range of J/I combinations. Comparing equations (10) and (6), we see that the site-response uncertainties for the GI approach are identical to those for the traditional spectral ratios. One also finds that the siteresponse estimates themselves are equivalent, provided the geometric-attenuation correction is either negligible, or has been applied to the traditional spectral ratios as well. The main advantage of using the GI approach arises when not all events are recorded at all sites. In this case, the GI estimates are more efficient (have less uncertainty) than those obtained via traditional spectral ratios. For example, consider the case where the reference site has recorded every event. The situation where the GI approach will be most beneficial will occur when all J events are recorded at all

Various Site-Response Estimation Techniques

1133 where N~ is the amplitude spectrum of a sample taken immediately before the arrival of the wave of interest. This approach, which has subsequently been applied by others (Boatwright et al., 1991a; Seekins and Boatwright, 1994; Margheriti et al., 1994), is designed to down-weight records that have a poor observed signal-to-noise ratio, while at the same time requiring all data to have a minimum uncertainty of 0.5. Consider the case of a clean data set in that N J O k < 0.5 for all k. Then all the data will be weighted equally, which would seem appropriate given a lack of a p r i o r i information. The data uncertainty, on the other hand, has been specified as a o = 0.5 which, as Andrews acknowledges, is an arbitrary value. Since the resulting site-and source-effect uncertainties depend directly on this choice (equations 10 and 11), they will be somewhat arbitrary as well. The observed signal Ok is presumably composed of the true signal, say Rk, and the noise for which we have a sample, Ark. If the true signal and noise are uncorrelated, then the observed signal is expected to be Ok = [R~ + /V~k] -5, and the noise-to-observed-signal ratio will be

sites except one (say the i = 1 site, S~), which records only one event (say the j = 1 event, E0. From equation (6) we know that the uncertainty of In S sR (the traditional spectral ratio for site 1) will be a sR = ~ a o. By examining the unit covariance matrix over a wide range of J/1, we found that the uncertainty for In S~x (the generalized inversion estimate for site 1) is given by ~GI ~
s~

a2o +

"

)--~

-_

(12)

This expression can be understood intuitively from the following. Since the site response at site 1 is obtained as In $1 = In O11 - In El, the term in brackets can be understood as a2o + a z e~, where ae~ is obtained from equation (11) with I - 1 substituted for I (since only I - 1 sites recorded all of the events). With equation (12), and with ~R = ~/2 ao, the relative uncertainties between the two approaches for In S~ can be expressed as
o' '

[~Sl
asf -

1 +
J(I--

1) /J

"

(13)
Ok / R2 + N~kJ k~'*:tV-~-+ 1 __--< 1. Therefore, Andrews' data weighting scheme implicitly specifies that
1

Again, this expresses the relative uncertainties in the case where one site has recorded only one event, but all other (I - 1) sites have record all J events. For a given I and J, this represents the maximum benefit the GI approach can have over the traditional spectral ratios. For the more realistic case where each site records some number of events, the relative uncertainties will be somewhere between unity and that given by equation (13). For example, in our East Bay study we recorded events over five sites so the GI approach uncertainties will be somewhere in the range 0.8asf < Ors, z =< aSl R. Furthermore, equation (13) implies that for arbitrary I and J the maximum possible advantage (for the case of an infinite number of events and sites) is O~s~ 1 = a s f / ] 2 ~0.7asSa R. Therefore, in general, 0.7 ~R < Ofs[=< ~ f , (14)

~aokN

1.

(16)

which shows that the generalized inversion approach to siteresponse estimation offers a slight, but perhaps insignificant, improvement over traditional spectral-ratio estimates. The improvement can be greater if the reference site has not recorded every event. An estimate of the observational uncertainties (ao) must be available to apply the weighted least-squares inversion. In general, one would expect the data uncertainty to vary with the distribution of sources and receivers, with the type of spectral-estimation technique (i.e., smoothing) applied, and with frequency. Andrews (1986) applied the following:

Since a recording that is composed of complete noise will be assigned an uncertainty of 1, which is just twice that assigned to noise-free data, it will be important that such data are manually removed before applying this technique. Figure 2b shows the results obtained for two of our sites (S1 and $3) when the generalized inversion approach is applied with Andrews' weighting scheme to the entire East Bay data set. The resulting site-response estimates are very similar to those obtained from the traditional spectral ratios (Fig. 2a). However, the uncertainties are larger and therefore more conservative. An alternative approach to that introduced by Andrews is to give each observation a unit weight and subsequently estimate the observational uncertainty directly from the data
as

ao =

K-

(J+ 1-

z2

il 5

(17)

aok = max \Oh

'

(Press et al., 1988), where Z2 is as given in equation (8) and [K - (J + I - 1)] represents the number of observations minus the degrees of freedom lost. The model-parameter uncertainties can then be obtained as the square-root of the diagonal elements in the model covariance matrix given in equation (9). This approach is analogous to the uncertainty

1134 estimates of the traditional spectral ratios, except here the data from all sites are examined simultaneously. Figure 2c shows the results obtained using this scheme which, as expected, are very similar to those obtained using the traditional spectral-ratio approach. Finally, we present a GI weighting scheme that represents a hybrid between the previous two. In general, an observed signal will contain two types of noise: ambient seismic noise completely unrelated to the earthquake; and signal-generated noise [such as that produced by scattering that causes the path effects to vary between sites (Field et al., 1992)]. The presence of signal-generated noise was the motivation for Andrews to specify a minimum uncertainty. Assigning equal weights to the data is consistent with assuming the signal-generated noise is approximately equal between sites. Ambient seismic noise will presumably corrupt the source- and site-effect estimates, so its influence should be minimized. From equations (15) and (16) we see that this is achieved to a limited degree by the approach of Andrews. For the unit-weight case, however, the site-response estimates will represent relative noise levels at frequencies where the ambient noise dominates. Therefore, we introduce our preferred weighting scheme as that where data are simply removed from consideration if

E.H. Field and K. H. Jacob situation where the traditional spectral-ratio estimates might be preferable is when there are significantly different levels of signal-generated noise between sites (which could occur if the distance to the reference site varies among the stations). Another is where the response at some sites is intrinsically more variable than it is at others. In both of these situations, the spectral ratios will successfully identify these peculiarities by producing different estimates of o-o among the different sites. However, such differences will be smeared out in the GI approach where all data are lumped into one estimate of ao. Since this effect was found to be minor in our study, the GI estimates obtained by applying our preferred weighting scheme (Fig. 2d) represent our preferred spectral-ratio estimates of site response in the East Bay.

