See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/277686340

CPT-based stratigrahpic profiling using the wavelet transform modulus

maxima method

Article in Canadian Geotechnical Journal · May 2015

DOI: 10.1139/cgj-2015-0027

CITATIONS READS

77 2,366

4 authors:

Jianye Ching Jiun-Shiang Wang

National Taiwan University National Center for Research on Earthquake Engineering
257 PUBLICATIONS 7,068 CITATIONS 11 PUBLICATIONS 173 CITATIONS

SEE PROFILE SEE PROFILE

C.Hsein Juang Chih-sheng Ku

Clemson University I-Shou University
291 PUBLICATIONS 10,909 CITATIONS 21 PUBLICATIONS 618 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jianye Ching on 26 December 2015.

The user has requested enhancement of the downloaded file.

CPT-based stratigraphic profiling using the wavelet transform modulus maxima method

Jianye Ching1, Jiun-Shiang Wang2, C. Hsein Juang3, and Chih-Sheng Ku4

ABSTRACT

In this paper, a stratigraphic profiling approach is proposed based on the soil behavior type index Ic ob-

tained from the cone penetration test (CPT). The basic idea of this approach is simple: the layer

boundaries can be identified as the points at which a change occurs in the Ic profile. It is shown that

these change points can be easily identified using the wavelet transform modulus maxima (WTMM)

method. This method is able to accurately pinpoint the locations of change points in the Ic profile and

to produce graphs and plots that fit well with engineers’ methods of visualization and intuition. More-

over, by virtue of the Fast Fourier Transform (FFT), the computation is very fast. Case studies show

that the WTMM method is effective for the detection of change points in the Ic profile. It is also capa-

ble of detecting thin soil layers.

Key words: site investigation; underground stratification; CPT; soil behavior type; wavelet transform

INTRODUCTION

Underground stratigraphic profiling is a necessary step in site investigation. The primary purpose of

stratigraphic profiling is to identify the boundaries between homogeneous soil layers. In recent years,

stratigraphic profiling techniques based on the cone penetration test (CPT) have attracted significant

attention. The CPT parameters, such as cone resistance, sleeve friction, and pore water pressure, can

1
(Corresponding author) Professor, Dept. of Civil Engineering, National Taiwan University, Taipei, Taiwan. Email:
[email protected]. Tel: +886-2-33664328.
2
Graduate student, Dept. of Civil Engineering, National Taiwan University, Taipei, Taiwan.
1
generally reveal the physical and mechanical behaviors of soils (Ku et al. 2010). Moreover, the CPT

can produce nearly continuous (0.01 ~ 0.05 m vertical interval) records at an affordable cost and within

a reasonable time. This makes the CPT an ideal tool for stratigraphic profiling.

CPT stratigraphic profiling is an important topic that has received considerable attention. Some

of the previous studies on this topic have adopted the concept of probability or fuzzy sets (e.g., Zhang

and Tumay 1999; Jung et al. 2008; Wang et al. 2013), and others have adopted clustering analysis in

the (Qtn, Fr, Bq) space (or subspace) (e.g., Hegazy and Mayne 2002; Facciorusso and Uzielli 2004; Liao

and Mayne 2007; Das and Basudhar 2009). Other studies have attempted to identify homogenous

layers using statistical analysis (e.g., Phoon et al. 2003; Uzielli 2004; Uzielli et al. 2008) or to identify

layer boundary locations using statistical analysis (e.g., Wickremesinghe and Campanella 1991). The

Bayesian method developed by Wang et al. (2013) appears to be the most rigorous, because it

determines the layer profile (including the number of soil layers and the locations of the boundaries)

based on the largest posterior probability. However, it is a sophisticated method that requires advanced

knowledge of Bayesian decision analysis and non-convex optimization, and it is also computationally

intensive.

Although each of the aforementioned studies has its own theoretical grounds, many of them are

more or less “abstract” methods, in the sense that the stratification problem is converted to an abstract

domain or an abstract concept. For practicing engineers, the most “concrete” method of investigating

underground stratification is to consider the profile of the soil behavior type (SBT) index. This SBT

index Ic (Robertson and Wride 1998; Roberston 2009) is well known among engineers and has been

proven to be effective in discriminating soil types (e.g., Ku et al. 2010). Intuitively, a boundary

between homogeneous soil layers can be taken to be a depth at which Ic changes abruptly. Such

boundaries can be identified by finding the locations at which the trend in Ic changes abruptly. This

3
Glenn Professor of Civil Engineering, Glenn Dept. of Civil Engineering, Clemson University, SC, USA.
4
Professor, Dept. of Civil Engineering, I-Shou University, Kaohsiung, Taiwan.
2
concept is illustrated in Figure 1, where the dashed lines represent two Ic trends. At a depth of 4.6 m,

the trend changes abruptly. In this paper, this abrupt change in trend is referred to as a “jump”. Jumps

are actual layer boundaries. In contrast, small-scale oscillations around the same trend occur because

of the spatial variability within a homogeneous soil layer. This is illustrated by the dashed rectangular

box in Figure 1, where Ic changes locally because of spatial variability, producing many small-scale

oscillations. However, such small-scale oscillations are not actual layer boundaries. Jumps (layer

boundaries) are typically observable at all scales (small, medium, and large scales), whereas the small-

scale oscillations produced by spatial variability are observable only at small scales. Figure 1 shows

such an example of a jump. It can be challenging to discriminate between jumps and the small-scale

oscillations produced by spatial variability. The human eye may be able to perform this task

subjectively, but there are strong motivations, both scientific and practical, to develop an algorithm that

is able to accomplish this task objectively and automatically.

In this study, such an algorithm based on the wavelet transform modulus maxima (WTMM)

method (Mallat 1999) is demonstrated. Using the WTMM method, discriminating between jumps

(layer boundaries) and small-scale oscillations in the Ic profile becomes feasible because of the

“multiscale zoom” ability of the wavelet transform. As will be shown later, the wavelet transform

characteristics of these two types of signals are very different. The WTMM method is intuitive

because it operates directly on the Ic profile, which is well known to engineers. This enables the

WTMM method to conform to engineers’ intuition and methods of visualization. More importantly,

the WTMM method functions quite satisfactorily: it is able to effectively identify layer boundaries

without misidentifying small-scale oscillations as spurious layer boundaries. It is also effective in

detecting thin soil layers. Furthermore, the locations of the layer boundaries can be accurately

pinpointed using the “WTMM ridge”. It is worth noting that the layer boundaries identified by the

WTMM method are boundaries for “soil behaviors”, e.g., sands versus clays. Because the evaluation is

3
based on the Ic profile, this method cannot detect boundaries between different geological units with

similar Ic values, e.g., normally consolidated versus over-consolidated clays.

The MATLAB codes for the proposed WTMM method for CPT-based stratigraphic profiling

are available upon request from the first author. The structure of this paper is as follows. First, the

SBT index Ic proposed by Robertson and Wride (1998) and Roberston (2009) is reviewed. Next, the

WTMM method is developed. Finally, the analysis results for two case histories are presented. The

results obtained using various stratification methods are also compared and discussed.

