Computers and Geotechnics 105 (2019) 37–50

Contents lists available at ScienceDirect

Computers and Geotechnics

journal homepage: www.elsevier.com/locate/compgeo

Research Paper

Site variability analysis using cone penetration test data T

⁎
Rodrigo Salgado, Eshan Ganju , Monica Prezzi
Lyles School of Civil Engineering, 550 W. Stadium Avenue, West Lafayette, IN 47907, United States

A R T I C LE I N FO A B S T R A C T

Keywords: This paper presents a comprehensive methodology for quantification of site variability in a way that can be used
Cone penetration test in geotechnical practice. Site variability is separated into horizontal and vertical variability. To quantify vertical
Site variability variability in a CPT sounding, a vertical variability index is proposed based on the complexity of the soil profile,
Spatial statistics the overall coefficient of variation for the depth range of interest, and intra-layer variability measures. To
Random fields
quantify horizontal site variability, a horizontal variability index is proposed based on the similarity of
Horizontal variability
soundings performed at the same site. The method is illustrated using CPT data from five sites.
Vertical variability

1. Introduction
Site investigation is an essential component of every construction
project. A thorough site investigation aims to identify the stratigraphy,
locate ground water level and estimate the range of physical and me-
chanical properties of the in situ soil layers. Due to the spatial variability
of natural soil deposits, uncertainty in estimates of soil properties for a
site is inevitable. Although this uncertainty cannot be eliminated, it can
be quantified. If reasonably quantified, this uncertainty can be ac-
counted for in reliability analysis or can be used to select resistance
factors for use in Load and Resistance Factor Design (LRFD) [1–5].
Properties at different locations on a site are correlated but the
correlation becomes increasingly weak with increasing distance be-
tween the points considered [6–8]. The spatial variability of cone pe-
netration test (CPT) variables (e.g., cone resistance qc or sleeve re-
sistance fs) has been studied using different measures of variability,
such as the coefficient of variation (COV) and the scale of fluctuation
(SF) (distance within which data points are significantly correlated)
[9–11]. During the site investigation phase, if the variability of the CPT
parameters is high, additional CPTs can be performed to better char-
acterize soil properties; on the other hand, if variability is low, it may
be possible to reduce the number of CPT soundings from what had been
originally planned. Such decisions should of course be made with
caution, accounting for the natural variation of the soil profile and
geology of the area.
Paikowsky [10] suggests that site variability can be approximately
categorized by the COV of strength parameters of the soil layers in the
profile representative of the site. With this approach, first the bearing
layers of the representative soil profile of the site are identified. For
each of the bearing layers in the soil profile, the average strength
parameter is obtained (e.g., the standard penetration test (SPT) blow
count NSPT). The COV of the average strength parameters of a re-
presentative soil profile is then calculated, and this forms the basis of
the variability assessment. Based on the calculated COV, the entire site
is characterized as low (COV < 25%), medium (25% ≤ COV < 40%)
or high (COV ≥ 40%) variability. Difficulties with the application of
such a general approach based on the SPT include the very limited
amount of data available for statistical treatment (values are only
available at certain tested depths along the profile and, at each tested
depth, only a single value is obtained). In comparison to SPT data, CPT
data is considerably more reliable and provides a richer dataset of
measurements along depth for a profile. These two features of a CPT
dataset make it more amenable to statistical variability assessment
methods in comparison to an SPT dataset [12].
Rigorous theoretical treatment of spatial variability has been an
important topic of recent research [1–3,12–31] but remains difficult to
apply for a variety of reasons, including difficulties with determination
of key variables, such as the scale of fluctuation. The major focus of the
research in geospatial statistics has been to model a site as a random
field and then precisely infer the random field parameters (coefficient
of variance, scale of fluctuation) at the site. To find values for random
field parameters, researchers have used moment estimation techniques
[32,33], maximum likelihood estimation techniques [7,34,35] and
Bayesian techniques [28,36]. While the current research community is
very active in the pursuit of more accurate descriptions of sites as
random fields, research is needed for more immediate implementation
of methods of site variability characterization based on current practice
in CPT interpretation. This is important, for example, in LRFD-based

⁎
Corresponding author.
E-mail addresses: rodrigo@ecn.purdue.edu (R. Salgado), eganju@purdue.edu (E. Ganju), mprezzi@purdue.edu (M. Prezzi).

https://doi.org/10.1016/j.compgeo.2018.08.001
Received 3 December 2017; Received in revised form 18 June 2018; Accepted 3 August 2018
0266-352X/ © 2018 Elsevier Ltd. All rights reserved.
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

codes that attempt to set resistance factors as a function of site varia- 500
bility. In this paper, knowledge of spatial statistics is applied to develop Clayey Silty

Normalized cone resistance qc/pA

a methodology to quantify site variability via variability indices com- Sand

puted using CPT data. For the development of such a methodology, it Clean Sand or Silty Sand

was necessary to have a robust and logical algorithm capable of gen- 100
erating a soil profile from the CPT data recorded at the site. The

Silt
or
modified soil behavior type (SBT) charts and algorithms proposed by

nd

y
Sa

la
Very Stiff

C
[11,37] are used for this purpose. The methodology was developed in

y
ye
Clay

ilt
la

S
C
such a way as to be useful to practitioners wishing to gauge uncertainty

or
Stiff Cla

y
la
y

C
in a systematic and reproducible manner that may also be used in the
10

y
nd
Medium

Sa
future in LRFD code development, with values of resistance factors, Clayey Silt Stiff Clay
partial factors or factors of safety selected based on site variability
Soft Clay
measures. To demonstrate the use of the methodology, CPT data for five
Sensitive Clay
sites in the United States, taken from the United States Geological Very Soft Clay
Survey (USGS) CPT database [38], are analyzed using the proposed
method. 1
0 1 2 3 4 5 6
2. Soil profile generation from CPT data
Friction ratio (%)
Quantification of soil variability at a site requires identification of (a)
soil layers with different characteristics in a soil profile [14,39]. Typi-
1000
cally, soil profiles are inferred from CPT data using soil behavior type
Gravelly Sand to Sand
(SBT) charts. An SBT chart serves as a simple signal transfer function
that converts cone resistance-sleeve resistance pairs to “soil behavior”

