219

Propagation of Data Uncertainty in Surface Wave Inversion

Carlo G. Lai, Sebastiano Foti and Glenn J. Rix

European Centre for Training and Research in Earthquake Engineering, EUCENTRE,
Via Ferrata, 1 Pavia, 27100, Italy
Email: carlo.lai@eucentre.it
Politecnico di Torino, Department of Structural and Geotechnical Engineering,
corso Duca degli Abruzzi, 24 Torino, TO 10129, Italy
Email: sebastiano.foti@polito.it
Georgia Institute of Technology, School of Civil and Environmental Engineering,
Atlanta, Georgia 30332 USA
Email: glenn.rix@ce.gatech.edu

ABSTRACT

Although in recent years surface wave methods have undergone significant development that has
greatly enhanced their capabilities, little effort has been spent to determine the uncertainty associated
with surface wave measurements. The objective of this study is to determine how the uncertainty of the
experimental data is mapped into the uncertainty of the shear wave velocity profile via the inversion
algorithm. The methodology developed in this study for estimating the uncertainty of the shear wave
velocity profile from surface wave measurements is based on the assumption that the experimental data
are normally distributed. The validity of this hypothesis was experimentally verified using data gathered
at two sites in Italy where surface wave tests were performed using linear arrays of multiple receivers.
The experimental dispersion curve measured at the site was subsequently inverted to obtain the
expected shear wave velocity profile together with an estimate of the associated standard deviation. The
final results show that uncorrelated noise has a very little influence on multistation surface wave tests,
confirming their robustness for applications in noisy environments.

Introduction sequently they used these estimates to assess the uncertainty

of the shear wave velocity profile using an approach based
An essential aspect of any measurement technique is on multiple repetition of the inversion procedure which gives
the ability to estimate the uncertainty of the measured data scattered values of the coefficient of variation (Marosi and
and to determine how this uncertainty is projected onto the Hiltunen, 2004b). A study to assess the influence of several
derived parameters. This is particularly important in geo- factors on uncertainties in surface wave multistation tests
physical testing where the parameters of interest are often using Monte Carlo numerical simulations and repeated field
inferred from complex inversion procedures. In surface wave tests is reported by O’Neill (2004). He found coefficients of
testing the measured data are represented by the Rayleigh variation for Rayleigh wave phase velocities increased non-
dispersion and attenuation curves, whereas the derived linearly for decreasing frequencies and proposed a Lorentzian
parameters are the variation with depth of the shear modulus distribution for the low-frequency range.
(or equivalently, the shear wave velocity) and shear damping In this paper some of the algorithms introduced by Lai
ratio at small strains (Rix et al., 2000; Lai et al., 2002). and Rix (1998) are implemented to estimate the uncertainty
Despite the importance of this topic, few studies have associated with surface wave testing using real data. In
been attempted to quantify the uncertainty associated with particular the propagation of uncertainty through the inver-
the results obtained from surface wave testing. Tuomi and sion process is formulated on the basis of Occam’s algorithm
Hiltunen (1996) used reliability analysis to determine the (Constable et al., 1987). This approach is computation-
coefficient of variation (i.e., the ratio, for a random variable, ally efficient because it is based on the partial derivatives
between standard deviation and expected value) of a shear obtained during the final iteration of the inversion process.
wave velocity profile from a set of 30 dispersion curves The focus is restricted only to applications of surface wave
measured with the two-station SASW method. Again, for the testing concerned with stiffness profiling, although the same
two-station procedure, Marosi and Hiltunen (2004a) report concepts apply to measurements of the shear damping ratio.
experimental values of the coefficient of variation for phase The analysis of the uncertainty is conducted with reference to
velocity typically around 1.5% with samples normally dis- experimental dispersion curves determined with the frequency-
tributed for frequencies in the range 20–150 Hz. Sub- wavenumber ( f-k) multi-station technique.