SOIL BEHAVIOR TYPE INDEX

Robertson (1990, 2009) developed the CPT soil classification chart. There are two main parameters in

the classification: the normalized cone resistance Qtn and the normalized sleeve friction Fr (Fr is in units

of %) (Robertson 2009):

Q tn   q t   v0  Pa   C N (1)

Fr  %   100  f s  q t   v0  (2)

where qt is the (corrected) cone resistance; fs is the sleeve friction; Pa = 101.3 kN/m2 is one atmosphere

of pressure; v0 and v0 are the effective and total overburden stresses, respectively; and CN is the

correction factor for the effective overburden stress:

C N   Pa v0 
n
(3)

The exponent n varies from 0.5 to 1 depending on the soil behavior type (n = 1 for clay-like soils and n

= 0.5 for sand-like soils) (Robertson 2009):

n  min  0.381 Ic  0.05   v0 Pa   0.15, 1.0  (4)

Idriss and Boulanger (2008) suggested that an upper bound of 1.7 should be applied to CN:

C N  min  Pa v0  , 1.7 

n
(5)
 

4
 Based on the soil classification chart, Robertson and Wride (1998) proposed the following soil

behavior type (SBT) index, denoted by Ic:

Ic  3.47  log  Q tn    log  Fr   1.22 

2 2
(6)

The logarithm indicated by the symbol ‘log’ in this paper is the logarithm to base 10. Table 1

summarizes the soil behavior types corresponding to various Ic values for the Robertson-Wride Ic

formulation. The SBT index defined in Eq. (6) does not account for the pore water pressure recorded

in the piezocone penetration test (CPTU). Been and Jeffery (1992) proposed an SBT index that

considers the pore water pressure recorded as the normalized excessive pore pressure Bq = (the

measured excessive pore pressure behind the cone)/(qt - v0). However, only the Robertson-Wride

formulation is considered in this paper.

WAVELET TRANSFORM MODULUS MAXIMA

This section briefly introduces the continuous wavelet transform (CWT) and the wavelet transform

modulus maxima (WTMM). We will minimize the technical details and focus on their significance.

One appealing feature of the CWT is that it is able to zoom in on a signal at various scales. This

feature makes it possible to discriminate between small-scale oscillations and jumps.

Continuous wavelet transform (CWT)

Let us begin with the so-called wavelet function (z), where z is the coordinate. For our purposes, z is

the depth. A wavelet function (z) has an average of zero (the integral of (z) from - to  must be

zero). A popular choice for (z) is the mth derivative of the Gaussian function (m is a positive integer):

 1 d m exp   z 2 2  
m

 z   (7)
2 dz m

Figure 2 shows the wavelet functions defined in Eq. (7) for m = 1 (the Gaussian derivative) and m = 2

(the Mexican hat). Both are functions oscillating around zero. In this study, the Gaussian derivative
5
(m = 1) is used as the wavelet function. This wavelet function is more suitable for identifying jumps

than are other wavelet functions at larger orders of m. The reason for its greater suitability is explained

in the Appendix.

A set of wavelets is a collection of functions generated by scaling and shifting (z). A wavelet

of location u and scale s is denoted by u,s(z). It is related to (z) as follows:

zu
 u,s  z   s 1 2     (8)
 s 

where the factor s-1/2 ensures that the norm (size) of u,s(z) is a fixed constant. u,s(z) with a small

scale s is a rapidly oscillating version of (z) with a narrow support region, whereas u,s(z) with a large

scale s is a slowly oscillating one with a large support region. Figure 3 shows several examples. The

set of wavelets, i.e., {u,s(z): s  R+, u  R}, contains wavelets of various scales and locations.

Similar to the Gaussian probability density function with mean = u and standard deviation = s, the

Gaussian derivative wavelet is centered at z = u and has a width of approximately 4s.

Wavelet transform modulus maxima (WTMM)

The continuous wavelet transform of the SBT index profile Ic(z) is denoted by WIc(u,s):

WIc  u,s    Ic (z)   u,s (z)dz

D
(9)
0

where D is the total depth of the Ic profile and WIc(u,s) is the inner product of Ic(z) with u,s(z) (the

projection of Ic(z) onto u,s(z)). Therefore, WIc(u0,s0) contains the information of Ic(z) at location z =

u0 and scale s = s0. In reality, the Ic profile consists of discrete data points with a vertical spacing of z

= 0.01 ~ 0.05 m. Let Ic(nz) be a discrete Ic data point at depth z = nz (n = 0, 1, … D/z). First of all,

the discrete Ic data are augmented into a dyadic sequence of length 2p, where p is chosen such that 2p-1

 D/z < 2p. This can be achieved by augmenting the original Ic sequence with repeated values of

Ic(D). The augmented Ic sequence is Ic(0), Ic(z), …, Ic([2p-1]z). The wavelet u,s(z) is also

6
discretized into a sequence over length 2p: u,s(0), u,s(z), …, u,s([2p-1]z). Both the Ic and u,s

sequences are converted into the frequency domain using the Fast Fourier Transform (FFT):

2p 1
Îc  k    Ic  nz  exp  i2kn 2p  k  0,..., 2p  1 (10)
n 0

2p 1
ˆ u,s  k     u,s  nz  exp  i2kn 2p  k  0,..., 2p  1 (11)
n 0

where i = (-1)0.5 and  = 2/2p is the frequency interval (in radians) in the frequency domain. The

inner product in Eq. (9) at discrete locations u = 0, z, …, (2p-1)z can be efficiently calculated using

the inverse FFT:

2p 1
WIc  nz,s    ˆIc  k   ˆ u,s  k   exp  i2kn 2p  n  0,..., 2p  1
*
(12)
k 0

where the superscript * denotes the complex conjugate.

The two-dimensional graph of the absolute value of WIc(u,s) plotted in (u, s) space is called the

wavelet transform modulus spectrum. Figure 4 shows the wavelet transform modulus spectrum for a

function that contains one jump. Figure 4a shows the idealized Ic(z) (z = 0.02 m) that contains a

single jump. Figure 4b shows its wavelet transform modulus spectrum. This spectrum is calculated

using the Gaussian derivative (m = 1) as the wavelet function. The horizontal axis is the logarithm of s

(log to base 10, s in meters). The finest scale, denoted by sf, should be chosen to be z. The coarsest

scale, denoted by sc, should be chosen to be D/4. The coarsest scale sc is chosen to be D/4 rather than

D because the Gaussian derivative wavelet has a width of approximately 4s, and therefore, a wavelet

with s = D/4 already covers the entire depth D. For the case depicted in Figure 4, the logarithm for the

finest scale is log(sf) = log(z) = log(0.02)  -1.7 and the logarithm for the coarsest scale is log(sc) =

log(D/4) = log(20.48/4)  0.71.

At each scale s, there is clearly a local maximum in the modulus |WIc(u,s)| (see Figure 4c).

These local maxima are called the wavelet transform modulus maxima (WTMM). A proposition in
7
Mallat (1999) (Proposition 6.1) states that if the wavelet function is chosen to be an mth derivative of

the Gaussian function (Eq. 7), then a local maximum at a large s is guaranteed to propagate to another

local maximum at a smaller s. In other words, the line connecting the local maxima across all scales is

uninterrupted from large to small scales. In this paper, such a non-interrupted line of WTMM is called

a WTMM ridge. The WTMM ridge for the jump depicted in Figure 4a is shown as the dashed lines in

Figure 4b and 4c.

An interesting aspect of the wavelet transform modulus spectrum is that each jump will produce

a WTMM ridge. Moreover, if we plot the log(s) versus log(|WIc(u,s)|) relationship along the WTMM

ridge for a jump, the gradient of this relationship will be (approximately) equal to 0.5. The reason why

this log(s) versus log(|WIc(u,s)|) relationship has a gradient of 0.5 is rather technical. A detailed

explanation is provided in the Appendix. Figure 5 shows the wavelet transform modulus spectrum for

an artificial Ic(z) profile to a depth of 20.48 m (z = 0.02 m; 0.02 m  210 data points = 20.48 m).

