STATISTICAL TO FUZZY APPROACH TOWARD CPT

SOIL CLASSIFICATION
By Zhongjie Zhang,1 P.E., Member, ASCE, and Mehmet T. Tumay,2 P.E., Fellow, ASCE

ABSTRACT: A soil engineering classification derived from the cone penetration test (CPT) involves the un-
certainty of correlation between soil composition and soil mechanical behavior. This uncertainty results in
overlaps of different soil types in currently used CPT soil classification systems. Accordingly, two statistical soil
classification criteria, region estimation and point estimation, are suggested to address this problem. Further, a
new fuzzy subset approach is introduced to develop a truly independent CPT soil engineering classification, and
to establish a transition between the new fuzzy approach and conventional soil classifications by utilizing local
Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

site- and project-specific calibrations. CPT results conducted at the National Geotechnical Experimentation Site
at Texas A&M University are used to demonstrate this new CPT soil engineering classification methodology.

INTRODUCTION from preliminary data reduction on CPT sounding data (Zhang

and Tumay 1996). The results published in the previous re-
A soil engineering classification from the cone penetration
search indicate empirically that U has a statistical correlation
test (CPT) relies on the physical response of soils to cone
with soil types defined by a soil type index (SI ) presented in
intrusion, which indicates the mechanical properties of geo-
Table 1, and each soil type has a pronormal frequency distri-
logic media tested, such as the strength, compressibility, and
bution of index U. Therefore, U is treated or approximated as
so on. Due to complicated environmental conditions, the cor-
a random variable, with assumed normal distributions to be
relation between soil composition and mechanical properties
determined by using the results from the 1996 paper. The sta-
will never be a simple one-to-one correspondence, as dis-
tistical analysis of the U data was accomplished by using
cussed previously by the writers (Zhang and Tumay 1996).
STATGRAPHICS software (Statgraphics 1989).
The CPT soil classification charts currently used do not pro-
The validation of the normal distribution assumption is
vide an accurate prediction of soil types defined by composi-
checked first by two distribution fitting tests. Table 2 shows
tional indices, but only serve as a guide to soil behavior type
the results from chi-square and Kolmogorov-Smirnov one-
(Douglas and Olsen 1981; Campanella and Robertson 1988).
sample tests. Dashes in the table signify that the size of the
Because of the way in which current CPT classification charts
corresponding sample is too small to perform the test. These
were developed, it is sometimes unavoidable to incorrectly
results indicate with confidence that all the soil samples follow
identify soil types. Fig. 1 displays a CPT sounding data with
the normal distributions. Fig. 2 exhibits all the normal density
its boring log profile, the corresponding Robertson and Cam-
functions of the seven soil types.
panella (1983a,b) soil classification, and Olsen and Mitchell
(1995) soil classifications at the National Geotechnical Exper-
REGION ESTIMATION (ZHANG 1994)
imentation Site (NGES) at Texas A&M University (Simon and
Briaud 1996). The somewhat questionable agreement between Region estimation is an approach to classify soils, similar
the carefully documented boring log and the soil types derived to a conventional soil classification. In this approach, the index
from current CPT soil classification charts shows disparities U-axis is divided into several regions. The soil type predicted
that would have to be resolved by additional CPT tests or for a given soil sample would depend on a U region within
costly soil borings. This inability to correctly identify the sed- which the U value of the sample falls. Each U region, how-
iments, specifically in transition zones, and the absence of a ever, will correspond to not just one soil type, but several ones,
quantifiable parameter in mixed/interbedded strata (i.e., clayey, each with different probabilities. In other words, it directly
silty, sandy; pockets/seams) are the obvious shortcomings. The addresses the probability of the misidentification of soil types
probability that soil types could be incorrectly identified by in situ.
current CPT classification methodologies requires a new soil The U regions are determined using the following assump-
classification approach to address this probability explicitly so tions:
that the user will not be misled. Nonconventional approaches,
such as statistical and fuzzy methodologies, are therefore in- 1. The sample density functions presented in Fig. 2 are the
vestigated here to meet this need and to help users to correctly real density functions of soil types.
interpret and use CPT soil classification. 2. All seven soil types identified in Table 1 are equally im-
portant when the determination of boundary values is
NORMAL DISTRIBUTION OF SOIL CLASSIFICATION considered and performed between adjacent soil types.
INDEX U
The nonconventional approaches presented here are based The boundary values Ui,i⫹1 for soil types i and i ⫹ 1, as
upon CPT soil engineering classification index U that results defined in Table 1, are derived from the condition
1 1 ⫺ Fi (Ui,i⫹1) = Fi⫹1(Ui,i⫹1) (1)
Engr., Louisiana Dept. of Transp. and Devel., Baton Rouge, LA
70804. where Fi and Fi⫹1 = cumulative distribution functions for soil
2
Assoc. Dean for Res. and Prof. of Civ. and Envir. Engrg., Coll. of
types i and i ⫹ 1 (i = 1, 2, . . . , 6). Eq. (1) gives
Engrg., Louisiana State Univ., Baton Rouge, LA 70803.
Note. Discussion open until August 1, 1999. To extend the closing ␮i⫹1␴i ⫹ ␮i ␴i⫹1
date one month, a written request must be filed with the ASCE Manager Ui,i⫹1 = (2)
of Journals. The manuscript for this paper was submitted for review and ␴i ⫹ ␴i⫹1
possible publication on May 13, 1996. This paper is part of the Journal
of Geotechnical and Geoenvironmental Engineering, Vol. 125, No. 3, where ␮i , ␴i , and ␮i⫹1, ␴i⫹1 = means and standard deviations
March, 1999. 䉷ASCE, ISSN 1090-0241/99/0003-0179 – 0186/$8.00 ⫹ for soil types i and i ⫹ 1, respectively (i = 1, 2, . . . , 6). The
$.50 per page. Paper No. 13267. boundary values calculated by (2) will divide the U axis into
JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999 / 179

