992

ARTICLE
Sample size determination in geotechnical site investigation
considering spatial variation and correlation
Yu Wang, Zheng Guan, and Tengyuan Zhao
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21

Abstract: Site investigation is a fundamental element in geotechnical engineering practice, but only a small portion of geoma-
terials is sampled and tested during site investigation. This leads to a question of sample size determination: how many samples
are needed to achieve a target level of accuracy for the results inferred from the samples? Sample size determination is a
well-known topic in statistics and has many applications in a wide variety of areas. However, conventional statistical methods,
which mainly deal with independent data, only have limited applications in geotechnical site investigation because geotechnical
data are not independent, but spatially varying and correlated. Existing design codes around the world (e.g., Eurocode 7) only
provide conceptual principles on sample size determination. No scientific or quantitative method is available for sample size
determination in site investigation considering spatial variation and correlation of geotechnical properties. This study performs
an extensive parametric study and develops a statistical chart for sample size determination with consideration of spatial
variation and correlation using Bayesian compressive sensing or sampling. Real cone penetration test data and real laboratory
test data are used to illustrate application of the proposed statistical chart, and the method is shown to perform well.

Key words: geotechnical site investigation, sample size, Bayesian method, compressive sensing, random field.

Résumé : L'étude du site est un élément fondamental de la pratique de l’ingénierie géotechnique, mais seule une petite partie
des géomatériaux est échantillonnée et testée au cours de l’étude du site. Cela conduit à une question de détermination de la
For personal use only.

taille de l’échantillon : combien faut-il d’échantillons pour atteindre un niveau cible d’exactitude pour les résultats déduits des
échantillons? La détermination de la taille des échantillons est un sujet bien connu en statistique et a de nombreuses applica-
tions dans de nombreux domaines. Cependant, les méthodes statistiques conventionnelles, qui traitent principalement de
données indépendantes, n’ont que des applications limitées dans l’étude géotechnique de sites, car les données géotechniques
ne sont pas indépendantes, mais varient dans l’espace et en corrélation. Les codes de conception existants dans le monde
(Eurocode 7, par exemple) ne fournissent que des principes conceptuels pour la détermination de la taille de l’échantillon.
Aucune méthode scientifique ou quantitative n’est disponible pour la détermination de la taille de l’échantillon dans l’étude du
site, compte tenu de la variation spatiale et de la corrélation des propriétés géotechniques. Cette étude réalise une évaluation
paramétrique approfondie et développe un tableau statistique pour la détermination de la taille de l’échantillon en prenant en
compte la variation et la corrélation spatiales à l’aide de la détection ou de l’échantillonnage bayésien en compression. Les
données des tests de pénétration au cône réel et les données des tests de laboratoire réels sont utilisées pour illustrer
l’application du diagramme statistique proposé, et il est démontré que la méthode donne de bons résultats. [Traduit par la
Rédaction]

Mots-clés : étude géotechnique du site, taille de l’échantillon, méthode bayésienne, détection de compression, champ aléatoire.

Introduction
Site investigation is a fundamental element in geotechnical en-
gineering practice, and interpretation of site investigation data
leads to the expected ground condition (e.g., spatial distribution
of different soil types) and design profiles of geotechnical proper-
ties that are subsequently used in geotechnical analyses and de-
signs. Without proper site investigation or data interpretation to
produce reliable input parameters, the subsequent geotechnical
analyses and designs are much less meaningful, and the geotech-
nical construction projects might be subjected to significant risk.
Clayton (2001) conducted a survey in the UK and found that about
42% of the problems that occurred in geotechnical construction
projects were caused by uncertainty associated with site investi-
gation. During geotechnical investigation of a site, only a small
portion of geomaterials is examined, and site investigation data
are often sparsely measured (e.g., Mayne et al. 2002; Phoon 2017;
Wang et al. 2017; Orr 2017; Zhao et al. 2018). This results in estima-
tion error in the interpretation results and leads to a question of
sample size determination: how many samples are needed to
achieve a target level of accuracy for the results inferred from the
samples?
Sample size determination is a well-known topic in statistics
(e.g., Krejcie and Morgan 1970) and has many applications in a
wide variety of areas (e.g., quality control, surveys, and polls).
However, conventional statistical methods, which mainly deal
with independent data, only have limited application in geotech-
nical site investigation, because geotechnical data are not inde-
pendent, but spatially correlated. Note that soils are natural
geomaterials, whose properties are affected by many spatially
varying but correlated factors during the geological process, such

Received 13 July 2018. Accepted 2 October 2018.

Y. Wang,* Z. Guan, and T. Zhao. Department of Architecture and Civil Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong.
Corresponding author: Yu Wang (email: yuwang@cityu.edu.hk).
*Y. Wang currently serves as an Editorial Board Member; peer review and editorial decisions regarding this manuscript were handled by C. Lake.
Copyright remains with the author(s) or their institution(s). Permission for reuse (free in most cases) can be obtained from RightsLink.