Normalized cone resistance

types [11,37]. Many SBT classification charts have been proposed over
the years [40–53]. While any SBT chart in the literature could be used Clean Sand to Silty Sand
100
as a simple signal transfer function to develop a soil profile from CPT
:
data, in this paper the soil profile generation algorithms proposed by res ilt
xtu dy S
d Mi a n
[11,37] were used to generate soil behavior profiles from the CPT data. an o S ay
S
nd
t Cl
Sa lty
The algorithms [11,37] use modified versions of the Tumay chart [48] lty o Si
Si lt t
and Robertson chart [50], shown in Fig. 1(a) and (b), to convert CPT Si
ay
10 Cl
data into a soil profile that is amenable to the site variability algorithms Clay to Silty Clay
introduced in the present paper. These algorithms were used because
(1) they generate coherent soil profiles in which layers are classified Sensitive Fine Grained
using intrinsic and state-variable based descriptors and (2) occurrence
Organic Clay
of soil layers thinner than what can be properly sensed by the CPT
probe [54,55] is handled in a logical manner. To properly sense a soil 1
0.1 1 10
layer, the cone penetrometer must penetrate the layer to a certain ex-
tent to develop the shaft resistance fs and tip resistance qc corre- Normalized friction ratio
sponding to that layer; this length of penetration is called the devel-
(b)
opment distance. Also, while within a soil layer, the CPT probe will
start sensing the next soil layer before even reaching it; the distance Fig. 1. Modified SBT charts used for soil profile generation: (a) modified Tumay
from the next layer at which this happens is called sensing distance. chart [11,37,48]. (b) Modified Robertson chart [11,37,50].
Both the sensing and development distances are affected by the density
of the soil layers and the dimensions of the probe and are of the order of
2.2–5.4 cone diameters [4–6]. If the layer is too thin in comparison to Obtain CPT data
the probe dimensions, the cone resistance and sleeve friction will be as
much affected by the overlying and underlying layers as by the thin
layer itself, with the result that they are not representative of the layer.
Generate soil behavior profile using
Thin layers cannot therefore be characterized as individual layers using
CPT data
the CPT; accordingly, layers thinner than 150 mm are consolidated into
the adjacent thicker layers.
Calculate Vertical Variability Index (VVI) of
3. Site variability assessment algorithm
site
Fig. 2 shows the overarching procedure proposed to quantify site
variability, which consists of the following steps: (1) soil profile gen-
eration, (2) quantification of vertical variability, (3) quantification of Calculate Horizontal Variability Index (HVI) of site
horizontal variability and (4) integration of vertical and horizontal
variability into a site variability rating system.
Soil profiles can be obtained using SBT charts, as discussed in the
Classify site according to Site Variability
previous section. Once a soil profile is established, the vertical varia-
Rating (SVR) system
bility index VVI (which reflects variability in qc, fs, layering and other
factors for each CPT sounding) and the horizontal variability index HVI Fig. 2. Site variability assessment procedure based on computation of vertical
(which is based on the cross-correlation [11,56] between cone re- variability index (VVI) and horizontal variability index (HVI).
sistance logs, cone resistance trend differences and the spacing between

38
R. Salgado et al.
every pair of CPTs) are calculated. Considering both indices, the
variability of the site is established according to a site variability rating
system. Calculation of these indices is discussed next.
3.1. Vertical Variability Index (VVI)
The vertical variability index VVI of a CPT sounding is calculated
based on an analysis of the soil profile combined with the CPT data. The
vertical variability index of a sounding has three components: (1) the
intra-layer variability index (VVI)IL, (2) the log variability index (VVI)log
and (3) the cone resistance vertical variability index (VVI)qc.
The (VVI)IL attempts to capture the vertical variability of the soil
profile arising from the inherent variations in the qc and fs values within
soils layers and is calculated from the scale of fluctuation (SF) and
coefficient of variation (COV) of qc and fs of each soil layer in the soil
profile being considered. The (VVI)log attempts to capture the vertical
variability of the soil profile due to the number of different soil types
present in the profile per unit length (NDL) and how different those
layers are from each other (quantified by the Diversity Factor DF).
Lastly, the (VVI)qc attempts to capture the variability arising from the
presence of extremely dissimilar soil layers in the soil profile as re-
flected in the qc values along the entire profile. (VVI)qc is quantified by
the coefficient of variation (COV) of the cone resistance qc of the
sounding. Fig. 3 shows the various components of VVI for a CPT
sounding. These components of vertical variability are combined to
obtain the VVI of the sounding as:
VVI = w log (VVI ) log + w IL (VVI ) IL + wqc (VVI )qc (1)
where wlog, wIL and wqc are the weights assigned to each of the three
components of the VVI (0.2, 0.2 and 0.6 were used, respectively). The
choice of weights for the individual components of VVI is based on a
subjective assessment of the relevance of each component in the VVI.
We have verified by analysis of many sites that the weights selected do
provide good discrimination across soundings and across sites. Fur-
thermore, these weights may be refined considering new data and
further research.
The approach followed to compute the VVI of a sounding is con-
sistent with the current CPT interpretation practice of first classifying
soils in one of the types shown in Fig. 1(a) and (b). An additional ad-
vantage of doing this is that statistics can then be applied within layers
of what can be considered a single material at a comparable state within
the layer. Any alternative approach will mix different materials or the
same material at different states within the same dataset, considerably
complicating extraction of variability information from it.
The following sections describe the various steps in the calculation
procedure (see Fig. 4) of the VVI for individual CPT soundings and the
procedure used to de-trend CPT data for estimation of scale of fluc-
tuation and coefficient of variation within soil layers.
Intra-layer variability index
(VVI)IL
Vertical Variability Log variability index
Index (VVI) (VVI)log
Coefficient of variation of
qc (VVI)qc
Fig. 3. Components of vertical variability index.
39
Computers and Geotechnics 105 (2019) 37–50
3.2. De-trending of CPT data
The CPT data must first be detrended to make it stationary [2,5,32].
Detrending the data means to subtract the mean or best-fit line from the
data. Conceptually, detrending a plot of qc versus depth would aim to
eliminate the effect of depth (and, indirectly, vertical effective stress)
on qc, which would make the resulting, normalized cone resistance only
a function of the soil density for a normally consolidated soil. Given
that variations of OCR within a given soil layer are not expected to be
large, detrending can be expected to produce a plot of a function of
density versus depth, which is desirable for geospatial statistical treat-
ment. The algorithm used to detrend the qc data is shown in Fig. 5. The
depth vs qc data is first detrended using a power function of the form:
qc = α + βz 0.5 (2)
where z is the depth, and α and β are regression coefficients. The ex-
ponent of z is fixed at 0.5 because qc depends approximately on the
square root of the vertical effective stress [57] all other things being
equal. To check the suitability of the fit, the coefficient of determination
(R2) of the fit is calculated. If R2 of the fit is less than 0.85, the power
function in Eq. (2) is replaced by a polynomial of the smallest degree
(maximum degree of 2 [35]) that ensures that R2 is above the set
threshold. For the fs data, a simpler approach is adopted in which a
polynomial is directly used to fit the fs data. Again, the degree of the
polynomial (maximum degree of 2) is chosen as the minimum value
that achieves R2 value of 0.85.
After an attempt to detrend the CPT data, the Kendall Tau Test
(KTT) is carried out to assess if the detrended data is stationary or not.
The KTT uses the test statistic τ to measure the association of ranked
data [58]. The KTT statistic is given by:
P−Q
τ= n
(n−1) (3)
2
where n is the number of observations, P is the number of observation
pairs in natural order and Q is the number of observations in reverse
natural order. The numbers of observations in natural and reverse
natural order are calculated in the following manner: (1) arrange the
observations (Xi, Yi) in columns such that X increases from the top to
the bottom, (2) compare each Y value with all the Y values appearing
below it: a pair of Y values is in natural order if the Y below is greater
than the Y above, (3) calculate number P of pairs in natural order and
the number Q of pairs in reverse natural order, (4) calculate τ using Eq.
(3). The value of τ ranges from +1 to −1.
If in a set with n data points, Y increases monotonically with X, then
n
the number of pairs in natural order P will be equal to 2 (n−1) , while the
number of pairs in reverse natural order Q will be equal to 0, making τ
equal to 1. However, if Y decreases monotonically with X, then the
n
number of pairs in reverse natural order Q will be equal to 2 (n−1) and
Scale of fluctuation (SF)
Coefficient of variation (COV)
Diversity factor (DF)
Number of layers of different soil
types per unit length (NDL)
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

Read CPT data for generated soil profile Calculate number of different layers per unit
length (NDL)

Calculate Diversity Factor (DF) and multiply it to NDL to

Calculate a combined scale of fluctuation
obtain the non-normalized Log Factor
normalized COV (SNC) of qc and fs for soil
layer i

Standardize Log Factor with respect to its

minimum and maximum values to obtain
(VVI)log

i> number Obtain standardized (VVI)qc for the entire soil

FALSE of soil profile
i++ layers in
soil profile

VVI = wIL(VVI)IL + wlog(VVI)log + wqc(VVI)qc

TRUE

Calculate (VVI)IL as a weighted average of the

Save VVI
combined standardized SNCs of all soil
layers, where the weights are in proportion to
the respective layer thickness

Fig. 4. Calculation steps for VVI for individual CPT sounding.