Parameterized Source- and Path-Effects Inversion Estimates In the previous section, the source and path effects were estimated from observations obtained at a reference site. In practice, however, there may not always be an adequate reference site available. In this section, we examine site-response estimates obtained from a parameterized source- and path-effects inversion. By excluding the bedrock-site observations in this analysis, we provide a consistency test between these non-reference-site-dependent estimates and the spectral-ratio estimates of site response. Several parameterized source- and path-effect inversion schemes have been presented (e.g., Masuda and Suzuki, 1982; Scherbaum, 1990; Frankel et al., 1990; Boatwright et al., 1991b; Fletcher and Boatwright, 1991; Anderson and Humphrey, 1991; Lindley and Archuleta, 1992), each involving various assumptions regarding the source-spectral shape, the frequency dependence of attenuation, and the distance dependence of geometric attenuation. The previous studies that included a site-response term did so primarily to remove this effect in order to enhance the source- and/or path-effect estimates. Here we reverse this emphasis because we are most interested in the site-response estimates themselves and less so in the source and path parameters. Our inversion scheme follows closely that of Boatwright et al. (1991b), which we only outline here since they give a thorough discussion. For particle velocity observations, the source term in equation (1) is parameterized as

Ok

1 3

(18)

(or any other reasonable value). All of the remaining data are given unit weight, and the uncertainties are subsequently estimated from equations (8) and (17) as before for the unitweight case. In our application, N, is obtained from the ambient noise observed before the P-wave arrival. This approach effectively minimizes the influence of ambient seismic noise while providing nonarbitrary error estimates. The results obtained by applying this strategy are shown in Figure 2d. The site-response estimates are again similar to those obtained from the other procedures. The confidence limits, on the other hand, are somewhat different. Below --0.5 Hz, where ambient noise is problematic, the uncertainties obtained by our preferred scheme are appropriately larger than those produced by the other techniques. Therefore, it can be concluded for this study that Andrews' weighting scheme yields relatively conservative estimates of site response. The exception is at frequencies where ambient noise dominates, in which case Andrews' scheme underestimates the uncertainties. These conclusions are specific for our network configuration and spectral analysis technique; different conclusions may result from other studies. The comparison in Figure 2 also shows that, as expected, the GI approach offers a negligible improvement over traditional spectral ratios. The exception is the improvement produced at low frequencies by removing bad data in our final weighting scheme. However, this could have easily been applied to the traditional spectral ratios as well. One

2 rcfU: Ej(U~ - ,/1 + ( y f j

(19)

where U is the low-frequency displacement spectral level and fc is the corner frequency of the source. Equation (19) is Boatwright's (1978) approximation for the Brune (1970) far-field shear-wave spectrum. The path effect in equation (1) is modeled as

Various Site-Response Estimation Techniques


1

1135 fc and attenuation values need not be obtained provided we get an appropriate pair of values. To the extent that the corner frequencies span a range of values, the trade-off will be somewhat resolved if the inversion is performed over the entire data set simultaneously. A trade-off also exists between anelastic and geometric attenuation (Frankel et aL, 1990). However, as demonstrated by Kato et al. (1995), this has little influence on the site-response estimates. Because the number of model parameters is large (2J + 2 + IN), the inversion is divided into two parts. To begin with, all of the site-response terms are set to zero (In Si(f,) = 0 for all i and n). Then, the following two steps are repeated until the solution converges in the least-squares sense: 1. After correcting the data at each site with the current siteresponse estimate for that site (In O5 = lnOk - In $3, the following equation is solved iteratively for the 2 . / + 2 source and path parameters: In g ( f . ) + In Pk(f") = In O~(f.), where k, as before, represents the kth observation and i and j are implicit functions of k. 2. New site-response estimates are obtained, at each frequency independently, from the average residual between the observations and the previously determined source and path effects at each site:
K

P~j(f) = -- e- =a"+r,je),

r,j

(20)

where l/r Uis a geometrical spreading factor, TUis the shearwave travel time, Q is a frequency-independent regional quality factor, and t* is a near-receiver attenuation term. With this parameterization, the problem is reduced to solving the set of observations for 2J source parameters (one U and fc for each event), two attenuation parameters (Q and t*), and the I frequency-dependent site-response terms [S~(f)]. If the observed spectra are given at N discrete frequencies,f,, then there are 2J + 2 + INunknowns to solve for. Substituting equations (19) and (20) into equation (2), it is found that the dependence on fc is nonlinear. The nonlinearity is removed by approximation with the first two terms of its Taylor series expansion. Starting from assumed initial values, the changes in source corner frequencies needed to minimize the difference between actual and predicted observations are then solved for (in the least-squares sense). This process is repeated until additional iterations do not significantly improve the solution. A well-known problem with this source- and path-effects inversion is the trade-off between the corner frequency (fc) and attenuation (t* or 1/Q). That is, an observed spectrum is often equally well fit by a lowfc and t* (or I/Q), or by a highfc and t* (or I/Q). This trade-off, which is exemplified in Figure 3, is especially exacerbated by band limitations imposed by seismic noise and the instrument response. With respect to site-response estimation, however, the trade-off is somewhat beneficial. It means that the true

In S;(f,) = ~ [In Ok(f.) - (In Ej(f,) - In G(f"))],


k=l

-2-

where k and K are implicit functions of i (representing the kth observation and the total number of events recorded at site i, respectively). Our implementation differs from that of Boatwright et al. (1991b) in terms of assigning data weights and determining the site-response uncertainties. Whereas they applied the scheme of Andrews (1986) discussed previously (equation 15), we simply reject all data that have observed signalto-noise ratios less than 3 (the condition in equation 18), and give the remaining data equal weight. The site-response uncertainties are then estimated as they were for the traditional spectral ratios using equations (4) and (5), except where the reference-site observations in those equations (ln OR) are replaced with the final source- and path-effect terms determined here (ln Ej + In Pu)- While this approach may not be strictly valid in that we have not accounted for the uncertainties in the estimated source and path parameters, it should nevertheless provide rough uncertainty estimates. Furthermore, it does not involve the degree of arbitrariness discussed earlier for the Andrews (1986) weighting scheme. In the generalized-inverse spectral-ratio section we found that there was an unconstrained degree of freedom, which meant we could only determine the site response rel-

-3

.,.=q

-5
0.5

i 1.0

2.5

5.0

1 10.0

25.0

frequency (Hz)
Figure 3. An example of the trade-off between corner frequency and attenuation (t*). The solid line is the spectrum of a magnitude 4.2 event that occurred at 01:27 on Julian day 298 (Table 1). The dotted and dashed lines represent fits to the data (using equations 19 and 20 with Q-1 = 0) from two widely different sets of parameters. Each of these fit the data equally well.