There are three jumps, A, B, and C, in the Ic(z) profile (Figure 5a). These three jumps are non-

uniformly spaced: there is a thin layer between A and B and a thick layer between B and C. A

stratification method based on a fixed moving window (e.g., the T ratio method proposed by

Wickremesinghe and Campanella (1991)) is not suitable in this case because a wide window may

smear the thin layer between A and B, whereas a narrow window may produce spurious layer

boundaries. However, the WTMM method is robust to such non-uniformly spaced jumps. It

successfully detects all three WTMM ridges corresponding to these jumps (Figure 5b) because the

wavelet transform analyzes the Ic profile at multiple scales. Moreover, if we trace the ridges to the

finest scale sf (= z = 0.02 m), we can identify the locations of the jumps. The log(s) versus

log(|WIc(u,s)|) relationships along these three WTMM ridges are plotted in Figure 5c. They all have a

gradient of approximately 0.5. The magnitude of log(|WIc(u,s)|) is largest for jump C because C has

the largest jump amplitude. There is also an additional ‘D’ ridge. This ridge arises because of the Fast

8
Fourier Transform. In the FFT, the Ic(z) function is assumed to be periodic (period = 20.48 m in the

present case). In the present case, a “false jump” occurs at the boundary (z = 0) because Ic(z = 0)  Ic(z

= 20.48). Thus, this D ridge is not associated with an actual jump.

There is also a noise region in the artificial Ic(z) profile shown in Figure 5a. This is included to

mimic the small-scale oscillations observed in Ic for natural soils. In the noise region, there are many

short WTMM ridges that are produced by these small-scale oscillations. However, these noise ridges

have very different characteristics from the ridges associated with the three jumps A, B, and C:

1. The noise ridges are significantly shorter than the ridges for A, B, and C. This is because the small-

scale oscillations in the noise region are significant only at small scales. Therefore, these noise

ridges only exist at small scales and do not extend to large scales; thus, they appear as short ridges.

By contrast, the jumps A, B, and C are significant at all scales. Therefore, the three corresponding

ridges exist at all scales (long ridges).

2. At a fixed scale, the log(|WIc(u,s)|) magnitude for the noise ridges is significantly less than that for

the three long ridges. This is because the small-scale oscillations have smaller amplitudes than do

A, B, and C.

3. The log(|WIc(u,s)|) magnitude for the noise ridges diminishes toward - (or |WIc(u,s)| diminishes to

zero) at large scales, but that for the three long ridges does not. This is because the inner product of

a noise signal with a large-scale wavelet is close to zero, as illustrated in Figure 6. The product of a

noise signal (Figure 6b) and a large-scale wavelet (Figure 6a) is a function with many cycles of

oscillations around zero (Figure 6c). The inner product is equal to the sum of these areas. The

shaded areas in Figure 6c are positive, whereas the unshaded areas are negative. The sum of these

areas (inner product) is close to zero because of the cancellation between the positive and negative

areas. Nonetheless, such a cancellation effect is weak if the wavelet is of a small scale (Figure 6d)

9
 because the product of the noise signal and the small-scale wavelet is a function with very few

cycles of oscillations (Figure 6f).

The three characteristics listed above cause the noise ridges to look quite different from the

ridges produced by jumps. Now, consider the following two parameters of a ridge:

1. The logarithm of the maximum scale, log(smax), where smax is the maximum scale (in units of meters)

to which a ridge can extend (see Figure 7). This logarithm of smax is denoted by L. In essence, L

quantifies the length of a ridge. It is recommended that the coarsest scale sc (= D/4) should be

taken as the upper bound on smax for the previously discussed reason (a wavelet with s = D/4

already covers the entire depth D).

2. The average log(|WIc(u,s)|) magnitude along the ridge. Let log(s1), log(s2), …, log(smax) be

logarithms of increasing scales (the initial scale is s1 = 0.05 m). log(s1), log(s2), …, log(smax) are

equally spaced, namely, log(si+1) - log(si) = log(si) - log(si-1) =  ( = 0.025). The average

log(|WIc(u,s)|) magnitude can be computed as the average of log(|WIc(u,s1)|), log(|WIc(u,s2)|), …,

log(|WIc(u,smax)|). This average magnitude is denoted by M. In essence, M quantifies the height of

a ridge. The initial scale s1 should be taken to be 0.05 m rather than 0.01 or 0.02 m because z =

0.05 m for many CPT datasets. These CPT data do not have sufficient resolution to allow

|WIc(u,s)| to be computed at s = 0.01 or 0.02 m.

In reality, jumps and small-scale oscillations (noise) coexist. Figure 8a shows three jumps

coexisting with small-scale oscillations. The corresponding WTMM ridge plot is shown in Figure 8b.

Each ridge produces a point in L-M space. The horizontal coordinate of this point is its L value,

whereas the vertical coordinate is its M value. Figure 8c shows the L-M plot for all ridges. It is clear

that the L-M points for the noise ridges reside in the lower left region of the plot, where both L and M

are small, whereas the L-M points for the three long ridges (A, B, C) reside in the upper right region,

10
where both L and M are large. Following the three long ridges to the finest scale sf (= z) will yield

the locations of the three jumps.

The issue of transition zones

Using numerical analyses, Ahmadi and Robertson (2005) demonstrated that depending on the strength

of the soil and the in situ stress, the cone in a CPT can sense a layer boundary at a distance of up to 15

cone diameters ahead and behind. The distance is longer for stiff soils and under low effective stress

and is shorter for soft soils and under high effective stress. This behavior introduces a further

complication into the WTMM method: a layer boundary may produce a gradual jump in Ic(z) rather

than a sharp jump. Nonetheless, Figure 9 shows that a gradual jump can still produce a long WTMM

ridge that has large L and M values (even buried in noise). However, it is anticipated that the layer

boundary location that can be identified from a long transition zone will be less accurate than that

identified from a short transition zone.

Probability of a jump

In this paper, we propose that a linear combination of the L and M parameters that characterize a ridge

can be used to effectively determine the probability that the ridge is associated with a jump. Recall that

L and M quantify the length and height of a ridge, respectively. If aL + bM is large (a and b are

positive), then the ridge is more likely to be associated with a jump. Otherwise, it is more likely to be

associated with a small-scale oscillation. This “jump probability” can be calibrated based on the L-M

data points derived from real CPT sounding records, as described below.

Fifty real CPT sounding records were collected. Table 2 summarizes the basic information

concerning these real records. The Ic(z) profiles for all real records were discretized using z = 0.02 m,

and the discrete Ic data were augmented into a dyadic sequence of length n = 212 = 4096. Figure 10a

shows the Ic profile for a real case in Yuanlin (Taiwan), and Figure 10b shows the corresponding

WTMM ridge plot. In the Ic(z) profile, there are 5 possible jumps at depths of 2.73 m, 3.43 m, 5.37 m,

11
12.77 m, and 15.03 m. These possible jumps were identified by eye. The WTMM ridges

corresponding to these jumps were then determined (solid lines in Figure 10b), and their L-M data

points are plotted as circles in Figure 10c. To match the number of jumps, 5 ridges that were possibly

produced by small-scale oscillations were also selected by eye. Their L-M data points are plotted as

crosses in Figure 10c. The same exercise was performed for all fifty real Ic profiles. In total, 406 L-M

data points for jumps and 406 L-M data points for small-scale oscillations were identified (see Figure

11a). As shown in Figure 11a, it is difficult to identify a deterministic boundary that separates the

jumps and small-scale oscillations. We cannot determine with certainty whether a WTMM ridge

corresponds to a jump or a small-scale oscillation. The uncertainty of the decision model in judging

whether a WTMM ridge corresponds to a jump or a small-scale oscillation is quantified by Eq. 13.

