Rock Mech Rock Eng

DOI 10.1007/s00603-016-0957-5

ORIGINAL PAPER

Evaluating Variability and Uncertainty of Geological Strength

Index at a Specific Site
Yu Wang1 Adeyemi Emman Aladejare1

Received: 7 October 2015 / Accepted: 14 March 2016

Springer-Verlag Wien 2016

Abstract Geological Strength Index (GSI) is an important parameter for estimating rock mass properties. GSI can
be estimated from quantitative GSI chart, as an alternative
to the direct observational method which requires vast
geological experience of rock. GSI chart was developed
from past observations and engineering experience, with
either empiricism or some theoretical simplifications. The
GSI chart thereby contains model uncertainty which arises
from its development. The presence of such model uncertainty affects the GSI estimated from GSI chart at a specific
site; it is, therefore, imperative to quantify and incorporate
the model uncertainty during GSI estimation from the GSI
chart. A major challenge for quantifying the GSI
chart model uncertainty is a lack of the original datasets
that have been used to develop the GSI chart, since the GSI
chart was developed from past experience without referring
to specific datasets. This paper intends to tackle this
problem by developing a Bayesian approach for quantifying the model uncertainty in GSI chart when using it to
estimate GSI at a specific site. The model uncertainty in the
GSI chart and the inherent spatial variability in GSI are
modeled explicitly in the Bayesian approach. The Bayesian
approach generates equivalent samples of GSI from the
integrated knowledge of GSI chart, prior knowledge and
observation data available from site investigation. Equations are derived for the Bayesian approach, and the

& Adeyemi Emman Aladejare

aaladejar2-c@my.cityu.edu.hk
Yu Wang
yuwang@cityu.edu.hk
1

Department of Architecture and Civil Engineering,

City University of Hong Kong, Tat Chee Avenue, Kowloon,
Hong Kong

proposed approach is illustrated using data from a drill and

blast tunnel project. The proposed approach effectively
tackles the problem of how to quantify the model uncertainty that arises from using GSI chart for characterization
of site-specific GSI in a transparent manner.
Keywords Geological strength index  Bayesian
approach  Model uncertainty  Prior knowledge 
Probabilistic characterization  Site characterization

1 Introduction
The Geological Strength Index (GSI) is a rock mass
characterization system which was introduced in the mid90s (Hoek et al. 1995). The GSI system concentrates on the
description of two factors, rock structure and block surface
conditions. It was developed for estimation of the rock
mass strength and deformation parameters. These parameters are to serve as inputs for numerical analysis or closedform solutions for designing tunnels, slopes, or foundations
in rocks (e.g., Marinos and Hoek 2000; Marinos et al.
2007; Carter 2010; Morelli 2015). As shown in Fig. 1, GSI
is the first point of entry into the HoekBrown criterion,
and unless this index is well understood and correctly
applied, the reliability of the estimated properties is open to
question (Hoek et al. 2013).
GSI can be estimated through observational or quantitative characterization (see Fig. 1). Observational characterization of GSI using descriptive input can be performed
if a rock mass is exposed in outcrops, surface and underground excavations, and if sufficient core is available even
in boreholes. GSI is then estimated by visualizing the
competence of the rock mass and assigning the appropriate
character directly on the standard GSI chart shown in

123

Y. Wang, A. E. Aladejare
Fig. 1 Data entry stream for
using the HoekBrown system
for estimating rock mass
properties for numerical
analysis (modified after Hoek
et al. 2013)

Geological observations

Quantitative input
based on established
rock mass indices

Descriptive input

Observational method

Quantitative method

OR

GSI Characterization

Laboratory testing of
intact rock samples

Hoek-Brown criterionengineering properties

of rock masses

Parameters required for

numerical analysis

Verification and
modification through
in situ monitoring and
back analysis

Fig. 2 (Hoek and Marinos 2000). The standard GSI chart in

Fig. 2 has two basic parameters, the condition of discontinuities and the blockiness of the rock mass. These geological parameters are easy for geologists to evaluate in the
field. However, there is limitation to observational characterization of GSI when engineers and engineering staff
are assigned to collect data. This is because engineers are
less comfortable assigning a GSI range based solely on
visual assessment. This leads to the development of
quantitative approach for characterizing GSI by different
researchers (e.g., Sonmez and Ulusay 1999; Cai et al. 2004;
Russo 2007; Hoek et al. 2013).
Quantitative method of GSI characterization was
developed to address the limitation of observational characterization. Engineers without strong geological backgrounds are less comfortable assigning a GSI range based
solely on visual assessment. As a consequence, some
quantified scales were included on the standard GSI

123

Numerical analysis of
overstress and remedial
measures

In situ stresses
Groundwater
Damage Factor
Excavation sequence
G

chart to aid its application in a more quantitative manner

(Carter and Marinos 2014). In the quantitative characterization of GSI, GSI is estimated as a function of the ratings
of scales on the GSI chart. These are scales derived from
rock mass rating (RMR), rock tunneling quality index (Q),
rock mass index (RMi) or other rock mass classifications
(e.g., Sonmez and Ulusay 1999; Cai et al. 2004; Russo
2007). For example, Hoek et al. (2013) proposed quantified
scales on the GSI chart, published by Hoek and Marinos
(2000), to include two scales: Joint Condition (JCond89) for
representing surface condition and Rock Quality Designation (RQD) for block interlocking. Based on the addition of
ratings of JCond89 and RQD (i.e., GSI = 1.5JCond89 ? RQD/2), the corresponding GSI is determined accordingly.
Note that the quantified GSI chart was developed from past
observations and engineering experience, and it inevitably
contains various uncertainties, such as observation scatterness and measurement error.

Evaluating Variability and Uncertainty of Geological Strength Index at a Specific Site

VERY POOR
Slickensided, highly weathered surfaces with
soft clay coatings or infillings

POOR
Slickensided, highly weathered surfaces with
compact coatings or fillings of angular

FAIR
Smooth, fresh moderately weathered and altered
f

GOOD
Rough, slightly weathered, iron-stained
f

From the lithology, structure and surface

conditions of the discontinuities, estimate the
average value of GSI. Do not try to be too
precise. Quoting a range from 33 to 37 is more
realistic than stating that GSI = 35. Note that
the table does not apply to structurally
controlled failures. Where weak planar
structural planes are present in an unfavourable
orientation with respect to the excavation face,
these will dominate the rock mass behaviour.
The shear strength of surfaces in rocks that are
prone to deterioration as a result of changes in
moisture content will be reduced if water is
present. When working with rocks in the fair to
very poor categories, a shift to the right may be
made for wet conditions. Water pressure is
dealt with by effective stress analysis.