1136 ative to a reference site. By specifying the frequency dependence of the source and path effects in this section, the unconstrained function of frequency is reduced to a frequency-independent scaling factor (Boatwright et al., 1991b). That is, we can still multiply all low-frequency source spectrum levels (Uj) by some amount, as long as we divide all of the site-response terms by the same frequencyindependent value. In the formulation presented here, we have effectively specified that the site amplification averaged over all frequencies and sites be unity. As described below, this scaling factor can conceivably be constrained from a priori information. Apart from the unresolved scaling factor, the parameterized inversion should produce good site-response estimates provided that, on average, equations (19) and (20) are adequate representations of the source and path effects, respectively, and that the global minimum is obtained in the two-step inversion. One might be concerned, however, that the global minimum will not be obtained if strong site resonances are present (as in our East Bay observations). In this situation, the resonant frequencies might strongly influence the final corner-frequency estimates, which will in turn influence the site-response estimates. This potential problem can be investigated by testing the method against other more established methods. The results obtained from the parameterized inversion are presented in Figure 4b for sites S 1 and $3 (analogous results were obtained for the other two sites, $2 and $5). The reference site ($4) data were excluded from the inversion, and only frequencies between 0.4 and 30 Hz were considered in the analysis. A maximum of three iterations over the comer frequencies were allowed in step 1 before step 2 was performed. Steps 1 and 2 were repeated until the variance reduction was less than 1% (this took 10 cycles for the results shown here). The final solutions are independent of the initial corner frequencies (set at 5 Hz for the results shown here). We also found, as did Boatwright et aL (1991b), that since our data do not span a sufficient range of distances, we could not resolve differences between the regional and local attenuation terms in equation (20). The observations are equally well fit by a t* of 0.086 sec or a Q of 370. The final comer frequencies range between 3.0 and 30 Hz. However, because of the trade-offs mentioned previously, we have low confidence in the strict reliability of these values. Fortunately, the site-response estimates are not strongly influenced by these trade-offs. For comparison, our preferred GI spectral-ratio estimates of site response are replotted (from Fig. 2d) in Figure 4a. Apart from the unconstrained frequency-independent scaling factor, the site-response estimates obtained from the parameterized inversion agree surprisingly well with the spectral-ratio estimates. Incidentally, we also found an equally good comparison if the parameterized inversion is applied to each site separately, Since the spectral-ratio estimates are relative to the reference site, this agreement implies that the site response at the reference site is negligible.

E.H. Field and K. H. Jacob Figure 5a shows the average ratio of the reference-site ($4) observations to the source and path effects predicted from the parameterized inversion (where the reference-site data were excluded). That the result is essentially flat up to 10 Hz implies, simultaneously, that the reference site has a negligible site response and that the parameterized source- and path-effect models are adequate over that frequency range (unless, of course, these are both wrong in an opposite manner). Interestingly, the peak near 16 Hz may represent a reference-site resonance. As stated previously, the parameterized inversion siteresponse estimates are determined up to a frequency-independent scaling factor, which can conceivably be constrained using any available a priori information. For example, Boatwright et aL (1991b) applied geotechnical information and a one-dimensional approximation to determine the absolute site-response level over particular frequency bands. It has been found, however, that uncertainties in geotechnical information can translate into large uncertainties in site-response predictions (Seed et al., 1988; Field and Jacob, 1993a). Furthermore, as mentioned previously for our East Bay observations, one-dimensional predictions can sometimes predict the frequency-dependent character of site response while exhibiting a frequency-independent discrepancy in the absolute amplitude (Mueller, 1986; Field et aL, 1992; Graves, 1992). Therefore, using geotechnical information to constrain the scaling factor will likely introduce uncertainty and perhaps some bias as well. In principle, the site response should approach unity at low frequencies where the wavelengths are too large to be influenced by the sediments. Therefore, the scaling factor could easily be constrained if the observations are reliable at these low frequencies. Unfortunately, this is not the case in our East Bay study because of seismic noise contamination and the fact that L-22 sensors are known to be unreliable below - 0 . 5 Hz (Reidesel et aL, 1990). A final way in which the scaling factor might be constrained is by using any independent seismic-moment estimates available for the events [see Boatwright et aL (1991 b) for the relationship between the low-frequency spectral level (U) and the seismic moment (M0)]. However, any constraint based on this method will presumably have a considerable degree of uncertainty as well. The important point here is that, although constraining the scaling factor remains problematic, the parameterized inversion successfully identifies the frequency dependence of site response at our East Bay sites. R e c e i v e r - F u n c t i o n - T y p e Estimates Lermo et aL (1993 ) recently presented evidence that site response can be estimated by taking horizontal- to verticalcomponent ratios of the shear-wave spectra (rather than dividing by reference site spectra). This technique was originally introduced by Nakamura (1989) to analyze ambient seismic noise recordings (discussed in the next section). However, Lermo et al. (1993) applied the method to earthquake data ob-

Various Site-Response Estimation Techniques

1137

a) preferred spectral-ratio estimates

,
~t~

20 10l ~l
Il

&, ,~,~,
tz
I"

$1

S3 d

4.02.01.00.40.2

b) parameterized inversion estimates


20-

;=~

4.0, , ~,, ,-,r 2 . 0 - " ~ , : , , , , -,


/ ~ dlI ,

t
!
I

%t

"'

0.40.2

"~',~,4, / ~ , ~'i'~'~,,i ' ~ ,


I I I I I I

c) receiver-function type estimates


2010I I~

S1
tl
^

$3

~1~.~ 4 . 0 2.01.0~ 0.40.2

d) Nakamura's estimates
2010 ,~ 4.0
z "~ xx

SI

$3

;-~ 2.0
1.0 0.4 0.2 0.4

0.8

2.0

4.0

8.0

20

0.4

0.8

2.0

4.0

8.0

20

frequency (Hz)

frequency (Hz)

Figure 4. Comparisonof the site-response estimates obtained from the non-reference-site-dependent techniques (for S1 and $3). The dashed lines represent 95% confidence limits of the mean. (a) The preferred spectral-ratio estimates (reproduced from
Fig. 2d). (b) The site-response estimates obtained from the parameterized inversion

(where all data except those for the reference site were included). (c) Receiver-functiontype estimates (average horizontal- to vertical-component spectral ratios of aftershock shear waves). (d) Nakamura's estimate (average horizontal- to vertical-component spectral ratios of ambient noise). These noise data were collected using Kinemetrics (SH1 and SV-1) 5-sec velocity sensors.