Can. Geotech. J. 56: 992–1002 (2019) dx.doi.org/10.1139/cgj-2018-0474 Published at www.nrcresearchpress.com/cgj on 3 October 2018.
 Wang et al.
as the properties of their parent materials, weathering and ero-
sion processes, transportation agents, and sedimentation condi-
tions. Geotechnical properties are therefore spatially varying and
correlated (e.g., Lacasse and Nadim 1996; Phoon and Kulhawy
1999; Wang et al. 2016). To deal with spatially varying and corre-
lated data, a suite of specialized statistical methods called geo-
statistics has been developed, originally for ore estimation.
Geostatistics generally require a large number of data points mea-
sured at preferably regular intervals for proper construction of
semi-variogram (e.g., Wang et al. 2017). Geotechnical data are,
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21
however, often sparsely measured during site investigation, and
the number of measured data points may not be sufficient for
proper application of geostatistics (e.g., Baecher and Christian
2003).
Currently there is no scientific or quantitative method to deter-
mine sample size in geotechnical site investigation. Existing de-
sign codes around the world only provide conceptual principles
on sample size determination. For example, clause 2.4.2.4 in
Eurocode 7 (CEN 2007) states that “The necessary number of spec-
imens to be tested shall be established depending on the homo-
geneity of the ground, the quality and amount of comparable
experience with the ground and the geotechnical category of the
problem.” No detailed or quantitative method has been provided
on how to determine the sample size in engineering practice. A
straightforward way to bypass this problem is to increase the
sample size, or even, to test the geomaterials at every location in
a site. However, this substantially increases investigation costs,
and the manpower and time required. Even if additional cost and
manpower are available, access may not be possible for a long
For personal use only.
period of site investigation, particularly in urban areas where the
ground surface is already occupied with buildings and ongoing
activities.
Some attempts to tackle the sample size determination prob-
lem have been reported in geotechnical literature. For example,
statistical methods have been developed for determining the re-
quired sample size for rock strength tests (Yamaguchi 1970; Gill
et al. 2005; Cui et al. 2017) or construction of compacted soil liners
(Benson et al. 1994), but the assumption of independent samples is
still adopted (i.e., no consideration of spatial correlation) in these
studies. More recently, Fenton et al. (2015) and Liza et al. (2017)
developed statistical methods to determine the sample size for
quality control of cement-based solidification–stabilization and to
check whether or not the sample average of the hydraulic conduc-
tivity satisfies the regulatory hydraulic conductivity. Their meth-
ods rely on random finite element simulation of flow and do not
aim at design profiles of geotechnical properties (i.e., spatial vari-
ation of geotechnical properties), which is the objective of site
investigation. In addition, the effect of sampling location has also
been explored (e.g., Goldsworthy et al. 2007).
This paper aims at developing a rational method and statistical
chart for sample size determination in geotechnical site investi-
gation, with consideration of spatial variation and correlation in
soil property. An extensive parametric study is performed to de-
velop a relationship between sample size and interpretation ac-
curacy with consideration of spatial variation and correlation.
After this introduction, development of the relationship is de-
scribed, followed by development of the statistical chart for sam-
ple size determination. Examples of using the proposed statistical
chart are also provided.
Development of relationship between sample size
and interpretation accuracy considering spatial
variation and correlation
To determine the sample size required to achieve a target level
of accuracy on the inferred geotechnical profiles (e.g., spatial vari-
ation of a soil property along depths), a relationship between the
desired level of accuracy and sample size should be established
993
Fig. 1. Framework for establishing relationship between sample
size and accuracy in geotechnical site investigation. [Colour online.]
Step 1: Suppose that there is a soil property profile, all
data points of which are known
Step 2: Obtain a number, M (e.g., M = 5), of measured data
points from this given soil property profile
Step 3: Interpolate the M measured data points to infer a
Stage 1: for a complete interpreted soil property profile using Bayesian
given soil compressive sampling (BCS)
property profile
Step 4: Compare the interpreted soil property profile with
the original one, and use relative error to evaluate the
accuracy of results obtained from M measured data points
Step 5: Repeat Steps 2-4 for many different M values
Step 6: Establish a relationship between sample size M and
relative error for a given soil property profile
Step 7: Suppose that there is a given site where many (e.g.,
50) sets of complete soil property profiles are available.
Repeat Steps 1 to 6 for each set of the soil property profiles
Stage 2: for a
given site
in this site.
Step 8: Calculate statistics (e.g., mean) of relative error at
different sample sizes, and establish a relationship between
sample size and relative error statistic for the given site
Step 9: Suppose that there are a wide variety of typical sites
commonly encountered in geotechnical site investigation.
Repeat the Steps 7 to step 8 for each of the sites.
Stage 3: for a wide
variety of sites
Step 10: Summarize the results for all the sites above
and develop a relationship between normalized sample
size and relative error statistic
first. Figure 1 shows a conceptual framework for developing such
a relationship in geotechnical site investigation. As shown in
Fig. 1, the framework mainly involves three stages: (1) the relation-
ship between the desired level of accuracy and sample size for a
given soil property profile, (2) the relationship for a given site, and
finally (3) the relationship for a wide variety of typical sites com-
monly encountered in geotechnical site investigation. Note that
Fig. 1 only deals with sampling along depth, and this study focuses
on one-dimensional cases.
In stage 1, it is instructive to firstly consider the required sample
size for one soil property profile (i.e., steps 1–6 in Fig. 1). Suppose
that there is a soil property profile that is completely known (i.e.,
all data points in the profile are known). To mimic the data mea-
surement and interpretation in geotechnical site investigation, a
number, M (e.g., M = 5), of data points is obtained from the soil
property profile and these points are used as measurement data
points to interpret a complete soil property profile using interpo-
lation methods, such as Bayesian compressive sampling (BCS)
(Wang and Zhao 2017). Because the original soil property profile is
completely known, the soil property profile interpreted from the
M measurement data points may be compared with the original
soil property profile to evaluate the accuracy of the interpreted
profile (i.e., steps 2–4 in Fig. 1). Let X = [X(D1), X(D2), …, X(DN)]T be a
vector representing the original soil property X profile along
depths D with a total number N of data points, and let a vector X̂ =
Published by NRC Research Press
 994
[X̂(D1), X̂(D2), …, X̂(DN)]T denotes the soil property profile inter-
preted from the M measurement data points. The accuracy of the
inferred geotechnical profile X̂ can be quantified by comparing a
relative error, RE, between X and X̂, which is expressed as
兺 size determination with consideration of spatial variation and
冪
N
[X(Di) ⫺ X̂(Di)]2
㛳X ⫺ X̂㛳 i⫽1
(1) RE ⫽ ⫽ × 100%
㛳X㛳
兺 In the proposed framework, a key element is BCS that is used to
N
X(Di)2
i⫽1
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21
where the symbol “㛳 · 㛳” denotes an operator of vector norm, which
is defined as the square root of the sum of squared value of all
elements, as shown on the right-hand side of eq. (1). Note that the
definition of RE in eq. (1) is similar to the normalized root-mean-
square error in statistics if the original soil property profile X is
treated as the mean. To investigate the influence of different sam-
ple sizes (i.e., different M values) on the accuracy of the inferred
geotechnical profile, steps 2–4 described above are repeated for
many different M values (e.g., M = 10), and the corresponding REs
are calculated (i.e., step 5 in Fig. 1). When the sample size M is
small, the soil property profile interpreted from such sparse mea-
surement data might not be accurate, and the RE is relatively
large. As M increases, the accuracy of the inferred geotechnical
profile improves, and the RE gradually decreases. When all data
points in the original soil property profile are treated as measure-
ment data points (i.e., M = N), the inferred geotechnical profile
converges to the original soil property profile, and the RE ap-
proaches zero. The relationship between RE and M for a given soil
For personal use only.
property profile can be established after steps 2–4 are repeated for
many different M values (i.e., step 6). This relationship quantifies
the variation of the estimation error as sample size changes for a
given soil property profile.
In stage 2 (i.e., steps 7–8 in Fig. 1), the study is extended from a
given soil property profile (i.e., stage 1) to a given site. Suppose
that there is a given site where many (e.g., 50 or 100) sets of
complete soil property profiles are available. The procedure devel-
oped in stage 1 (i.e., steps 1–6) may be used repeatedly to each set
of the complete soil property profiles in this site, resulting in
many sets of relationships between RE and M. Statistical analysis
can be performed to obtain the mean of RE, ␮RE, at different M
values for all the relationships obtained. This establishes a rela-
tionship between the interpretation accuracy (e.g., using ␮RE) and
sample sizes for a given site.
In stage 3 (i.e., steps 9–10 in Fig. 1), the study is further extended
from a given site (i.e., stage 2) to a wide variety of typical sites
commonly encountered in geotechnical site investigation. The
procedure developed in stage 2 (i.e., steps 7–8) may be used repeat-
edly for each typical site, resulting in many sets of relationships
between ␮RE and M. Because well-investigated and well-documented
sites are rare in reality, random field simulation (e.g., Vanmarcke
1977) is adopted in this study to generate a large number of typical
sites as a parametric study in stage 3. Three statistical parameters
are commonly needed in a random field simulation of spatial varia-
tion of a soil property: (i) mean ␮ of the soil property, (ii) variance or
coefficient of variation, COV, and (iii) correlation length ␭, which
reflects quantitatively spatial correlation of the soil property at
different depths. Typical ranges of ␮, COV, and ␭ for a variety of
soil properties have been reported in the literature (e.g., Phoon
and Kulhawy 1999; Cao et al. 2016), and these typical ranges are
used in the random field simulation in the parametric study to
generate a large number of typical sites that mimic spatially vary-
ing and correlated soil properties in real site conditions. The pro-
cedure developed in stage 2 (i.e., steps 7–8) is used repeatedly to