Read qc data for the generated soil

profile
i=1

Fit power function to qc data of layer

i, and calculate corresponding
regression coefficients
j=1 Fit polynomial of degree j to qc data
j ++ of layer i and calculate
corresponding regression
Detrend qc data of layer i using coefficients
chosen function
FALSE

Is detrended
data for layer i Is j = 3?
stationary FALSE
according to
KTT and fit R2 TRUE
> 0.85
FALSE i ++
Retain data detrended using quadratic
TRUE polynomial. Inform user that fit fails KTT
Is i = and/or R2 check for layer i
number of Save detrended
layers in soil data for layer i
profile?
Save detrended
data for entire soil
TRUE
profile
Fig. 5. Detrending of qc data.

the number of pairs in natural order P will be equal to 0, making τ equal by [58] is calculated as:
to −1. A τ value close to zero indicates stationarity of the data, i.e.,
absence of any observable global trends. 3τ n (n−1)
χ=
As τ may not be precisely 0 for a data set which can be considered 2(2n + 5) (4)
stationary, its value can be compared with a critical value associated
with a confidence level (e.g., 95%) to assess stationarity. For a large The value of χ approximately follows a normal distribution with a
number of observations (n > 40), the critical value of χ recommended mean of 0 and variance of 1. To assume stationarity with 95%

40
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

Trendline
Data
Captured
by SF

Captured Crossing
by SF

Depth
Depth

distance
Captured
by COV

Captured
by COV

CPT data
Trendline
Cone resistance Cone resistance
(a) (b) Cone resistance
Fig. 6. Idealized plots of qc vs. depth: (a) small SF with low COV and (b) large Fig. 7. CPT data with fitted trendline illustrating the crossing distance.
SF with high COV.

where Δ = the average crossing length, Δi = individual crossing lengths

confidence level, the χ value should be less than 1.96. In this research, a
95% confidence level is adopted to check for stationarity. After de-
trending and stationarity check, the various components of the VVI are
calculated.
3.3. Calculation of intra-layer variability index (VVI)IL
Variability within a soil layer appears as either the amplitude of the
oscillations of the dependent variable (qc or fs) around the trend, which
is captured by the coefficient of variation COV, or by the frequency with
which these oscillations occur, which is captured by the scale of fluc-
tuation SF. Fig. 6(a) and (b) show idealized plots of qc vs. depth. The
scale of fluctuation of qc in Fig. 6(a) is smaller than that in Fig. 6(b)
(which would be suggestive of higher variability), but its coefficient of
variation is also lower (suggesting lower variability).
This example emphasizes that to quantify variability within a layer,
we must consider both the SF and the COV. A single number SNC may
be defined to do this:
COV
SNC =
( )
SF
LR
where LR is the reference length (1 m, 3.281 ft or equivalent value in
other units). The SF is normalized by the reference length LR to make it
dimensionless.
After a trendline is fit to the data, the trendline tends to cross the
line joining the data multiple times, as shown in Fig. 7. The vertical
distance between two crossings is termed the crossing distance. All the
points between two crossings are on one side of the trend line and can
be said to be positively correlated. Since the scale of fluctuation also
indicates a length within which data are strongly correlated, the
crossing distances can be used to estimate the correlation length. The
scale of fluctuation may be estimated using Vanmarcke’s simplified
method [59]:
2 Maximum and minimum values of SF for qc and fs (modified after [32]).
SF ≈ Δ
π (6)
and nc = number of crossings. In cases where Vanmarcke’s simplified
method is not applicable, when there is just one crossing in the soil
layer, the SF is calculated using integration of the autocorrelation
function ACF [56]. The SF is taken as twice the area under the positive
part of the ACF.
Researchers have calculated SF for different soil parameters and for
different soil types. Honjo and Kuroda [60] found the vertical scale of
fluctuation to be about 4 m for soft alluvial clay using an exponential
correlation model. Researchers [61] have reported a value of 1.5 m for
vertical scale of fluctuation of qc for lightly overconsolidated and highly
sensitive clays using a spherical correlation model. Researchers [62],
working with natural deposits of Keswick clay in Australia, found that
the SF of qc ranged from 63 to 255 mm, with a mean of 151 mm. These
authors calculated the SF using autocorrelation functions after de-
trending CPT data with quadratic polynomials. Table 1, modified after
[32], shows ranges of SF values in the vertical direction for CPT data for
different soil types reported in the literature [62–64]. The review found
SF of qc to range from 0.1 to 3.0 m for sands and clays. Bombasaro and
Kasper [65] reported SF for marine and fluvial sands and clays to range
from 0.07 m to 0.78 m. While these SF values were calculated using
(5)
three different correlation models and correspond to these specific sites,
these can be used to provide a general range for limits of SF for qc and fs.
In addition to the SF, COV is also calculated using the trendline. If a
function x = f(z) represents the trendline for the relationship between
two random variables X and Z (X could represent qc or fs and Z could
represent depth, for example), the COV of X can be calculated using:
xi − f (z i) 2
COVX ≈
n
∑i = 1 ( f (z i))
(n−1) (8)
th
where n = number of data points, xi is the i value of X in the data set,
corresponding to the ith value of Z (zi).
Table 1
Property Notation Soil type Testing method SF (m)

with Cone resistance qc Sand, clay CPT 0.1–3.0

Cone resistance qc Marine clays CPT 0.3–0.4
nc − 1
1 Sleeve resistance fs Sand CPT 1.3
Δ=
nc−1
∑ Δi Sleeve resistance fs Deltaic soils CPT 0.3–0.4
i=1 (7)