1138

E.H. Field and K. H. Jacob

tained at various sediment sites in Mexico and found that the frequency and amplitude of the fundamental resonant peak could be identified (higher mode peaks were not visible). They also presented some theoretical reasoning as to why the method might work for shear-wave observations. However, we have had difficulty in reproducing their modeling results.

a) parameterized inversion
2010 4.0 2.0 1.0 0.4
0.2
I I I I I

$4

o~,=1

b) horizontal/vertical
20 10

$4
t~ J J

4.0
2.0 --

1.00.40.2
I I t I I

c) Nakamura's estimate 20 -I $4
4.0 2.0 -i ~:~ 1.0

' 1
. . . . .
i

." "':'""

.,,,,,, ,,., ,,,, , ~

0.4~ .
0.2| 0.4

'

~
, I ,

'" ,

0.8

2.0

4.0

8.0

20

frequency (Hz)
Figure 5. The average ratio of reference-site ($4) observations to the source and path effects determined from the parameterized inversion (where the reference-site data were excluded). (b) Receiver-functiontype estimate (average horizontal- to vertical-component spectral ratio of aftershock shear waves) for the reference site ($4). (c) Nakamura's estimate (average horizontal- to vertical-component spectral ratios of ambient noise) observed at the reference site ($4). Pre-event noise samples from the aftershock data were used for this estimate because noise data from the Kinemetrics sensors were not collected at this site.

Perhaps the simplest theoretical justification of the method is by analogy with receiver-function studies. Langston (1979) applied a method of determining the velocity structure of the crust and upper mantle from teleseismically recorded P waves. The vertical component is assumed to be relatively uninfluenced by the local structure, whereas the radial component contains P- to S-wave conversions from structural discontinuities below the site. Therefore, by deconvolving the vertical component from the radial, an estimate of the impulse response function, or receiver function, below the site is obtained. Many receiverfunction studies have since been applied to study the Earth's crust and upper mantle (e.g., Owens et al., 1984; Owens and Crosson, 1988). By analogy, one might expect the method to apply to smaller-scale features of sedimentary layers. Here, the vertical component of shear waves could be composed primarily of S- to P-wave conversions, as has been observed in borehole studies (Takahashi et al., 1992), which may be relatively uninfluenced by the sediments (at least, perhaps, over the lower frequencies where the most prominent shear-wave resonances occur). If this were true, then an estimate of the site response could be obtained by deconvolving, or dividing in the frequency domain, the vertical component from the horizontal component. Plotted in Figure 4c are the average horizontal- to verticai-component shear-wave ratios computed for sites S 1 and $3. These estimates clearly reproduce the four resonant peaks at S1 and the three resonant peaks at $3 observed in the preferred spectral-ratio estimates (Fig. 4a). This is in contrast to the study by Lermo et al. (1993) where only the fundamental resonant frequency was revealed. The horizontal- to vertical-component ratios in Figure 4c, however, underpredict the spectral-ratio estimates by a frequency-independent factor of --0.6. This is also in contrast to the study by Lermo et aL (1993) who found good amplitude agreement (at the fundamental resonant peak). Furthermore, Lermo et aL (1993) found less variability among the horizontal- to vertical-component ratios, compared to traditional spectral ratios, whereas the uncertainties in Figure 4c are similar to those in Figure 4a. As an additional test of the method, in Figure 5b we present the average horizontal- to vertical-component ratios computed for our reference site $4. The result is relatively flat over the entire frequency range, with a slight peak near 8 Hz. Conclusions similar to those made for S 1 and $3 can be made for the other two sediment sites ($2 and $5). We also computed horizontal- to vertical-component ratios for P-wave windows, but found little agreement with those based on the shear waves. Therefore, we conclude that horizontal- to vertical-component ratios of shear-wave spectra, but not P-wave spectra, reveal the overall frequency dependence of site response at our East Bay sites. Nakamura's Estimate As introduced by Kanai (1957), an economically attractive means of estimating site response involves using am-

Various Site-Response Estimation Techniques

1139 which is relatively fiat and near unity over the entire frequency band. These observations contribute to a growing body of evidence that Nakamura's method is effective in identifying the fundamental resonant frequency of sedimentary deposits. Discussion and Conclusions We have found that when similar data weighting schemes are applied, generalized-inversion spectral ratios are very similar to traditional spectral-ratio estimates of site response. The exception occurs when some stations have recorded significantly more events than others. In this case, the generalized-inversion estimates will have less uncertainty than traditional spectral ratios, but by no more than a factor of 0.7 if all or nearly all of the events were recorded at the reference site (which will usually be the case). This improvement is probably insignificant since the uncertainties are approximate anyway. Traditional spectral ratios may sometimes be more advantageous in that they are capable of revealing any differing levels of signal-generated noise or intrinsic site-response variability among the stations. The generalized-inversion estimates, on the other hand, lump the data from all sites together into one shared estimate of the data uncertainty and thereby mask any differences among the sites. We found this effect to be minor, however, in the data presented here. We have also found that the choice of data weighting scheme can significantly influence the uncertainty estimates of generalized-inversion spectral ratios. Obtaining reliable uncertainties is important with respect to identifying a nonlinear response, matching observations with the results of theoretical modeling, or when incorporating the results into assessments of seismic hazard. We advocate estimating the observational uncertainty directly from the data after eliminating all data for which the observed signal-to-ambient seismic noise ratio is less than a factor of 3 (or so). This approach will reduce the effects of ambient seismic noise while at the same time avoiding the degree of arbitrariness present in the scheme introduced by Andrews (1986). Equations (6) and (14) give the uncertainties of spectralratio site-response estimates in terms of the number of events recorded and the data uncertainty (ao). These expressions can therefore be used for deployment planning purposes. If one has an idea of the noise level in the data, then equations (6) and (14) indicate how many events need to be recorded in order to obtain site-response estimates of a specified level of precision. In our preferred generalized-inversion spectral ratios, we found that go varies from - 0 . 2 to ~0.5 approximately linearly between 0.5 and 20 Hz [which can be compared with the range of 0.5 to - 1.0 for all frequencies specified explicitly by the weighting scheme of Andrews (1986)]. In general, the value of o o will vary among studies depending on the amount of smoothing applied to the data and on the distribution of stations relative to the reference site. Nevertheless, if we take ao = 0.3 as a typical value, and specify