each typical site in the parametric study, resulting in many sets of
relationship between ␮RE and M for different site conditions. The
large number of ␮RE and M relationships obtained in the paramet-
ric study will be summarized in a dimensionless plot, and a rela-
Can. Geotech. J. Vol. 56, 2019
tionship between the normalized ␮RE and normalized M will be
developed from the dimensionless plot. The normalized ␮RE and
normalized M relationship provides a link between the accuracy
of the interpretation results and sample size, and this relationship
will be used subsequently to develop a statistical chart for sample
correlation of soil properties.
interpret a complete soil property profile from M measurement
data points (Wang and Zhao 2017). A brief review of BCS method is
provided in the next section.
Brief review of Bayesian compressive sampling
(BCS) method
The BCS method is a probabilistic version of compressive sens-
ing (CS), which is a novel sampling method that can reconstruct a
signal from sparse measurements on that signal (e.g., Candès et al.
2006; Donoho 2006; Tropp and Gilbert 2007; Ji et al. 2008, 2009;
Wang and Zhao 2016; Zhao and Wang 2018a). A signal may be
defined loosely as a quantity that exhibits variation with time or
space, such as a variation of soil property along depth. Therefore,
spatial variation of a soil property with depth can be interpreted
from sparse and limited measurement data using BCS method.
The BCS method is based on the fact that natural signals have
clear trends and are compressible. The term “compressible”
means that a signal can be concisely represented by a weighted
summation of a limited number of basis functions, such as wave-
let functions (e.g., Candès et al. 2006; Donoho 2006; Tropp and
Gilbert 2007).
Mathematically, let f = [f1, f2, …, fN]T be a soil property profile
with a length of N, which denotes spatial variation of a soil prop-
erty with depths. The measurement data drawn from f is defined
as y = [y1, y2, …, yM]T, which is a column vector with a length of M,
where M < N. The soil property profile interpreted from the BCS
method is denoted as f̂ ⫽ 关f̂1, f̂2, ..., f̂N兴T, which can be expressed as
(2) f̂ ⫽ B␻s
where B is an orthonormal matrix with a dimension of N × N, each
column of which represents a basis function; ␻s is a weight col-
umn vector with a length of N, corresponding to N basis functions
in B. Note that matrix B is independent of the soil property profile
f, and that B can be constructed using discrete wavelet transform
(DWT) (e.g., Daubechies 1992).
Wang and Zhao (2017) developed a Bayesian framework to
statistically reconstruct the soil property profile f̂ from sparsely
measured data. Using the Bayesian framework, the posterior
distribution of ␻s is derived as a multivariate Student’s t distribu-
tion with a degree of 2cn, where cn = M/2 + c, with c being a very
small nonnegative number. The mean and covariance matrix of
␻s are expressed as (Wang and Zhao 2017)
(3) ␮␻s ⫽ HATy
dnH
(4) COV␻s ⫽
cn ⫺ 1
where H = (ATA + D)−1, A = ⌿B (where ⌿ is a measurement matrix
that represents the locations of components of y in f), and D is a
diagonal matrix with diagonal components ␣i (i = 1, 2, …, N) that
can be obtained by maximizing the likelihood of y (i.e., p(y); cn =
M/2 + c, dn ⫽ d ⫹ 共yTy ⫺ ␮␻sT⌯⫺1␮␻s兲/2, where c and d are very small
nonnegative number (e.g., c = d = 10−4). As f̂ = B␻s, the mean and
covariance of f̂ are expressed as
Published by NRC Research Press
 Wang et al.
Table 1. Summary of real CPT data used in illustration example.
Min. Max. No. of
depth depth Thickness data
Soil type CPT No. (m) (m) (m) points
Keswick clay CD1⬃CD50 2.01 4.56 2.55 512
(5) ␮f̂ ⫽ B␮␻s
(6) COVf̂ ⫽ BCOV␻sBT
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21
The BCS method is able to not only provide an expected (or
mean) soil property profile interpreted from sparse measurement
data, but also quantify estimation error of the inferred geotechni-
cal profile (e.g., Zhao et al. 2018). Wang and Zhao (2017) showed
that, as the number M of the measured data points increases, the
BCS mean soil property profile gradually converges to the original
soil property profile and the statistical uncertainty gradually re-
duces and approaches a negligible value when all data points in
the original soil property profile are measured. Note that mea-
surement error has been explicitly modeled in BCS as a Gaussian
random variable with a zero mean and unknown variance, which
is subsequently integrated through a marginalization in the
Bayesian framework. Because BCS has the unique capability of
interpolating sparsely measured data and quantifying the evolu-
tion of interpolation error as the number of sparse measurements
increases, BCS is used in this study to develop the relationship
For personal use only.
between interpretation accuracy and sample size. In the next sec-
tions, stages 1 and 2 of the proposed framework are illustrated
using BCS and real cone penetration test (CPT) data.
Illustration of stage 1 using real CPT data
To illustrate stages 1 and 2 of the proposed framework, 50 sets of
real CPT tip resistance qc data are taken from Jaksa et al. (1999) and
used in this study. Because CPT data are almost continuous
and have a measurement interval as low as 0.005 m, the high-
resolution qc profile may be used to compare with the qc profile
reconstructed from BCS method with a limited number M of mea-
sured qc data points as input, as described in stages 1 and 2 of the
proposed framework. A large number of CPTs was carried out in a
relatively homogeneous, stiff and overconsolidated clay layer in
Adelaide, Australia (Jaksa et al. 1999). The clay is known as Kes-
wick clay, and details of the 50 sets of CPT data are summarized in
Table 1. The 50 sets of CPT tip resistance qc data used in this section
were downloaded from the ISSMGE TC304 database website (http://
140.112.12.21/issmge/Database_2011.htm). According to Jaksa and
Kaggwa (1994), the depth of Keswick clay is from about 1.5 to about
5 m below the ground surface at this site. To guarantee that all
CPT data used in this study come from the Keswick clay, CPT data
ranging from 2.01 to 4.56 m were used in this section. The CPT qc
data were recorded with a sampling interval of 0.005 m, and the
total number of data points for each set CPT data is N = 512. In
addition, the 50 sets of CPTs were performed with a small hori-
zontal interval of 0.5 m, and each of the 50 sets of CPT data may be
considered as a random sample of a statistically homogeneous
one-dimensional random field that represents spatial variations
of qc with depths in the homogenous Keswick clay.
To illustrate stage 1 in Fig. 1 (i.e., steps 1–6), different numbers of
measured data points (i.e., sample size), M = 5, 10, 15, 20, 25, 30, 35,
40, 45, 50, 60, and 80, are taken from the CPT No. CD 5 and used as
input to the BCS method for interpreting the complete qc data
profile. For simplicity, the measured data points are taken with an
equal interval. Figure 2 shows the interpretation results for M = 10,
15, 20, 60. In this figure, the solid lines represent the original qc
profile, and the open circles represent the measurement data
taken from the original qc profile and used as input to the BCS
995
method. The dashed lines indicate the mean qc profiles inter-
preted from the BCS method, and the statistical uncertainty of the
interpreted qc profiles is represented by an interval with two dot-
ted lines that are the mean qc profile plus or minus one standard
deviation of qc obtained from the BCS method. The corresponding
relative error, i.e., RE, for each M value (e.g., M = 10, 15, 20, 60) are
calculated using eq. (1) as RE = 8.29%, 6.28%, 3.71%, and 1.90%,
respectively. Figure 2 shows that RE decreases significantly as the
number M of measured data points increases. In other words, the
accuracy of the interpreted qc profile from BCS method improves
as M increases. When M increases to 60, out of a total number of
data points N = 512, the mean qc profile interpreted from BCS
method almost completely overlaps with the original complete qc
profile. In addition, the interval within two dotted lines decreases
as M increases, and the statistical uncertainty of interpreted qc
data profile also decreases with the increase of M.
The REs for different number of measurement data points are
summarized in Fig. 3 (i.e., step 6 in Fig. 1). The vertical axis in Fig. 3
represents the RE value, while the horizontal axis denotes the
number of measured data points or sample size. It shows that the
RE decreases from about 10% to about 1% as the number of mea-
surement data increases from 5 to 80. Figure 3 establishes a rela-
tionship between sample size M and RE for the qc profile of CPT
No. CD5.
Illustration of stage 2 using real CPT data
To illustrate stage 2 of the proposed framework, steps 1–6 in
stage 1 are repeated for each of the 50 sets of CPT data in the
Keswick clay layer. For each set of the CPT qc profile, different
number M of measured data points, i.e., M = 5, 10, 15, 20, 25, 30, 35,
40, 45, 50, 60, and 80, are taken from the respective original qc
profile and used repeatedly as input to the BCS method to inter-
pret the corresponding complete qc profiles. For example, Fig. 4
shows the original qc profiles for CPT No. CD1 to CD11 by gray solid
lines, together with the complete qc profiles interpreted from BCS
method using M = 10, 20, and 60 measured data points taken from
each set of CPT data as input, respectively. The interpreted qc
profiles are shown by dashed lines, dotted lines, and dashed dot-
ted lines for M = 10, 20, and 60, respectively. Similar to Fig. 2, Fig. 4
shows that, as M increases, the interpreted qc profiles gradually
approach the respective original qc profiles. Similar to Fig. 3, a
relationship between RE and M may be obtained for each set of
CPT data.
Figure 5 summarizes the RE and M relationships for the 50 sets
of CPT data in the Keswick clay layer. Note that the horizontal axis
of Fig. 5 is sampling interval, ds. Because thickness of the Keswick
clay layer in which the 50 sets of CPTs were performed is known
and the measured data points are taken with an equal interval,
sampling interval ds may be determined as the thickness (e.g.,
2.55 m in this illustration) divided by (M – 1). For example, M = 5,
10, 15, 20, 25, 30, 35, 40, 45, 50, 60, and 80 leads to ds = 0.64, 0.28,
0.18, 0.13, 0.11, 0.09, 0.08, 0.07, 0.06, 0.05, 0.04, and 0.03 m, respec-
tively. Figure 5 shows REs for the 50 sets of CPT data with ds = 0.64,
0.28, 0.18, 0.13, 0.11, 0.09, 0.08, 0.07, 0.06, 0.05, 0.04, and 0.03 m,
respectively. It is obvious in Fig. 5 that RE increases as ds increases
or M decreases. In addition, variability of RE also increases as ds
increases. At each ds value, statistical analysis is performed to
estimate the mean, ␮RE, of the 50 RE values obtained for the
50 sets of CPT data. A variation of ␮RE as a function of ds is shown
in Fig. 6. The mean RE, ␮RE increases from about 1.7% to about
11.2% as the sampling interval increases from 0.03 to 0.64 m.
Figure 6 establishes a relationship between the interpretation
accuracy (e.g., ␮RE) and sample size (i.e., sampling interval ds) for
the given Keswick clay layer.
In an extensive parametric study reported in the next section,
stages 1 and 2 illustrated in these two sections are repeated for a
wide variety of sites typically encountered in geotechnical prac-
Published by NRC Research Press
 996 Can. Geotech. J. Vol. 56, 2019