VERY GOOD
Very rough, fresh unweathered surfaces

GEOLOGICAL STRENGTH INDEX FOR

JOINTED ROCKS (Hoek and Marinos, 2000)

SURFACE CONDITIONS

Fig. 2 Standard chart for GSI

estimates from geological
observations (modified after Hoek
and Marinos 2000)

DECREASING SURFACE

STRUCTURE
INTACT or MASSIVE Intact rock
specimens or massive in situ rock with
few widely spaced discontinuities

90

N/A

N/A

VERY BLOCKY Interlocked, partially

disturbed mass with multi-faceted angular
blocks formed by 4 or more joint sets
BLOCKY/DISTRIBUTED/SEAMY
folded with angular blocks formed by
many intersecting discontinuity sets.
Persistence of bedding planes or schistosity
DISINTEGRATED poorly interlocked,
heavily broken rock mass with mixture
of angular and rounded rock pieces
LAMINATED/SHEARED Lack
of blockiness due to close spacing of weak
schistocity or shear planes

The various uncertainties (e.g., observation scatterness

and measurement error) contained in the quantified GSI
chart results in model uncertainty, when using the GSI
chart for GSI estimation. This model uncertainty adds up to
the inherent spatial variability already present in GSI.
Inherent spatial variability arises from various geological
processes that rocks have undergone in their geological
histories. While the inherent spatial variability is independent of the state of knowledge about GSI and cannot be
reduced as the knowledge improves (Baecher and Christian
2003; Wang et al. 2016), the model uncertainty arises as a
result of imperfect representation of reality, which may be
due to simplifications and idealizations that have been
made, purposely or due to lack of knowledge (e.g., Ditlevsen 1981) and can be eliminated or reduced as knowledge
improves. The model uncertainty may be reflected in terms

DECREASING INTERLOCKING OF ROCK PIECES

80
BLOCKY Well interlocked undisturbed
rock mass consisting of cubical blocks
formed by three intersecting discontinuity
sets

70
60
50
40
30
20
10
N/A

N/A

of a bias or a random variation in GSI which is predicted by

the GSI chart. This can make the GSI estimated from the
chart, to be conservative or overestimated for a specific
site. This makes it important to properly quantify and
incorporate model uncertainty when using GSI chart to
estimate GSI for a specific site.
Model uncertainty has been studied in geotechnical
engineering (e.g., Ronold and Bjerager 1992; Phoon and
Kulhawy 1999; Wang and Cao 2013; Wang and Aladejare
2015). These past researches quantified model uncertainty
as the residuals between the observed and estimated values
from models, using the data which were originally used to
construct such models. This is possible when the data
which were originally used to construct such models are
available. However, GSI chart was developed based on
engineering experience and past observations, without

123

DECREASING SURFACE QUALITY

STRUCTURE

40

BLOCKY Well interlocked

undisturbed rock mass made up of
cubical blocks formed by three sets
of intersecting joints

DISINTEGRATED

poorly
interlocked, heavily broken rock
mass with mixture of angular and
rounded rock pieces

35

70

25

50
20

40

15

30

10

20
5

10
45

123

30

60

RQD/2

BLOCKY, DISTRIBUTED/SEAMY
folded with angular blocks formed
by many intersecting joint sets.
Persistence of bedding planes or
schistosity

80

DECREASING INTERLOCKING

VERY BLOCKY Interlocked,

partially disturbed rock mass,
multi-faceted
angular
blocks
formed by 4 or more joint sets

referring to specific datasets. Then, quantifying the model

uncertainty associated with the GSI chart under such condition becomes a challenge. This leads to the problem of
how to properly and rationally quantify and incorporate the
model uncertainty during characterization of GSI from GSI
chart, without the original datasets that have been used to
develop the GSI chart.
This paper addresses the problem of how to quantify the
model uncertainty during GSI characterization by integrating the GSI chart with Bayesian approach to explicitly
and properly consider the model uncertainty in chart-based
GSI characterization. The paper starts with an overview of
estimation of GSI from RQD and Joint Condition, followed
by probabilistic modeling of model uncertainty in GSI
chart and probabilistic modeling of inherent spatial variability in GSI. Next, the probability density function (PDF)
of GSI is derived based on the prior knowledge and sitespecific observation data under a Bayesian framework. The

POOR
Slickensided, highly weathered surfaces with
compact coatings or fillings of angular fragments

This chart applies to tunnels of about 10m span and

slopes < 20 m high. For larger caverns and slopes
consider reducing GSI to account for decreasing
block interlocking.

GOOD
Rough, slightly weathered, iron-stained surfaces

For intact or massive rock with GSI > 75, check for
brittle spalling potential. For sparsely jointed rock
with GSI > 75, failure will be controlled by
structurally defined blocks or wedges. The HoekBrown criterion should not be used for either of
these conditions.

VERY GOOD
Very rough, fresh unweathered surfaces

From the lithology, structure and observed

discontinuity surface conditions, estimate the
average GSI based on the descriptions in the row
and column headings. Alternatively, from logged
RQD values and Joint Condition rating (from
Bieniawski, 1989), estimate
based on the scales attached to the chart
axes.

FAIR
Smooth, fresh moderately weathered and altered surfaces

GEOLOGICAL STRENGTH INDEX (GSI) FOR

JOINTED BLOCKY ROCK MASSES

SURFACE CONDITIONS

Fig. 3 Quantification of GSI by

joint condition and RQD
(modified after Hoek et al.
2013)

VERY POOR
Slickensided, highly weathered surfaces with soft
clay coatings or infillings

Y. Wang, A. E. Aladejare

40

35

30

25

20

15

10

Bayesian framework integrates prior knowledge from

engineering judgement with site-specific observation data.
Then, Markov Chain Monte Carlo (MCMC) simulation is
used to transform the integrated knowledge into a large
number of GSI samples, and their statistics and probability
distribution are subsequently evaluated. The approach
developed in this paper is illustrated using data from a drill
and blast tunnel project.

2 Estimation of GSI From RQD and Joint

Condition
Hoek et al. (2013) proposed a quantitative GSI chart, which
makes use of joint condition and RQD data, and it is presented in Fig. 3. They proposed quantified scales on the
original GSI chart to include joint condition JCond89
rating, defined by Bieniawski (1989) as parameter for

Evaluating Variability and Uncertainty of Geological Strength Index at a Specific Site

horizontal scale, and RQD defined by Deere (1963) for

vertical scale. The relationship on the chart in Fig. 3 has
been expressed as (Hoek et al. 2013):
GSI 1:5JCond89 RQD=2

where RQD is a measure of the degree of jointing or

fracture in a rock mass, measured as a percentage of the
drill core in lengths of 10 cm or more. JCond89 is the joint
condition rating of the rock mass. The chart in Fig. 3 has a
minimum and maximum values of JCond89 as 0 and 45,
respectively, and minimum and maximum values of RQD/
2 as 0 and 40, respectively. This implies that the minimum
and maximum values of GSI obtainable from the chart are
0 and 85, respectively [see Eq. (1) and Fig. 3]. When
Fig. 3 is compared with the original GSI chart in Fig. 2, it
is worthwhile to note that the upper and lower rows of the
original GSI chart in Fig. 2 has been removed in Fig. 3,
and the quantitative GSI chart is only applicable to the 4
center rows of the original GSI chart in Fig. 2. In other
words, the methods developed in this paper are not applicable to intact rock specimens, massive in situ rock with
few widely spaced discontinuities, or very weak laminated/
sheared rock masses.