J. Geotech. Geoenviron. Eng. 1999.125:179-186.

Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

FIG. 1. CPT Data with Its Boring Log Profile and CPT Classifications for National Geotechnical Experimentation Site at Texas A&M
University (NGES/Texas A&M)

TABLE 1. Soil Types from Unified Soil Classification System

(Standard 1987)
Soil type
Symbol index (SI ) Typical names
(1) (2) (3)
GP 1 Poorly graded gravels, gravel-sand mixtures, little
or no fines
SP 2 Poorly graded sands, gravelly sands, little or no
fines
SM 3 Silty sands, poorly graded sand-silt mixtures
SC 4 Clayey sands, poorly graded sand-clay mixtures
ML 5 Inorganic silts and very fine sands, rock flour,
silty or clayey fine sands with slight plasticity
CL 6 Inorganic clays of low to medium plasticity,
gravelly clays, sandy clays, silty clays, lean
clays
CH 7 Inorganic clays of high plasticity, fat clays

seven regions, R1, R2 , . . . , R7 , which constitute the base of

region estimation as shown in Table 3. Table 4 gives the prob- FIG. 2. Normal Density Functions for Seven Soil Types
abilities qi, j that an i type of soil falls in a j region of U. These
qi, j values are determined by Table 5 presents the probabilities pi, j with which a j region
contains an i type of soil. Probability pi, j is determined by
qi, j = Fi (Uj, j⫹1) ⫺ Fi (Uj⫺1, j) (3)
qi, j

冘
Here, i, j = 1, 2, . . . , 7; U0,1 = ⫺⬁; U7,8 = ⫹⬁; and pi, j = 7 (5)

冘
7
qi, j
i=1
qi, j = 1 (4)
j=1 Here, i, j = 1, 2, . . . , 7, too. Table 5, with the primary clas-
180 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999

J. Geotech. Geoenviron. Eng. 1999.125:179-186.

 TABLE 2. Results of Distribution Fitting Tests

Chi-Square Test K-S Test

Distribution
Soil Sample function Standard Degrees Significance Significance
type size (see Fig. 2) Mean ␮ deviation ␴ Estimate of freedom level Dplus Dmi level
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
GP 6 F1 2.97 0.824 — — — 0.116 0.174 0.993
SP 115 F2 2.86 0.649 4.07 7 7.7E-1 0.035 0.065 0.708
SM 69 F3 2.45 0.964 54.43 7 1.9E-9 0.167 0.173 0.032
SC 18 F4 1.89 0.544 — — — 0.115 0.126 0.939
ML 17 F5 0.85 0.472 — — — 0.141 0.097 0.889
CL 77 F6 0.22 0.754 6.36 6 3.8E-1 0.066 0.053 0.887
CH 85 F7 ⫺0.55 0.861 23.90 9 4.5E-3 0.071 0.101 0.355
Note: Dplus = maximum positive deviation of empirical cumulative distribution over hypothesized cumulative distribution function; Dmi = maximum
Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

negative deviation of empirical cumulative distribution over hypothesized cumulative distribution function.