bient seismic noise to evaluate sediment amplification. Many publications have appeared describing various successes (e.g., Ohta et al., 1978; Kagami et al., 1982, 1986; Celebi et al., 1987; Field et al., 1990) and failures (e.g., Udwadia and Trifunac, 1973; Borcherdt and Gibbs, 1976) of the method. The main assumption is that the response of sediments to noise sources is in some way related to that for incident seismic waves. This assumption was recently given theoretical support by Field and Jacob (1993b). They computed the theoretical response of a horizontally stratified sedimentary layer to ambient noise sources represented as a random distribution (in time, space, and orientation) of impulsive point forces located on the Earth's surface. They found that the expected horizontal-component spectrum of ambient noise contains its most prominent peak at the fundamental resonant frequency predicted for incident shear waves. This agrees with the results of several empirical studies that have determined the fundamental resonant frequency from ambient noise observations (e.g., Ohta et al., 1978; Celebi et al., 1987; Lermo et al., 1988; Field et al., 1990; Hough et al., 1991). In practice, the biggest challenge in the ambient noise method is in isolating the source (and possibly path) effects from the site effects. This is usually done by assuming the source spectra are white over the frequency range of interest (e.g., Celebi et al., 1987), or by dividing the sediment-site noise spectrum by that observed at a nearby bedrock site (e.g., Lermo et al., 1988; Field et al., 1990). Perhaps the most effective technique, however, is that introduced by Nakamura (1989). He hypothesized that the vertical component of ambient seismic noise is relatively uninfluenced by the sediments, and can therefore be used to remove the source effects from the horizontal components. He then showed, as have others since (e.g., Omachi et aL, 1991; Lermo et al., 1992; Field and Jacob, 1993b; Field et al., 1994), how such horizontal- to vertical-component noise ratios can be used to identify the fundamental resonant frequency of sediments. These observations have been given theoretical support by Lermo et al. (1992, 1994), based on the assumption that ambient noise represents Rayleigh-wave motion, and by the less-restrictive modeling of Field and Jacob (1993b), described above, where expected horizontalto vertical-component noise ratios were found to exhibit a sharp peak at the fundamental S-wave resonant frequency. Additional theoretical support has also been presented by Lachet and Bard (1995). Figure 4d shows the average horizontal- to verticalcomponent spectral ratios of ambient seismic noise observed at S1 and $3. These noise recordings were obtained using Kinemetrics SH-1 velocity sensors, which have a natural period of 5 sec. Although amplitudes are underpredicted, the most prominent peaks in Figure 4d correspond precisely to the fundamental resonant frequencies observed in the preferred spectral-ratio estimates (Fig. 4a). Plotted in Figure 5c is the average horizontal- to vertical-component spectral ratio of ambient noise observed at the reference site ($4),

1140 that we desire a 10% precision on the mean spectral-ratio estimate (with respect to the 95% confidence limits), then equation (6) and a bit of algebra imply that - 8 0 events need to be recorded. In assigning confidence limits, we have implicitly assumed that the estimates are lognormally distributed. Using our preferred spectral-ratio results, we applied a chi-square test (Press et al., 1988) to investigate whether the residuals between actual and predicted observations are indeed consistent with this distribution. The results are shown in Figure 6a. The diamonds represent the value of chi-square obtained for each frequency, and the solid line represents the region above which the null hypothesis (that the distributions are consistent) is to be rejected at the 95% confidence level. Care was taken to exclude frequencies where not enough data were available to make a fair test (see figure caption for details). Of the remaining data, 324 out of 337 frequencies (or 96%) passed the chi-square test. In fact, that 4% failed is consistent with the notion that --5% would be expected to fail if taken from a true lognormal distribution. Shown in Figure 6b is the observed cumulative distribution for the data at all frequencies combined (crosses connected by the dotted line), which agrees well with the assumed distribution (solid line). Therefore, for the data analyzed in this article, and the particular spectral estimation technique applied, the lognormal distribution appears to be consistent with the observations and therefore the confidence limits should be reliable. In this study, we were unable to determine how much of the observed uncertainty comes from noise in the data and how much of it comes from an intrinsic variability of the site response itself. As mentioned previously, this distinction is important with respect to hazard assessment. If the variability comes from random differences between the bedrock site recording and the true input, then we would want to minimize this source of uncertainty by applying the standard deviation of the mean (which goes down as more observations are obtained). If the scatter in spectral ratios is produced by an intrinsic variability in the site response itself, then we would want to account for this by applying the standard deviation (which represents the scatter of individual estimates). We plan to address this issue in a separate study of Landers aftershock data collected in the Coachella Valley near Palm Springs, California. In addition to several sediment sites, three different bedrock sites were occupied on different sides of the valley. These should allow us to constrain the variability of input motion among the various sites, and thereby determine the relative contribution of this and any intrinsic site-response variability. All of the three non-reference-site-dependent techniques that we tested here were found to reveal useful information regarding site response at our East Bay sites. The frequency dependence of the parameterized-inversion siteresponse estimates agree remarkably well with the sedimentto-bedrock spectral ratios. This is somewhat surprising given the simplicity of the source- and path-effect models used (equations 19 and 20, respectively), especially since po-

E . H . Field and K. H. J a c o b

stcritical Moho reflections are thought to be significant at the epicentral distances of the aftershocks in this study (Sommerville and Yoshimura, 1990). Because of the trade-off between corner frequency and attenuation (exemplified in Fig. 3), and between geometric and anelastic attenuation (Frankel et aL, 1990; Kato et al., 1995), we have low confidence in the strict reliability of theft and Q values determined in the inversion. Fortunately, the site-response estimates are relatively robust with respect to these trade-offs. In principle, independent estimates of geometric or anelastic attenuation

1512-

(a)

9~.~ ,.~ 6. ..:


. .
. " :

. . . . . .
.

. . . : . :
o . .

. . . . . .

r..)

:."
.* . a , ~

".
.* "-" o o ":

.. " ~**

~ .."~.