3 Probabilistic Modeling of Model Uncertainty

in GSI Chart
As the GSI chart was developed based on past observations
and engineering experience, it inevitably contains various
uncertainties, such as observation and measurement error.
This results in model uncertainty when using the GSI
chart for GSI estimation. The model uncertainty represents
the deviation between predicted and true values, which will
be random and varying from case to case (Ronold and
Bjerager 1992). For given site-specific JCond89 and RQD
values, such as JCond89 16:65 and RQD = 40,
GSI = 45 can be roughly expected from Eq. (1). However,
the presence of model uncertainty may cause the estimate
from the GSI chart to be significantly different from the
true value of GSI. Figure 4 illustrates the effect of model
uncertainty in the GSI chart produced by Hoek et al.
(2013). GSI0 represents the expected GSI value of 45 at
JCond89 16:65 and RQD = 40 using Eq. (1), while
points GSImin and GSImax in Fig. 4 represent the lowest and
highest possible true values of GSI, respectively. The
greater the magnitude of the model uncertainty, the more
the deviation between the expected GSI0 value and true
value of GSI which may be any value between GSImin and
GSImax. Since model uncertainty has such influence on the
estimation of GSI from GSI chart, it should be considered

explicitly when using the chart to estimate GSI. The model

uncertainty can be incorporated into the GSI estimation by
adding a model uncertainty term, e, to Eq. (1), to account
for deviation of the estimated GSI from the true value of
GSI as expressed:
GSI 1:5JCond89 RQD=2 e

where e is the model uncertainty of using the GSI chart to

predict values of GSI. e can be taken as a Gaussian random
variable with a zero mean and standard deviation re (e.g.,
Cetin et al. 2002; Moss et al. 2006).
The standard deviation re is an indicator of the magnitude of model uncertainty associated with the chart-based
estimation of GSI. Previous studies (e.g., Ronold and
Bjerager 1992; Phoon and Kulhawy 1999; Wang and Cao
2013; Wang and Aladejare 2015) quantified re of models
using the model data that were originally used to construct
such models. Because no specific dataset is referred during
development of the GSI chart, to quantify model uncertainty in this study, re is modeled as a random variable
varying between two extreme cases. Figure 5a, b illustrate
the modeling of re at the two extreme cases. Figure 5a
illustrates that as re decreases, the peak of the normal
distribution of GSI estimated from GSI chart increases. The
peak of GSI distribution increases with decreasing re until
only one point of GSI0 is obtained at re 0 (i.e.,
remin 0. At this point, there is no uncertainty associated
with the GSI chart in Fig. 3 or measurement data, and GSI
can be estimated with full certainty by the measured values
of JCond89 and RQD. On the other hand, Fig. 5b illustrates
that as re increases, the peak of the normal distribution of
GSI estimated from GSI chart decreases. The peak of GSI
distribution decreases from its maximum at GSI0 when
re 0 to a nearly uniform distribution when re reaches its
maximum value as shown in Fig. 5b. At this point, the GSI
chart provides no valuable information for GSI characterization. In other words, there is no preferred GSI value, and
the probability of different GSI values being the true value
of GSI is more or less the same (i.e., follows a uniform
distribution). re for such a uniform distribution is taken as
the upper bound value remax , which can be calculated as
(e.g., Ang and Tang 2007; Wang et al. 2013):
remax

q
GSImax  GSImin 2 =12

The lower bound of GSI is GSImin = 0 while the upper

bound of GSI is GSImax = 85 (see Sect. 2). The upper
bound value remax is then calculated from Eq. (3) as 24.5.
This range of re (i.e., re 2 0; 24:5) represents the
boundary of the standard deviation re of the model
uncertainty.

123

DECREASING SURFACE QUALITY

GSImax

BLOCKY Well interlocked

undisturbed rock mass made up of
cubical blocks formed by three sets
of intersecting joints

BLOCKY, DISTRIBUTED/SEAMY
folded with angular blocks formed
by many intersecting joint sets.
Persistence of bedding planes or
schistosity
DISINTEGRATED

poorly
interlocked, heavily broken rock
mass with mixture of angular and
rounded rock pieces

35

70
30

DECREASING INTERLOCKING

VERY BLOCKY Interlocked,

partially disturbed rock mass,
multi-faceted
angular
blocks
formed by 4 or more joint sets

40

80

60
25

RQD/2

STRUCTURE

POOR
Slickensided, highly weathered surfaces with
compact coatings or fillings of angular fragments

This chart applies to tunnels of about 10m span and

slopes < 20 m high. For larger caverns and slopes
consider reducing GSI to account for decreasing
block interlocking.

GOOD
Rough, slightly weathered, iron-stained surfaces

For intact or massive rock with GSI > 75, check for
brittle spalling potential. For sparsely jointed rock
with GSI > 75, failure will be controlled by
structurally defined blocks or wedges. The HoekBrown criterion should not be used for either of
these conditions.

VERY GOOD
Very rough, fresh unweathered surfaces

From the lithology, structure and observed

discontinuity surface conditions, estimate the
average GSI based on the descriptions in the row
and column headings. Alternatively, from logged
RQD values and Joint Condition rating (from
Bieniawski, 1989), estimate GSI = 1.5JCond89 +
RQD2 based on the scales attached to the chart
axes.

FAIR
Smooth, fresh moderately weathered and altered surfaces

GEOLOGICAL STRENGTH INDEX (GSI) FOR

JOINTED BLOCKY ROCK MASSES

SURFACE CONDITIONS

Fig. 4 Illustration of the effect

of model uncertainty during GSI
quantification from GSI chart

VERY POOR
Slickensided, highly weathered surfaces with soft
clay coatings or infillings

Y. Wang, A. E. Aladejare

50
20

40

15

30
10

20
5

10
45

40

35

30

25

20

15

GSImin
10
5

0
0

4 Probabilistic Modeling of Inherent Spatial

Variability in GSI
Geotechnical materials such as rocks are natural materials,
and their properties are affected by various geologic processes
(e.g., Baecher and Christian 2003; Mitchell and Soga 2005;
Wang and Aladejare 2015; Wang et al. 2016). These properties therefore vary spatially, and such inherent spatial
variability is independent of the state of knowledge regarding
geotechnical properties and cannot be reduced as the
knowledge improves (Baecher and Christian 2003). For
example, GSI has been modeled to follow a normal distribution in literature (e.g., Carter and Marinos 2014; Morelli
2015). Therefore, to model explicitly the inherent spatial
variability, GSI is taken to follow a normal distribution with a
mean lGSI and standard deviation rGSI, which is expressed as
(e.g., Ang and Tang 2007; Wang and Aladejare 2015):

123

GSI lGSI rGSI z

where rGSI represents the inherent spatial variability of the

GSI; while z is a standard Gaussian random variable. In the
next section, the PDF of GSI is derived to incorporate both
the inherent spatial variability and model uncertainty.

5 Probability Density Function for Site-Specific

GSI
The Geological Strength Index is modeled by a random
variable GSI which follows a normal distribution with a
mean lGSI and standard deviation rGSI. The distribution
model parameters lGSI and rGSI can be estimated using prior
knowledge of GSI and site-specific observation data. Prior
knowledge include, but are not limited to, maps and surveys,

Evaluating Variability and Uncertainty of Geological Strength Index at a Specific Site

PlGSI ; rGSI jData is the integrated knowledge on model

parameters lGSI and rGSI based on prior knowledge and
site-specific JCond89 and RQD data. In the next section,
Bayesian framework is developed to calculate
PlGSI ; rGSI jData.