JEEG, June 2005, Volume 10, Issue 2, pp. 219–228

220
Journal of Environmental and Engineering Geophysics

The article is organized as follows. Section 2 describes curves is simpler. The interested reader is referred to the rich
the methodology for assessing the uncertainty of the exper- literature on the subject (see for instance McMechan and
imental dispersion curve. Section 3 illustrates the procedure Yedlin, 1981; Gabriels et al., 1987; Tselentis and Delis,
used to determine how the uncertainty of the experimental 1998; Park et al., 1999; Yilmaz and Doherty, 2003).
dispersion curve is mapped into uncertainty of the shear Perhaps the most widely used multi-station processing
wave velocity profile at a site. It will be shown that this technique to determine the experimental dispersion curve
procedure is intimately related to the algorithm adopted for is the frequency-wavenumber ( f-k) method (Gabriels et al.,
the inversion of surface wave data. Finally, Sections 4 1987). The f-k method can be implemented by transforming
through 6 present the results obtained from the application the signals detected by the receivers from the time-offset
of the methodology to a set of experimental data gathered at (t, x) domain to the frequency-wavenumber ( f, k) domain by
two sites in Italy. means of a 2D Fourier transform or by using spectral
estimation techniques (Zywicki and Rix, 1999; Rix et al.,
Uncertainty in Surface Wave Testing 2001). The dispersion curve is then estimated from the
(absolute) maxima in the frequency-wavenumber spectrum
Uncertainties in any test are related to a number of of the surface wavefield (Tselentis and Delis, 1998).
different factors. In indirect measurements such as surface The estimation of the uncertainty is not trivial when the
wave tests, it is important to distinguish between uncertain- experimental dispersion curve is determined using transform-
ties related to the model used for the interpretation and un- based methods. Indeed, difficulties arise in properly
certainties in the measurements. The former are caused, for quantifying how the data error is propagated through a series
example, by near-field effects and inadequacy of the hori- of complex data processing steps from the acquisition of the
zontally layered model. These aspects are not covered in the signals in the time-offset domain (t, x) to the calculation of
present paper. Strobbia and Foti (2005) proposed a statistical the dispersion curve.
test to detect the presence of model errors based on the The uncertainty associated with the experimental
regression of phase vs. offset. dispersion curve in multi-station surface wave testing is more
Data uncertainties are due mainly to noise in the re- easily determined by a direct measurement of the statistical
corded signals and to geometrical uncertainties related to the distribution of both primitive and derived surface wave data.
location and tilting of the receivers. The influence of several Even though this approach requires considerably more time
sources of uncertainty has been studied by O’Neill (2003) and effort to be implemented, it does not involve simplifying
using numerical simulations. He reports minimal influence assumptions (e.g., small variance of the raw data) and is
from geophone tilt and coupling, while positional errors, exempt from major technical difficulties. This is the approach
static shift and additive Gaussian noise introduce larger un- followed in the case-histories presented in this paper.
certainties in the experimental dispersion.
Uncertainties in recorded signals are associated with Mapping the Uncertainty by the Inversion Algorithm
coherent noise and uncorrelated noise. The latter is externally
generated noise (environmental noise) and can be studied via Once the uncertainty of the experimental dispersion
the statistical distribution of the recorded signals if many curve VR(x) has been estimated, the final task is to determine
repetitions of the test in a given configuration are available. how this uncertainty is mapped by the inversion algorithm
Coherent noise is due to events generated by the seismic into uncertainty of the expected shear wave velocity profile
source (e.g., near-field effects). The proposed approach can VS(z). The solution of this problem is not unequivocal and
be used to estimate the effect of uncorrelated noise on depends on the algorithm used for the inversion of the non-
surface wave measurements and its propagation in sub- linear relation VR ¼ VR(VS) where VR is a nF 3 1 vector of
sequent interpretation. Rayleigh phase velocities VR(x) associated with nF different
The simplest procedure for determining the apparent frequencies, and VS is a nL 3 1 vector whose components are
dispersion curve from the signals recorded by an array of the unknown shear wave velocities of a nL-layer soil deposit.
receivers is performing a linear regression of the displace-
ment (or particle velocity, acceleration or transfer function) Occam’s Algorithm
phases measured at each receiver location. This procedure can In this work the inversion of the non-linear relation
be extended to estimate data uncertainty in the experimental VR ¼ VR(VS) has been performed using a local search al-
dispersion curve (Lai and Rix, 1998). An implementation of gorithm where the stationary point in the solution space is
such an approach is reported by Strobbia and Foti (2005). sought with a constrained, least-squares technique known as
Other, more refined techniques include transform- Occam’s algorithm (Constable et al., 1987; Parker, 1994).
based methods in which the data measured in the time-offset The strategy is to find the smoothest profile (in the sense
domain are transformed into a different domain where the of the first Fréchet derivative) of shear wave velocities sub-
identification of the loci of points defining the dispersion ject to the constraint of a specified misfit between the
 221
Lai et al.: Propagation of Uncertainty