Instead of asserting that a ridge corresponds to a jump with perfect certainty, we assign a probability

(pjump) to reflect this uncertainty using Eq. 13, which was obtained by applying logistic regression to

the L-M data points shown in Figure 11a:

1
p jump  P  jump   (13)
1  exp    4.398  L  3.429  M  5.025  

where pjump is the probability that the ridge under consideration is associated with a jump in the Ic(z)

profile. The contours for the pjump values are plotted in Figure 11a. Note that pjump depends only on X

= (4.398L + 3.429M + 5.025), and Figure 11b illustrates this dependency: pjump increases with

increasing X. In this paper, a change point with pjump > 0.5 is considered to be a jump because such a

value implies that the probability that it is a jump is greater than the probability that it is a small-scale

oscillation [P(jump) > P(small-scale oscillation)]. This simple rule is often used in the field of decision

science. Other choices can also be made for the pjump threshold. Increasing this threshold to a higher

value (e.g., 0.8) will reduce the number of jumps identified. The benefit of doing so is to reduce the

probability of misidentifying small-scale oscillations as jumps, at the cost of missing some actual

jumps. Although this jump probability equation was calibrated based on cases with z = 0.02 m and n
12
= 4096, this equation is also applicable to other scenarios (e.g., z = 0.05 m and n = 512) because the

calculation results for L and M do not depend on z and n.

APPLICATION OF THE WTMM METHOD TO CASE HISTORIES

The NGES at Texas A&M University

The CPT data collected at the National Geotechnical Experimentation Site (NGES) at Texas A&M

University (clay site) (Briaud 2000) are analyzed in this section. The same case history has also been

analyzed by Zhang and Tumay (1999) and Wang et al. (2013) for CPT-based underground

stratification. From the ground surface to a depth of 15 m, this site comprises a sequence of very stiff

to hard clays with silt seams. The ground water table is located at a depth of approximately 6 m.

Figure 12 shows the CPT data together with the Ic profile (Figure 12c). The data interval z is equal to

0.02 m. Figure 12d shows the (Qtn, Fr) data points in the Robertson soil classification chart. It is clear

that the soil behavior types are primarily 3, 4, 5, 6, 8, and 9.

The wavelet transform modulus spectrum for the Ic(z) profile was calculated, and the traces of

all WTMM ridges are shown in Figure 13. The original Ic(z) profile shown in Figure 12c was

augmented into a record of length n = 210 = 1024 (see Figure 13a), yielding a total depth of D = nz =

10240.02 = 20.48 m. Figure 13b shows the wavelet transform modulus spectrum together with the

traces of all ridges, whereas Figure 13c shows the L-M plot for all ridges. The contour lines in Figure

13c correspond to pjump. There are 7 ridges (A, B, …, F, G) with pjump > 0.5. However, ridge G is a

false ridge originating from the periodicity assumption in the FFT. Therefore, this ridge was discarded.

The L-M plot shows the remaining 6 ridges as circles. They all reside in the upper right portion of the

L-M plot. These ridges were deemed to have been caused by jumps in the Ic(z) profile. By following

these ridges to the finest scale, sf = z = 0.02 m, the locations of the layer boundaries could be

determined.

13
 Figure 14a shows the soil stratification pattern determined based on the 6 ridges. The

horizontal dashed lines in Figure 14a represent the layer boundary locations that were identified by

following the 6 ridges to the finest scale, sf = z = 0.02 m. The number marked on the left-hand side of

each dashed line represents the pjump value of each ridge. To visualize the pjump values more clearly, the

dashed lines with pjump close to 1 are shown in black, whereas those with pjump close to 0.5 are shown in

light gray. In essence, the dashed lines shown in black have higher probabilities of being actual layer

boundaries than do those shown in light gray. Figure 14b shows the probabilities that each identified

homogeneous layer belongs to various SBT zones of the Robertson chart. Each of these probabilities

was simply determined as the proportion of the (Qtn, Fr) data that lies in SBT zone n (SBTn). For

instance, the soil between the ground surface and a depth of 0.78 m (Layer I) was identified as a

homogeneous layer. There are 35 (Qtn, Fr) data points within this homogeneous layer; 23 of them lie in

SBT6, and 12 of them lie in SBT8. Therefore, the probability that this layer belongs to SBT6 or SBT8

is 66% or 34%, respectively. Moreover, the dominant zone is SBT6.

Careful inspection of Figure 14b reveals the presence of neighboring homogeneous layers with

identical dominant SBT zones. For instance, Layers II and III have the same dominant SBTn (SBT3).

Although the WTMM method identifies a jump between Layers II and III, Layers II and III do not

necessarily belong to distinct SBT zones. This can occur if the Ic values on both sides (above and

below) of the jump do not differ considerably. Neighboring layers with identical dominant SBT zones

may be merged into a single homogeneous layer. For the NGES case depicted in Figure 14a, Layers II

and III may be merged into a single layer, and Layers VI and VII may be merged into another single

layer.

Procedure of the WTMM method for CPT stratigraphic profiling

To summarize, the procedure of the WTMM method for CPT stratigraphic profiling is as follows:

14
1. Convert the CPT data into the Ic(z) profile. To facilitate the FFT, augment the Ic(z) data into a

dyadic length of n = 2p. Compute the wavelet transform modulus spectrum |WIc(u,s)| using Eqs.

(10), (11), and (12) and find all WTMM ridges.

2. Based on the log(s) versus log(|WIc(u,s)|) relationship along each WTMM ridge, determine the (L,

M) parameters for each ridge. Note that when determining L, the maximum scale smax should have

an upper bound defined by the coarsest scale, sc = D/4 (where D is the total depth of the augmented

Ic data). When determining M, the initial scale s1 should be taken to be 0.05 m.

3. For each WTMM ridge, compute its pjump using Eq. (13). Identify all ridges with pjump > 0.5.

Follow these ridges to their finest-scale locations. These locations are the (tentative) layer

boundaries. The soil layers between these boundaries are the (tentative) homogeneous layers.

Determine the dominant SBTn for each layer.

4. (Optional) Merge neighboring soil layers with identical dominant SBTn results.

Figure 15c shows the final stratification results for the WTMM method applied to the CPT

sounding profile shown in Figure 12. The two horizontal dashed lines are the layer boundaries that are

removed in Step 4 above (Step 4 is optional). The dominant SBTn for each homogeneous soil layer is

annotated (e.g., SBT6 for the top layer) in Figure 15c. The descriptions of the SBTn indices are

provided in Table 1. For comparison, Figure 15a shows the soil profile obtained based on the boring

log, and Figure 15b shows the Ic profile based on the CPT. Figures 15d-h show the stratification results

obtained based on the fuzzy method proposed by Zhang and Tumay (1999), the Bayesian method

proposed by Wang et al. (2013), the clustering method proposed by Hegazy and Mayne (2002) and

Liao and Mayne (2007), and the T ratio method proposed by Wickremesinghe and Campanella (1991),

respectively. For the results of Zhang and Tumay (1999) (Figure 15d), the soil type membership (HPS,

HPC, or HPM) for each homogeneous layer is shown. For the results of Wang et al. (2013) (Figure

15e), the most probable SBTn is shown. For the clustering method and the T ratio method (Figure 15f

and Figures 15g and h), the dominant SBTn is shown.