PDF

as decreases,
peak increases

6 Bayesian Framework for Determining Model

Parameters

GSImin

GSImax

GSI0

(a)

The updated knowledge on model parameters lGSI and

rGSI is denoted as PlGSI ; rGSI jData, which is expressed
as (e.g., Ang and Tang 2007; Wang and Aladejare 2015):
PlGSI ; rGSI jData KPDatajlGSI ; rGSI PlGSI ; rGSI
7

PDF

as increases,
peak decreases

GSImin

GSI0

GSImax

(b)

n 1:5J 0:5R GSI  e

Fig. 5 Illustration of modeling of re for two extreme cases. a Extreme

case I: re 0. b Extreme case II: remax

local experience, engineering judgment, visual observations,

and published reports and studies (Wang et al. 2010; Cao
et al. 2016). The prior knowledge used in this study is the
engineering judgement on the possible ranges of lGSI and
rGSI of GSI. For a given set of prior knowledge and sitespecific observation data, there are many possible combinations of lGSI and rGSI. Each combination of lGSI and rGSI
has its corresponding occurrence probability given the data,
PlGSI ; rGSI jData. Using the Theorem of Total Probability
(e.g., Ang and Tang 2007), the PDF of GSI for a given set of
prior knowledge and site-specific data is expressed as:
ZZ
PGSI jData
PGSI jlGSI ; rGSI
 PlGSI ; rGSI jDatadlGSI drGSI

where K = ($P(Data|lGSI, rGSI)P(lGSI, rGSI))-1 is a normalizing constant that does not depend on lGSI and rGSI.
Data is the site-specific observation points of JCond89 and
RQD. For convenience purpose, JCond89 and RQD will be
represented by J and R, respectively, in this section. P(Data|lGSI, rGSI) is the likelihood function, which
reflects the model fit with the Data.
To derive the likelihood function, Eq. (2) is transformed
as:

in which Data denotes site-specific JCond89 and RQD data.

Explanation on Data is given in full details in Sect. 6.
PGSI jlGSI ; rGSI is the PDF of normally distributed GSI
for a given set of mean lGSI and standard deviation rGSI,
and it is expressed as:


1
1 GSI  lGSI 2
PGSIjlGSI ; rGSI p
exp 

6
2
rGSI
2prGSI

where n = (1.5J ? 0.5R) represents the estimated GSI

from the GSI chart. As previously defined, e is the model
uncertainty of using the GSI chart to estimate site-specific
GSI. Combining Eqs. (4, 8) leads to:
n 1:5J 0:5R lGSI rGSI z  e

If the inherent spatial variability is assumed to be

independent of the model uncertainty (i.e., z is independent
of e), n = (1.5J ? 0.5R) is a Gaussian random variable
with a mean of lGSI and standard deviation of
q
rGSI 2 re 2 . The likelihood function, P(Data|lGSI,
rGSI),
of
observing
Data
(i.e.,
ni 1:5J 0:5Ri ; i 1; 2; 3. . .; ns ) then follows a
Gaussian distribution with a mean of lGSI and standard
q
deviation of
rGSI 2 re 2 . Note that P(Data|lGSI,
rGSI) is a function of re , and it can be obtained by
marginalizing over the range of re to properly consider
the model uncertainty e (e.g., Sivia and Skilling 2006):
Z
PDatajlGSI ; rGSI PDatajlGSI ; rGSI ; re Pre dre
10
where Pre is the PDF of re . When there is no prevailing
knowledge on the PDF of re ; Pre is taken as a uniform

123

Y. Wang, A. E. Aladejare

distribution varying between two extreme cases discussed

in Sect. 3. Therefore, Pre is calculated as
Pre 1=remax  remin . Note that the range of re has
previously been given in Sect. 3. PDatajlGSI ; rGSI ; re in
Eq. (10) is expressed as:
ns
Y

1
pq
i1
2p rGSI 2 re 2
9
8
>
>
=
< 1 1:5J 0:5R  l
GSI 2
i
 exp  q 
>
>
;
: 2
rGSI 2 re 2

PData jlGSI ; rGSI ; re

11

The likelihood function P(Data|lGSI, rGSI) in Eq. (7) is

obtained by using Eqs. (10, 11).
P(lGSI, rGSI) is the prior distribution of lGSI and rGSI,
which reflects the engineering experience prior to obtaining
information at the site. The prior distribution can be simply
taken as a joint uniform distribution of lGSI and rGSI with
respective minimum values of lGSImin and rGSImin and
respective maximum values of lGSImax and rGSImax, and it is
expressed as (Cao et al. 2016):

PlGSI ; rGSI

8
<
:

lGSImax

1
1

 lGSImin rGSImax  rGSImin
0

PGSI jData K

The steps involved in using the proposed approach for sitespecific GSI characterization are listed below:

12

others

PGSIj lGSI ; rGSI

 PDatajlGSI ; rGSI PlGSI ; rGSI dlGSI drGSI

13
Equation (13) is a product of the normalizing constant
RR
K and the integral term defined as I
PGSI jlGSI ;
rGSI PDatajlGSI ; rGSI PlGSI ; rGSI dlGSI drGSI and it

123

7 Implementation Procedure

for lGSI 2 lGSImin ; lGSImax  and rGSI 2 rGSImin ; rGSImax 

The joint uniform distribution only requires the possible

ranges of the model parameters to completely define it. These
likely ranges require relatively limited prior knowledge, i.e.,
reasonable ranges of rock property of interest. From the Hoek
et al. (2013) chart in Fig. 3, the lower bound value of GSI is 0,
hence lGSImin 0 while the upper bound of GSI is 85, thus
lGSImax 85. Considering the non-negative physical meaning
of standard deviation of GSI, the lower bound of rGSI is set at
rGSImin 0. The upper bound of GSI, rGSImax is calculated from
the range of GSI values using six-sigma rule as rGSImax
GSImax  GSImin =6 85  0=6 14:2.
Using the updated knowledge of the model parameters
given by Eq. (7), the PDF of geological strength index, GSI
[i.e., Eq. (5)] is rewritten as:
ZZ

gives the PDF of GSI for a given set of prior knowledge

and site-specific observation data.
Since the derived PDF in Eq. (13) may be too complicated and difficult to express analytically and explicitly. Markov chain Monte Carlo (MCMC) simulation is
used to generate a large number of Bayesian equivalent
samples of GSI from the PDF of the random variable GSI
in Eq. (13) (e.g., Wang and Cao 2013; Wang and
Aladejare 2015). Conventional statistical analysis of the
samples is performed to estimate the statistics (e.g., mean
and standard deviation) of GSI, and the PDF and Cumulative distribution function (CDF) of GSI estimated
through a histogram and cumulative frequency diagram,
respectively.

1.
2.

3.

4.

5.

Obtain a set of JCond89 and RQD data from a site.