experimental and theoretical dispersion curves. The adop- smoothing parameter l must be chosen so that the specified
tion of this strategy is motivated by the observation that residual error e2(*) is obtained. The computer program for
inversions performed with classical least-squares techniques the inversion of surface wave data using Occam’s algorithm
often lead to physically unreasonable profiles of model is freely available (see Rix and Lai, 2000).
parameters. By enforcing global smoothness and regularity
in the solution, one minimizes the dependence of shear wave Uncertainty Estimation of Shear Wave Velocity Profile
velocity profile upon the assumed number of layers and It can be shown that if an inverse problem is non-
imposes a constraint on the solution of the Rayleigh inverse linear, a Gaussian distribution of the errors in the data will in
problem that like many other inverse problems is inherently general be mapped into a distribution of the uncertainty of
ill-posed (Lai and Rix, 1998). the model parameters that is non-Gaussian (Tarantola, 1987;
Implementation of the Occam’s algorithm requires Menke, 1989). In principle, this fact poses serious difficul-
a definition of smoothness or its converse roughness of a ties when estimating the uncertainty of model parameters
candidate solution that in our case is the VS profile. In a derived from the inversion of data affected by measurement
stratified soil profile, roughness may be defined by the fol- errors. First, in non-Gaussian probability distributions, cen-
lowing expression: tral estimators like the expectation and the maximum
T likelihood value do not coincide, and more importantly,
R1 ¼ ð›VS Þ ð›VS Þð1Þ they do not necessarily represent the most sensible estimator
where › is an nL 3 nL matrix defining the two-point-central of the distribution (Menke, 1989).
finite difference operator (Constable et al., 1987). The error The same may be said about the covariance as an
misfit e2 between the measured and the predicted Rayleigh estimator of dispersion. Another critical problem of dealing
phase velocities may be written as follows: with non-Gaussian distributions is that there are in general
no analytical expressions for both the solution and the asso-
e2 ¼ ½WVR exp 2 WVR theo
T ½WVR exp 2 WVR theo
ð2Þ ciated uncertainty. In this case a very general and rigorous
where VRexp is an nF 3 1 vector of experimental Rayleigh approach to solve the problem would be a thorough explo-
phase velocities, VRtheo is an nF 3 1 vector of theoretical ration of the space of model parameters, either systematically
modal or apparent Rayleigh phase velocities and W is a by defining a grid, or randomly as in Monte Carlo methods
diagonal nF 3 nF matrix: (Tarantola, 1987). The main drawback of this approach is
that it is computationally time-consuming.
W ¼ diagf1=r1 ; 1=r2 ; . . . ; 1=rnF g ð3Þ However, if the inverse problem is not too non-linear,
especially around the point of maximum likelihood in the
containing the uncertainties associated with the experimen-
model space (i.e., the probability distribution of model
tal data. The solution of the non-linear Rayleigh inverse
parameters around the point of maximum likelihood is not
problem consists of finding a vector VS that minimizes
too different from a Gaussian function), it is possible to use
R1 with the constraint that the residual error e2 be equal to
a simplified approach to estimate the uncertainty of the
a specified value e2(*) that is acceptable in light of the un-
model parameters.
certainties. The method of Lagrange multipliers is employed
If in solving the non-linear inverse problem VS ¼
to solve this constrained optimization problem resulting in
VS1(VR), it is assumed that: a) the uncertainty of the
(Menke, 1989; Tarantola, 1987):
experimental dispersion curve follows a Gaussian distri-
T 1 T
VS ¼ ½l›T › þ ðWJVS0 Þ WJVS0
ðWJVS0 Þ bution (this assumption is indeed reasonable from an
 W½JVS0 VS0 þ ðVR 2VR0 Þ
ð4Þ experimental viewpoint as shown later), b) the inverse
problem VS ¼ VS1(VR) is only moderately non-linear
where l is the Lagrange multiplier, which may be interpreted around its solution, and c) the relation VR ¼ VR(VS) is
as a smoothing parameter, and VR0 are the modal or apparent inverted using the method of maximum likelihood, then the
Rayleigh phase velocities obtained from the solution of the uncertainty associated with the expected shear wave
Rayleigh forward problem with VS ¼ VS0. The term JVS0 is velocity profile can be approximated by the following
the nF 3 nL Jacobian matrix whose elements are the partial formula (Tarantola, 1987; Menke, 1989):
derivatives of the modal or apparent Rayleigh phase
velocities with respect to the shear wave velocities of the Cov½VS
’ ½ðJS T ðCov½VR
Þ1 JS Þ1 JS T ðCov½VR
Þ1
last# 
layers. These partial derivatives were computed from the 1 1
Cov½VR
½ðJS T ðCov½VR
Þ JS Þ 
solution of the Rayleigh forward problem using closed-form,
analytical expressions derived from the variational principle JS T ðCov½VR
Þ1
last # T ð5Þ
of Rayleigh waves (Lai and Rix, 1998).
Equation 4 is used iteratively to refine the estimated where Cov[VR] is an nF 3 nF matrix of covariances of the
shear wave velocity profile VS0 until convergence. The experimental Rayleigh phase velocities at nF frequencies x. It
222
Journal of Environmental and Engineering Geophysics