" ,

~ ."

3-

" ** $

.:..

** ~

..:

.
I

"....-. .

.*~

..:. .....,....,.

. . . , . ' o
.
I

...
I

...

0.0

215

5.0

7.5

10.0

12.5

15.0

17.5

20.0

frequency (Hz)
1.00..... (b) i ; ; ; : ~ J ;

0.75 -

0.50-

0.25 o 0.00 -4
I I I i I I I

-3

-2

-1

normalized residual Figure 6. (a) Values of the chi-square test at each frequency (diamonds). These were computed by standardizing the residuals of the logarithmic data (i.e., the quantity in brackets in equation 8) and sorting these into six bins. Any frequency that had less than five data points in any bin was excluded from further consideration. The dashed line represents the threshold above which the null hypothesis (that the two distfibution are consistent) is to be rejected at the 95% confidence level (i.e., Zo2.o5 for v = 6 - 3). (b) The observed cumulative distribution of residual data for all frequencies combined (crosses connected by the dotted line), obtained by sorting the normalized data
into 40 bins. T h e solid line represents the expected c u m u l a t i v e distrubtion if the observations are lognorreally distributed.

Various Site-Response Estimation Techniques

1141

for the region, and any seismic moment estimates available from, for example, empirical Green' s function studies, could also be applied as a priori information to improve the inversion results. Constraining the unresolved frequency-independent scaling factor remains problematic. In practice, one should apply as much a priori information as possible for this purpose (e.g., geotechnical data and theoretical modeling, independent seismic moment estimates of the aftershocks, or that site response should become negligible at long periods). However, given the uncertainties associated with these approaches, the scaling factor may not be resolvable to better than a factor of 2 or so. Nevertheless, the parameterizedinversion scheme of Boatwright et al. (1991b) appears capable of providing the frequency dependence of site response at locations where a nearby reference site is not available. Additional tests of the method will be needed to determine whether or not our results are typical. The receiver-function-type estimates (horizontal- to vertical-component spectral ratios) applied to shear waves also reveal the frequency dependence of site response at the East Bay sediment sites (although these estimates disagree with the spectral-ratio estimates by a frequency-independent scaling factor of -0.6). This implies that the local geology is relatively transparent to the motion observed on the vertical component. Equivalent estimates computed for the reference site ($4) are relatively fiat and close to unity, giving additional credibility to this non-reference-site-dependent method. This result suggests that investigations based only on the vertical component, such as the coda-wave studies of Phillips and Aki (1986), Mayeda et al. (1991), S u e t al. (1992), and Koyanagi et al. (1992), may underestimate the site response at sediment sites. In contrast to the results obtained for S waves, when applied to P waves, the receiverfunction-type estimates do not compare favorably with the spectral-ratio estimates. The third non-reference-site-dependent technique that we investigated is Nakamura's estimate, obtained by taking horizontal- to vertical-component ratios of ambient seismic noise. The ambient noise approach is attractive given the relative ease and economy of data collection. Nakamura's estimate clearly reveals the fundamental resonant frequency of the sediment sites, which is similar to the results of several other studies (e.g., Omachi et al., 1991; Lermo et aL, 1992; Field and Jacob, 1993b; Field et al., 1994a, Lermo et al., 1994). These observations have also been bolstered by the results of theoretical modeling (Lermo et al., 1992, 1994; Field and Jacob, 1993b; Lachet and Bard, 1995). One might dismiss Nakamnra's method since it is apparently capable of identifying only the fundamental resonant frequency. However, as discussed by Field et aL (1994a), this may be the only resonant frequency of concern in hazard assessment if the higher modes are significantly reduced by increased damping during strong ground motion. We conclude for this study that two of the three nonreference-site-dependent techniques (the parameterized-in-

version and receiver-function-type estimates) were successful in identifying the frequency dependence of site response, and the third (Nakamura's method) was successful in identifying the fundamental resonant frequency. The frequencyindependent amplitude discrepancy between the various estimates remains problematic. If we take the spectral-ratio estimates as the most reliable, then all others underpredict the site response [including the one-dimensional model predictions shown in Field et al. (1992), which differ by a frequency-independent scaling factor of -2]. One possibility is that the sediment-to-bedrock spectral ratios are biased upward by a systematic focusing below the sediment sites, and/ or a defocusing below the bedrock site. However, we have no evidence of this. Nevertheless,that the frequency dependence of site response can be obtained from any one of a number of methods is promising. If these results are corroborated by other studies, then these estimates will be useful for making site-specific hazard assessments in areas that lack an adequate reference site. Acknowledgments
We thank William Menke, Noel Barstow, and Feng Su for reviewing this article, and John Boatwright for providing computer software and a review of this article, and for answering many questions during this study. We also thank Sue Hough, Paul Friberg, Robert Busby, and all others who participated in the field work following the Loma Prieta earthquake. This study was supported by the National Center for Earthquake Engineering Research Grant Numbers NCEER-93-1002 and NCEER-94-1002. Partial support for the data collection was provided by the Incorporated Research Institutions for Seismology (IRIS) award 0133 scope DA. Lamont-Doherty Earth Observatory Contribution Number 5323.

References
Aki, K. (1988). Local site effects on strong ground motion, Proc. Earthquake Eng. Soil Dyn. II, 103-155. Anderson, J. G. and J. R. Humphrey (1991). A least-squares method for objective determination of earthquake source parameters, Seism. Res. Lett. 62, 201-209. Andrews, D. J. (1986). Objective determination of source parameters and similarity of earthquakes of different size, in Earthquake Source Mechanics, S. Das, J. Boatwright, and C. H. Scholz (Editors), American Geophysical Union, Washington, D.C., 259-268. Beresnev, I. A., W.-L. Wen, and Y. T. Yeh (1994). Seismological evidence for nonlinear-elastic ground behavior during large earthquakes, Soil Dyn. Earthquake Eng. (in press). Boatwright, J. (1978). Detailed spectral analysis of two small New York State earthquakes, Bull. Seism. Soc. Am. 68, 1117-1131. Boatwright, J., L. C. Seekins, and C. S. Mueller (1991a). Ground motion amplification in the Marina, Bull. Seism. Soc. Am. 81, 1980-1997. Boatwright, J., J. B. Fletcher, and T. E. Fumal (1991b). A general inversion scheme for source, site, and propagation characteristics using multiply recorded sets of moderate-sized earthquakes, Bull. Seism. Soc. Am. 81, 1754-1782. Bonamassa O. and C. S. Mueller (1989). Source and site response spectra from the aftershock seismograms of the Whittier Narrows, California, earthquake, Seism. Res. Lett. 59, 23. Borcherdt, R. D. (1970). Effects of local geology on ground motion near San Francisco Bay, BulL Seism. Soc. Am. 60, 29-61. Borcherdt, R. D. and J. F. Gibbs (1976). Effects of local geological con-