15
 The clustering method adopted here is not exactly identical to those adopted in Hegazy and

Mayne (2002) and Liao and Mayne (2007). In the latter two studies, the CPTU pore pressure

coefficient Bq was considered. By contrast, only Qtn and Fr were considered in the clustering analysis

presented herein for the NGES example (i.e., Bq was not considered). Essentially, log(Qtn) and log(Fr)

were the only two features considered, and the number of clusters (nc) was chosen to be 40. A

clustering analysis with nc << 40 (e.g., nc = 30) is not able to identify all important homogeneous layers.

There are also two horizontal dashed lines indicated in the results of the clustering method (Figure 15f).

Their significance is the same as that of those in Figure 15c: the dominant SBT zones identified for the

two soil layers above and below the dashed line are identical. The T ratio method adopted here is

based on the T ratio on the Ic profile. A moving window of width = 1 or 2 m was applied to the Ic

profile, and the sample averages in the upper and lower halves of the window were computed. Other

window widths were not found to perform significantly better than 1 or 2 m. The normalized

difference between these two sample averages was computed. This normalized difference is called the

T ratio. The locations (spikes) at which the absolute values of the T ratio reach a local maximum are

potential layer boundaries (Figures 15g and h). For demonstration, we adopted a T ratio threshold of

10 to identify the layer boundaries. This threshold is consistent with that adopted in Wickremesinghe

and Campanella (1991).

The stratification results for all four methods (Figure 15c-h) are similar, but the following

notable differences exist in the results and procedures of each method:

1. The WTMM method, the clustering method, and the T ratio method with a narrow (1 m) window

appear to be more sensitive to the detailed CPT profile than are the other methods. All three

methods are able to identify the two layer boundaries near the depths of 5 m and 13 m (the two

horizontal dashed lines in Figures 15c and f). Among these methods, the clustering method is very

sensitive: it identifies numerous distinct clusters in the depth ranges of [0 m, 1 m] and [6.5 m, 8 m].

According to Liao and Mayne (2007), these two regions contain either outliers or very thin layers.
16
 However, the WTMM method still identifies two homogeneous soil layers for these two regions,

and the random oscillations present in those regions are treated as small-scale oscillations within

these homogeneous layers. The T ratio method with a 1 m window is also sensitive. It identifies 7

layer boundaries, including a spurious layer boundary at a depth of 3.7 m (Figure 15g). However,

it misses the potential layer boundary that is present near a depth of 8 m (there is a T value spike

near 8 m, but it does not exceed the specified threshold of 10). The T ratio method with a 2 m

window is less sensitive than that with a 1 m window. It identifies only 4 layer boundaries (Figure

15h).

2. Among all methods, the Bayesian method (Wang et al. 2013) and the clustering method operate on

the Qtn and Fr profiles when identifying layer boundaries. In particular, the Bayesian method

attempts to determine the ‘most probable’ number of layers and layer boundary positions, whereas

the clustering method attempts to divide the (Qtn, Fr) data points into ‘clusters’ in the two-

dimensional space of (Qtn, Fr). The fuzzy method (Zhang and Tumay 1999) operates on an abstract

normally distributed soil classification index. By contrast, the WTMM and T ratio methods operate

directly on the Ic profile when finding layer boundaries, which is preferable for two reasons. First,

the Ic profile is well known and concrete to engineers. Second, it is quite intuitive to identify layer

boundaries as jumps in the Ic profile.

3. The WTMM and T ratio methods are able to accurately pinpoint the locations of the layer

boundaries with minimum effort. In the WTMM method, the locations of the layer boundaries can

be found by tracing the WTMM ridges to the finest scale. In the T ratio method, the locations

correspond to the local maxima in the absolute value of the T ratio. The Bayesian method finds the

locations by solving a non-convex optimization problem, which is a non-trivial task when there are

many layers. In the clustering method, judgment is required when determining the exact locations

of the layer boundaries.

Lukang, Taiwan case

17
A CPT dataset collected in Lukang Township in Changhua County (Taiwan) is analyzed in this section.

This case was included in the calibration database (Table 2). The site is located on reclaimed land.

The backfill material used for the reclamation comprises approximately 4~5 m of hydraulic fill. The

groundwater table is within approximately 2 m of the current ground surface. Near the CPT sounding

location (approximately 3 m away), a borehole was drilled and samples were collected and tested for

their index properties. Figure 16 shows the CPT data together with the Ic profile (Figure 16c). The

data interval z is 0.05 m. Figure 16d shows the (Qtn, Fr) data points on the Robertson soil

classification chart. The SBTn is quite broad, ranging primarily from zone 3 to zone 6. There are also

a few data points located in zones 1 and 2. The wavelet transform modulus spectrum for the Ic(z)

profile and the traces of all WTMM ridges are shown in Figure 17b, and Figure 17c shows the L-M

plot for all ridges. There are numerous noise ridges, but only a few of them are shown in Figure 17b,

where they appear as dashed ridges. There are 18 ridges with pjump > 0.5. By following these ridges to

the finest scale, sf = z = 0.05 m, the locations of the layer boundaries can be identified. By virtue of

the FFT, the computation for the WTMM method is very fast. For this case, which contains many CPT

data points and many ridges, the entire WTMM analysis required only a few seconds, including the

time needed to identify all ridges and produce all figures and plots.

Figure 18b shows the Ic profile and the locations of the layer boundaries that correspond to the

18 identified ridges (horizontal dashed lines). Figure 18c shows the final stratification results obtained

using the WTMM method. The horizontal dashed lines in Figure 18c are the layer boundaries that

were removed in Step 4 above. Ultimately, 16 homogeneous soil layers were identified using the

WTMM method. The first 16 m consists of primarily SBT6 (sands), with a thin SBT1 layer (sensitive,

fine grained). Below 16 m are located sand-silt-clay interlayers. In this case, a nearby SPT borehole

was drilled (3 m away from the CPT sounding). Figure 18a shows the soil profile obtained based on

the boring log. This soil profile is generally consistent with the WTMM results.

18
 Because an excessive number of potential soil layers are identified, the application of the

Bayesian method (Wang et al. 2013) to the entire dataset is computationally challenging. To apply to

the entire case, the Bayesian method would require solving high-dimensional integrals (dimension > 30

for 15 potential layers) and also would require solving a high-dimensional non-convex optimization

problem (dimension > 14 for 15 potential layers). Therefore, it can be difficult to attain the global

optimum using this method. Instead, the entire stratification problem was divided into several

segments (e.g., 0~20 m and 20~40 m), and each segment was solved separately using the Bayesian

method. The combined stratification results are shown in Figure 18d. The results are similar to the

stratification results obtained using the WTMM method (Figure 18c). The minor differences between

them include the following:

1. The locations of the layer boundaries identified using the Bayesian method are not as accurate as

those identified using the WTMM method. In the WTMM method, the layer boundaries are

identified by tracing the ridges to the finest scale. Therefore, the boundaries identified using the

WTMM method correspond to the exact locations at which the Ic profile abruptly changes. By

contrast, in the Bayesian method, the boundary locations are determined by solving a potentially

high-dimensional optimization problem (where the layer thicknesses are the variables to be

optimized). It can be challenging to precisely determine the optimal layer thicknesses.

2. The WTMM method identified a thin SBT3 layer near a depth of 32 m, whereas the Bayesian

method did not.

3. The Bayesian method identified a thin SBT5 layer near a depth of 3 m and another thin SBT3 layer

near a depth of 18 m, whereas the WTMM method did not.