is assumed that the data VR(x) are statistically uncorrelated, The v2 distribution is well known and is tabulated in
and hence the matrix Cov[VR] is diagonal with the non-zero most books on statistics. v2 is small if the observed and the
elements equal to Var[VR]. Cov[VS] is a nL 3 nL diagonal expected frequencies are nearly the same and thus a small
matrix whose elements are the variances of the estimated value of v2 is indicative of a good correspondence between
shear wave velocities (VS)j with j ¼ 1, nL. Finally, the term JS measured and presumed statistical distributions. It is stan-
is the nF 3 nL Jacobian matrix whose elements (JS)kj are dard practice to assume that the data do not follow a given
defined by the partial derivatives (JS)kj ¼ ›(VR)k/›(VS)j. The distribution if values greater or equal than X2 occur with
subscript last# outside the brackets of Eq. (5) indicates that the a frequency of less than 5% (Menke, 1989; Spiegel et al.,
terms inside the parentheses (in essence the Jacobian matrix 2000). The threshold upper bound of X2 is established from
JS) should be computed with respect to the final iteration in the the v2 distribution. For the problem at hand, k ¼ 5 and thus
solution of the non-linear problem VR ¼ VR(VS). m ¼ 2. With this value of m a v2 distribution table gives
If the relation VR ¼ VR(VS) is inverted using Occam’s v0:05 2 ¼ 5.99 (Spiegel et al., 2000).
algorithm (as in the present work), then the uncertainty of
the expected shear wave velocity profile can be estimated Field Measurements
through the following expression, which was derived using
the concept of generalized inverse (Menke, 1989; Lai and Sites Description and A Priori Information
Rix, 1998): The first site is located in Saluggia (VC) in the
T 1 T northern part of Italy, close to the Dora Baltea River and
Cov½VS
’ ½ðl›T › þ ðWVR JS Þ WVR JS Þ ðWVR JS Þ WVR
last#  part of a large flat area of fluvial sediments. The shallow
T 1
Cov½VR
½ðl›T › þ ðWVR JS Þ WVR JS Þ  sediments are composed of gravels and gravelly sands with
T fine sand and clayey silt in the form of lenses (Foti, 2000).
ðWVR JS Þ WVR
last# T ð6Þ
The water table is between 2 and 3 meters below the ground
where the term WVR is a nF 3 nF diagonal matrix defined by surface. A cross-hole test was also performed in the vicinity
WVR ¼ diagf1/(r VR )1, 1/(rVR )2, . . . , 1/(rVR )nF g where
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi of the surface wave testing alignment (Foti, 2000).
(rVR )k ¼ þ VarðVR Þk with k ¼ 1, nF and represents the The second site is located in Florence (near Via
standard deviations associated with the experimental data Novoli) in Central Italy. The site is mostly comprised of
VR (The parameter l and the symbol › have been defined alluvial soil deposits consisting of thin layers of clayey silt
previously). and silty sand underlain by thick formations of coarse-
grained soils, mainly sandy gravel with silt and occasional
boulders. The water table is located at an average depth of
Testing the Assumption of Gaussian Distribution of
6 m below the ground surface.
Experimental Data
An extensive investigation was performed at the two
A fundamental assumption in deriving Eqs. (5) and
sites in Saluggia and Florence for geotechnical character-
(6) is that the experimental dispersion curve exhibits
ization and included soil borings, SPT and CPT penetration
Gaussian statistics. It is therefore of primary importance to
tests. From the retrieved data it was possible to construct two
verify the validity of this hypothesis with respect to the
litho-stratigraphic profiles that were used to guide the
population of independent measurements.
inversion of surface wave data. Soil layering (that is number
A standard criterion to validate the assumption that
and thickness of soil layers) is a fundamental piece of infor-
a set of experimental data follows a Gaussian distribution is
mation required by the inversion algorithm, and it constitutes
the v2-test (‘chi-squared test’: Snedecor and Cochran, 1980;
a type of a priori information that is kept fixed during the
Menke, 1989). It can be shown (Spiegel et al., 2000) that for
inversion, reducing the number of unknowns and hence the
two discrete statistical distributions formed by k number
overall uncertainties. The choice of soil layering certainly
of classes and by observed (oi) and expected (ei) absolute
affects the results of an inversion analysis and ultimately
frequencies, the quantity X2 defined by (Snedecor and
the predicted shear wave velocity profile. However, the
Cochran, 1980):
influence of this choice in the current study has not been
Xk
ðoi  ei Þ
2 investigated. Other input parameters required by the inver-
X2 ¼ ; ð7Þ sion algorithm include an assumed density of 1,800 kg/m3
i¼1
ei
and Poisson’s ratio of 0.2 and 0.45 above and below the
which expresses a sort of normalized distance (in the sense water table, respectively, at both sites.
of the L2 norm) between observed and expected distribu-
tions, follows a v2 distribution with m ¼ k  1  m degrees Testing Equipment and Procedure
of freedom (Menke, 1989). The parameter m depends on the Surface wave data were collected using a multi-
number of constraints of a specific type of distribution. For channel seismograph and a set of 24, 4.5 Hz vertical geo-
a Gaussian distribution m ¼ 2. phones, which were arranged in a linear array at a constant
 223
Lai et al.: Propagation of Uncertainty