TABLE 3. Division of Seven Regions over U -Axis sification frame in Table 3, consists of the region estimation
Between regions Boundary value, Ui, i ⫹1 1 ⫺ Fi (Ui, i ⫹1) approach to identify different soil types with certain probabil-
(1) (2) (3) ity.
Table 5 can be rearranged by grouping soil types, as shown
R1 and R2 2.91 0.471
R2 and R3 2.70 0.401
in Table 6. This new table indicates that sandy and gravelly
R3 and R4 2.01 0.354 soils generally fall in regions 1, 2, and 3; silty soils in regions
R4 and R5 1.33 0.152 4 and 5; and clayey soils in regions 6 and 7. If the silty soils
R5 and R6 0.61 0.306 are further divided and merged with the sandy and clayey soils
R6 and R7 ⫺0.14 0.316 separately, the boundary value between regions 4 and 5 can
reasonably be taken as the dividing point. These results are
summarized and depicted in Fig. 3.
TABLE 4. Probability with Which Different Soil Types Fall in
Each Region
Region estimation assumes that different points of a region
have exactly the same statistical property so that the whole
Probability q i, j over Regions region is treated in exactly the same way. This assumption
Region Region Region Region Region Region Region sometimes is far from reality, since points in a region can have
Distribution R1 R2 R3 R4 R5 R6 R7 a big difference in probabilities corresponding to different soil
(1) (2) (3) (4) (5) (6) (7) (8) types. From a theoretical point of view, the validation of the
F1 0.5285 0.0991 0.2274 0.1214 0.0214 0.0021 0.0000
region estimation therefore is somewhat limited. This problem
F2 0.4715 0.1274 0.2821 0.1098 0.0090 0.0002 0.0000 may be resolved by another approach called point estimation,
F3 0.3200 0.0811 0.2452 0.2323 0.0940 0.0238 0.0036 where every point in a U region is treated distinctively.
F4 0.0308 0.0380 0.2848 0.4944 0.1428 0.0090 0.0001
F5 0.0000 0.0000 0.0040 0.1478 0.5421 0.2873 0.0187 POINT ESTIMATION
F6 0.0002 0.0003 0.0060 0.0642 0.2353 0.3781 0.3159
F7 0.0000 0.0000 0.0009 0.0134 0.0754 0.2261 0.6841 Point estimation is an approach to predict soil types directly
[Subtotal] 1.3510 0.3459 1.0504 1.1833 1.1200 0.9266 1.0224
by probability, where each probable U value is evaluated in-
dividually. The fundamental question for this approach to an-
TABLE 5. Probability with Which Each Region Receives Dif-
ferent Soil Types
Probability pi, j over Regions
Soil Region Region Region Region Region Region Region
type R1 R2 R3 R4 R5 R6 R7
(1) (2) (3) (4) (5) (6) (7) (8)
GP 0.3912 0.2865 0.2165 0.1026 0.0191 0.0023 0.0000
SP 0.3490 0.3683 0.2686 0.0928 0.0080 0.0002 0.0000
SM 0.2369 0.2345 0.2334 0.1963 0.0839 0.0257 0.0035
SC 0.0229 0.1098 0.2711 0.4178 0.1275 0.0097 0.0001
ML 0.0000 0.0000 0.0038 0.1249 0.4840 0.3100 0.0183
CL 0.0000 0.0009 0.0057 0.0542 0.2100 0.4080 0.3090
CH 0.0000 0.0000 0.0009 0.0113 0.0675 0.2441 0.6691

TABLE 6. Simplified Results from Table 5

Probability pi, j over Regions

Soil Region Region Region Region Region Region Region
type R1 R2 R3 R4 R5 R6 R7
(1) (2) (3) (4) (5) (6) (7) (8)
GP, SP, SM 0.9771 0.8893 0.7185 0.3971 0.1110 0.0282 0.0035
SC, ML 0.0229 0.1098 0.2749 0.5427 0.6115 0.3197 0.0184
CL, CH 0.0000 0.0009 0.0066 0.0655 0.2775 0.6521 0.9781
GP, SP, SM,
SC 1.0000 0.9991 0.9896 0.0896 0.2358 0.0379 0.0036
ML, CL, CH 0.0000 0.0009 0.0104 0.1904 0.7615 0.9621 0.9964
FIG. 3. Region Estimation Chart for CPT Soil Classification

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999 / 181

J. Geotech. Geoenviron. Eng. 1999.125:179-186.

 冘冕
u
swer is this: Given a specific U value, what are the probabil- si

ities that the corresponding soil sample belongs to a different F(u, si) = f (x, y) dx (12)
y=1 ⫺⬁
soil type?
A probabilistic model of two dimensions has been derived or
for point estimation. One dimension is the continuous random