Fig. 2. Interpreted qc profiles from BCS method for different number M of measured data points: (a) M = 10; (b) M = 15; (c) M = 20; (d) M = 60 for
CPT No. CD5. [Colour online.]

qc (MPa) qc (MPa)
1.5 2.0 2.5 3.0 1.5 2.0 2.5 3.0
2.0 2.0

2.5 2.5
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21

3.0 RE=8.29% 3.0 RE=6.28%

Depth (m)

Depth (m)
3.5 3.5

4.0 4.0

4.5 4.5
Original qc profile Original qc profile
5.0 BCS mean qc profile 5.0 BCS mean qc profile
BCS mean qc profile ± 1 σ BCS mean qc profile ± 1 σ
Measurement data y Measurement data y
5.5 5.5

(a) (b)
qc (MPa) qc (MPa)
1.5 2.0 2.5 3.0 1.5 2.0 2.5 3.0
For personal use only.

2.0 2.0

2.5 2.5

3.0 3.0
RE=3.71% RE=1.90%
Depth (m)

Depth (m)

3.5 3.5

4.0 4.0

4.5 4.5
Original qc profile Original qc profile
5.0 BCS mean qc profile 5.0 BCS mean qc profile
BCS mean qc profile ± 1 σ BCS mean qc profile ± 1 σ
Measurement data y Measurement data y
5.5 5.5

(c) (d)

tice. Because well-characterized and well-documented sites are
rare in reality, random field simulation is used to generate a large
number of typical sites that mimic spatially varying and corre-
lated soil properties in real site conditions.
Parametric study using random field simulation
data (stage 3)
Random field simulation of soil property profiles
Random fields have been successfully used to model spatially
correlated geotechnical properties during the last several decades
(e.g., Vanmarcke 1977; Vanmarcke et al. 1986; Phoon and Kulhawy
1999; Fenton and Griffiths 2008; Bong and Stuedlein 2018; Wang
et al. 2018; Zhao and Wang 2018b). In this study, soil property
profiles from a wide variety of statistically homogenous and sta-
tionary soil layers are simulated using random fields. Consider,
for example, a one-dimensional stationary Gaussian random field
X(D), where D is depth and X is a random variable representing a
soil property of interest with a mean ␮ and a standard deviation ␴.
The spatial correlation between X(D) at different depths is mod-
eled by a correlation function ␳(⌬D). A single exponential correla-
tion function (SECF) is adopted in this study. For a SECF, the
correlation function ␳(⌬D) between X(Di) and X(Dj) at the respec-
tive depth of Di and Dj may be expressed as
(7) ␳(⌬D) ⫽ exp(⫺2|⌬D|/ ␭)
in which ␭ is the correlation length, ⌬D is the separation distance
between Di and Dj (= |Di – Dj|).