Obtain the prior knowledge on the ranges of lGSI, rGSI
and re with respective minimum values of lGSImin ,
rGSImin and remin , and respective maximum values of
lGSImax , rGSImax and remax . In this study, the ranges have
been given in Sects. 3 and 6.
Choose intervals for ranges of lGSI, rGSI and re , to
calculate PData jlGSI ; rGSI and P(lGSI, rGSI) according to Eqs. (11, 12) accordingly.
Choose an initial state for the Markov Chain of GSI, such
as the mean of the prior knowledge of lGSI [e.g.,
(lGSImin ? lGSImax )/2], and use the MCMC simulation to
generate a large number of the GSI samples from Eq. (13).
Estimate lGSI and rGSI of the equivalent GSI samples
generated through MCMC simulation, and estimate the
PDF and CDF of GSI from the histogram and
cumulative frequency diagram of the equivalent samples, respectively.

Although the proposed Bayesian approach seems

mathematically complicated, it is rather straightforward
and can be programmed and implemented as a user function or toolbox in computer software such as MATLAB.

Evaluating Variability and Uncertainty of Geological Strength Index at a Specific Site

8 Illustrative Example

Cumulative distribution function (CDF) of

GSI

This section illustrates the proposed Bayesian approach

using a set of real-life field data from a drill and blast
tunnel project. These data are obtained from Hoek et al.
(2013). The data includes 72 mapped GSI obtained through
the observational method and corresponding 72 calculated
GSI obtained using Eq. (1) through the quantitative
method. Figure 6 shows the distribution of these GSI data
points obtained from observational and quantitative methods by open circles and shaded triangles, respectively. Note
that the quantitative chart proposed by Hoek et al. (2013)
represents a recent attempt at quantification of the original
Hoeks chart using parameters for rock mass fracturing and
joint conditions that are well established in literature. This
quantification implies a modest correction to the original
charts structure, making the iso-GSI lines, which were
originally hand drawn, parallel and equally spaced. This
geometric correction may introduce some divergences
between the GSI values estimated with the two charts
(observational and quantitative, see Fig. 6), and may supplement the inherent model uncertainty which, at the end,
is contained in both charts.
While some of the calculated GSI values from the
quantitative method will be used in the proposed approach,
the 72 mapped GSI data points obtained through the
observational method will be used to compare and validate
the results from the proposed approach. Note that the
Bayesian equivalent sample approach only needs a relatively limited number of observation data (Wang and Cao
2013). Therefore, only some of the 72 observation data will
be used as input in this illustrative example. This feature is
appealing and beneficial, particularly when the project size
is relatively small and the observation data (i.e., JCond89
and RQD) are sparse from the project. The calculated GSI

1
Direct Calculation with
Eq. (1)

0.8

Observational Method

0.6

0.4

0.2

10

20

30

40

50

60

70

Geological Strength Index (GSI)

Fig. 6 Probability distribution of GSI samples used in the illustrative

example

values obtained using GSI = 1.5JCond89 ? RQD/2

become the input data in this study because the input data
in this study has been modeled as 1.5JCond89 ? RQD/2
[see Sect. 6 and Eq. (11)].
For small to medium size rock engineering projects,
often only a few rock cores are available for measuring
JCond89 and RQD for a site. Thus, the input data (i.e.,
JCond89 and RQD) are often limited for site-specific estimation of GSI through the quantitative method. To simulate limited input data for estimating GSI from JCond89 and
RQD in real-life engineering practice, three data sets
having limited number of (GSI = 1.5JCond89 ? RQD/2)
data points are randomly selected from the 72 calculated
GSI values to perform probabilistic characterization of
GSI. The first set has five data points of
(GSI = 1.5JCond89 ? RQD/2) and is named Data set I, the
second set has 10 data points of (GSI = 1.5JCond89 ?
RQD/2) and is named Data set II while the third set has 30
data points of (GSI = 1.5JCond89 ? RQD/2) and is named
Data set III. The data points of (GSI = 1.5JCond89 ?
RQD/2) in each of the data sets are listed in Table 1. The
data points of (GSI = 1.5JCond89 ? RQD/2) in each data
set are used in the proposed approach to generate 30,000
equivalent GSI samples. The results of GSI obtained from
the proposed approach are discussed and compared with
the mapped GSI from the observational method and calculated GSI from Eq. (1), as described in the following
subsections.
8.1 Statistics of GSI
Figure 7a shows a scatter plot for the 30,000 Bayesian
equivalent samples of GSI generated using Data set I as the
input data, Fig. 7b shows a scatter plot for the 30,000
Bayesian equivalent GSI samples generated from Data set
II, while Fig. 7c shows a scatter plot for the 30,000
Bayesian equivalent GSI samples generated from Data set
III. The summary of the statistics of the 30,000 Bayesian
equivalent GSI samples in Fig. 7ac is presented in the
third column of Table 2ac, respectively. Table 2ac also
include statistics of the mapped GSI using observational
method and calculated GSI using Hoek et al. (2013) GSI
chart/equation in second and fourth columns, respectively.
The mean and standard deviation of the mapped GSI
using observational method are calculated as 44.36 and
10.35, respectively. In Table 2a, the mean and standard
deviation of Bayesian equivalent GSI are 44.13 and 11.42,
respectively, while those of the calculated GSI are 44.07
and 13.02, respectively. In Table 2b, the mean and standard deviation of Bayesian equivalent GSI are 44.66 and
10.60, respectively, while those of the calculated GSI are
44.85 and 15.80, respectively. In Table 2c, the mean and
standard deviation of Bayesian equivalent GSI are 44.61

123

Y. Wang, A. E. Aladejare

No.

Data set I

Data set II

Data set III

53.93

40.16

45.25

29.51

51.80

45.57

32.46

34.43

62.46

59.34

58.36

29.51

5
6

45.08

61.31

63.93

62.46

54.43

85

Geological Strength Index (GSI)

Table 1 Data points of (GSI = 1.5JCond89 ? RQD/2) in three data

sets

76.5
68
59.5
51
42.5
34
25.5
17
8.5

7
8

27.38
60.98

26.07
40.00

28.52

59.34

Sample number

10

23.11

27.54

(a)

0
0

49.84

85

12

28.52

76.5

13

23.11

14

57.87

15

34.43

16

62.95

17

23.61

18

61.97

19

37.70

20

57.21

21

30.00

22
23

27.38
32.46

24

65.57
63.93

26

28.03

27

65.41

28

61.31

29

58.36

30

24.59

and 10.00, respectively, while those of the calculated GSI

are 44.95 and 15.81, respectively. In Table 2ac, the mean
and standard deviation of Bayesian equivalent GSI are
close to the mean and standard deviation of the mapped
GSI. The small relative difference between the mean and
standard deviation of the mapped GSI from observational
method and that of the Bayesian equivalent GSI suggests
that the proposed Bayesian approach properly characterizes
the GSI at this site. Also, there is reduction in the standard
deviation values, when the same data used for direct GSI
calculation are used in the Bayesian characterization of
GSI, even at input data points of 30 which is considered the
minimum number of data points to give a reasonable
statistics of standard deviation (Walpole et al. 1998). This
is because the calculated GSI values contain both inherent
spatial variability and model uncertainty which propagated

123

10000

15000

20000

25000

30000

68
59.5
51
42.5
34
25.5
17
8.5
0
0

5000

10000

15000

20000

25000

30000

25000

30000

Sample number

(b)
85

Geological Strength Index (GSI)

25

Geological Strength Index (GSI)