1142

E . H . Field and K. H. Jacob

ditions in the San Francisco Bay region on ground motions and the intensities of the 1906 earthquake, Bull. Seism` Soc. Am. 66, 467-500. Borcherdt, R. D., G. Glassmoyer, A. Der Kiureghian, and E. Cranswick (1989). Results and data from seismologic and geologic studies following earthquakes of December 7, 1988 near Spitak, Arineuia, S.S.R., U.S. GeoL Surv. Open-File Rept. 89-163A. Borcberdt, R. D. and G. Glassmoyer (1992). On the characteristics of local geology and their influence on ground motions generated by the Loma Prieta earthquake in the San Francisco Bay region, California, Bull. Seism` Soc. Am. 82, 603--641. Brnne, J. N. (1970). Tectonic stress and spectra of seismic shear waves from earthquakes, J. Geophys. Res. 75, 4997-5009. Celebi, M., C. Dietel, J. Prince, M. Onate, and G. Chavez (1987). Site amplification in Mexico City (determined from 19 September 1985 strong-motion records and from recordings of weak motions), in Ground Motion and Engineering Seismology, A. S. Cakmak (Editor), Elsevier, Amsterdam, 141-152. Chin, B. and IC Aki (1991). Simultaneous study of the source, path, and site effects on strong ground motion during the 1989 Loma Prieta earthquake: a preliminary result on pervasive nonlinear site effects, Bull. Seism` Soc. Am. 81, 1859-1884. Darragh, R. B. and A. F. Shakal (1991). The site response of two rock and soil station pairs to strong and weak ground motion, Bull. Seism` Soc. Am. 81, 1885-1899. Field, E. H., S. H. Hough, and IC H. Jacob (1990). Using microtremors to assess potential earthquake site response: a case study in Flushing Meadows, New York City, Bull. Seism. Soc. Am. 80, 1456-1480. Field, E. H., K. H. Jacob, and S. H. Hough (1992). Earthquake site response estimation: a weak-motion case study, Bull. Seism. Soc. Am. 82, 2283-2307. Field, E. H. and K. H. Jacob (1993a). Monte-Carlo simulation of the theoretical site response variability at Turkey Flat, California, given the uncertainty in the geoteehnically derived input parameters, Earthquake Spectra 9, 669-701. Field, E. H. and K. H. Jacob (1993b). The theoretical response of sedimentary layers to ambient seismic noise, Geophys. Res. Lett. 20, 2925-2928. Field, E. H., A. C. Clement, V. Aharonian, P. A. Friberg, L. Carroll, T. O. Babaian, S. S. Karapetian, S. M. Hovanessian, and H. A. Abramian (1994a). Earthquake site response study in Ginmri (formerly Leuinakan), Armenia using ambient noise observations, Bull. Seism. Soc. Am. (in press). Field, E. H., S. E. Hough, K. H. Jacob, and P. A. Friberg (1994b). Earthquake site response in Oakland, California near the failed Nimitz Freeway, U.S. Geol. Soc. Profes. Pap. 1551-A, 169-180. Field, E. H. (1994). Earthquake site response estimation, Ph.D. Thesis, Columbia University, New York. Fletcher, J. B. and J. Boatwright (1991). Source parameters of Loma Prieta aftershocks and wave propagation characteristics along the San Francisco Peninsula from a joint inversion of digital seismograms, Bull. Seism. Soc. Am. 81, 1783-1812. Frankel, A., A. McGarr, J. Bicknel, J. Moil, L. Seeber, and E. Cranswick (1990). Attenuation of high-frequency shear waves in the crust: measurements from New York, South Africa, and southern California, J. Geophys. Res. 95, 17441-17457. Graves, R. W. (1992). Modeling site response effects within the Marina District basin, San Francisco, California, Seism. Res. Lett. 63, 34. Hartzell, S. H. (1992). Site response estimation from earthquake data, Bull Seism. Soc. Am. 82, 2308-2327. Hough, S. E., R. D. Borcberdt, P. A. Friberg, R. Busby, E. Field, and K. H. Jacob (1990). The role of sedimeut-induced amplification in the collapse of the Nimitz freeway during the October 17, 1989 Loma Prieta earthquake, Nature 344, 853-855. Hough, S. E., E. H. Field, and K. H. Jacob, (1991). Using microtremors to assess site-specific earthquake hazard, Proc. of the 4th Int. Conf. on Seis. Zonation, Stanford University, Stanford, California, 25-29 August, Vol. 3, 585-592.