Published by NRC Research Press

 Wang et al.
Fig. 3. Relationship between number, M, of measured data points
and relative error for CPT No. CD5.
15
Relative error, RE (%)
10
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21
5 dimensionless parameters, normalized mean relative error ␮RE/
0 with respect to the spatial correlation (i.e., correlation length).
0 10 20 30 40 50 60 70 80
Number of measured data points, M
Let X = [X(D1), X(D2), …, X(DN)]T be a vector of X(D) at N different
depths. X is a Gaussian vector with a mean vector ␮IN×1 and cova-
riance matrix C = ␴2R. IN×1 is column with N components in which
all equal to one. R is a correlation matrix for X, and each compo-
For personal use only.
nent of R is calculated using eq. (7). Then X can be expressed as
(8) X ⫽ ␮IN×1 ⫹ V兹diag[L]Z
where V (= [v1, …, vN]) is eigenvector matrix of the covariance matrix
C; diag[L] is a diagonal matrix, with diagonal elements being eigen-
values of C; and Z is a standard Gaussian vector (e.g., Au and Wang
2014). In essence, eq. (8) is a Karhunen–Loève expansion of the ran-
dom vector X, and random field samples of X may be generated
using eq. (8) (e.g., Zhang and Ellingwood 1994; Huang et al. 2001;
Wang et al. 2018).
Parametric study
To simulate a wide variety of typical sites, typical ranges of the
three statistical parameters (i.e., ␮, COV, and ␭) for random field
modeling of soil properties are considered and used in the para-
metric study. For example, Phoon and Kulhawy (1999) and Cao
et al. (2016) summarized the typical ranges of ␮ and COV for a wide
variety of soil properties, including undrained shear strength,
friction angle, natural water content, plastic and liquid limits,
unit weight, relative density, and field measurements (e.g., CPT;
vane shear test, VST; standard penetration test, SPT). Generally
speaking, the mean values range from 0.5 to 700, while the COV
varies from 2% to 90%. The typical values of vertical correlation
length ␭ for different soil properties range from 0.8 to 12.7 m (e.g.,
Phoon and Kulhawy 1999). Hence, these typical ranges of random
field parameters are used in the parametric study. Table 2 sum-
marizes the parameters used in the parametric study. The ␮ val-
ues range from 0.5 to 1000, while the COV values range from 1% to
100%. The values of vertical correlation length ␭ vary from 0.5 to
20 m. The thicknesses of the soil layers vary from 2.0 to 51.1 m,
which cover the typical ranges of thickness for a homogeneous
soil layer in geotechnical practice. In total, 504 random fields are
simulated. For each random field, 100 random field samples are
generated, and the total number of data points for each random
field sample is N = 512. Then, 12 different sample sizes (i.e., M = 5,
10, 15, 20, 25, 30, 35, 40, 45, 50, 60, and 80) are adopted for each
random field sample. The numbers of measurement data points
997
ranging from 5 to 80 are used as input to BCS for interpreting the
complete soil property profiles.
In stage 3, for each of the 504 random fields, 100 random field
samples are generated, and steps 1–6 are repeatedly performed for
each random field sample at 12 different sample sizes (i.e., step 7),
resulting in 504 × 100 × 12 = 604 800 relative errors. Then, step 8 is
repeatedly performed to estimate the mean relative error for each
M value and random filed, leading to 504 × 12 = 6048 ␮RE. In next
section, the parametric study results (i.e., 6048 data pairs of ␮RE
and M) are summarized for developing a dimensionless relation-
ship between the normalized sample size and ␮RE.
Dimensionless relationship between sample size and
relative error
The parametric study results are summarized in Fig. 7 with two
COV and normalized sampling interval ds/␭. The vertical axis in
Fig. 7 (i.e., ␮RE/COV) represents the interpretation accuracy (i.e.,
mean RE) normalized with respect to the variability (i.e., COV) of
the soil property concerned. The horizontal axis in Fig. 7 (i.e., ds/␭)
represents the sample size (i.e., sampling interval) normalized
90 100 Figure 7 shows the 6048 data pairs of ␮RE/COV and ds/␭ by open
triangles. The data pairs in Fig. 7 reveal a relationship that the
value of ␮RE/COV increases as the value of ds/␭ increases from 0 to
about 3, and that the ␮RE/COV value approaches to an asymptotic
value when ds/␭ is larger than 3.
When the normalized sampling interval ds/␭ is close to 0, a
small positive intercept is observed along the vertical axis of ␮RE/
COV. There are two possible scenarios leading to this small inter-
cept. In the first scenario, the sampling interval ds approaches 0,
which means that the soil properties are sampled and measured
at every location. The small intercept in the vertical axis of ␮RE/
COV reflects the nugget effect or the difference in soil properties
at neighboring locations (e.g., Matheron 1963). In the second sce-
nario, the correlation length ␭ is very large, which indicates that
the soils are uniform within the layer and that the soil properties
at different locations within the soil layer are perfectly correlated
among each other. Therefore, the entire soil property profile can
be represented by a single sample and measurement. The small
intercept in the vertical axis of ␮RE/COV in this scenario repre-
sents the measurement error, which is also often referred to as
nugget effect in geostatistics. Note that measurement error has
been probabilistically modeled in BCS as a Gaussian random vari-
able with a zero mean and unknown variance, which is subse-
quently integrated through a marginalization in the Bayesian
framework. More discussion on the nugget effect is provided sub-
sequently in this section.
On the contrary, when the normalized sampling interval ds/␭ is
large (e.g., larger than 3), the sample size is quite small and the
sampling interval is much larger than the distance within which
the measurement data are significantly correlated. In this case,
the ds/␭ has negligible effect on the ␮RE/COV, which plots more or
less horizontally over different ds/␭ values in Fig. 7. When ds/␭ is
larger than 3, reducing sampling interval (i.e., ds) or increasing
sample size (i.e., M) does not improve the level of accuracy for the
interpretation results. This suggests that there is a minimum sam-
ple size threshold or maximum sampling interval threshold for
geotechnical site investigation. When the sample size is smaller
than the minimum sample size threshold or the sampling interval
is larger than the maximum sampling interval threshold (e.g., 3␭),
changing the sample size or sampling interval has negligible in-
fluence on the level of accuracy for the interpretation results.
Note that the data pattern in Fig. 7 is very similar to a semi-
variogram in geostatistics, which shows variation of a half of au-
tocovariance with lag distance. Figure 8 shows a schematic of a
typical semi-variogram where the experimental semi-variogram
data (i.e., open circles) are fitted to an exponential function with a
Published by NRC Research Press
 998 Can. Geotech. J. Vol. 56, 2019

Fig. 4. Interpreted qc profiles from BCS method for CPT No. CD1 to CD11 with different number M of measured data points.

qc (MPa) qc (MPa) qc (MPa) qc (MPa) qc (MPa)

1.5 2.0 2.5 3.0 3.5 1.5 2.0 2.5 3.0 3.5 1.5 2.0 2.5 3.0 3.5 4.0 1.5 2.0 2.5 3.0 1.0 1.5 2.0 2.5 3.0
2.0 2.0 2.0 2.0 2.0

2.5 2.5 2.5 2.5 2.5

Depth (m)
Depth (m)

Depth (m)

Depth (m)

Depth (m)
3.0 3.0 3.0 3.0 3.0
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21

3.5 3.5 3.5 3.5 3.5

4.0 4.0 4.0 4.0 4.0

4.5 4.5 4.5 4.5 4.5

CPT No. CD1 CPT No. CD2 CPT No. CD3 CPT No. CD4 CPT No. CD6
5.0 5.0 5.0 5.0 5.0
qc (MPa) qc (MPa) qc (MPa) qc (MPa) qc (MPa)
1.5 2.0 2.5 3.0 1.5 2.0 2.5 3.0 1.5 2.0 2.5 3.0 1.5 2.0 2.5 3.0 1.5 2.0 2.5 3.0 3.5
2.0 2.0 2.0 2.0 2.0

2.5 2.5 2.5 2.5 2.5

Depth (m)
Depth (m)

Depth (m)

Depth (m)

Depth (m)
3.0 3.0 3.0 3.0 3.0

3.5 3.5 3.5 3.5 3.5

For personal use only.

4.0 4.0 4.0 4.0 4.0

4.5 4.5 4.5 4.5 4.5

CPT No. CD7 CPT No. CD8 CPT No. CD9 CPT No. CD10 CPT No. CD11
5.0 5.0 5.0 5.0 5.0
Original qc profile Interpreted qc profile for: M=10; M=20; M=60

Fig. 5. Variation of relative error with sampling interval from 50 Fig. 6. Relationship between mean relative error and sampling interval.
sets of real CPT data.
15
30
M=80 M=60
M=50 M=45
M=40 M=35
Mean relative error, PRE (%)

M=30 M=25
10
M=20 M=15
Relative error, RE (%)

20 M=10 M=5

5
10

0
0.0 0.2 0.4 0.6 0.8
0 Sampling interval, ds (m)
0.0 0.2 0.4 0.6 0.8
Sampling interval, ds (m)
the lag distance where the value of semi-variogram reaches 95% of
its sill value. The semi-variogram contains a small positive inter-
nugget component (i.e., solid line). As shown in Fig. 8, the semi- cept, ng, at the vertical axis, which is referred to as nugget effect in
variogram monotonically increases as lag distance increases, and geostatistics for modeling measurement errors and spatial varia-
it approaches an asymptotic value, defined as sill, s, in geostatis- tion within the shortest sampling interval (e.g., Matheron 1963).
tics, when lag distance is larger than an effective range, a (e.g., Given that Fig. 7 exhibits a pattern that is similar to an expo-
Webster and Oliver 2007). The effective range is usually taken as nential semi-variogram, an exponential function is used in a re-

Published by NRC Research Press

 Wang et al. 999

Table 2. Summary of parameters used in parametric study. Fig. 9. Statistical chart for sampling interval determination.
Parameter Variation 3
Best-fit curve-1σ
Thickness (m) 2.0, 5.0, 10.2, 20.4, 30.7, 40.9, and 51.1
Mean, ␮ 0.5, 20, 40, 50, 60, 80, 100, and 1000 Best-fit curve+1σ
COV (%) 1, 5, 10, 20, 30, 40, 50, and 100 Best-fit curve

Normalized sampling interval, ds/O

Correlation length, ␭ (m) 0.5, 1, 2, 3, 5, 7, 9, 11, 13, 15, 18, and 20
Number of measurement 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 60,
data points, M and 80 2
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21

Fig. 7. Dimensionless plot of normalized mean relative error versus

normalized sampling interval. [Colour online.]