11

5000

76.5
68
59.5
51
42.5
34
25.5
17
8.5
0
0

5000

10000

15000

20000

Sample number

(c)
Fig. 7 Scatter plots of the Bayesian equivalent GSI samples. a Data
set I with 5 data points of (GSI = 1.5JCond89 ? RQD/2) as input
data. b Data set II with 10 data points of (GSI = 1.5JCond89 ? RQD/
2) as input data. c Data set III with 30 data points of
(GSI = 1.5JCond89 ? RQD/2) as input data

through Eq. (1), when the GSI values are estimated directly
from Eq. (1). The corresponding standard deviation values
therefore contain both inherent spatial variability and
model uncertainty, and they are obviously larger than the

Evaluating Variability and Uncertainty of Geological Strength Index at a Specific Site

Table 2 Summary of statistics of GSI
Statistics

Mapped GSI from

observational method

Bayesian equivalent GSI

Calculated GSI with Eq. (1)

(a) Data set I with 5 data points of (GSI = 1.5JCond89 ? RQD/2) as input data
Number of samples

72

30000 (Based on 5 input data points

and prior knowledge)

Mean
Standard deviation

44.36
10.35

44.13
11.42

5
44.07
13.02

(b) Data set II with 10 data points of (GSI = 1.5JCond89 ? RQD/2) as input data
Number of samples

72

30000 (Based on 10 input data points

and prior knowledge)

10

Mean

44.36

44.66

44.85

Standard deviation

10.35

10.60

15.80

(c) Data set III with 30 data points of (GSI = 1.5JCond89 ? RQD/2) as input data
Number of samples

72

30

44.36

30000 (Based on 30 input data points

and prior knowledge)
44.61

Mean
Standard deviation

10.35

10.00

15.81

inherent spatial variability itself. In contrast, the model

uncertainty is formulated and considered explicitly in the
Bayesian approach, and its corresponding standard deviation values mainly reflect the inherent spatial variability in
GSI (Wang et al. 2016). It has been shown that incorporating the model uncertainty in the formulation of the
Bayesian approach significantly improves the estimation of
GSI from quantified chart. This can be observed when
comparing the results obtained from the Bayesian approach
and direct quantification through quantified GSI chart, with
the mapped GSI which is from another approach (i.e.,
observational method using the original GSI chart). Note
that the Bayesian approach was formulated using the
quantified chart, and the comparison of the results from the
Bayesian approach with those from observational method
(i.e., original GSI chart) was to validate the results from the
proposed approach using other method apart from the
quantified GSI chart.
Furthermore, as the input data increase, the standard
deviation of the Bayesian equivalent GSI reduces and
almost equal to the standard deviation of the mapped GSI.
This indicates that the proposed approach improves the
estimates of GSI from the Hoek et al. (2013) chart by
incorporating the model uncertainty during the chart-based
GSI characterization. The Bayesian approach proposed in
this study is able to explicitly incorporate the model
uncertainty in the GSI chart during site-specific characterization of GSI. This will be very useful in rock engineering, where GSI chart is frequently used for site-specific
estimation of GSI values for input during rock design and
construction purposes.

44.95

8.2 Probability Distribution of GSI

Figure 8ac show the probability distributions of the
30,000 Bayesian equivalent GSI samples by solid line,
when Data sets I, II and III are used in probabilistic characterization of GSI, respectively. The PDF line is plotted
on the histogram of the mapped GSI by the observational
method in the figures. It can be seen that for the three cases
(i.e., Fig. 8ac), the PDFs of the Bayesian equivalent GSI
fit relatively well with the histogram of the mapped GSI.
This shows that the information contained in the Bayesian
equivalent samples is consistent with that of the mapped
GSI obtained from the observational method.
The CDFs of the Bayesian equivalent GSI samples
generated from Data sets I, II and III are plotted in
Fig. 9ac by solid line, respectively. The figures also
include the CDFs of the mapped GSI from the observational method and the calculated GSI using Eq. (1) plotted
by open circles and shaded triangles, respectively. In the
figures, the CDF of the Bayesian equivalent GSI plots
more closely to the CDF of the mapped GSI from the
observational method than that of the calculated GSI
using Eq. (1). Such agreement suggests that the information in the Bayesian equivalent GSI samples is consistent with that obtained from 72 data points of mapped
GSI from the observational method. Also, it is observed
that the distribution of Bayesian equivalent GSI becomes
more consistent with the mapped GSI in Fig. 9c, followed
by Fig. 9b and lastly Fig. 9a. This is due to the increase in
number of input data points from 5, 1030 as reflected in
Data sets I, II and III.

123

0.04

25

0.035

Frequency

20

Bayesian Equivalent
Sample Approach

0.03
0.025

15

0.02
10

0.015
0.01

0.005
0

Probability density function

Observational Method

0
0

8.5

17

25.5

34

42.5

51

59.5

68

76.5

85

Cumulative distribution function (CDF)

of GSI

Y. Wang, A. E. Aladejare

Geological Strength Index (GSI)

Direct Calculation
with Eq. (1)

Observational
Method

0.6

0.4

0.2

(a)

8.5

17

34

Bayesian Equivalent
Sample Approach

0.035
0.03
0.025
0.02

10

0.015
0.01

0.005
0

0
8.5

17

25.5

34

42.5

51

59.5

68

76.5

85

Geological Strength Index (GSI)

(b)
25

Cumulative distribution function (CDF) of

GSI

0.04

Probability density function

Observational Method

15

0.05

10

0.02

0.01

0
17

25.5 34

42.5 51

59.5 68

0
76.5 85

Geological Strength Index (GSI)

(c)
Fig. 8 Probability density function (PDF) for GSI. a Data set I with 5
data points of (GSI = 1.5JCond89 ? RQD/2) as input data. b Data set
II with 10 data points of (GSI = 1.5JCond89 ? RQD/2) as input data.
c Data set III with 30 data points of (GSI = 1.5JCond89 ? RQD/2) as
input data

The Bayesian equivalent GSI samples contain combined

information from the GSI chart, site-specific data (i.e.,
JCond89 and RQD) and prior knowledge (e.g., local
experience, engineering judgment, visual observations,
published reports and studies, and available regressions,
etc.). In this example, the integrated information from the
GSI chart, prior knowledge and 5, 10 and 30 site-specific
data (for Data sets I, II and III, respectively) is consistent

123

68

76.5

85

76.5

85

76.5

85

Direct Calculation with

Eq. (1)
0.8

Bayesian Equivalent
Sample Approach
Observational Method

0.6

0.4

0.2

8.5

17

25.5

34

42.5

51

59.5

68

(b)
Cumulative distribution function (CDF) of
GSI

0.03

Probability density function

0.04

15

8.5

59.5

Geological Strength Index (GSI)

Bayesian Equivalent
Sample Approach

51

Observational Method

20

42.5

(a)

0.045

20

Frequency

25.5

Geological Strength Index (GSI)

25

Frequency

Bayesian Equivalent
Sample Approach

0.8

Direct Calculation
with Eq. (1)
Bayesian Equivalent
Sample Approach

0.8

Observational
Method

0.6

0.4

0.2

8.5

17

25.5

34

42.5

51

59.5

68

Geological Strength Index (GSI)

(c)
Fig. 9 Validation of the probability distribution of the GSI from
Bayesian equivalent samples. a Data set I with 5 data points of
(GSI = 1.5JCond89 ? RQD/2) as input data. b Data set II with 10
data points of (GSI = 1.5JCond89 ? RQD/2) as input data. c Data set
III with 30 data points of (GSI = 1.5JCond89 ? RQD/2) as input data

with that of the 72 data points of the mapped GSI. Based on

the GSI chart, limited JCond89 and RQD data [i.e., 5, 10
and 30 (1.5JCond89 ? RQD/2) values in Data sets I, II and
III, respectively] and prior knowledge (i.e., reasonable

Evaluating Variability and Uncertainty of Geological Strength Index at a Specific Site

ranges of model parameters), the Bayesian approach provides a reasonable estimate of the statistical distribution of
GSI. Such probabilistic characterization normally requires
a large amount of data from site investigation, which are
not often available at a specific project because of cost,
man power, and time.