In the clustering method (Hegazy and Mayne 2002; Liao and Mayne 2007) adopted here, only

the (Qtn, Fr) parameters were considered. Numbers of clusters of nc = 40, 80, and 200 were considered,

and Figures 18e-g show the corresponding analysis results. Using nc = 80 (Figure 18f), the

homogeneous soil layer corresponding to SBT6 within the first 16 m could be readily identified.
19
Additionally, two deeper homogeneous soil layers corresponding to SBT6 and SBT6 could be

identified. However, many regions were identified as outliers/thin layers, especially in regions with

large spatial variability (e.g., 17~31 m). Decreasing nc to 40 did not yield improved stratification

results (Figure 18e) because most regions were reduced to a single cluster. Increasing nc to 200 also

did not improve the stratification results (Figure 18g) because an even greater number of regions were

identified as outliers/thin layers.

Figures 18h-j show the analysis results obtained using the T ratio method with 1 m, 2 m, and 5

m windows, respectively. As seen from the results for a 1 m window, this method was able to identify

very thin layers near depths of 4.5 m and 20~22 m (Figure 18h). However, there are many spikes in

the results. Some spikes are even evident in the depth range of 5~16 m, although this depth range may

contain only a single soil layer (see the Ic profile shown in Figure 18b). Thus, these spikes may be

spurious ones. Increasing the window size to 2 m or 5 m generally suppressed the occurrence of

spurious spikes (Figures 18i and j), but the ability to identify thin layers was greatly reduced.

A common feature of the clustering method and the T ratio method is the need to prescribe a

parameter that is similar to the ‘scale s’ in the wavelet transform. In the clustering method, nc controls

the scale. A large nc means that a large number of clusters is allowed. Hence, a cluster can be formed

from only a few data points in (Qtn, Fr) space. This enables the clustering method to explore the CPT

profile in greater detail (at smaller scales). Thus, a large nc corresponds to a small-scale analysis. In

the T ratio method, the window size controls the scale. A narrow window corresponds to a small-scale

analysis. The key feature of both the clustering method and the T ratio method is that the scale (nc or

window size) is prescribed and cannot vary. Therefore, a trade-off always exists. A small-scale

analysis (large nc or narrow window) is sensitive to fine details and can detect thin layers or outliers.

However, it may also yield spurious layer boundaries. Conversely, a large-scale analysis (small nc or

wide window) is insensitive to fine detail and can suppress the identification of spurious layer

boundaries. However, it cannot detect thin layers or outliers. Such ‘fixed-scale’ analyses may work
20
well in cases in which the layer boundaries are approximately equally spaced. Unfortunately, the

Lukang example is not of this type; in this case, the layer boundaries are non-uniformly spaced. The

Bayesian method is immune to the fixed-scale issue discussed above because it is based on a different

strategy. However, the interpretation of the Bayesian method is abstract to engineers, and it is

computationally intensive to perform.

The Lukang example clearly demonstrates the advantages of the WTMM method. First, the

WTMM method is multiscale in nature. It does not adopt a fixed scale but rather considers a large

range of scales. Therefore, the impact of the trade-off discussed above is minimized in the WTMM

method. The WTMM method can ‘see’ details by virtue of the wavelet transform spectrum at small

scales, whereas it can suppress spurious layer boundaries by virtue of the wavelet transform spectrum

at large scales. As seen in Figure 18c, the WTMM can detect thin layers, yet it does not produce

spurious layer boundaries. Moreover, there is no need to prescribe any fixed window size or number of

clusters. Finally, its computation can be performed very rapidly, and the interpretations of its analysis

processes and results are more concrete to engineers.

CONCLUSIONS

A stratigraphic profiling approach based on the soil behavior type index Ic determined from the CPT is

proposed. The basic concept is simple: layer boundaries can be identified as jumps in the Ic profile.

The locations of these jumps can be accurately pinpointed using the wavelet transform modulus

maxima (WTMM) method. The computation is very fast because of the use of the FFT. Numerical

studies show that the WTMM method is effective in discriminating between jumps and small-scale

oscillations and is also quite effective in detecting thin layers. These results indicate that direct

analysis of the Ic profile may be sufficient to solve the CPT stratification problem.

The clustering method proposed by Hegazy and Mayne (2002) and Liao and Mayne (2007) and

the T ratio method proposed by Wickremesinghe and Campanella (1991) both use a fixed number of
21
clusters or a fixed window size. Such methods are subject to a trade-off that is essentially equivalent to

the well-known bias-variance trade-off in statistics. Using the T ratio method as an illustration, a

narrow window is beneficial for detecting thin soil layers but will produce spurious layer boundaries,

whereas a wide window results in the opposite behavior: it will suppress spurious layer boundaries but

will not be able to detect thin layers. Therefore, such methods are not suitable for cases in which the

layer boundaries are non-uniformly spaced. Additionally, a common issue facing both methods is the

selection of the number of clusters or the window size, which is, in itself, a challenging task. The

Bayesian method proposed by Wang et al. (2013) performs well. It is immune to the trade-off

described above because it adopts a different analysis strategy. Instead, the primary limitation of the

Bayesian method is that it is computationally intensive because it requires solving high-dimensional

integrals and a high-dimensional non-convex optimization problem. Thus, stratification problems with

many potential layers can become computationally challenging. Note that in the presented

comparisons with the Bayesian method, we focus only on its most probable stratification result.

However, the Bayesian method offers more than merely the most probable result. It also provides a

means of progressively identifying statistically homogenous layers by gradually zooming in on local

differences with improved resolution (Cao and Wang 2013). This aspect is not emphasized in the

comparisons presented herein.

In our opinion, the WTMM method is an ideal method for stratigraphic profiling. The impact

of the aforementioned bias-variance trade-off is minimized in the WTMM method because it is

multiscale in its nature. It is sufficiently sensitive to detect thin soil layers and yet does not detect

many spurious layer boundaries. The wavelet transform characteristics of actual layer boundaries and

those of spurious boundaries due to small-scale oscillations are very different. The computation of this

method is very rapid because of the use of Fast Fourier Transform. Any case can be analyzed within a

few seconds. Finally, the WTMM method can produce stratification results in the form of graphs and

plots that more closely conform with engineers’ methods of visualization and intuition. Unlike
22
previous methods that operate in abstract domains or rely on abstract concepts, the WTMM method

operates directly on the Ic profile, which is well known and concrete to practicing engineers.

However, the WTMM method has the following potential limitations and restrictions. First, it

can be used only to analyze a single parameter, which means that either only one source of information

can be taken into account or multiple sources of information must be combined into a single parameter

such as Ic. Several previous methods, such as the Bayesian method, are capable of simultaneously

analyzing multiple sources of information to further reduce uncertainties. Second, the WTMM method

can be sensitive to spurious jumps in the CPT data. Such spurious jumps can be produced, for instance,

by changes in drainage conditions (e.g., pauses caused by the addition of more rods). This

phenomenon will significantly affect pore pressure measurements (u2) in clay layers, but this issue is

minor in sand layers. Third, in a scenario with highly inclined layers, the CPT transition zone between

two adjacent soil layers will become longer. It is anticipated that the layer boundary locations

identified based on such long transition zones will be less accurate. Finally, the layer boundaries

identified using the WTMM method are boundaries in “soil behavior”, e.g., boundaries between sand-

like versus clay-like soils. The method cannot detect boundaries between different geological units

with similar Ic values, e.g., normally consolidated versus over-consolidated clays, because it is based

on the Ic profile. One possible solution to address this shortcoming is to analyze the qt profile; however,

that approach is not pursued here.