Figure 1. Sample of experimental data at Saluggia Figure 2. Experimental results at Saluggia testing site:
testing site: (a) Shot gather (b) Frequency-wavenumber (a) Ensemble of 15 independent dispersion curves (b)
spectrum. Coefficient of variation of surface wave phase velocities.

experimental measurements and the uncertainty associated

with the mean dispersion curve.
receiver spacing of 3 m for Saluggia and 2 m for Florence.
A 1.3 kN weight dropped from a height of about 3 m was
used as the active source. The distance of the source from Results Obtained at Saluggia Site
the first receiver was 4 m at both testing sites.
Ensembles of 15 and 11 repetitions of the experiment Figure 1 shows an example of experimental data in
were performed at Saluggia and Florence, respectively. For terms of shot gather and the corresponding f-k spectrum,
each one of these independent measurements, the signals while Fig. 2a reports the ensemble of 15 independent
detected by the geophones were transformed into the dispersion curves. The corresponding coefficient of variation
frequency-wavenumber domain using a 2D Fast Fourier ranges from 1.1 to about 5% for frequencies less than 12.5
Transform algorithm. The corresponding dispersion curves Hz, while at higher frequencies the uncertainty is lower with
were determined from the loci of the absolute maxima of a coefficient of variation ranging from 0.23 to about 1.25%
the frequency-wavenumber spectra (Tselentis and Delis, (see Fig. 2b). Thus, as far as the uncertainty is concerned, the
1998). This procedure yielded ensembles of 15 dispersion dispersion curve appears to be separated in two regions: the
curves for Saluggia and 11 for Florence, which were subse- first at low frequency characterized by moderate values of
quently used to determine the statistical distribution of the the uncertainty and a region at higher frequencies where
224
Journal of Environmental and Engineering Geophysics

Figure 3. Fourier spectra of external background

noise at Saluggia testing site during the surface wave
measurement session.