41
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

Phoon and Kulhawy [29] report COV of qc for different soil types. Read CPT data for soil layer
The COV of qc, according to these authors, is quite large, ranging from
less than 20% to 60%. The reported data suggest that the COV of qc is
greater in sand (20–60%) than in clay (20–40%). These ranges are Discard data for 2-cone-diameter length from top and bottom
based on a literature survey carried out by [63] that relies on the data of each soil layer to consider sensing and development
collected and analyzed by multiple researchers [66–75]. In addition to lengths of cone
the ranges presented by Phoon and Kulhawy [29], Foye [5] reported
COV of qc to be about 8% for a reasonably uniform sand layer, and Foye
et al. [2] reported the COV of qc to be about 6% for a reasonably uni- Detrend qc and fs data using trend function
form clay layer. Jaksa et al. [62] reported COV of qc of a slightly over
consolidated clay (Keswick clay) to be about 30%. SF and COV values
depend on how the soil layers are divided and how the data are de- Calculate scale of fluctuation of both qc and fs
trended, and therefore are significantly affected by how authors define
layers within which the COV would be calculated and, indeed, if data
from different soils from different profiles were combined to perform Calculate coefficient of variation of both qc and fs
the calculations [76]. Conceptually, COV values must be defined for a
specific layer of a given material, say sand. If the range of COV values
for sand is then to be determined, one would obtain data for sand layers Calculate scale-of-fluctuation normalized COV (SNC) of both
from many different soil profiles to make that assessment, even if dif- qc and fs
ferent sand would add some variability to the assessment. General
combinations of data without first this focus on proper determination of
what the individual layers are, should be avoided. Additionally, in
Standardize the SNC of both qc and fs by their respective
defining layers, it is important that dissimilar materials (say, dense sand
maximum and minimum values
and medium dense sand) not be considered part of the same layer, as
this will considerably distort the statistics. A fine-tuned soil classifica-
tion or soil “behavior” chart or algorithm is essential so that layers can
be properly defined [11,37]. Calculate a combined standardized SNC as a
Table 2 shows the maximum and minimum SF, COV and SNC values weighted average of standardized SNC of qc and fs
of qc and fs used in this paper for standardization purposes. Regardless
of soil type, the maximum value of SF of both qc and fs was taken as 2 m; Fig. 8. Calculation procedure for scale of fluctuation-normalized coefficient of
the minimum value of SF of qc was taken as 50 mm, while that of fs was variability (SNC) for a soil layer.
taken as 40 mm. The maximum value of COV for qc was taken as 10%,
while that of fs was taken as 12%. The minimum COV of qc and fs was average using weights of 0.8–0.2, respectively, to obtain the SNC of the
taken as the theoretical minimum, i.e., 0%. We assumed a slightly soil layer. The chosen weight reflects the notion that qc captures the in
higher maximum COV value for fs in comparison to that of qc and situ variability more precisely than fs. The SNC for the entire soil profile
slightly lower minimum SF value for fs in comparison to that of qc to is calculated as the weighted average of the SNCs of individual soil
account for the fact that fs data are observed to be more variable than qc layers, with ratio of layer thickness to profile length taken as the
data. Values presented in Table 2 are based on values of COV and SF weight.
observed in the literature and the range of values obtained from ana-
lysis of CPT data collected from test sites located in the state of Indiana
[11] and the USGS CPT database [38]. These limits may be updated 3.4. Calculation of log vertical variability index (VVI)log
further as additional data become available.
To calculate the maximum SNC value for qc, the maximum COV is The (VVI)IL, discussed in the previous section, considers the intra-
divided by the minimum SF (normalized by the reference length of layer variability of the soil profile through the standardized SNC. The
1 m). When processing a CPT sounding, SNC values are capped by the overall variability of the soil profile also depends on inter-layer varia-
maximum specified SNC value. Limits for SF and COV adopted in this bility; this fact is addressed by considering:
research are found to work well. For most soundings processed with this
approach, when SF or COV values are outside the limits, they are (1) the number of layers along the soil profile per unit length of soil
usually very close, within 10%, of these limits. This adopted cap is profile;
infrequently hit, with a frequency of approximately 1 in 20 layers (2) the ratio of the number of different soil layer types to the total
analyzed. number of soil layers;
The flowchart of the calculation of SNC of a soil layer is presented in
Fig. 8. First, the COV, SF and SNC of qc and fs are calculated for each soil and
layer in the profile. Then, the SNCs are standardized with respect to the
minimum and maximum values given in Table 2. After standardization, (3) the dominance of any soil group (the broad soil groups are shown in
the SNCs of qc and fs of each layer are combined through a weighted Table 3 for the modified Tumay and Robertson charts).

Table 2 The log vertical variability index (VVI)log is calculated from the
Minimum and maximum SF, COV and SNC values used for calculation of intra- number of soil layers of different types in the soil profile and a diversity
layer vertical variability index VVIIL. factor DF, which accounts for the dominance of any soil group. The
qc fs flowchart in Fig. 9 shows the calculation procedure for the DF. To
calculate the DF, first the layers in the soil profile are tagged as sand,
SF (m) COV (%) SNC SF (m) COV (%) SNC clay or mixed (see Table 3). Then, the dominance factor Di of each one
of the three soil groups is calculated by dividing the total number of soil
Min. 0.05 0 0 0.04 0 0
Max. 2 10 10/0.05 = 200 2 12 12/0.04 = 300 layers of that soil group by the total number of layers in the soil profile.
The Di value of a group i (sand, clay or mixed) indicates the fraction of

42
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

Table 3 case, respectively. A low DF value indicates high dominance of a soil

Soil group classification for the modified Tumay chart and modified Robertson group and therefore low diversity and vertical variability within the soil
chart used to group soils for calculation of dominance factor Di and diversity profile. The (VVI)∗log, before normalization, is calculated as:
factor DF.
(VVI )∗log = DDF × NDL (11)
Soil Group Soil Types (modified Tumay Soil Types (modified Robertson
chart) chart) where NDL is the number of different soil layers per unit length. As in
the case for (VVI)IL, the objective is to standardize the value of (VVI)log
Sand Clean sand or silty sand Gravelly sand to sand
(very loose to very dense) (very loose to very dense) to have it range from 0 to 100%. The minimum and maximum value of
Clean sand to silty sand the (VVI)log depends on the minimum and maximum values of DF and
(very loose to very dense) NDL. The theoretical maximum value of NDL is 6.7 (= reference
Clay Sensitive clay Organic clay length/minimum layer thickness = 100 cm/15 cm). While theoretically
Very soft clay Sensitive fine grained
Soft clay
possible, the maximum value of NDL was found to be too extreme and
Medium stiff clay therefore was reduced from 6.7 to 1 for both the modified Tumay chart
Stiff clay and the modified Robertson chart to obtain more realistic limits. The
Very stiff clay reduction was based on an extensive analysis of CPT sounding available
Mixed Soil Clayey silt Silty sand to sandy silt
in the USGS CPT database.
Clayey sand or silt Clayey silt to silty clay
Clayey silty sand Clay to silty clay The minimum value of NDL is chosen based on the length of the soil
Sandy clay or Silty clay profile being analyzed. Minimum NDL occurs when the entire profile
consists of just one thick soil layer. The calculated minimum values of
NDL for different profile lengths are shown in Fig. 11. The minimum
Read CPT data for soil layer value of NDL ranges from 1 layers/m for a profile length of 1 m to 0.03
layers/m for a profile length of 30 m.