ACKNOWLEDGMENTS

We would like to express our gratitude to the WaveLab development team at Stanford University. The

team members include (in alphabetical order) Dr. J. Buckheit, Dr. S.S. Chen, Dr. D. Donoho, Dr. X.

Huo, Dr. I. Johnstone, Dr. O. Levi-Tsabari, Dr. J. Scargle, and Dr. T. Yu. The MATLAB codes used in

this study were modified based on the MATLAB codes provided in WaveLab 850 (Buckheit and

23
Donoho 1995). We would like to thank Dr. Y. Wang for conducting the Bayesian analysis of the

second case (Lukang). We would also like to thank Dr. P.K. Robertson for his valuable comments.

REFERENCES

Ahmadi, M.M. and Robertson, P.K. 2005. Thin layer effects on the CPT qc measurement. Canadian

Geotechnical Journal, 42(5), 1302-1317.

Been, K. and Jefferies, M.G. 1992. Towards systematic CPT interpretation. Proceedings of the Wroth

Memorial Symposium, 121-134. London: Thomas Telford.

Briaud, J.L. 2000. The National Geotechnical Experimentation Sites at Texas A&M University: clay

and sand. A summary. In National Geotechnical Experimentation Sites (Geotechnical Special

Publication 93), ASCE, 26-51.

Buckheit, J. and Donoho, D.L. 1995. WaveLab and reproducible research. In Wavelets and Statistics,

A. Antoniadis, Ed., Springer, Berlin, New York.

Cao, Z. and Wang, Y. 2013. Bayesian approach for probabilistic site characterization using cone

penetration tests. ASCE Journal of Geotechnical and Geoenvironmental Engineering, 139(2), 267-

276.

Das, S.K. and Basudhar, P.K. 2009. Utilization of self-organizing map and fuzzy clustering for site

characterization using piezocone data. Computers and Geotechnics, 36, 241-248.

Facciorusso, J. and Uzielli, M. 2004. Stratigraphic profiling by cluster analysis and fuzzy soil

classification. Proceedings of the 2nd International Conference on Geotechnical Site

Characterization ISC-2, Porto.

Fenton, G.A. 1999. Random field modeling of CPT data. ASCE Journal of Geotechnical and

Geoenvironmental Engineering, 125(6), 486-498.

Hegazy, Y.A. and Mayne, P.W. 2002. Objective site characterization using clustering of piezocone

data. ASCE Journal of Geotechnical and Geoenvironmental Engineering, 128(12), 986-996.

24
Idriss, I. M. and Boulanger, R. W. 2008. Soil liquefaction during earthquakes. Monograph MNO-12,

Earthquake Engineering Research Institute, Oakland, CA.

Jaffard, S. 1991. Pointwise smoothness, two-microlocalization and wavelet coefficients. Publicacions

Matemàtiques, 35, 155-168.

Jung, B.-C., Gardoni, P., and Biscontin, G. 2008. Probabilistic soil identification based on cone

penetration tests. Géotechnique, 58(7), 591-603.

Ku, C.S., Juang, C.H., and Ou, C.Y. 2010. Reliability of CPT Ic as an index for mechanical behaviour

classification of soils. Geotechnique, 60(11), 861-875.

Liao, T. and P.W. Mayne 2007. Stratigraphic delineation by three-dimensional clustering of piezocone

data. Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards, 1(2),

102-119.

Mallat, S. 1999. A Wavelet Tour of Signal Processing, 2nd Ed. Academic Press, San Diego, CA.

Phoon, K.K., Quek, S.T., and An, P. 2003. Identification of statistically homogeneous soil layers using

modified Bartlett statistics. ASCE Journal of Geotechnical and Geoenvironmental Engineering,

129(7), 649-659.

Robertson, P.K. 1990. Soil classification using the cone penetration test. Canadian Geotechnical Jour-

nal, 27, 151-158.

Robertson, P.K. 2009. Interpretation of cone penetration tests – a unified approach. Canadian Geotech-

nical Journal, 46, 1337-1355.

Uzielli, M. 2004. Variability of stress-normalized CPT parameters and application to seismic

liquefaction initiation analysis. Ph.D. thesis. University of Florence, Italy.

Uzielli, M., Simonini, P., and Cola, S. 2008. Statistical identification of homogeneous soil units for

Venice lagoon soils. Proceedings of the 3rd International Conference on Site Characterization,

Taiwan.

25
Wang, Y., Huang, K., and Cao, Z. 2013. Probabilistic identification of underground soil stratification

using cone penetration tests. Canadian Geotechnical Journal, 50, 766-776.

Wickremesinghe, D.S. and Campanella, R.G. 1991. Statistical methods for soil layer boundary location

using the cone penetration test. Proceedings of the 6th International Conference on Applications of

Statistics and Probability in Civil Engineering (ICASP 6), Mexico City, 636-643.

Zhang, Z. and Tumay, M.T. 1999. Statistical to fuzzy approach toward CPT soil classification. ASCE

Journal of Geotechnical and Geoenvironmental Engineering, 125(3), 179-186.

APPENDIX log(s) VERSUS log(|WIc(u,s)|) RELATIONSHIP ALONG A WTMM RIDGE

Each jump in Ic(z) may produce a WTMM ridge. In the following, two questions will be answered: (a)

why is the gradient of the WTMM ridge that is caused by a jump approximately equal to 0.5? and (b)

which wavelet function is the best choice for our purposes?

Gradient of the WTMM ridge caused by a jump

A theorem proposed by Jaffard (1991) states that for a change point with “Lipschitz regularity” equal

to , the log(s) versus log[|WIc(u,s)|] relationship along the corresponding WTMM ridge has a gradient

of 0.5+ if an adequate wavelet function is chosen:

log  WIc (u,s)      0.5     log  s  (14)

where  is the Lipschitz regularity of the change point and  is the intercept. The meaning of the

Lipschitz regularity is rather technical (see Definition 6.1 in p. 164 of Mallat 1999). Here, we simply

state its significance. The Lipschitz regularity  measures the regularity of a change point. A larger 

indicates that the change point is more regular (smoother). For a jump (a discontinuity), its Lipschitz

regularity is  = 0. For a cusp (the integral of a jump), its Lipschitz regularity is  = 1. The three

jumps shown in Figure 5a have a Lipschitz regularity of  = 0. Figure 5c shows the log(s) versus

26
log[|WIc(u,s)|] relationships for the WTMM ridges of these jumps. It is clear that the gradient of these

relationships is indeed approximately 0.5+ = 0.5.

Choice of the wavelet function (z)

The above theorem proposed by Jaffard (1991) is based on the following premise: an adequate wavelet

function must be chosen. For a wavelet to “see” a change point, it must have sufficient “vanishing

moments”. A wavelet function (z) has m vanishing moments if the inner product (projection)

between (z) and a polynomial zk of degree k < m is always zero:



z k  (z)dz  0 k  0,1,..., m  1 (15)

It can be shown that the mth derivative of the Gaussian function (Eq. 7) has exactly m vanishing

moments. The basic principle is that a (z) with m vanishing moments can only “see” a change point

with a Lipschitz regularity  that is less than m. For instance, the jumps shown in Figure 5a have a

Lipschitz regularity of  = 0, and the chosen (z) is the Gaussian derivative (m = 1). Therefore, the

chosen (z) is adequate. It can “see” the jumps and can produce their WTMM ridges. If the change

point of interest is a cusp ( = 1) but the Gaussian derivative (m = 1) is still chosen as the (z), then

this choice is inadequate. It cannot “see” the cusp and cannot produce a WTMM ridge. This does not

mean that a (z) with many vanishing moments (e.g., m = 10) is always a good choice. If a jump in

the SBT index profile Ic(z) corresponds to a layer boundary but a cusp does not, then the best (z) is

indeed the Gaussian derivative (m = 1) because this (z) will see only the jumps and not the cusps.