the uncertainty is lower. The threshold frequency between

these two regions is located at about 12.5 Hz.
The same pattern has been reported by Tuomi and
Hiltunen (1996) with respect to the variation of the uncer-
tainty of phase angle in two-station surface wave testing.
Similar results have also been obtained with numerical
simulations of surface wave tests using full P-SV wavefield
synthetics, supported by multi-station surface wave tests
with various source and receiver combinations over co-
incident lines (O’Neill, 2004).
In Fig. 2b the coefficient of variation of surface phase Figure 4. Statistical distribution of experimental
velocity shows the presence of three outliers at frequencies surface wave phase velocities at Saluggia testing site:
ranging from 11.5 and 12.5 Hz. This larger uncertainty in (a) Relative frequency distribution at f = 11.5 Hz,
the experimental data is caused by the presence of external, compared with the values predicted by a Gaussian
background noise whose Fourier spectra exhibit peaks in distribution (b) Results of v2-test.
this frequency range (see Fig. 3).
The relative and cumulative frequency distribution of The expected experimental dispersion curve and its
the experimental phase velocities was computed for each associated uncertainty were inverted using Occam’s algo-
frequency and arranged in 5 homogeneous classes. The ele- rithm. Figure 5a and Table 1 report the expected shear
ments of each class were then compared with those obtained wave velocity profile obtained from the inversion together
for a Gaussian distribution with mean and standard deviation with an estimate of the standard deviation computed using
derived from the experimental data. Figure 4a shows an Eq. (6). As expected, the uncertainty of the VS(z) profile
example of the relative frequency distribution of phase veloci- yielded by the inversion algorithm through the mapping
ties for f ¼ 11.5 Hz. The comparison between the experi- Var(VR) 7! Var(VS) is very low with a coefficient of
mental and the theoretical distributions appears reasonable. variation ranging from 0.3 to 0.4%, apart from the shallowest
Figure 4b shows the plot of the quantity X2 computed layer. This is a natural consequence of: a) the stability of
with the experimental data set from Saluggia as a function of Occam’s algorithm, b) the very low values of the uncertainty
frequency. From the plot it can be noted that all the points fall associated with the experimental dispersion curve (see Fig.
below v0:05 2 , confirming that Gaussian distribution is a reason- 2b) at frequencies greater than 12.5 Hz (with a coefficient of
able assumption for this dataset over the whole frequency variation varying from 0.3 to 1.2%), and c) the damping
range of interest. This finding is not in agreement with the effect produced by the smoothing parameter l in Eq. (6) that
results of O’Neill (2003) showing a much broader distribution, tends to yield smooth profiles with low variance. The higher
best fitted with a Lorentzian curve in the low-frequency range. value of uncertainty associated with the first layer is likely
 225
Lai et al.: Propagation of Uncertainty

Table 1. Shear wave velocity profile obtained for

Saluggia site (std = standard deviation; CoV = coef-
ficient of variation).

Thickness Vs mean Vs std Vs CoV

Layer (m) (m/s) (m/s) (%)
1 0.9 167 2.1 1.2
2 1.5 201 0.5 0.3
3 2.1 303 1.2 0.4
4 2 314 1.2 0.4
5 3.4 337 1.2 0.4
6 5.9 467 1.6 0.4
7 6 607 1.8 0.3
8 — 691 3.0 0.4

algorithm after the seventh and final iteration. The

theoretical dispersion curve is in reasonable agreement with
the curve measured experimentally.
Figure 5a also shows the results of a cross-hole test
(CHT) performed at the same site. The VS(z) profile deter-
mined with the surface wave method compares well with
the values of shear wave velocities measured with the CHT
test up to a depth of about 18 m. At depths greater than 18 m
the two methods differ by as much as 100 to 200 m/s. The
results shown in Fig. 5a provide an opportunity to
emphasize the difference between accuracy and reliability
in experimental measurements (Lai et al., 2001). Accuracy
may be defined as the probability that the expected value
of a measurement (in this case E[VS(z)]) determined with
a certain technique is equal to the true value. Accuracy of a
measurement may be affected by bias and systematic errors
(Tuomi and Hiltunen, 1996). It is frequently difficult to
assess accuracy because the true value of a parameter cannot
be known with certainty. The present study focuses on
reliability, which represents an estimate of the dispersion or
scatter of the measurement around its expected value (in this
case Var[VS(z)]). The reliability of the surface wave mea-
Figure 5. Inversion of surface wave data at Saluggia surement at the Saluggia site is excellent as indicated by the
testing site: (a) Expected shear wave velocity profile and low values of the coefficient of variation.
estimate of the standard deviation including comparison
with cross-hole test results; (b) Comparison between
expected experimental and theoretical dispersion curves Results Obtained at Florence Site
at the final iteration.
Figure 6 shows an example of experimental data in
terms of shot gather and the corresponding f-k spectrum,
due to the lack of information in the high-frequency range while Fig. 7a reports the ensemble of 11 independent
that makes this layer not well resolved. dispersion curves. The experimental dispersion curve can be
It is also important to note the reduced influence that subdivided into two regions in a manner similar to the
the noise concentrated at 12 Hz has on the final results, Saluggia site: a high-frequency region with very low values
showing the robustness of the method in noisy environments. of the coefficient of variation and a low-frequency region
Figure 5b illustrates the match between the experi- characterized by higher values (see Fig. 7b). The threshold
mental dispersion curve and the corresponding theoretical frequency between the two regions is 10.5 Hz. In the high-
curve (fundamental mode) yielded by the inversion frequency region the coefficient of variation ranges from
226
Journal of Environmental and Engineering Geophysics