冕冘
u si
variable of soil index U. The other is a discrete random vari-
able representing soil types, as defined in Table 1. Both vari- F(u, si) = f (x, y) dx (13)
⫺⬁ y=1
ables are defined in the soil media considered. Discussing a
hypothetical case can enhance understanding of this probabi- and x and y are used in place of u and si to observe the rule
listic model. of integration and avoid possible confusion. The correspond-
Assume there is a box containing balls with different colors. ing marginal distribution function of soil classification index
The number of colors is N and the number of balls is infinite U is then
for each color. If a ball is taken out of the box, the probability,

冕冘
u M
Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

P(color), for a specific color selected is 1/N; i.e., P(color) =

1/N. Fm (u) = f (x, si) dx (14)
⫺⬁ si=1
In addition to the prior assumptions, suppose that a quantity
x can be measured for each ball, and this quantity is a contin- or

冘冕
uous random variable X with a specific distribution for each M u
color. Now, if a ball is taken out of the box revealing a cor-
responding x value, the probability of a specific color being Fm (u) = f (x, si) dx (15)
si=1 ⫺⬁
selected will be modified by the x value, since balls of different
colors will carry x values with different probabilities. Then, The marginal distribution function of SI has the form

冘冕 冘
what will be the modification of this x value on the probability si ⫹⬁ si
that a ball of a specific color is selected, or P(color兩X = x) = ?
Fm (si) = f (u, y) du = q( y) (16)
The solution to this problem, according to the multiplication y=1 ⫺⬁ y=1
rule in probability theory (Hoel et al. 1971), will be
with
P[(X = x) 艚 (color)]

冕
P(color兩X = x) = (6) ⫹⬁
P(X = x)
q( y) = f (u, y) du; y = 1, 2, . . . , si (17)
⫺⬁
Since X is a continuous random variable, P(X = x) = 0 exists.
So (6) needs to be changed to Under certain mathematical assumptions, (12) is actually
P[(x ⱕ X < x ⫹ ⌬x) 艚 (color)]
equivalent to (13), and (14) is equivalent to (15). Substituting
P(color兩X = x) = lim (7) (12) into (10) and rearranging produces
⌬ x→0 P(x ⱕ X < x ⫹ ⌬x)
P[(u ⱕ U < u ⫹ ⌬u) 艚 (SI = si)]
Now, imagine that the big box is the earth and the countless

冘 冋冕 冕 册
si u ⫹⌬u u
balls are the soil media. The continuous random variable X is
the soil classification index U and the different colors stand = f (x, y) dx ⫺ f (x, y) dx
y=1 ⫺⬁ ⫺⬁
for different soil types to be identified in a site investigation.
Then, the solution to statistically predict a soil type based on
冘 冋冕 冕 册
si⫺1 u ⫹⌬u u

a single U observation should be ⫺ f (x, y) dx ⫺ f (x, y) dx

y=1 ⫺⬁ ⫺⬁ (18)
P[(u ⱕ U < u ⫹ ⌬u) 艚 (soil type)]
P(soil type兩U = u) = lim (8) Also, substituting (15) into (11) yields
⌬u →0 P(u ⱕ U < u ⫹ ⌬u)

冘冕
M u ⫹⌬u
If a soil type is represented by a discrete numerical random P(u ⱕ U < u ⫹ ⌬u) = f (x, si) dx (19)
variable SI (soil type index from Table 1), (8) will be si=1 u

P[(u ⱕ U < u ⫹ ⌬u) 艚 (SI = si)] Now, the two-dimensional density function f (u, si) in the
P(SI = si兩U = u) = lim (9)
⌬u →0 P(u ⱕ U < u ⫹ ⌬u) previous equations has to be determined. From the definition
of F(u, si), we have
Here, SI will take values of 1, 2, . . . , M. M is the number of
soil types in a specific soil classification system. Uppercase SI F(u, si ⫺ 1) = P(U ⱕ u, SI ⱕ si ⫺ 1) (20)
or U signifies a variable; lowercase si or u is the value of that and
variable.
Suppose F(u, si) is a two-dimensional probability distribu- F(u, si) = P(U ⱕ u, SI ⱕ si) (21)
tion function of U and SI, and Fm (u) is its marginal distribution
function of U. By definition, the probabilities of events in (9) Subtracting (20) from (21) gives
can be rewritten in terms of F(u, si) and Fm (u). That is F(u, si) ⫺ F(u, si ⫺ 1) = P(U ⱕ u, SI ⱕ si)
P[(u ⱕ U < u ⫹ ⌬u) 艚 (SI = si)] = F(u ⫹ ⌬u, si) ⫺ P(U ⱕ u, SI ⱕ si ⫺ 1) = P[(U ⱕ u) 艚 (SI = si)] (22)
⫺ F(u, si) ⫺ F(u ⫹ ⌬u, si ⫺ 1) ⫹ F(u, si ⫺ 1) (10) since SI is a discrete random variable. Therefore
and P[(U ⱕ u) 艚 (SI = si)] = F(u, si) ⫺ F(u, si ⫺ 1)
P(u ⱕ U < u ⫹ ⌬u) = Fm (u ⫹ ⌬u) ⫺ Fm (u)
冘冕 冘冕 冕
si u si⫺1 u u
(11)
= f (x, y) dx ⫺ f (x, y) dx = f (x, si) dx
In probability theory, the two-dimensional distribution func- y=1 ⫺⬁ y=1 ⫺⬁ ⫺⬁