1.5 Parametric study result(total No.6048)

Best-fit curve 1
Best-fit curve+1V
Normalized mean relative error, PRE /COV

Best-fit curve-1V

1.0
0
0.0 0.5 1.0 1.5
Equation of best-fit curve: Target normalized relative error, RET /COV
x
y=0.157+0.975 u {1-exp(- )}
0.657
0.5 y=PRE /COV, x=ds/O where y = ␮RE/COV and x = ds/␭. The best-fit ng value in eq. (9) is
0.157, and the sill s = 0.157 + 0.975 = 1.132. The effective range a = 1.8
R2=0.969 can be estimated from the best-fit line in Fig. 7. A dotted line and
a dashed dotted line are also included in Fig. 7 to represent the
best-fit curve plus or minus one standard deviation, respectively.
For personal use only.

The equations for the dotted and dashed dotted lines are ex-
0.0 pressed, respectively, as
0 1 2 3 4 5 6
Normalized sampling interval, ds/O (10) 关 共 0.657
y ⫽ 0.209 ⫹ 0.975 1 ⫺ exp ⫺
x
兲兴
y ⫽ 0.105 ⫹ 0.975关1 ⫺ exp共⫺
0.657 兲兴
Fig. 8. Illustration of an exponential semi-variogram model with x
fitting parameters. (11)

Fitted semi-variogram model The relationship between the normalized mean RE and normal-
Experimental semi-variogram ized sampling interval shown in eqs. (9)–(11) establishes a connec-
tion between the sample size and level of accuracy for the
interpretation results. This relationship may be further used to
develop statistical chart for sample size determination with con-
sideration of spatial variation and correlation of soil properties, as
Semi-variogram, γ(h)

described in next section.

Development of statistical chart

x
y n g  c1 u {1- exp(- )} As shown in Fig. 9, a statistical chart is developed for determi-
Sill, s

r nation of sample size along vertical direction in a single statisti-

y γ ( h) cally homogeneous soil layer. The chart is developed from the
x h parametric study results and Fig. 7. Similar to Fig. 7, Fig. 9 is
plotted in accordance with two normalized dimensionless param-
s n g c1
eters, a target normalized relative error (i.e., RET/COV) in the hor-
Effective range, a izontal axis and a normalized sampling interval in the vertical
axis (i.e., ds/␭). The target relative error, denoted as RET, is the
Nugget, ng desired level of accuracy for the interpretation results that a user
0 would like to have. When replacing ␮RE by RET in eqs. (9)–(11) and
Lag distance, h (m)
switching the vertical and horizontal axis of Fig. 7, a statistical
chart is generated, as shown in Fig. 9. Similar to Fig. 7, a dotted
line and a dashed dotted line are also included in Fig. 9 to repre-
gression analysis to best fit the 6048 data points in Fig. 7. The sent the best-fit curve minus or plus one standard deviation, re-
best-fit line is shown by a solid line in Fig. 7, and corresponding spectively. Figure 9 shows that, as the RET value decreases, or a
best-fit equation, with a coefficient of determination R2 = 0.969, is stringent level of accuracy is required for the interpretation re-
expressed as sults, the RET/COV value decreases, and the normalized sampling
interval decreases as well. As the sampling interval decreases, the
(9) 关
y ⫽ 0.157 ⫹ 0.975 1 ⫺ exp ⫺ 共 x
0.657 兲兴 required number of samples increases. Figure 9 provides a quan-
titative tool for geotechnical practitioners to rationally determine