Classify soil type of each layer into one of 3.5. Calculation of the cone resistance (qc) vertical variability index (VVI)qc
three soil groups (sand, clay and mixed)
The (VVI)log does not fully reflect the variability in a sounding. For
example, the DF component would treat the transition of a loose sand
For each soil group i calculate dominance layer to a soft clay layer the same way it would the transition of a very
factor Di
dense layer to a very soft clay layer. What is needed, thus, is a con-
sideration of the most important source of variability across layers,
which is that of qc. To address this, a direct calculation of COV of qc of
Calculate diversity factor DF the entire soil profile is made. To standardize the COV of qc, thereby
obtaining (VVI)qc, a maximum COV of qc for the length of soil profile
Fig. 9. Flowchart for calculation of diversity factor DF used to capture the
under consideration is used.
dominance of different soil layers in the profile. To calculate the maximum value [COV(qc)]max of the COV of the
cone resistance for a sounding, an idealized, highly variable soil profile
is considered, as shown in Fig. 12. To maximize the COV(qc), the
the soil profile that belongs to the group:
idealized soil profile must have high standard deviation and low mean
Ni value of qc. A soil profile containing mostly a soft clay layer and a thin
Di =
N (9) sand layer would satisfy this requirement. The thickness of this clay
layer can change, which enables the analysis of soil profiles of various
where N is the number of layers in a soil profile and Ni is the total
thicknesses. In the soil profile considered, the sand layer thickness re-
number of layers of the soil group i in the soil profile (i = clay (c), sand
mains the same (0.5 m) for all soil profile thicknesses considered. The
(s), mixed-soils (ms)).
plot of maximum COV of qc vs the length of soil profile is shown in
If, in a soil profile, all soil layers belong to a single group (sand, clay
Table 4 and Fig. 13.
or mixed soils), the Di of that group will be 1; therefore, the soil profile
To calculate the cone resistance for the sand layer, the equation
can be considered to have low variability. If all three groups are equally
proposed by [57] was used, assuming relative density of 75%, critical-
represented in the soil profile, the Di values of all the three groups will
state friction angle of the sand equal to 30° and a saturated unit weight
be 1 , and the soil profile can be considered to have high variability. To
3 of 20 kN/m3. To calculate the cone resistance for the clay layer, the
quantify this, consider the plane shown in Fig. 10. Any triplet (c, ms, s)
following equation was used:
of dominance factors (Di) obtained from a soil profile will exist in the
qc = Nk su + σv (12)
1 1 1
(
Dc + Dms + Ds = 1 plane. The point 3 , 3 , 3 would correspond to a )
soil profile in which all three groups are equally represented. The fur- where the cone factor Nk is taken as 15 [77,78], su is the undrained
1 1 1
( )
ther the (c, ms, s) triplet is from 3 , 3 , 3 , the less diverse the soil profile shear strength of the clay layer and σv is the total vertical stress at the
depth of interest. To calculate su, a rate of increase of su with respect to
( 1 1 1
)
is. The distance D∗ from 3 , 3 , 3 can be used as a tool to assess the
vertical effective stress σv′ of 0.35 was assumed. A saturated unit weight
variability due to soil layering. A D∗ value of 0 would indicate equal
of 17 kN/m3 was assumed for the clay layer. To achieve reasonable
representation of the three soil groups in the profile, while a D∗ value of
values of qc in clay layers that are more commonly found in day-to-day
2
3
(distance from
1 1 1
, ,
3 3 3 ( )
to the point where the plane practice, an overburden pressure of 150 kPa was applied and removed
Dc + Dms + Ds = 1 intersects any one of the three axis) would indicate to make the clay layers overconsolidated. This choice was motivated by
presence of layers from only one soil group. The DF is assigned a value the fact that, for profiles in the USGS CPT database comprising clay
of 100 when D∗ = 0 and a value of 1 when D∗ = 2 . For points in layers overlaying sand layers, the clay layers were found to have qc in
3
between, simple interpolation is used: the range of 0.5–5 MPa. In rare cases, the qc values of the clay layers
(10) were found to be lower than 0.5 MPa, but these cases consisted of sites
DF = −121.25D∗ + 100
with deep clay layers with little to no sand. For all profiles, the ground
The value of DF ranges from 1 to 100 for the least and most variable water table was assumed to be at ground level. Results of maximum

43
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

Ds Ds

(0.0,1)
(0,0,1)
Dc + Dms + Ds = 1

(c, ms, s)
D*
D*
(0,1,0) (1,0,0)
Dms ,
(0.1,0)
,
Dc Dms
(1,0,0)
Dc

(a) (b)
Fig. 10. Schematic representation of Dc + Dms + Ds = 1 plane used to calculate diversity factor DF from the calculated dominance factors Di (c, ms, s) and the
distance D* of dominance factors from central point ( 1 1 1
, ,
3 3 3 ): (a) side view (b) view normal to plane.
1.0
0.9
Minimum NDL (layers/m)

0.8
0.7
0.6
0.5
0.4
0.3 Clay
Variable
0.2
depth
0.1
0.0
0 5 10 15 20 25 30
Profile length (m)

Fig. 11. Minimum number of different layers per unit length NDL as a function
of profile length assuming only one layer in entire profile.

COV for different soil profile length are presented in Fig. 13; an ap-
propriate value may be chosen based on the length of the soil profile.
Sand 0.5 m
3.6. Horizontal Variability Index (HVI)

The horizontal variability index is calculated in terms of how well qc

correlates across soundings. Fig. 14 shows how the HVI is obtained. Fig. 12. Idealized soil profile comprising clay layer overlying a sand layer of
First, pairs of CPT soundings are formed from the available CPTs. For fixed thickness (0.5 m), used for calculation of maximum COV of qc [COV
every pair of CPTs, two quantities are computed for the cone resistance: (qc)]max.
(1) a measure of the difference between the trends of cone resistance qc
with depth (denoted here as |Δqc,avg|) and (2) the cross-correlation
coefficient of qc. To quantify the trend difference between the qc values Table 4
for a pair of soundings, the following steps are followed: Maximum COV of qc obtained from idealized soil profiles
of different lengths having clay at the top followed by
sand layers (of fixed thickness) below.
(1) calculate the absolute value |Δqc,avg| of the difference of qc between
the two soundings at each depth for the segment length considered; Length of soil profile COV (%)
(2) calculate the average |Δqc,avg| of all the values obtained in step 1;
3 142
(3) divide the resulting average value by the maximum credible dif- 4 173
ference |Δqc,avg|max between qc trends for the length of soil profile 5 181
considered. 7 190
10 190
15 182
This maximum credible difference |Δqc,avg|max is obtained by con-
20 170
sidering two idealized soil profiles, one with a very soft clay layer 30 146
throughout, and one with a sand with 85% relative density throughout.

44
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

220 11

Maximum average qc difference (MPa)

200 10
180
9
160
Maximum COV of qc

140 8

120 7
100 6
80
5
60
4
40
20 3
0 2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Length of soil profile (m) Length of soil profile (m)

Fig. 13. Maximum COV of qc [COV(qc)]max versus length of soil profile. Fig. 15. Maximum average qc difference versus length of soil profile.

The resulting differences in average qc as a function of soil profile Cxy

ρxy =
length are shown in Fig. 15. The differences between the cone re- σx σy (15)
sistance trends can be expressed by the function f0:
where σx and σy are the standard deviations of X and Y. The cross-
|Δqc,avg | correlation coefficient takes values in the −1 to +1 range. To obtain an
f0 = ρ0 (|Δqc,avg |) = index that takes values between 0 and 1 or 0 and 100%, a function f1(ρ)
|Δqc,avg |max (13) is defined:

After f0 is quantified, the next step is to calculate the cross-corre- ρxy + 1

f1 (ρ) =
lation coefficient between the soundings. Calculation of the cross-cor- 2 (16)
relation coefficient requires first the calculation of the cross-covariance. The only missing component in the HVI calculation is the distance
The cross-covariance Cxy of two signals X and Y, each containing N between the soundings. A high cross-correlation value and small trend
points, can be calculated as: difference of a CPT pair indicates high correlation and similarity be-
N −1 tween the two CPTs, and this suggests low variability in the horizontal
1 direction for the site. Two nearby CPTs will naturally have a high cross
Cxy =
N
∑ (x i−x )(yi −y )
i= 0 (14) correlation value and a small cone resistance trend difference; there-
fore, we need to take the spacing s between CPT soundings into account
where x and y are the mean values of qc for the soundings. when calculating the HVI from trend difference and cross-correlation
The cross-correlation coefficient ρxy is then calculated as: coefficient values. The spacing s is calculated using the geographical

For the CPT pair calculate cross correlation

Obtain CPT data
coefficient of qc, (ȡxy)

Form all possible CPT pairs
i =1
For ith CPT pair calculate average of the
absolute value of the difference of qc values
(at each depth) across the depth of
sounding, (ȡ0)
Normalize calculated cross correlation coefficient
of qc, to obtain values between 0 and 1, f1 (ȡxy)
= (1 + ȡxy )/2
Calculate horizontal variability function value for the
CPT pair, f2 = 1-e-(spacing between the CPT pair)/4

Divide resulting value by maximum

credible difference between cone
Calculate f of ith pair as:
resistance trends for the depth of soil
f = [0.8×(1-f0) + 0.2×f1]× f2
profile considered, term it f0(ȡ0)

i++ i > number

of possible Calculate HVI as Save site HVI
FALSE CPT pairs
TRUE

Fig. 14. Flowchart for calculation of horizontal variability index HVI.