9 Sensitivity Study
To further test the validity and potentiality of the proposed
approach for more generic applicability, sensitivity study is
performed using additional input datasets. The sensitivity
study
is
performed
by
randomly
drawing
(GSI = 1.5JCond89 ? RQD/2) data points from 72 data
points of (GSI = 1.5JCond89 ? RQD/2) obtained from
Hoek et al. (2013). 10 data sets each of 5, 10, 15, 20, 25 and
30 data points of (GSI = 1.5JCond89 ? RQD/2) are generated. This results in a total of 60 different data sets of
input data points of (GSI = 1.5JCond89 ? RQD/2). As an
example, Fig. 10 shows 10 sets of the input data with 10
(GSI = 1.5JCond89 ? RQD/2) values in each data set (i.e.,
ns = 10 in each data set). Then, using each set of the 60
data sets of input data and the prior knowledge already
given in Sects. 3 and 6 (i.e., remin 0, remax 24:5,
lGSImin 0, lGSImax 85, rGSImin 0, rGSImax 14:2),
30,000 equivalent samples of GSI are generated for each of
the 60 data sets, respectively. Thus, a total of 60 sets of the
estimates of mean and standard deviation of GSI denoted
by l*GSI and r*GSI, respectively, were obtained. Also, for
each set of the input data obtained using Eq. (1), estimates
of mean and standard deviation of GSI were also obtained.
This leads to another 60 sets of l*GSI and r*GSI.
Hypothesis tests are used to compare the results of
l*GSI and r*GSI from both Bayesian approach and direct

(GSI=1.5JCond89+RQD/2) values

70
65
60
55
50
45
40
35
30
25
20

10

Data set ID

Fig. 10 Ten sets of 10 input data points of (GSI = 1.5JCond89 ?

RQD/2)

calculation obtained through Eq. (1) with the mean (lGSI)

and standard deviation (rGSI) of mapped GSI from
observational method. The acceptance regions of estimates
l*GSI and r*GSI at significance level of are calculated as
follows:
rGSI
rGSI

1
lGSI U1
14
/=2 p \lGSI  lGSI U1/=2 p
ns
ns
r
r
c/=2;ns 1
c1/=2;ns 1

15
rGSI \rGSI 
rGSI
ns  1
ns  1
-1
where U-1
/2 and U1-/2 are the values of an inverse standard Gaussian CDF at /2 and 1 - /2, respectively;
c/=2;ns 1 and c1/=2;ns 1 are the values of a Chi-squared
statistic with ns-1 degrees of freedom at the levels of /2
and 1 - /2, respectively. Estimates of the mean and
standard deviation of GSI are presented in the subsequent
subsections.

9.1 Estimates of the GSI Mean

Figure 11a is a plot of l*GSI from both Bayesian approach and
direct calculation against the number ns of input data in each
data set. The l*GSI values estimated from Bayesian samples
are plotted by open circles, while those from directly calculated GSI through Eq. (1) are plotted by open triangles.
Figure 11a also includes the mean value of the mapped GSI
(i.e., 44.36), obtained from observational method, and the
acceptance region of l*GSI at the 5 % significance level plotted
by a dashed line and solid lines, respectively. It can be
observed from Fig. 11a that open circles plot more closely to
the dashed line than open triangles. In addition, it is observed
that even when mean of the input data ranges from 38.05 to
55.00 for 5 input data points, the estimates of GSI mean from
the Bayesian approach (i.e., 42.9750.63) are still close to the
mean value of the mapped GSI. This means that l*GSI values
estimated from the Bayesian samples are better estimates of
the mean GSI at the site than those estimated from directly
calculated GSI using Eq. (1).
It can also be observed from Fig. 11a that the differences between the l*GSI values obtained from the Bayesian
approach and those from direct calculation become relatively small as the number of input data increases. This
shows that as the input data increases, the statistical
uncertainty that results from insufficient site information
tends to diminish. When only five input data points were
used, the Bayesian approach significantly reduces its statistical uncertainty. This results in significant improvement
on the estimations of lGSI from Bayesian approach over
direct calculation. This is particularly useful in rock
mechanics and engineering, where the number of rock
property data obtained during site characterization is often
small and its statistical uncertainty is substantial.

123

Y. Wang, A. E. Aladejare
60

55

Mean of GSI

50

45

40
Bayesian Equivalent Sample Approach
Direct Estimates from Equation 1

35

Mapped Values from Observational Method

Acceptance Region at the 5% Significance Level

30

10

15

20

25

30

35

Number of data

(a)
18

Standard deviation of GSI

16
14
12

direct calculation from Eq. (1). This kind of improvement

is particularly noticeable when the number of input data ns
is small (say five for instance).
It can also be observed from Fig. 11b that as number of
input data ns increases, the r*GSI values estimated from
Bayesian samples gradually approach the standard deviation value of the mapped GSI. Whereas, r*GSI values estimated directly from Eq. (1) are larger than the standard
deviation of the mapped GSI obtained from observational
method, and plot farther away from the standard deviation
of the mapped GSI even when the number of input data ns
is large at 30. This is because direct estimation of GSI
using Eq. (1) involves the use of input data, which contains
inherent variability and Eq. (1) which contains model
uncertainty. Therefore, estimates of r*GSI values from
directly calculated GSI samples contain both inherent
variability and model uncertainty. These r*GSI values from
directly calculated GSI samples are obviously larger than
the inherent variability. On the other hand, the model
uncertainty is considered and formulated explicitly in the
Bayesian approach, and its r*GSI values mainly reflect the
inherent variability in GSI.