Published by NRC Research Press

 1000
sample size with consideration of desired level of accuracy and
spatial variation and correlation of soil properties.
To use the statistical chart in Fig. 9, users must first specify their
target relative error RET. Target relative error is determined de-
pending on the importance and risk of the project (e.g., failure
consequences), as well as complexity of local geology. Then, the
coefficient of variation, COV, and vertical correlation length, ␭, of
the soil properties concerned is estimated based on existing
knowledge of the site and soil properties concerned. Such existing
knowledge includes, but is not limit to, local experience on soil
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21
properties and sites under similar conditions, geologic and topo-
graphic maps, and the typical range of COV and ␭ for different soil
properties reported in the literature (e.g., Phoon and Kulhawy
1999; Cao et al. 2016). With the estimated COV value and specified
RET value, the RET/COV in the horizontal axis of Fig. 9 is calculated
accordingly. Then, Fig. 9 is used to determine the corresponding
ds/␭ value in the vertical axis. Finally, the sampling interval ds is
calculated using the ds/␭ value obtained from Fig. 9 and the esti-
mated ␭ value. The required sample size M is calculated from the
ds value and the thickness of soil layer.
Application examples
To demonstrate application of the proposed statistical chart,
real CPT qc data and laboratory test result (i.e., liquid limit) are
used as illustrative examples in this section. Note that CPT data
are used in the first illustrative example, because CPT is per-
formed in a nearly continuous manner. The CPT data then can be
used for validating the interpretation results of measurement
For personal use only.
data, the sample size of which is determined from the statistical
chart. The proposed quantitative method and statistical chart are
equally applicable to measurements on other soil properties, such
as liquid limit data shown in the second example.
Real CPT qc data example
A CPT (i.e., test number CPT-1) was carried out in a cohesive soil
layer in eastern Houston, Texas, USA, and its qc data were down-
loaded online from ISSMGE TC304dB database website (http://
140.112.12.21/issmge/Database_2011.htm) and used in this illustrative
example. According to Stuedlein et al. (2012), the water table is
about 1.8 m below the ground surface, and the site contains a
thick layer of stiff, slightly silty clay. The CPT qc data ranging from
5.10 to 15.32 m are used in this example. The qc data were recorded
at 20 mm interval, so the total number of data points for this CPT
qc data profile is 512.
Suppose that the target relative error, RET, in this application is
10%. The COV and vertical ␭ of the qc data in this site are estimated
as 15% and 2.0 m, respectively, based on local experience (e.g.,
nearby CPT data). These ranges are also within the typical ranges
of 5%⬃40% for COV and 0.1⬃2.2 m for vertical ␭ for the clay CPT qc
data reported in the literature (e.g., Phoon and Kulhawy 1999).
Then, the value of RET/COV is calculated as 0.67. Using the best-fit
curve in the statistical chart, the normalized sampling interval
ds/␭ = 0.50 is calculated. Because ␭ is estimated as 2.0 m, the
sampling interval is then calculated as ds = 1.00 m.
The thickness of soil property profile is 10.22 m, and the num-
ber of measurement data points is estimated as M = 10.22/1.00 +
1 ≈ 11. Subsequently, 11 data points with an equal interval are
taken from the CPT data, and they are used as measurement data
and input for the BCS method to interpret the complete soil prop-
erty profile. As shown in Fig. 10, the measurement data are shown
by open circles. The solid line represents the original CPT qc pro-
file, and the dashed line indicates the mean qc profile obtained
from BCS. The statistical uncertainty of the interpreted qc profiles
is indicated by the interval within two dotted lines, which are the
mean qc profile plus or minus one standard deviation. The RE
between the original and mean qc profiles is 9.98%, which is quite
close to the target relative error of 10%. In other words, the statis-
Can. Geotech. J. Vol. 56, 2019
Fig. 10. Comparison between original qc profile and qc profile
interpreted from BCS method for real CPT data example. [Colour
online.]
qc (MPa)
5 10 15 20 25 30 35
6
RE=9.98%
8 RET=10%
10
Depth (m)
12
14
16 Original qc profile
BCS mean qc profile
18 BCS mean qc profile ± 1 σ
Measurement data y
tical chart provides reasonable results and achieves the target
level of accuracy for this site.
Liquid limit data example
In this subsection, liquid limit, wL, data measured from the
laboratory are used to illustrate application of the proposed sta-
tistical chart. Consider, for example, determination of sample size
for quantifying spatial variation of liquid limit along depth within
a silty clay layer in the construction site of Taipei City Hall Station,
Taiwan (Chin et al. 1994). The site contains a thick layer of silty
clay from a depth of about 2.9–30.0 m.
Suppose that the target relative error is 12%. The COV and ver-
tical ␭ of the liquid limit in this silty clay layer are estimated as 18%
and 5.0 m, respectively. These two values are approximately the
mean of the typical ranges of 6%⬃30% for COV and 1.6⬃8.7 m for
vertical ␭ of the clay liquid limit reported in the literature (e.g.,
Phoon and Kulhawy 1999). Then, the RET/COV = 0.67 is calculated.
Using the best-fit curve in the statistical chart, ds/␭ = 0.5 is calcu-
lated. Because ␭ = 5 m, the sampling interval ds = 2.5 m is esti-
mated. The thickness of silty clay layer is 27.1 m, and the number
of measurement data points is estimated as M = (27.1/2.5) + 1 ≈ 12.
Subsequently, 12 samples within the silty clay layer (i.e., from a
depth of about 2.9 to 30.0 m) shall be obtained from a borehole
with an equal sampling interval for carrying out liquid limit tests
in the laboratory and quantifying spatial variation of the liquid
limit along depth.
Chin et al. (1994) happened to report that a total number of 39
soils samples (i.e., T1–T39) was obtained from a borehole within
this silty clay layer for carrying out liquid limit tests in the labo-
ratory. Figure 11 shows the 39 wL values measured along depth by
open triangles. To validate the performance of the proposed sta-
Published by NRC Research Press
 Wang et al.
Fig. 11. Comparison between original and profile interpreted from
BCS method for liquid limit data example. [Colour online.]
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21
For personal use only.
tistical chart in this illustrative example, 12 wL values are taken
from the 39 wL values with an approximately equal sampling
interval, and they are shown in Fig. 11 by solid circles. The 12 wL
values obtained are used as the measurement data, i.e., input to
BCS for interpolating a high-resolution liquid limit profile with
256 data points. Figure 11 also shows the mean liquid limit profile
interpreted from BCS by a dashed line and the mean plus or minor
one standard deviation profiles by two dotted lines. It is observed
from Fig. 11 that the dashed line follows a trend quite similar to
that of the open triangles (i.e., all 39 wL values), and that most of
the 39 wL data points (i.e., the open triangles) fall within the dot-
ted lines. The relative error RE between the 39 original liquid limit
values and mean liquid limit values at the corresponding depths
obtained from BCS is calculated as 11.64%, which is very consistent
with the target relative error of 12%. The consistency suggests that
the sample size determined by the proposed statistical chart (i.e.,
Fig. 9) achieves the target level of accuracy for the results inter-
preted from the given number of measurements. In summary, the
proposed statistical chart performs well in the illustrative exam-
ples for both in situ and laboratory test data.
Conclusion
This paper performed an extensive parametric study and devel-
oped a quantitative method and statistical chart for sample size
determination in geotechnical site investigation with consider-
ation of spatial variation and correlation of soil properties. The
extensive parametric study was performed to establish a quanti-
tative relationship between the sample size M and the correspond-
ing level of accuracy for the results interpreted from the M
number of measured data points. The parametric study results
were summarized using a dimensionless plot between the nor-
malized mean relative error ␮RE/COV and normalized sampling
interval ds/␭. The dimensionless plot showed such a relationship
that the normalized mean relative error increases as the normal-
ized sampling interval increases from 0 to about 3, and that the
1001
normalized mean relative error approaches an asymptotic value
when the normalized sampling interval is larger than 3. A nugget
effect was observed from the dimensionless plot that represents
measurement errors and spatial variation in soil properties at
neighboring locations. There also exists a minimum sample size
threshold or maximum sampling interval threshold for geotech-
nical site investigation. When the sample size is smaller than the
minimum sample size threshold or the sampling interval is larger
than the maximum sampling interval threshold (e.g., 3␭), chang-
ing the sample size or sampling interval has negligible influence
on the level of accuracy for the spatial variation of soil properties
interpreted.
A statistical chart for sample size determination was further
developed from the dimensionless plot. Real CPT data and real
liquid limit data were used to illustrate the application of the