The Mexican hat wavelet function (m = 2) is not the optimal choice because it will see both the jumps

and the cusps, and we do not wish to see the cusps because they are not real layer boundaries. In this

study, we adopt the Gaussian derivative (m = 1) as the wavelet function.

27
Table 1 Soil behavior types corresponding to various Ic values.

Zone*
Soil behavior type index, Ic Soil behavior type
(SBTn)
Ic < 1.31 7 Gravelly sand
1.31 < Ic < 2.05 6 Sands – clean sand to silty sand
2.05 < Ic < 2.60 5 Sand mixture – silty sand to sandy silt
2.60 < Ic < 2.95 4 Silt mixture – clayey silt to silty clay
2.95 < Ic < 3.60 3 Clays
Ic > 3.60 2 Organic soils – peat
* Zonation in the soil behavior type chart developed by Robertson (1990)

Table 2 Basic information concerning the 50 real CPT sounding records used for calibration.

# of Depth
Country Location Site Soil description soundings (m)
Cumhuriyet Dist. Yakin St. Interbedded silts and silty clays 2 9
Karaosman Dist. Kuyudibi Ave./Yaprak St. Silty clays underlain by sands 1 11.2
Istiklal Dist. Boluk St. Silts with silty clays 3 6.2~12.6
Kurahmediye Dist. Meydan St. Silty clays to silts underlain by sands 2 7.2~10.9
Turkey Tigcilar Dist. Kavaklar Ave. Interbedded silts and sands 1 12
(Adapazari) Yenig Dist. Hasircilar St. Silts with clays 1 11.4
Semerciler Dist. Park Ave. Clays to silty clays with silty sands 1 9.2
Yenig Dist. Rak St. Interbedded silty sands and silty clays 1 10.4
Tigilar Dist. Kavaklar Ave. Clays to silts underlain by silty sands 1 10.5
Orta Dist. Ankara Ave. Clayey silts with silty clays 1 10.3
- Interbedded clays and silts 3 50~56
Kaohsiung - Silts with silty clays 1 36.2
- Silty sands underlain by silty clays 1 41
Lukang - Sands underlain by silts to silty clays 4 30~40
Lunwei - Silty sands with silts to silty clays 1 28.5
- Silty sands underlain by clays 2 22.3~40
- Interbedded silty sands and silty clays 1 33.2
Tainan
- Sands 1 31
- Clays underlain by silty sands 1 28.6
Taiwan
- Silty clays underlain by silty sands 2 18~20
Yuanlin - Interbedded silty sands and silty clays 2 25~30
- Silty sands to silty clays 2 18~25
Provincial Council Parking Silty sands 1 8.9
Wufeng Yue-The Rd. Silty sands to silts 1 11.5
- Clays to silts with silty sands 1 28.5
- Clays underlain by sands 1 18
Nantou - Silty clays with silts 1 20
Ming-Ju Rd./Chung-Sin Rd. Silts with silty clays 1 16
- Silty sands to sands 4 20~25
- Clays underlain by sands 1 9.2
Norway Oslo Main Airport - Sands 1 37
- Interbedded silty sands and silty clays 2 31~43
- Sands underlain by clays 1 33.6

28
Figure 1 Jump versus small-scale oscillations.

29
Figure 2 Wavelet functions: (a) the Gaussian derivative (m = 1); (b) the Mexican hat (m = 2).

Figure 3 Effect of u and s on u,s(z): (a) the effect of the location u; (b) the effect of the scale s.

30
Figure 4 Wavelet transform modulus spectrum for a function containing one jump: (a) the Ic(z)
function; (b) its |WIc(u,s)| spectrum (the dark region represents points with large |WIc(u,s)|); (c) the
|WIc(u,s)| versus u relationships at certain values of s. The line connecting the local maxima is the
WTMM ridge.

Figure 5 Wavelet transform modulus spectrum for a function containing three jumps and noise: (a) the
Ic(z) function; (b) its |WIc(u,s)| spectrum (the traces of the ridges are also plotted); (c) the log(s) versus
log(|WIc(u,s)|) relationships along the ridges. The dashed lines represent the noise ridges.

31
Figure 6 (a) A large-scale wavelet; (b) a noise signal; (c) the product of the large-scale wavelet and the
noise signal; (d) a small-scale wavelet; (e) the same noise signal as (b); (f) the product of the small-
scale wavelet and the noise signal.

32
Figure 7 Significances of log(s1), log(smax), and a log(si)-log(|WIc(u,si)|) data pair.

Figure 8 WTMM ridges for a function containing three jumps and noise: (a) the Ic(z) function (the
gray dashed line is the function without noise); (b) the log(s) versus log(|WIc(u,s)|) relationships along
the ridges; (c) the L-M plot.

33
Figure 9 WTMM ridges for a function containing three smoothed jumps and noise: (a) the Ic(z)
function (the gray dashed line is the function without noise); (b) the log(s) versus log(|WIc(u,s)|)
relationships along the ridges; (c) the L-M plot.

Figure 10 (a) Ic profile for a real case in Yuanlin (Taiwan); (b) the WTMM ridge plot; (c) the L-M plot.

34
Figure 11 (a) L-M data points for 50 real Ic profiles (circles represent jumps, whereas crosses represent
small-scale oscillations); (b) relationship between pjump and X.

Figure 12 NGES data: (a) the qt profile; (b) the fs profile; (c) the Ic profile; (d) the Robertson chart.

35
Figure 13 NGES analysis results: (a) the Ic profile; (b) the wavelet transform modulus spectrum and
the traces of all ridges; (c) the L-M plot.

36
Figure 14 NGES stratification results: (a) the soil stratification pattern determined based on the 6
ridges; (b) the probability that each homogeneous layer belongs to SBT zone i.

37
 * In Figure 15d, HPS stands for highly probable sandy soil, HPC stands for highly probable clayey soil, and HPM stands for highly probable mixed soil. The
profile shown in Figure 15d is based on the most probable fuzzy type among HPS, HPC, and HPM.

Figure 15 Comparisons for the NGES stratification results: (a) the boring log; (b) the Ic profile; (c) the results obtained based on the WTMM
method; (d) the results of Zhang and Tumay (1999); (e) the results of Wang et al. (2013); (f) the results obtained based on clustering with nc
= 40; (g) the results obtained based on the T ratio method with a 1 m window; (h) the results obtained based on the T ratio method with a 2
m window.
38
Figure 16 CPT data for the Lukang case: (a) the qt profile; (b) the fs profile; (c) the Ic profile; (d) the
Robertson chart.

39
Figure 17 Analysis results for the Lukang case: (a) the Ic profile; (b) the wavelet transform modulus
spectrum and the traces of all ridges; (c) the L-M plot.

40
View publication stats

Figure 18 Comparisons of stratification results for the Lukang case: (a) the boring log; (b) the Ic profile; (c) the results obtained based on the WTMM
method; (d) the results obtained based on the Bayesian method; (e-g) the results obtained based on clustering with nc = 40, 80, and 200, respectively; (h-j)
the results obtained based on the T ratio method with window size = 1 m, 2 m, and 5 m, respectively.

41