tion F(u, si) is defined as (23)

182 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999

J. Geotech. Geoenviron. Eng. 1999.125:179-186.

 On the other hand, the multiplication rule will give
P[(U ⱕ u) 艚 (SI = si)] = P(SI = si) ⭈ P[(U ⱕ u)兩(SI = si)] (24)
By definition

冕
u

P[(U ⱕ u)兩(SI = si)] = gsi (x) dx (25)

⫺⬁

where gsi (x) in (25) represents gsi (u), the conditional density
function of U for soil type si. Also, according to (16), P(SI =
si) in (24) is
P(SI = si) = Fm (si) ⫺ Fm (si ⫺ 1) = q(si) (26)
Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

It is the probability with which a soil type is met randomly in

situ, and is constant for soil type si.
Now, substitute (25) and (26) into (24), and then (24) into
(23); we get

冕 冕
u u
FIG. 4. CPT Point Estimation Chart for Soil Type CL
f (x, si) dx = q(si) gsi (x) dx (27)
⫺⬁ ⫺⬁
probabilities for soil type CL, assuming function q(si = 6)
Since u can be any real value, (27) exists only if takes values of 1/7, 1/3, and 1/2, respectively. These q(si)
f (u, si) = q(si)gsi (u) (28)
values are only for estimation purposes. With q(si) taking a
value of 1/7, the seven soil types considered in the present
This is the two-dimensional density function needed. study would generally have the same probability of being en-
Substitute (28) into (18) and (19). After some derivation, countered in situ. In the case that soil type CL may have higher
we have probabilities of being encountered than other soil types, the
results from q(si = 6) = 1/3 and 1/2 will provide a range for
P[(u ⱕ U < u ⫹ ⌬u) 艚 (SI = si)] estimation. Also, when q(si = 6) is equal to one of ␤ = 1/7,

冕 冕
u ⫹⌬u u ⫹⌬u 1/3, 1/2, the q(si) values for the other soil types are assumed
= f (x, si) dx = q(si) gsi (x) dx to be (1 ⫺ ␤)/6 for simplicity.
u u (29)
and FUZZY CLASSIFICATION

冘 冕 The results obtained so far have shown the statistical cor-

M u ⫹⌬u

P(u ⱕ U < u ⫹ ⌬u) = q(si) gsi (x) dx (30) relation between the soil engineering classification by CPT
si=1 u profile data and the Unified Soil Classification System
(USCS). In using CPT profile data to identify soil types, we
Therefore, the conditional probability determined by (9) will
actually classify geo-media according to their mechanical be-
be
havior characteristics. Like the USCS, this kind of classifica-

冕
u ⫹⌬u tion will also be useful for helping us study and understand
q(si) gsi (x) dx the engineering properties of in situ soils. Thus, there is a need
u for a truly independent CPT soil engineering classification to

冘 冕
P(SI = si兩U = u) = lim M u ⫹⌬u (31)
⌬u →0 be developed. In such a new system, the classification criteria
q(si) gsi (x) dx will not be borrowed or inherited from the USCS, but will
si=1 u
consider the accumulated engineering experience related to it.
Since ⌬u is an infinitesimal quantity, (31) can be rewritten as The fuzzy subset theory (Kaufmann 1975; Brown et al. 1985)
is an ideal tool to accomplish this task. The suggested fuzzy
q(si)gsi (u)⌬u approach will, contrary to a conventional soil engineering clas-

冘
P(SI = si兩U = u) = lim (32)
⌬u →0
M
sification, release the constraint of soil composition and put an
q(si)gsi (u)⌬u emphasis on the certainty: soil behavior (i.e., cone tip resis-
si=1
tance and local friction).
The limit in (32) is actually independent of ⌬u. Consequently, Three fuzzy soil types are defined in a CPT fuzzy soil clas-
we have sification. They are highly probable clayey soil (HPC), highly
probable mixed soil (HPM), and highly probable sandy soil
q(si)gsi (u) (HPS). Empirically, the three membership functions of HPS,