45
R. Salgado et al. Computers and Geotechnics 105 (2019) 37–50

Function of spacing between soundings f2

Table 5
1.0 Description of test sites.
Site No. Site name Number of soundings Sounding depth analyzed (m)
0.8
1 SNC 6 10
2 IMC-1 6 13
3 IMC-2 6 08
0.6
4 LAC 6 15
5 SMC 6 12
0.4

uncertainty in design. The protocol proposed in this paper has per-

0.2
formed very well for many sites to which it has been applied, as is
discussed next.
0.0
0 5 10 15 20 25 30
4. Application of algorithm
Spacing between soundings (m)

Fig. 16. Decay function f2 as a function of spacing between CPT sounding.
coordinates (latitude and longitude) of the CPTs in the pair and the
Haversine formula for the great circle distance between two points on
the surface of the Earth. A separation-decay function f2(s) is then de-
fined such that it decays towards zero as the spacing approaches zero:
−s
f2 (s ) = 1−e 4 (17)
Fig. 16 shows the plot of f2(s). It can be seen to be approximately
equal to 1 for spacing exceeding 15 m, which is considered sufficient for
the cross-correlation coefficient to render a strong indication of whe-
ther variability exists in the horizontal direction without any correc-
tion.
To calculate the HVI, a horizontal variability function f is formed to
consider both the variability measures (difference in cone resistance
trends and cross correlation) and the spacing between the CPT sound-
ings:
The CPT based site variability assessment algorithm is applied to
five CPT sites to assess its efficacy. CPT data for the sites were obtained
from the Unites States Geological Survey (USGS) CPT database [38].
Sites were chosen such that the vertical sampling during CPT was at
most 50 mm apart to have three CPT data points within the thinnest soil
layers (min layer thickness 150 mm). Sites with at least 6 soundings
were chosen. A description of each site is given in Table 5. The depths
to which the soil profiles were analyzed varied from 8 m to 15 m, de-
pending on the depth of available data. The computed variability in-
dices for individual profiles and the sites are presented in Table 6. For
the analysis, soil profiles were generated using the modified Tumay
chart and the soil profile generation algorithms proposed by Salgado
et al. [11] and Ganju et al. [34]. Soil profiles using these algorithms are
presented in Fig. 17.
Table 6
Vertical variability index, horizontal variability index and site variability rating
values for the sites analyzed.

f = f2 [0.8(1−f0 ) + 0.2f1 ] Site No. Site name Soundinga ID VVI Site VVI Site HVI SVRb
(18)
1 SNC SNC002 21 19 10 LL
The HVI is calculated based on the average of the horizontal
SNC003 18
variability function f(|Δqc,avg|, ρ, s) for all CPT pairs i at the site (it takes SNC004 13
values between 0 and 1): SNC009 21
SNC010 19
n
∑i = 1 f (|Δqc, avg |i , ρi , si ) SNC012 23
HVI = 1−
n (19) 2 IMC-1 IMC012 50 47 31 ML
IMC013 41
where n is the number of CPT pairs available at the site. IMC014 44
IMC015 44
IMC016 48
3.7. Site variability rating (SVR)
IMC017 55

The average VVI and HVI (both of which range from 0 to 100%) of a 3 IMC-2 IMC029 53 51 42 MM
IMC030 48
site define its variability. It is not possible to combine them numerically IMC031 51
because they are fundamentally different, but they can be used as a IMC033 52
numeric pair or, alternatively, to classify a site as being of low (L), IMC034 44
medium (M) or high (H) variability in the horizontal or vertical direc- IMC035 56
tion, depending on whether the HVI or VVI fall in the [0–33%), 4 LAC LAC005 36 39 47 MM
[33–66%) or [66–100%] range. We can then establish a rating, defined LAC006 53
LAC007 40
in terms of a string variable with two characters, each of which may
LAC008 45
assume the values, L, M or H. LAC009 26
Once a site has been classified using this system, it is possible to LAC010 30
develop design protocols whose measures of uncertainty reflect the 5 SMC SMC001 57 54 66 MH
variability of the site. Resistance factors in LRFD, for example, would SMC002 55
increase with decreasing site variability. The system takes the very SMC003 53
complex mathematics of spatial variability, which is difficult to apply SMC006 54
SMC007 43
given the nature and amount of data collected, and reduces it to an
SMC008 53
algorithm that can enable practicing engineers to take site variability
into account in a relatively simple way. It is important to stress that a
Sounding ID same as found in https://earthquake.usgs.gov/research/cpt/
rigorous quantification of variability at a site is likely not viable, but data/table/.
b
that should not be a reason not to have protocols for accounting for First letter corresponds to VVI and second letter corresponds to HVI.