proposed statistical chart. The statistical chart was shown to per-
form well in both illustrative examples.
Acknowledgements
The work described in this paper was supported by grants from
the Research Grants Council of the Hong Kong Special Adminis-
trative Region, China (Project No. 9042516 (CityU 11213117) and
Project No. 8779012 (T22-603/15N)). The financial support is grate-
fully acknowledged. The authors would like to thank the mem-
bers of the TC304 Committee on Engineering Practice of Risk
Assessment and Management of the International Society of Soil
Mechanics and Geotechnical Engineering for developing the da-
tabase 304dB used in this study and making it available for scien-
tific inquiry. The authors also wish to thank M. Jaksa and A.W.
Stuedlein for contributing this database to the TC304 compen-
dium of databases.
References
Au, S.-K., and Wang, Y. 2014. Engineering risk assessment with subset simula-
tion. John Wiley and Sons, Singapore.
Baecher, G.B., and Christian, J.T. 2003. Reliability and statistics in geotechnical
engineering. John Wiley and Sons, Hoboken, New Jersey.
Benson, C.H., Zhai, H., and Rashad, S.M. 1994. Statistical sample size for con-
struction of soil liners. Journal of Geotechnical Engineering 120(10): 1704–
1724. doi:10.1061/(ASCE)0733-9410(1994)120:10(1704).
Bong, T., and Stuedlein, A.W. 2018. Efficient methodology for probabilistic anal-
ysis of consolidation considering spatial variation. Engineering Geology, 237:
53–63. doi:10.1016/j.enggeo.2018.02.009.
Candes, E.J., Romberg, J.K., and Tao, T. 2006. Stable signal recovery from incom-
plete and inaccurate measurements. Communications on Pure and Applied
Mathematics, 59(8): 1207–1223. doi:10.1002/cpa.20124.
Cao, Z., Wang, Y., and Li, D. 2016. Quantification of prior knowledge in geotech-
nical site characterization. Engineering Geology, 203: 107–116. doi:10.1016/j.
enggeo.2015.08.018.
CEN. 2007. Eurocode 7: Geotechnical design — Part 1: General rules. EN 1997-1:
2007. European Committee for Standardization (CEN), Brussels, Belgium.
Chin, C., Crooks, J., and Moh, Z. 1994. Geotechnical properties of the cohesive
Sungshan deposits, Taipei. Geotechnical Engineering, 25(2).
Clayton, C.R. 2001. Managing geotechnical risk: improving productivity in UK
building and construction. Thomas Telford Publishing, London.
Cui, J., Jiang, Q., Li, S., Feng, X., Zhang, M., and Yang, B. 2017. Estimation of the
number of specimens required for acquiring reliable rock mechanical pa-
rameters in laboratory uniaxial compression tests. Engineering Geology,
222: 186–200. doi:10.1016/j.enggeo.2017.03.023.
Daubechies, I. 1992. Ten lectures on wavelets. Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, Pa., USA.
Donoho, D.L. 2006. Compressed sensing. IEEE Transactions on Information The-
ory, 52(4): 1289–1306. doi:10.1109/TIT.2006.871582.
Fenton, G.A., and Griffiths, D.V. 2008. Risk assessment in geotechnical engineer-
ing. John Wiley and Sons, New York, USA.
Fenton, G.A., Liza, R., Lake, C.B., Menzies, W.T., and Griffiths, D.V. 2015. Statis-
tical sample size for quality control programs of cement-based “solidifica-
tion/stabilization”. Canadian Geotechnical Journal, 52(10): 1620–1628. doi:10.
1139/cgj-2013-0478.
Gill, D.E., Corthésy, R., and Leite, M.H. 2005. Determining the minimal number
of specimens for laboratory testing of rock properties. Engineering Geology,
78(1–2): 29–51. doi:10.1016/j.enggeo.2004.10.005.
Goldsworthy, J.S., Jaksa, M.B., Fenton, G.A., Kaggwa, W.S., Griffiths, V., and
Poulos, H.G. 2007. Effect of sample location on the reliability based design of
pad foundations. Georisk: Assessment and Management of Risk for Engineered
Systems and Geohazards, 1(3): 155–166. doi:10.1080/17499510701697377.
Published by NRC Research Press
 1002
Huang, S., Quek, S., and Phoon, K. 2001. Convergence study of the truncated
Karhunen–Loeve expansion for simulation of stochastic processes. Interna-
tional Journal for Numerical Methods in Engineering, 52(9): 1029–1043. doi:
10.1002/nme.255.
Jaksa, M., and Kaggwa, W.S. 1994. A micro-computer based data acquisition
system for the cone penetration test. Department of Civil and Environmental
Engineering, University of Adelaide.
Jaksa, M., Kaggwa, W., and Brooker, P. 1999. Experimental evaluation of the
scale of fluctuation of a stiff clay. In Proceedings of the 8th International
Conference on the Application of Statistics and Probability, A.A. Balkema,
Sydney, Rotterdam. Vol. 1, pp. 415–422.
Ji, S., Xue, Y., and Carin, L. 2008. Bayesian compressive sensing. IEEE Transac-
tions on Signal Processing, 56(6): 2346–2356. doi:10.1109/TSP.2007.914345.
Can. Geotech. J. Downloaded from cdnsciencepub.com by FAC CIENCIAS EST PROFESIONALES on 11/03/21
Ji, S., Dunson, D., and Carin, L. 2009. Multitask compressive sensing. IEEE Trans-
actions on Signal Processing, 57(1): 92–106. doi:10.1109/TSP.2008.2005866.
Krejcie, R.V., and Morgan, D.W. 1970. Determining sample size for research
activities. Educational and Psychological Measurement, 30(3): 607–610. doi:
10.1177/001316447003000308.
Lacasse, S., and Nadim, F. 1996. Uncertainties in characterising soil properties. In
Uncertainty in the geologic environment: From theory to practice. Publica-
tion No. 201. Norwegian Geotechnical Institute, Oslo, Norway, pp. 49–75.
Liza, R., Fenton, G.A., Lake, C.B., and Griffiths, D.V. 2017. An analytical approach to
assess quality control sample sizes of cement-based “solidification/stabilization”.
Canadian Geotechnical Journal, 54(3): 419–427. doi:10.1139/cgj-2016-0218.
Matheron, G. 1963. Principles of geostatistics. Economic Geology, 58(8): 1246–
1266. doi:10.2113/gsecongeo.58.8.1246.
Mayne, P.W., Christopher, B.R., and Dejong, J. 2002. Manual on subsurface in-
vestigations. Nat. Highway Inst. Sp. Pub. FHWA NHI-01-031. Federal Highway
Administration, Washington, D.C.
Orr, T.L.L. 2017. Defining and selecting characteristic values of geotechnical
parameters for designs to Eurocode 7. Georisk: Assessment and Management
of Risk for Engineered Systems and Geohazards, 11(1): 103–115. doi:10.1080/
17499518.2016.1235711.
Phoon, K.K. 2017. Role of reliability calculations in geotechnical design. Georisk:
Assessment and Management of Risk for Engineered Systems and Geohazards,
11(1): 4–21. doi:10.1080/17499518.2016.1265653.
For personal use only.
Phoon, K.-K., and Kulhawy, F.H. 1999. Characterization of geotechnical variabil-
ity. Canadian Geotechnical Journal, 36(4): 612–624. doi:10.1139/t99-038.
Stuedlein, A.W., Kramer, S.L., Arduino, P., and Holtz, R.D. 2012. Geotechnical
characterization and random field modeling of desiccated clay. Journal of
Geotechnical and Geoenvironmental Engineering, 138(11): 1301–1313. doi:10.
1061/(ASCE)GT.1943-5606.0000723.
Tropp, J.A., and Gilbert, A.C. 2007. Signal recovery from random measurements
Can. Geotech. J. Vol. 56, 2019
via orthogonal matching pursuit. IEEE Transactions on Information Theory,
53(12): 4655–4666. doi:10.1109/TIT.2007.909108.
Vanmarcke, E.H. 1977. Probabilistic modeling of soil profiles. Journal of the
Geotechnical Engineering Division, ASCE 103(11): 1227–1246.
Vanmarcke, E., Shinozuka, M., Nakagiri, S., Schueller, G., and Grigoriu, M. 1986.
Random fields and stochastic finite elements. Structural Safety, 3(3–4): 143–
166. doi:10.1016/0167-4730(86)90002-0.
Wang, Y., and Zhao, T. 2016. Interpretation of soil property profile from limited
measurement data: a compressive sampling perspective. Canadian Geotech-
nical Journal, 53(9): 1547–1559. doi:10.1139/cgj-2015-0545.
Wang, Y., and Zhao, T. 2017. Statistical interpretation of soil property profiles
from sparse data using Bayesian compressive sampling. Géotechnique, 67(6):
523–536. doi:10.1680/jgeot.16.P.143.
Wang, Y., Cao, Z., and Li, D. 2016. Bayesian perspective on geotechnical variabil-
ity and site characterization. Engineering Geology, 203: 117–125. doi:10.1016/
j.enggeo.2015.08.017.
Wang, Y., Akeju, O.V., and Zhao, T. 2017. Interpolation of spatially varying but
sparsely measured geo-data: a comparative study. Engineering Geology, 231:
200–217. doi:10.1016/j.enggeo.2017.10.019.
Wang, Y., Zhao, T., and Phoon, K.-K. 2018. Direct simulation of random field
samples from sparsely measured geotechnical data with consideration of
uncertainty in interpretation. Canadian Geotechnical Journal, 55(6): 862–
880. doi:10.1139/cgj-2017-0254.
Webster, R., and Oliver, M.A. 2007. Geostatistics for environmental scientists.
John Wiley and Sons, Chichester, England.
Yamaguchi, U. 1970. The number of test-pieces required to determine the
strength of rock. International Journal of Rock Mechanics and Mining
Sciences & Geomechanics Abstracts, 7: 209–227. doi:10.1016/0148-9062(70)
90013-6.
Zhang, J., and Ellingwood, B. 1994. Orthogonal series expansions of random
fields in reliability analysis. Journal of Engineering Mechanics, 120(12): 2660–
2677. doi:10.1061/(ASCE)0733-9399(1994)120:12(2660).
Zhao, T., and Wang, Y. 2018a. Interpretation of pile lateral response from deflec-
tion measurement data: a compressive sampling-based method. Soils and
Foundations, 58(4): 957–971. doi:10.1016/j.sandf.2018.05.002.
Zhao, T., and Wang, Y. 2018b. Simulation of cross-correlated random field sam-
ples from sparse measurements using Bayesian compressive sensing. Me-
chanical Systems and Signal Processing, 112: 384–400. doi:10.1016/j.ymssp.
2018.04.042.
Zhao, T., Montoya-Noguera, S., Phoon, K.K., and Wang, Y. 2018. Interpolating
spatially varying soil property values from sparse data for facilitating char-
acteristic value selection. Canadian Geotechnical Journal, 55(2): 171–181. doi:
10.1139/cgj-2017-0219.
Published by NRC Research Press