冘
P(SI = si兩U = u) = M (33)
HPM, and HPC are determined based upon the data (relative
q(si)gsi (u) frequency of U ) given in Fig. 5 (Zhang and Tumay 1996). In
si=1
this representation, soils have been reorganized into three
This is the basic formula to perform the point estimation. groups (group 1, group 2, and group 3) that are directly related
Eq. (33) requires a series of conditional density functions to HPS, HPM, and HPC. The density functions of the three
gsi (u), si = 1, 2, . . . , M, which are the normal distributions of soil groups are approximated as

冋 冉 冊册
the soil classification index U for different soil types, as shown 2

in Fig. 2. These soil types are represented by the variable soil 1 1 u ⫺ 2.6575
fs (u) = exp ⫺ (34)
type index, SI, defined in Table 1. Also defined previously, 0.834586兹2␲ 2 0.834586

冋 冉 冊册
q(si) is the probability with which an si soil type is found 2

randomly in situ. Fig. 4 shows one example of the conditional 1 1 u ⫺ 1.35

fm (u) = exp ⫺ (35)
probabilities determined by (33). They are the conditional 0.724307兹2␲ 2 0.724307

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999 / 183

J. Geotech. Geoenviron. Eng. 1999.125:179-186.

Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

FIG. 5. Empirical CPT Soil Classification for Three General Groups: Group 1 — SM and SP; Group 2 — ML and SC; Group 3 — CL and
CH

will reflect the overall perspective of soil properties. Soils in

these three groups have fundamentally different engineering
properties, but with no sharp boundaries between them. The
changes are gradual from one soil type to another.
Three soil types HPS, HPM, and HPC with their fuzzy
membership functions compose a basic CPT fuzzy soil clas-
sification with several unique characteristics and properties.
First, it is a real behavior type of soil classification, where not
only the classification index but also the classification criteria
are based upon the measurements of soil behavior. There is
no place in this classification for misidentification of soil types,
i.e., the uncertainty from randomness.
Second, soil types in this classification include the infor-
mation about their relations with the corresponding composi-
tional soil types. It clearly states that these soil types corre-
spond to their counterparts in USCS with high probabilities,
but not always 100%. The emphasis is put on the certainty of
CPT data: soil behavior. Since ‘‘highly probable’’ is an am-
FIG. 6. Tentative CPT Fuzzy Soil Classification Chart biguous term, the variation of these soil types is described by

冋 冉 冊册
2 the fuzzy membership functions.
1 1 u ⫹ 0.1775 Third, this fuzzy soil classification includes an empirical
fc (u) = exp ⫺ (36)
0.86332兹2␲ 2 0.86332 summary of current knowledge about soil behavior. HPS gen-
erally has the properties of high strength, high permeability,
After normalization and other empirical modifications on (34), and low compressibility, which correspond to a higher tip re-
(35), and (36), the three fuzzy membership functions of HPS, sistance qc and a lower friction ratio Rf , and therefore a larger
HPM, and HPC are defined as (Zhang 1994) U value. HPC is supposed to have a lower strength, a lower

再
permeability, and a higher compressibility that is usually con-
1.0 for u > 2.6575

冋 冉 冊册
␮s (u) = 2
(37)
sistent with a lower qc and a higher Rf , and therefore a lower
1 u ⫺ 2.6575 U value. The engineering properties of HPM lie in between
exp ⫺ for u ⱕ 2.6575
2 0.834586 HPS and HPC. Thus, its U will take a value between the val-

冋 冉 冊册
2 ues of HPS and HPC.
1 u ⫺ 1.35 Fourth, this classification provides a tool to separate the de-
␮m (u) = exp ⫺ ; ⫺⬁ < u < ⬁ (38)
2 0.724307 scription of soils in situ from the simplification of the soil state

再 冋 冉 冊册
2 in situ. Here, the simplification means a conventional process
1 u ⫹ 0.1775 to define the stratigraphy of a testing site. By using this new
exp ⫺ for u ⱖ 0.1775
␮c (u) = 2 0.86332 (39) type of soil classification, the in situ soil situation can be de-
1.0 for u < 0.1775 picted continuously, and the identification of the stratigraphy
of that site can be postponed. With a condensed format, the
as shown in Fig. 6. fuzzy soil classification describes the situation of soil in situ
As shown in Fig. 6, these empirically defined functions will as it is. No simplification is performed in advance. A condi-
have a maximum value of 1 and an ‘‘S’’ or a bell-shaped curve tional truncation on fuzzy soil types (simplification) can be
so that they can approximately reveal the law of quantity conducted later as need occurs and some criteria are available.
change to quality change concerning soil composition and The resulting crisp soil types (CST ) are expressed by a fuzzy
properties. The three fuzzy membership functions as a whole classification with a group of ␣ threshold values
184 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999