46
R. Salgado et al.
The IMC-2 and SMC sites have the highest vertical variability in-
dices among the sites presented, which is evident from the fact that
each profile at these sites consists of sands, clays and silty soils with
relatively equal representation. Comparatively, the SNC site has the
lowest vertical variability indices, which can be attributed to the fact
that the soil profiles consist predominantly of medium stiff to very stiff
clays (SNC), soils which have very similar behavior. The highest hor-
izontal variability is observed in SMC, as half of the profiles at this site
Fig. 17. Soil profiles generated using algorithms proposed by [11,37] for (a) SNC, (b) IMC-1, (c) IMC-2, (d) LAC and (e) SMC sites for CPT data obtained from USGS
CPT database [38].
47
Computers and Geotechnics 105 (2019) 37–50
are dominated by stiff clays and the other half by silty clays and sands.
In comparison, all profiles of site SNC consist of medium stiff to very
stiff clays, which results in a very low horizontal variability index.
Based on the procedures outlined in this paper, the presented sites have
been given a quantitative assessment of variability ratings ranging from
LL for Low horizontal and Low vertical variability (e.g., site SNC) to MH
medium vertical variability and high horizontal variability (e.g., SMC).
R. Salgado et al.
Fig. 17. (continued)
5. Conclusions
There is an increasing demand on geotechnical engineers to better
quantify ground properties in ways that also include reference to their
variability. In geotechnical engineering design, an assessment of the
variability of soil properties, which are often obtained from a limited
number of in situ or laboratory tests, should enable engineers to produce
better designs. To assess the variability of soil properties at a site,
consideration of the correlation structure of soil properties is necessary.
In this paper, a site variability quantification methodology was devel-
oped that relies on cone penetration test data.
To develop a comprehensive methodology for site variability
48
Computers and Geotechnics 105 (2019) 37–50
assessment, consideration was given to inter- and intra-layer variability
in a soil profile, and both vertical and horizontal variability. First, soil
classification (SBT) charts were used to obtain the subsurface soil
profiles from CPT parameters (e.g., qc and fs).
A vertical variability index (VVI), which reflects the intra-layer
variability, the log variability and the COV of the cone resistance of the
sounding, was defined to quantify the vertical variability in a CPT
sounding. The site VVI was taken as the average of the VVIs for all CPT
soundings performed at a site. A site horizontal variability index (site
HVI), based on the cross-correlation between the cone resistances of
soundings, how much they differ, and the spacing between them, was
also developed to quantify the soil variability of a site in the horizontal
R. Salgado et al.
Fig. 17. (continued)
direction. A site variability rating (SVR) system, integrating the vertical
and horizontal site variability, was developed to provide an easy-to-use
overall site variability indicator. The methodology was demonstrated
using CPT data collected at five sites. The SVR maps can also prove to
be helpful with the planning of site investigations projects (a highly
variable site would require a greater number of soundings or more
closely spaced soundings). With the enlargement of databases and re-
finements to the methodology, it should develop into a tool that will
enable geotechnical engineers to perform planning, design and con-
struction activities with greater confidence.
Acknowledgement
The work presented in this paper was funded by the Joint
Transportation Research Program (JTRP) administered by the Indiana
Department of Transportation (INDOT) and Purdue University through
contract SPR-3408. The support of the Indiana Department of
Transportation (INDOT) and the Federal Highway Administration
(FHWA) are gratefully acknowledged. The authors would also like to
acknowledge the assistance of Farhan Rahman with preliminary work
leading to this research.
References
[1] Foye KC, Abou-Jaoude G, Prezzi M, Salgado R. Resistance factors for use in load and
resistance factor design of driven pipe piles in sands. J Geotech Geoenviron Eng
2009;135:1–13. https://doi.org/10.1061/(ASCE)1090-0241(2009)135:1(1).
[2] Foye KC, Salgado R, Scott B. Assessment of variable uncertainties for reliability-
based design of foundations. J Geotech Geoenviron Eng 2006;132:1197–207.
[3] Foye KC, Jaoude GA, Salgado R. Limits states design of deep foundations 2004.
https://doi.org/10.5703/1288284313262.This.
[4] Foye KC, Prezzi M, Salgado R. Developing resistance factors for design of piles in
sand. ASCE Proc. Georisk, Reston, VA: ASCE; 2011, p. 444–15.
[5] Foye KC. A rational, probabilistic method for the development of geotechnical load
and resistance factor design. Purdue University; 2005. doi: 3198165.
[6] Fenton GA, Griffiths DV. 3. Random fields. Risk assess. Geotech Eng, vol. 1. 1st ed.,
Hoboken, New Jersey: John Wiley & Sons, Inc.; 2008. p. 393.
[7] Baecher GB, Christian JT. Relaibility and statistics in geotechnical engineering. 1st
ed. West Sussex, England: John Wiley & Sons; 2003.
[8] Vanmarcke E. Random fields: analysis and synthesis. MIT Press. Cambridge, MA. 1st
ed., London, England: The MIT Press; 1983. p. 382.
[9] Griffiths DV, Fenton GA. Probabilistic methods in geotechnical engineering. 491th
49
Computers and Geotechnics 105 (2019) 37–50
ed. New York: Springer; 1997.
[10] Paikowsky SG. Load and resistance factor design (LRFD) for deep foundations.
Transp Res Board 2004.
[11] Salgado R, Prezzi M, Ganju E. Assessment of site variability from analysis of cone
penetration test data 2015. https://doi.org/10.5703/1288284315523.
[12] Vanmarcke EH, Fenton GA. Probabilistic site characterization at the national geo-
technical experimentation sites Reston, VA: American Society of Civil Engineers;
2003. 10.1061/9780784406694.
[13] Elkateb T, Chalaturnyk R, Robertson PK. An overview of soil heterogeneity:
quantification and implications on geotechnical field problems. Can Geotech J
2003;40:1–15. https://doi.org/10.1139/t02-090.
[14] Wang Y, Huang K, Cao Z. Probabilistic identification of underground soil stratifi-
cation using cone penetration tests. Can Geotech J 2013;50:766–76. https://doi.
org/10.1139/cgj-2013-0004.
[15] Cao Z, Wang Y. Bayesian approach for probabilistic site characterization using cone
penetration tests. J Geotech Geoenviron Eng 2013;139:267–76. https://doi.org/10.
1061/(ASCE)GT.1943-5606.0000765.
[16] Wang Y, Au SK, Cao Z. Bayesian approach for probabilistic characterization of sand
friction angles. Eng Geol 2010;114:354–63. https://doi.org/10.1016/j.enggeo.
2010.05.013.
[17] Fenton GA, Griffiths DV. Naghibi F. Future directions in reliability-based geo-
technical design. Geo-Risk 2017 GSP 2017;282:350–9.
[18] Kulhawy F. Foundation engineering, geotechnical uncertainty, and reliability-based
design. Geotech Saf Reliab GSP 2017;286:174–84.
[19] Zhu H, Zhang L. Characterizing geotechnical anisotropic spatial variations using
random field theory. Can Geotech J 2013;50:723–34. https://doi.org/10.1139/cgj-
2012-0345.
[20] Firouzianbandpey S, Griffiths DV, Ibsen LB, Andersen LV. Spatial correlation length
of normalized cone data in sand: case study in the north of Denmark. Can Geotech J
2014;857:844–57. https://doi.org/10.1139/cgj-2013-0294.
[21] Uzielli M, Vannucchi G, Phoon K. Random field characterisation of stress-nomalised
cone penetration testing parameters. Geotechnique 2005;55:3–20.
[22] Phoon K, Quek S-T, An P. Geostatistical analysis of cone penetration test (CPT)
sounding using the modified Bartlett test. Can Geotech J 2004;41:356–65. https://
doi.org/10.1139/t03-091.
[23] Stolle D, Guo P, Sedran G. Impact of random soil properties on stress – strain re-
sponse. Can Geotech J 2004;355:351–5. https://doi.org/10.1139/T03-083.
[24] Lloret-Cabot M, Fenton GA, Hicks MA. On the estimation of scale of fluctuation in
geostatistics. Georisk Assess Manag Risk Eng Syst Geohazards 2014;8:129–40.
https://doi.org/10.1080/17499518.2013.871189.
[25] Liu WWF, Leung YFAYF, Lo MKMK. Integrated framework for characterization of