10

10 Summary and Conclusions

8
6

Bayesian Equivalent Sample Approach

Direct Estimates from Equation 1
Mapped Values from Observational Method
Acceptance Region at the 5% Significance Level

4
2

10

15

20

25

30

35

Number of data

(b)
Fig. 11 Sensitivity study results. a Estimates of the mean of GSI.
b Estimates of the standard deviation of GSI

9.2 Estimates of the GSI Standard Deviation

Figure 11b is a plot of r*GSI from both Bayesian approach
and direct calculation against the number ns of input data in
each data set. The r*GSI values estimated from Bayesian
samples are plotted by open circles, while those from
directly calculated GSI through Eq. (1) are plotted by open
triangles. Figure 11b also includes the standard deviation
value of the mapped GSI (i.e., 10.35), obtained from
observational method, and the acceptance region of r*GSI at
the 5 % significance level plotted by a dashed line and
solid lines, respectively. In a similar trend to l*GSI, the
r*GSI values estimated from the Bayesian samples plot more
closely to the standard deviation value of the mapped GSI
from observational method than those estimated from
directly calculated GSI samples. This means that the
Bayesian approach provides better estimation of rGSI than

123

This paper integrated the GSI chart with Bayesian

approach to explicitly and properly quantify and incorporate the model uncertainty in the GSI chart and inherent
spatial variability of GSI. A Bayesian framework for
probabilistic characterization of GSI based on the quantified GSI chart has been developed. The proposed approach
considers the model uncertainty to be uniformly distributed
between two extreme cases: when the GSI chart has no
model uncertainty and when the GSI chart offers no useful
information for GSI characterization. The proposed
approach effectively modeled the model uncertainty in the
GSI chart during GSI characterization. The approach
simulates equivalent samples of GSI from the integrated
knowledge of the GSI chart, prior knowledge and sitespecific observation data of JCond89 and RQD.
Equations were derived for the Bayesian approach, and
the approach was illustrated using data from a drill and
blast tunnel project. In addition, a sensitivity study was
performed to test the validity and potentiality of the
Bayesian approach. It has been shown that the proposed
approach properly and explicitly considers the model
uncertainty associated with the chart-based GSI characterization. The proposed approach reduces the uncertainties
associated with GSI estimation from the GSI chart in a
rational and transparent manner. This is because the model
uncertainty involved in the GSI chart has been formulated
and considered explicitly in the Bayesian approach.

Evaluating Variability and Uncertainty of Geological Strength Index at a Specific Site

Therefore, the standard deviation of GSI from the Bayesian

approach mainly reflects the inherent spatial variability
alone, unlike the standard deviation of calculated GSI from
the GSI chart that includes inherent spatial variability and
model uncertainty. The proposed Bayesian approach has
been shown to perform well in a drill and blast tunnel
project.
Acknowledgements The authors would like to thank the two
anonymous reviewers for their constructive and valuable comments,
which helped to improve the quality of the paper. The work described
in this paper was supported by grants from the Research Grants
Council of the Hong Kong Special Administrative Region, China
(Project No. 9042172 (CityU 11200115) and Project No. 8730035
(CityU8/CRF/13G)) and a Strategic Research Grant from City
University of Hong Kong (Project No. 7004178). The financial supports are gratefully acknowledged.

References
Ang AH-S, Tang WH (2007) Probability concepts in engineering:
emphasis on applications to civil and environmental engineering.
Wiley, New York
Baecher GB, Christian JT (2003) Reliability and statistics in
geotechnical engineering. Wiley, Hoboken, p 605
Bieniawski ZT (1989) Engineering rock mass classification. Wiley
Interscience, New York
Cai M, Kaiser PK, Uno H, Tasaka Y, Minami M (2004) Estimation of
rock mass strength and deformation modulus of jointed hard rock
masses using the GSI system. Int J Rock Mech Mining Sci
41(1):319
Cao Z, Wang Y, Li D (2016) Quantification of prior knowledge in
geotechnical site characterization. Eng Geol 203:107116
Carter TG (2010) Applicability of classifications for tunneling
valuable for improving insight, but problematic for contractual
support definition or final design. In: Proceedings world
tunnelling conference (WTC 2010), 36th ITA Congress. Vancouver, Paper 00401, Session 6c, p 8
Carter TG and Marinos V (2014) Use of GSI for rock engineering
design. In: Proceedings 1st international conference on applied
empirical design methods in mining, Lima-Peru, 911th June, 19
Cetin KO, Kiureghian AD, Seed RB (2002) Probabilistic models for
the initiation of seismic soil liquefaction. Struct Saf 24:6782
Deere DU (1963) Technical description of rock cores for engineering
purposes. Felsmechanik und Ingenieurgeologie (Rock Mechanics and Engineering Geology) 1(1):1622
Ditlevsen O (1981) Uncertainty modeling with applications to
multidimensional civil engineering systems. McGraw-Hill,
New York
Hoek E, Marinos P (2000) Predicting tunnel squeezing problems in
weak heterogeneous rock masses. Tunnels and Tunnelling
International. Part 1November 2000, Part 2December, 2000

Hoek E, Kaiser PK, Bawden WF (1995) Support of underground

excavations in hard rock. A.A. Balkema, Rotterdam, p 215
Hoek E, Carter TG, Diederichs MS (2013) Quantification of the
Geological Strength Index Chart. Paper prepared for presentation
at the 47th US Rock Mechanics/Geomechanics Symposium held
in San Francisco
Marinos P, Hoek E (2000) GSIa geologically friendly tool for rock
mass strength estimation. In: Proceedings GeoEng2000 conference, Melbourne, Australia
Marinos P, Marinos V, Hoek E (2007) Geological Strength Index
(GSI). A characterization tool for assessing engineering properties for rock masses. In: Romana, Perucho, Olalla (eds)
Underground works under special conditions. Taylor and
Francis, Lisbon, pp 1321
Mitchell JK, Soga K (2005) Fundamentals of soil behavior. Wiley,
Hoboken
Morelli GL (2015) Variability of the GSI index estimated from
different quantitative methods. Geotech Geol Eng 33:983995
Moss RES, Seed RB, Kayen RE, Steward JP, Der Kiureghian A
(2006) CPT-based probabilistic and deterministic assessment of
in situ seismic soil liquefaction potential. J Geotech Geoenvironmental Eng 132(8):10321051
Phoon KK, Kulhawy FH (1999) Evaluation of geotechnical property
variability. Can Geotech J 36(4):625639
Ronold KO, Bjerager P (1992) Model uncertainty representation in
geotechnical reliability analyses. J Geotech Eng 118(3):363376
Russo G (2007) Improving the reliability of GSI estimation: the
integrated GSI-RMi system. In: Proceedings I.S.R.M. workshop
Underground Works under Special Conditions, Madrid,
pp 123130
Sivia DS, Skilling J (2006) Data analysis: a Bayesian tutorial. Oxford
University Press, New York
Sonmez H, Ulusay R (1999) Modifications to the geological strength
index (GSI) and their applicability to the stability of slopes. Int J
Rock Mech Mining Sci 36:743760
Walpole RE, Myers RH, Myers SL (1998) Probability and statistics
for engineers and scientists. Prentice Hall, Upper Saddle River
Wang Y, Aladejare AE (2015) Selection of site-specific regression
model for characterization of uniaxial compressive strength of
rock. Int J Rock Mech Min Sci 75:7381
Wang Y, Cao Z (2013) Probabilistic characterization of Youngs
modulus of soil using equivalent samples. Eng Geol
159:106118
Wang Y, Au SK, Cao Z (2010) Bayesian approach for probabilistic
characterization of sand friction angles. Eng Geol
114(34):354363
Wang Y, Huang K, Cao ZJ (2013) Probabilistic identification of
underground soil stratification using cone penetration tests. Can
Geotech J 50(7):766776
Wang Y, Cao ZJ, Li DQ (2016) Bayesian perspective on geotechnical
variability and site characterization. Eng Geol 203:117125

123