J. Geotech. Geoenviron. Eng. 1999.125:179-186.

 CSTj = [u兩␣j ⱕ ␮i (u)]; i = c, m, s (40)
Here, i and j = subscripts representing fuzzy and crisp soil
types, separately. These ␣ values can be determined for dif-
ferent engineering concerns. In this way, different users can
also accumulate and express their own engineering experience
mathematically. Here is an example. Suppose that CPT data
are used to check the liquefaction potential of a soil so that
some precautionary measures can be taken. This soil could be
defined, for instance, by the following criterion:
CSTliquefiable = [u兩␣liquefiable = 0.6 ⱕ ␮m (u)] (41)
where ␮m (u) is given by (38).
Last but not least, the CPT fuzzy soil classification will
Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

further serve a general communication purpose. An example

of this can be: Given U = 2.0, the corresponding soil layer
belongs to (‘‘looks like’’), depending upon the aspect that
needs to be emphasized,

• HPC with a degree of ␮c (2.0) = 0.04

• HPM with a degree of ␮m (2.0) = 0.67
• HPS with a degree of ␮s (2.0) = 0.73

This kind of description is used in daily life. For instance,

consider the following statement: A boy looks more like his
mother (with a degree 0.7, or 70%) than his father (with a
degree of 0.4, or 40%). People have the capability to under-
stand such a statement with little confusion, although it is a
fuzzy description. U = 2.0 is directly calculated from actual
cone measurements of tip resistance, qc and frictional resis-
tance, fs . Therefore, the engineering behavior of that soil layer
is certain.
FIG. 7. Probability Profile Related to NGES/Texas A&M CPT
Data
CPT SOIL CLASSIFICATION RESULTS
The results from using the new CPT soil engineering clas-
sification methodology, suggested in the present paper and pa-
pers published previously by the writers, will at least include
the following:

1. A soil classification index U profile from CPT sounding

data
2. Corresponding probability profiles to determine probable
USCS (compositional) soil types
3. A fuzzy soil type index profile (the values of the fuzzy
membership functions) to determine the soil types of the
fuzzy CPT soil classification

If required, some other criteria for different engineering con-

cerns can be used to produce additional profiles, and the whole
process can be routinely implemented and easily computer-
ized.
A typical application of the suggested methodology of using
a statistical to fuzzy approach in soil classification by CPT is
presented in Figs. 7 and 8. This example is based upon the
CPT data from the NGES at Texas A&M University, shown
in Fig. 1. Fig. 7 depicts a profile of related probabilities of
soil types encountered based upon the three-component region
estimation suggested in the present study. As the soil classi-
fication index U varies with the depth, the probabilities with
which the sediments are classified as a sandy soil (GP, SP,
SM), silty soil (SC, ML), or clayey soil (CL, CH) will also
vary. At some locations, sandy or clayey soils can be identified
clearly and reliably due to the very high probabilities. At other
locations, the U values can be statistically related to sandy,
silty, or clayey soils with different individual probabilities. In
other words, due to the fact that different soil types can some-
times share the same U values, the exact soil types cannot be FIG. 8. Fuzzy Soil Type Index Profile Related to NGES/Texas
verified without additional soil investigations (i.e., soil bor- A&M CPT Data

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999 / 185

J. Geotech. Geoenviron. Eng. 1999.125:179-186.

 ings). The significance of this information is that it alerts users (transition zones or mixed/interbedded strata, etc.) that might
about the existence of the problematic sediment layering (tran- otherwise be misinterpreted. Due to its flexible format, the
sition zones or mixed/interbedded strata, etc.) that otherwise fuzzy approach also adds value to fundamental classifications
would be neglected. by integration of other site- or project-specific behavior prop-
Fig. 8 presents a profile of the fuzzy soil type indices related erties necessary for versatile soil engineering predictions.
to the CPT data in Fig. 1. The indices describe and classify
the soil media in Fig. 1 from a different viewpoint. For in- ACKNOWLEDGMENTS
stance, the HPC index shows how the mechanical behaviors This research was supported by the Louisiana Transportation Research
of the soils in the profile correspond to the mechanical behav- Center (LTRC) under contract number 88-1 GT [state project number
ior of a typical clayey soil. It is a kind of comparison and 746-13-36, Louisiana HPR number 0010(12)] and partially by the Na-
classification without simplifying both within and between soil tional Science Foundation, MSS-9018249. LTRC is jointly sponsored by
types. The three variable indices provide a more comprehen- the Louisiana State University and the Louisiana Department of Trans-
portation and Development. The views reflected in this paper are not
sive perspective of soils encountered in Fig. 1. This is a fun-
necessarily those of the agencies and institutions with which the writers
Downloaded from ascelibrary.org by Ondokuz Mayis Universitesi on 11/11/14. Copyright ASCE. For personal use only; all rights reserved.