spatial variability of geological profiles. Can Geotech J 2016;54:47–58. https://doi.
org/10.1139/cgj-2016-0189.
[26] Phoon K, Kulhawy F. Evaluation of geotechnical property variability. Can Geotech J
1999;36:625–39. https://doi.org/10.1139/cgj-36-4-625.
[27] Salgado R, Kim D. Reliability analysis of load and resistance factor design of slopes.
J Geotech Geoenviron Eng 2014;140:57–73. https://doi.org/10.1061/(ASCE)GT.
1943-5606.0000978.
[28] Wang Y, Cao Z, Li D. Bayesian perspective on geotechnical variability and site
R. Salgado et al.
characterization. Eng Geol 2016;203:117–25. https://doi.org/10.1016/j.enggeo.
2015.08.017.
[29] Phoon K, Kulhawy F. Characterization of geotechnical variability. Can Geotech J
1999;36:612–24. https://doi.org/10.1139/cgj-36-4-612.
[30] Phoon K, Kulhawy F, Grigoriu MD. Development of a reliability-based design fra-
mework for transmission line structure foundations. J Geotech Geoenviron Eng
2003;129:798–806. https://doi.org/10.1061/(ASCE)1090-0241(2003)129:9(798).
[31] Popescu R, Deodatis G, Nobahar A. Effects of random heterogeneity of soil prop-
erties on bearing capacity. Probab Eng Mech 2005;20:324–41. https://doi.org/10.
1016/j.probengmech.2005.06.003.
[32] Uzielli M, Lacasse S, Nadim F, Phoon K. Soil variability analysis for geotechnical
practice. Charact Eng Prop Nat Soils 2006:1653–752. https://doi.org/10.1201/
NOE0415426916.ch3.
[33] Zhang LM, Dasaka SM. Uncertainties in geologic profiles versus variability in pile
founding depth. J Geotech Geoenviron Eng 2010;136:1475–88. https://doi.org/10.
1061/(ASCE)GT.1943-5606.0000364.
[34] DeGroot DJ, Baecher GB. Estimating autocovariance of in-situ soil properties. J
Geotech Eng 1993;119:147–66. https://doi.org/10.1061/(ASCE)0733-9410(1993)
119:1(147).
[35] Fenton GA. Estimation for stochastic soil models. J Geotech Geoenviron Eng
1999;125:470–85.
[36] Tian M, Li DQ, Cao ZJ, Phoon KK, Wang Y. Bayesian identification of random field
model using indirect test data. Eng Geol 2016;210:197–211. https://doi.org/10.
1016/j.enggeo.2016.05.013.
[37] Ganju E, Prezzi M, Salgado R. Algorithm for generation of stratigraphic profiles
using cone penetration test data. Comput Geotech 2017;90:73–84. https://doi.org/
10.1016/j.compgeo.2017.04.010.
[38] Holzer BTL, Noce TE, Bennett MJ. Maps and Documentation of Seismic CPT
Soundings in the Central, Eastern, and Western United States. Reston, VA; 2010.
[39] Li J, Cassidy MJ, Huang J, Zhang L, Kelly R. Probabilistic identification of soil
stratification. Géotechnique 2016;66:16–26. https://doi.org/10.1680/jgeot.14.P.
242.
[40] Begemann HKS. The friction jacket cone as an aid in determining the soil profile. In:
Proc 6th int conf soil mech found eng. Montreal: ICSMFE; 1965. p. 8–15.
[41] Sanglerat G, Nhim TV, Sejourne M, Andina R. Direct soil classification by static
penetrometer with special friction sleeve. In: Proc first Eur Symp penetration test;
1974. p. 5–7.
[42] Schneider JA, Randolph MF, Mayne PW, Ramsey NR. Analysis of factors influencing
soil classification using normalized piezocone tip resistance and pore pressure
parameters. J Geotech Geoenviron Eng 2008;134:1569–86.
[43] Robertson P. CPT-based Soil Behaviour Type (SBT) Classification System – an up-
date. Can Geotech J 2016. https://doi.org/10.1139/cgj-2016-0044. cgj-2016-0044.
[44] Robertson PK. Soil behaviour type from the CPT : an update. In: Second int symp
cone penetration test, Huntington Beach, California; 2010.
[45] Schneider JA, Hotstream JN, Mayne PW, Randolph MF. Comparing CPTU Q – F and
Q – Δ u 2 /σ v0 ′soil classification charts. Géotechnique Lett 2012;2:209–15.
https://doi.org/10.1680/geolett.12.00044.
[46] Schmertmann JH. Guidelines for cone penetration test: performance and design;
1978. Gainesville, Florida.
[47] Douglas BJ, Olsen RS. Soil classification using electric cone penetrometer. Symp
Cone Penetrat Test Exp 1981.
[48] Tumay MT. Field calibration of electric cone penetrometers in soft soils. Baton
Rouge; 1985.
[49] Senneset K, Sandven R, Janbu N. Evaluation of soil parameters from piezocone
tests. Transp Res Rec 1989;1235.
[50] Robertson PK. Soil classification using the cone penetration test. Can Geotech J
1990;27:151–8. https://doi.org/10.1139/t90-014.
[51] Larsson R, Mulabdic M. Piezocone tests in clay; 1991.
[52] Jefferies MG, Davis MP. Soil classification by the cone penetration test: discussion.
50
Computers and Geotechnics 105 (2019) 37–50
Can Geotech J 1991;28:1–178.
[53] Ramsey N. A calibrated model for the interpretation of cone penetration tests
(CPTs) in North Sea quaternary soils. Offshore Site Investig. Geotech. Divers.
Sustain., Society of Underwater Technology; 2002.
[54] Arshad MI, Tehrani FS, Prezzi M, Salgado R. Experimental study of cone penetration
in silica sand using digital image correlation. Géotechnique 2014;64:551–69.
https://doi.org/10.1680/geot.13.P.179.
[55] Xu X, Lehane BM. Pile and penetrometer end bearing resistance in two-layered soil
profiles. Géotechnique 2008;58:187–97. https://doi.org/10.1680/geot.2008.58.3.
187.
[56] Fenton GA, Griffiths DV. Risk assessment in geotechnical engineering. 10th ed.
Hoboken, New Jersey: John Wiley & Sons; 2008.
[57] Salgado R, Prezzi M. Computation of cavity expansion pressure and penetration
resistance in sands. Int J Geomech 2007;7:251–65. https://doi.org/10.1061/
(ASCE)1532-3641(2007)7:4(251).
[58] Daniel W. Applied nonparametric statistics. 2nd ed. Duxbury; 1990.
[59] Vanmarcke E. Probabilistic modeling of soil profiles. J Geotech Eng Div
1977;103:1227–46.
[60] Honjo Y, Kuroda K. A new look at fluctuating geotechnical data for reliability de-
sign. Soils Found 1991;31:110–20. https://doi.org/10.3208/sandf1972.31.110.
[61] Chiasson P, Lafleur J, Soulié M, Law KT. Characterizing spatial variability of a clay
by geostatistics. Can Geotech J 1995;32:1–10. https://doi.org/10.1139/t95-001.
[62] Jaksa MB, Kaggwa WS, Brooker PI. Experimental evaluation of the scale of fluc-
tuation of a stiff clay. In: Proc 8th int conf appl stat probab, Sydney; 1999.
[63] Phoon K, Kulhawy F, Grogoriu MD. Reliability-based design of foundations for
transmission line structures. Ithaca, New York: 1995. 10.1061/40790(218)17.
[64] Jaksa MB. The influence of spatial variability on the geotechnical design properties
of a stiff, overconsolidated clay. The University of Adelaide; 1995.
[65] Bombasaro E, Kasper T. Evaluation of spatial soil variability in the Pearl River
Estuary using CPTU data. Soils Found 2016;56:496–505. https://doi.org/10.1016/
j.sandf.2016.04.015.
[66] Reyna F, Chameau J. Statistical evaluation of CPT and DMT measurements at the
Heber Road site. Geotech Eng Congr 1991.
[67] Orchant C, Kulhawy F, Trautmann C. Reliability-based foundation design for
transmission line structures. Crit Eval In situ Test Methods 1988;2.
[68] Filippas O, Kulhawy F, Grigoriu M. Reliability-based foundation design for trans-
mission line structures. Uncertain Soil Prop Meas 1988.
[69] Spry M, Kulhawy F, Grigoriu M. Reliability-based foundation design for transmis-
sion line structures. Geotech Site Charact Strategy 1988.
[70] Campanella R, Wickremesinghe D, Robertson P. Statistical treatment of cone pe-
netrometer test data. In: Proc 5th int conf appl stat probab. Rotterdam: AA Balkema;
1987.
[71] Schultze E. Some aspects concerning the application of statistics and probability to
foundation structures. In: Proceeding 2nd int conf appl stat probab soil struct eng,
Aachen, Germany; 1975.
[72] Baligh MM, Vivatrat V, Ladd CC. Exploration and evaluation of engineering prop-
erties for foundation design of offshore structures. MIT; 1979.
[73] Joustra K, Comparative measurements on the influence of the cone shape on results
of soundings. In: Proc 1st Eur symp; 1974.
[74] Alonso E, Krizek R. Stochastic formulation of soil properties. Proc Int Conf Appl Stat
Probab Soil Struct Eng 1975.
[75] Kulhawy F, Birgisson B, Grigoriu MD. Reliability-based foundation design for
transmission line structures; 1992.
[76] Jaksa MB, Brooker PI, Kaggwa WS. Inaccuracies associated with estimating random
measurement errors. J Geotech Geoenviron Eng 1997;123:393–401. https://doi.
org/10.1061/(ASCE)1090-0241(1997)123:5(393).
[77] Kim KK, Prezzi M, Salgado R. Interpretation of cone penetration tests in cohesive
soils; 2006.
[78] Salgado R. The engineering of foundations. 1st ed. New York: McGraw-Hill; 2008.