Figure 6. Sample of experimental data at Florence Figure 7. Experimental results at Florence testing
testing site: (a) Shot gather (b) Frequency-wavenumber site: (a) Ensemble of 11 independent dispersion curves
spectrum. (b) Coefficient of variation of surface wave phase
velocities.
0.092 to about 0.53%, whereas in the low-frequency region
the variation is from 1.59 to 13.9% (see Fig. 7b). The highest
value of the coefficient of variation is that associated with the Concluding Remarks
lowest frequency (i.e., 5.6 Hz).
Using Occam’s algorithm, the experimental dispersion Although in recent years surface wave methods have
curve and associated uncertainty were successfully inverted received a great deal of attention by both the geophysical and
to determine the shear wave velocity profile illustrated in geotechnical communities, thus far very few attempts have
Fig. 8a and reported in Table 2. The error bars attached to been made to assess the level of confidence associated with
VS(z) represent the standard deviation of the expected shear this non-invasive seismic technique. This article illustrated
wave velocity profile with a coefficient of variation ranging a methodology to estimate the uncertainty in surface wave
from 0.2 to 3.7%. measurements due to ambient noise. The approach followed
To complete the results of the inversion, Fig. 8b to quantify the uncertainty in surface wave measurements
shows the agreement between the experimental dispersion makes use of the classical tools of discrete inverse theory and
curve and the corresponding theoretical curve (fundamental reliability analysis. With the assumption that the experimen-
mode) yielded by Occam’s inversion algorithm after the tal data are Gaussian distributed (which is shown to be
ninth and final iteration. reasonable), it is illustrated how uncertainties propagate from
 227
Lai et al.: Propagation of Uncertainty

Table 2. Shear wave velocity profile obtained

for Florence site (std = standard deviation; CoV =
coefficient of variation).

Thickness Vs mean Vs std Vs CoV

Layer (m) (m/s) (m/s) (%)
1 1.5 211 0.3 0.2
2 2 196 0.6 0.3
3 2.5 236 1.9 0.8
4 2.5 193 3.8 2.0
5 2.5 224 5.3 2.4
6 2.5 320 6.6 2.1
7 3 360 8.6 2.4
8 — 370 13.7 3.7

The threshold frequency at the two sites is 12.5 and

10.5 Hz, respectively.
— The uncertainty associated with the expected shear wave
velocity profiles determined using Occam’s algorithm is
low with a coefficient of variation less than 4%. This
result is a natural consequence of the stability of Occam’s
algorithm and the damping effect produced by the
smoothing parameter l in Eq. (6). The larger uncertainty
of the experimental dispersion curves at low frequencies
contributes to the observed increase of uncertainty with
depth of the expected shear wave velocity profiles.
— The presence of a few outliers in the experimental
dispersion curve has little influence on the results of the
inversion using Occam’s algorithm thereby demonstrat-
ing that the latter is not only stable but also robust.
— Finally, it is very important to note that the low
uncertainty in the final results shows the robustness of
surface wave methods in presence of even strong
ambient noise. This result is of particular interest for
applications in urban areas, where other geophysical tests
Figure 8. Inversion of surface wave data at Florence may encounter difficulties. This advantage of surface
testing site: (a) Expected shear wave velocity profile and wave tests is mainly associated with the fact that the
estimate of the standard deviation; (b) Comparison analysis is performed in the frequency domain rather
between expected experimental and theoretical disper- than in the time domain.
sion curves at the final iteration.

Acknowledgments
the experimental dispersion curve into levels of confidence of
This material is based upon work supported by the National
the shear wave velocity profile.
Science Foundation under Grant No. CMS-9616013. The authors
The methodology has been applied at two sites in Italy
would like to acknowledge the cooperation of Prof. L. Sambuelli,
where surface wave measurements were performed using
Dr. L.V Socco and Dr. C.L. Strobbia in performing surface wave
the multi-station method. The main findings obtained from
measurements at the two testing sites in Saluggia and Florence.
the study can be summarized as follows:
— The experimental dispersion curves appear to be
subdivided in two regions separated by a threshold References
frequency: a low-frequency region characterized by
higher values of uncertainty and a high-frequency region Constable, S.C., Parker, R.L., and Constable, C.G., 1987, Occam’s
characterized by lower values of coefficient of variation. inversion: A practical algorithm for generating smooth
228
Journal of Environmental and Engineering Geophysics