damental classification upon which other site- or project-spe- are affiliated.

cific classifications can be integrated by using (40). For
example, a local soil classification corresponding to the USCS APPENDIX. REFERENCES
can be developed based on local experience by calibrating the
␣ values in (40) with actual data accumulated from available Brown, C. B., Chameau, J. L., Palmer, R., and Yao, J. T. P. (1985). Proc.,
NSF Workshop on Civ. Engrg. Applications of Fuzzy Sets, Purdue Uni-
conventional soil boring logs and laboratory tests. versity, West Lafayette, Ind.
Campanella, R. G., and Robertson, P. K. (1988). ‘‘Current status of the
CONCLUSIONS piezocone test.’’ Proc., Penetration Testing 1988, ISOPT-1, De Ruiter,
ed., Orlando, Fla., 93 – 116.
The current paper presents a new methodology to classify Douglas, B. J., and Olsen, R. S. (1981). ‘‘Soil classification using the
soils by using CPT data. This methodology includes the sta- electric cone penetrometer.’’ Proc., Cone Penetration Testing and Ex-
tistical and fuzzy subset approaches, and intends to address perience, ASCE, Reston, Va., 209 – 227.
Hoel, P. G., Port, S. C., and Stone, C. J. (1971). Introduction to statistical
the problem of potentially misidentifying soil types inherent theory. Houghton Mifflin, Boston.
in using the existing CPT soil engineering classifications. Al- Kaufmann, A. (1975). Introduction to the theory of fuzzy subsets, I: Fun-
though there is still room for improvement by further accu- damental theoretical elements. Academic Press, New York.
mulation and calibration of in situ data at well-documented Olsen, R. S., and Mitchell, J. K. (1995). ‘‘CPT stress normalization and
sites, the fundamental structure of this methodology sets up a prediction of soil classification.’’ Proc., CPT ’95, Swedish Geotech-
logical connection between the site- and project-specific needs nical, Linkoping, Sweden, 2, 257 – 262.
Robertson, P. K., and Campanella, R. G. (1983a). ‘‘Interpretation of cone
of geotechnical engineers and the reality of the complex nature penetration tests. Part I: Sand.’’ Can. Geotech. J., 20(4), 718 – 733.
of geo-materials in the field. The fact that there is no simple Robertson, P. K., and Campanella, R. G. (1983b). ‘‘Interpretation of cone
one-to-one correlation between soil composition and mechan- penetration tests. Part II: Clay.’’ Can. Geotech. J., 20(4), 734 – 745.
ical properties dictates two alternatives: (1) Establish a statis- Simon, P. A., and Briaud, J.-L. (1996). ‘‘The national geotechnical ex-
tical correlation between CPT data and compositional soil perimentation sites at Texas A&M University: Clay and sand.’’ NGES-
types, such as the ones in USCS (a statistical approach); and TAMU-006 — December, Texas A&M University, College Station, Tex.
Standard test method for classification of soils for engineering purposes
(2) classify soils independently by the values of CPT data in — Annual book of ASTM standards; soil and rock, building stones, Vol.
order to evaluate the behavior of different soil types. A totally 04.08(Section 4); D2487-85. (1987). ASTM, West Conshohocken, Pa.
new soil classification would have the apparent disadvantage Statgraphics — Version 4.0: Statistical graphics system. (1989). Statistical
of disregarding the vast past experience related to a long Graphics Corp., STSC.
time – used methodology. However, the fuzzy approach can Zhang, Z. (1994). ‘‘Use of uncertainty methodology in identification and
build a bridge between the new and old by a local calibration classification of soils based upon CPT,’’ PhD dissertation, Louisiana
State University, Baton Rouge, La.
of ␣ threshold values obtained from actual field data (boring Zhang, Z., and Tumay, M. T. (1996). ‘‘The reliability of soil classification
logs and laboratory test results). The consequent statistical and derived from cone penetration test.’’ ASCE Spec. Conf. — Uncertainty
fuzzy classifications will make users aware of the existence in the geologic environment: From theory to practice, ASCE, Reston,
and interpretation of problematic stratification in geo-media Va., 383 – 408.

186 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / MARCH 1999

J. Geotech. Geoenviron. Eng. 1999.125:179-186.