Kagami, H., C. M. Duke, G. C. Liang, and Y. Ohm (1982). Observation of 1- to 5-second microtremors and their application to earthquake engineering, Part II: Evaluation of site effect upon seismic wave amplification due to extremely deep soils, Bull. Seism. Soc. Am. 72, 987998. Kagami,H., S. Okada, K. Shiono, M. Oner, M. Dravinski, and A. K. Mal (1986). Observation of 1- to 5-second microtremors and their application to earthquake engineering, Part III: A two-dimensional study of site effects in the San Femando Valley, Bull. Seism` Soc. Am. 76, 1801-1812. Kanai K. (1957). The requisite conditions for predominant vibration of ground, Bull. Earthquake Res. Inst. Tokyo Univ. 31, 457. Kato, K., K. Aid, and M. Takemura (1995). Site amplification from coda waves: validation and application to S-wave site response, Bull Seism` Soc. Am. 85, 467-477. King, J. L, and B. E. Tucker (1984). Observed variations of earthquake motion across a sedimeut-filled valley, Bull. Seism. Soc. Am` 74, 137151. Koyanagi S., K. Mayeda, and K. Aki (1992). Frequency-dependent site amplification factors using the S-wave coda for the island of Hawaii, Bull. Seism` Soc. Am. 82, 1151-1185. Lachet, C. and P.-Y. Bard (1995). Numericai and theoretical investigations on the possibilities and limitations of the "Nakamura's" technique, J. Phys. Earth (in press). Langston, C. A. (1979). Structure under Mount Rainier, Washington, inferred flom teleseismic body waves, J. Geophys. Res. 84, 4749-4762. Lermo, J., M. Rodriguez, and S. K. Singh (1988). The Mexico earthquake of September 19, 1985--natural period of sites in the valley of Mexico from microtremor measurements and from strong motion data, Earthquake Spectra 4, 805-814. Lermo, J., S. Francisco, and J. Chavez-Garcia (1992). Site effect evaluation using microtremors: a review (abstract), EOS 73, 352. Lermo, J. F., S. Francisco, and J. Chavez-Garcia (1993). Site effect evaluation using spectral ratios with only one station, Bull. Seism. Soc. Am. 83, 1574-1594. Lermo, J. F., S. Francisco, and J. Chavez-Garcia (1994). Are microtremors useful in site response evaluation? Bull. Seism` Soc. Am. 84, 13501364. Lindley, G. T. and R. J. Archuieta (1992). Earthquake source parameters and frequency dependence of attenuation at Coalinga, Mammoth Lakes, and Santa Cruz Mountains, California, J. Geophys. Res. 97, 14137-14154. Margheriti, L., L. Wennerberg, and J. Boatwright (1994). A comparison of coda and s-wave spectral ratios as estimates of site response in the southern San Francisco Bay area, Bull Seism` Soc. Am. 84, 18151830. Masuda, T. and Z. Suzuki (1982). Objectiveestimation of source parameters and local Q values by simultaneous inversion method, Phys. Earth Planet. Interiors 30, 197-208. Mayeda, K., S. Koyanagi, and K. Aki (1991). Site amplification from s-wave coda in the Long Valley Caldera region, California, Bull. Seism. Soc. Am. gl, 2194-2213. Menke, W. (1989). Geophysical Data Analysis: Discrete Inverse Theory, Academic Press, New York. Milne, J. (1898). Seismology, First Ed., Kegan Paul, Trench, Truber, London. Mueller, C. S. (t986). The influence of site conditions on near-source highfrequency ground motion: case studies from earthquakes in Imperial Valley, Ca., Coalinga, Ca., and Miramichi, Canada, Ph.D. Thesis, Stanford University, Stanford, California. Nakamura, Y. (1989). A method for dynamic characteristics estimation of subsurface using microtremor on the ground surface, QR Railway Tech. Res. Inst. 30, 1. Ohta, Y. H., Kagami, N. Goto, and K. Kudo (1978). Observation of 1- to 5-second microtremors and their application to earthquake engineering, Part I: Comparison with long-period accelerations at the TokachiOld earthquake of 1968, Bull. Seism` Soc. Am. 68, 767-779.

Various Site-Response Estimation Techniques

1143
Seekins, L. C. and J. Boatwright (1994). Ground motion, amplification, geology, and damage from the 1989 Loma Prieta earthquake in the city of San Francisco, Bull. Seism. Soc. Am. 84, 16-30. Singh, S. K., J. Lermo, T. Dominguez, M. Ordaz, J. M. Espinosa, E. Mena, and R. Quass (1988). The Mexico earthquake of September 19, 1985--a study of amplification of seismic waves in the valley of Mexico with respect to a hill zone site, Earthquake Spectra 4, 653673. Sommerville, P. and J. Yoshimura (1990). The influence of critical Moho reflections on strong ground motions recorded in San Francisco and Oakland during the 1989 Loma Prieta earthquake, Geophys. Res. Lett. 17, 1203-1206. Su, F., K. Aki, T. Teng, S. Koyanagi, and K. Mayeda (1992). The relation between site amplification factor and surficial geology in central California, Bull. Seism. Soc. Am. 82, 580--602. Takahashi, K., S. Ohno, M. Takemura, T. Ohta, Y. Sugawara, T. Hatori, and S. Omote (1992). Observation of earthquake strong-motion with deep borehole generation of vertical motion propagating in surface layers after S wave arrival, Proc. of the lOth World Conf. of Earthquake Eng., Vol. 3, 1245-1250. Thomson D. J. (1982). Spectrum estimation and harmonic analysis, Proc. IEEE, 70, 1055-1096. Udwadia, F. E. and M. D. Trifunac (1973). Comparison of earthquake and microtremor ground motions in E1 Centro, California, Bull Seism. Soc. Am. 63, 1227-1253. Lamont-Doherty Earth Observatory Columbia University Palisades, New York 10964 Manuscript received 10 June 1994.

Olson, A. H. (1987). A Chebychev condition for accelerating convergence of iterative tomographic methods: solving large least squares problems, Phys. Earth Planet. Interiors 47, 333-345. Omachi, T., Y. Nakamura, and T. Toshinawa (1991). Ground motion characteristics in the San Francisco Bay area detected by microtremor measurements, Proc. 2nd Int. Conf. on Recent Adv. in Geot. Earth Eng. and Soil Dyn., 11-15 March, St. Louis, Missouri: 1643-1648. Owens, T. J. and R. S. Crossen (1988). Shallow structure effects on broadband teleseismic P waveforms, Bull. Seism. Soc. Am. 78, 96-108. Owens, T. J., G. Zandt, and S. R. Taylor (1984). Seismic evidence for an ancient rift beneath the Cumberland Plateau, Tennessee: a detailed analysis of broadband teleseismic P waveforms, J. Geophys. Res. 89, 7783-7795. Park, J., C. R. Lindberg, and F. L. Vernon III (1987). Multitaper spectral analysis of high-frequency seismograms, J. Geophys. Res. 92, 1267512684. Press, W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling (1988). Numerical Recipes in C, Cambridge University Press, Cambridge, United Kingdom. Phillips W. S. and K. Aid (1986). Site amplification of coda waves from local earthquakes in central California, Bull. Seism. Soc. Am. 76, 627648. Reidesel, M. A., R. D. Moore, and J. A. Oreutt (1990). Limits of sensitivity of inertial seismometers with velocity transducers and electronic amplifiers, Bull. Seism. Soc. Am. 807 1725-1752. Scherbaum, F. (1990). Combined inversion for the three-dimensional Q structure and source parameters using microearthquake spectra, J. Geophys. Res. 95, 12423-12438. Seed, H. B., M. P. Romo, J. I. Sun, A. Jaime, and J. Lysmer (1988). The earthquake of September 19, 1985--relationships between soil conditions and earthquake ground motions, Earthquake Spectra 4, 687729.

Das könnte Ihnen auch gefallen