models from electromagnetic sounding data: Geophysics, Parker, R.L., 1994, Geophysical inverse theory: Princeton
32(3) 289–300. University Press, New Jersey, 386 pp.
Foti, S., 2000, Multistation methods for geotechnical character- Rix, G.J., Lai, C.G., and Spang, A.W. Jr., 2000, In situ mea-
ization using surface waves: Ph.D dissertation, Politecnico surement of damping ratio using surface waves: Journal of
di Torino, Italy. Geotechnical and Geoenvironmental Engineering, 126(5)
Gabriels, P., Snieder, R., and Nolet, G., 1987, In situ measure- 472–480.
ments of shear-wave velocity in sediments with higher- Rix, G.J., and Lai, C.G., 2000, Software package tools for surface
mode Rayleigh waves: Geophys. Prospect., 35, 187–196. wave analysis. Available at the WEB site: http://www.ce.
Lai, C.G., and Rix, G.J., 1998, Simultaneous inversion of Ray- gatech.edu/;grix/surface_wave.html#Software.
leigh phase velocity and attenuation for near-surface site Rix, G.J., Lai, C.G., Orozco, M.C., Hebeler, G.L., and Roma, V.,
characterization: Technical Report GIT-EE/GEO-98-2, 2001, Recent advances in surface wave methods for geotech-
Georgia Institute of Technology. nical site characterization: Proceedings of the XV Interna-
Lai, C.G., Foti, S., Godio, A., Sambuelli, L., Socco, V., and tional Conference on Soil Mechanics and Geotechnical
Strobbia, C., 2001, Geophysical investigations: Recent exe- Engineering–Istanbul, Turkey, August 27–31, 2001.
cution and interpretation procedures: Proceedings, XVIII Snedecor, G.W., and Cochran, W.G., 1980, Statistical methods,
Conferenze di Geotecnica di Torino, CGT 2001, November 7th ed.: Iowa State University Press, Ames, 507 pp.
20–22, 2001 (in Italian). Spiegel, M.R., Schiller, J.J., and Srinivasan, R.A., 2000, Pro-
Lai, C.G., Rix, G.J., Foti, S., and Roma, V., 2002, Simultaneous bability and statistics, 2nd ed.: McGraw-Hill, New York,
measurement and inversion of surface wave dispersion and 408 pp.
attenuation curves: Soil Dynamics and Earthquake Engi- Strobbia, C., and Foti, S., 2005, Multi-offset phase analysis
neering, 22(9–12) 923–930. of surface wave data (MOPA): submitted to J. Applied
McMechan, G.A., and Yedlin, M.J., 1981, Analysis of disper- Geophysics.
sive waves by wave field transformation: Geophysics, 46, Tarantola, A., 1987, Inverse problem theory: Methods for data
869–874. fitting and model parameter estimation: Elsevier, Amster-
Menke, W., 1989, Geophysical data analysis: Discrete inverse dam, 613 pp.
theory: Academic Press, San Diego, 260 pp. Tselentis, G.A., and Delis, G., 1998, Rapid assessment of S-wave
Marosi, K.T., and Hiltunen, D.R., 2004a, Characterization of profiles from the inversion of multichannel surface wave
SASW phase angle and phase velocity measurement dispersion data: Annali di Geofisica, 41, 1–15.
uncertainty: Geotechnical Testing Journal, ASTM, 27(2) Tuomi, K.E., and Hiltunen, D.R., 1996, Reliability of the SASW
March, 205–213. method for determination of the shear modulus of soils, in
Marosi, K.T., and Hiltunen, D.R., 2004b, Characterization of Uncertainty in geologic environment: From theory to practice,
SASW shear wave velocity measurement uncertainty: C.D. Shacklford, P.P. Nelson, and M.J.S. Roth (eds.), ASCE,
Journal of Geotechnical and Geoenvironmental Engineering, 1125–1237.
ASCE, 130(10) 1034–1041. Yilmaz, O., and Doherty, S.M., 2003, Seismic data analysis:
O’Neill, A., 2003, Full-waveform reflectivity for modelling, Processing, inversion, and interpretation of seismic data
inversion and appraisal of surface wave dispersion in (Investigations in Geophysics, No. 10). In-press.
shallow site investigations: Ph.D. dissertation, University Zywicki, D., and Rix, G.J., 1999, Frequency-wavenumber analysis
of Western Australia, Perth. of passive surface waves: Proc. Symp. on the Appl. of
O’Neill, A., 2004, Shear velocity model appraisal in shallow Geophysics to Environm. and Eng. Problems, Oakland,
surface wave inversion: Proc. SAGEEP 2004, 1544–1555. 75–84.
Park, C.B., Miller, R.D., and Xia, J., 1999, Multichannel analysis
of surface waves: Geophysics, 64, 800–808.