Efficient Probabilistic Back-Analysis of Slope Stability

Model Parameters
J. Zhang1; Wilson H. Tang, Hon.M.ASCE2; and L. M. Zhang, M.ASCE3

Abstract: Back-analysis of slope failure is often performed to improve one’s knowledge on parameters of a slope stability analysis
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.

model. In a failed slope, the slip surface may pass through several layers of soil. Therefore, several sets of model parameters need to be
back-analyzed. To back-analyze multiple sets of slope stability parameters simultaneously under uncertainty, the back-analysis can be
implemented in a probabilistic way, in which uncertain parameters are modeled as random variables, and their distributions are improved
based on the observed slope failure information. In this paper, two methods are presented for probabilistic back-analysis of slope failure.
For a general slope stability model, its uncertain parameters can be back-analyzed with an optimization procedure that can be imple-
mented in a spreadsheet. When the slope stability model is approximately linear, its parameters can be back-analyzed with sensitivity
analysis instead. A feature of these two methods is that they are easy to apply. Two case studies are used to illustrate the proposed
methods. The case studies show that the degrees of improvement achieved by the back-analysis are different for different parameters, and
that the parameter contributing most to the uncertainty in factor of safety is updated most.
DOI: 10.1061/共ASCE兲GT.1943-5606.0000205
CE Database subject headings: Landslides; Slope stability; Models; Parameters; Failure investigation; Probability; Reliability;
Bayesian analysis.
Author keywords: Landslides; Slope stability; Failure investigation; Probability methods; Reliability; Bayesian analysis.

Introduction analysis methods include: 共1兲 it provides a logical way to

A slope failure implies that the factor of safety of the slope at the
moment of failure is unity. Based on this information, back-
analysis is often carried out to improve knowledge on slope sta-
bility parameters. Stability parameters may include both soil
strength parameters and pore water pressure parameters at the
moment of slope failure. Both deterministic methods 共e.g., Wes-
ley and Leelaratnam 2001; Tiwari et al. 2005兲 and probabilistic
methods 共Luckman et al. 1987; Gilbert et al. 1998; Chowdhury et
al. 2004兲 have been used to perform back-analysis. The philoso-
phies behind deterministic and probabilistic back-analysis meth-
ods are different. While deterministic back-analysis methods
intend to find a set of parameters that would result in the slope
failure, probabilistic back-analysis methods recognize that there
might be numerous combinations of such parameters, but their
relative likelihoods are different, which can be quantified by prob-
ability distributions. Major advantages of probabilistic back-
incorporate information from other sources in the back-analysis
and 共2兲 it is capable of back-analyzing multiple sets of slope
stability parameters simultaneously. One possible disadvantage of
probabilistic back-analysis methods is that they are usually not as
easy to implement compared with deterministic back-analysis
methods.
The purpose of this paper is to suggest two efficient methods
for probabilistic back-analysis of slope failures, i.e., a method
based on optimization and a method based on sensitivity analysis.
To avoid the potential cumbersome programming work, a spread-
sheet template is developed to implement the back-analysis
method based on optimization. When the slope stability model is
approximately linear, multiple sets of stability parameters may be
back-analyzed even without resorting to the optimization process.
Two case studies are used to illustrate the suggested methods and
implementation ideas.

1 Methods of Probabilistic Back-Analysis

Ph.D. Candidate, Dept. of Civil and Environmental Engineering, The
Hong Kong Univ. of Science and Technology, Clear Water Bay, Hong
Kong. E-mail: zjxac@ust.hk
2
Chair Professor, Dept. of Civil and Environmental Engineering, The
Hong Kong Univ. of Science and Technology, Clear Water Bay, Hong
Kong 共corresponding author兲. E-mail: wtang@ust.hk
3
Associate Professor, Dept. of Civil and Environmental Engineering,
The Hong Kong Univ. of Science and Technology, Clear Water Bay,
Hong Kong. E-mail: cezhangl@ust.hk
Note. This manuscript was submitted on April 20, 2008; approved on
July 23, 2009; published online on December 15, 2009. Discussion period
open until June 1, 2010; separate discussions must be submitted for indi-
vidual papers. This paper is part of the Journal of Geotechnical and
Geoenvironmental Engineering, Vol. 136, No. 1, January 1, 2010.
©ASCE, ISSN 1090-0241/2010/1-99–109/$25.00.
Let g共␪ , r兲 denote a slope stability model 共such as a model based
on a limit equilibrium method兲, where ␪⫽vector denoting uncer-
tain input parameters and r⫽vector denoting input parameters
without uncertainty. In the following text, r is dropped from the
slope stability model for simplicity. The uncertain input param-
eters ␪ may include both soil strength parameters and pore water
pressure parameters. In practice, one may already have some prior
knowledge on ␪ before performing the back-analysis based on
either geotechnical tests or engineering experiences. For simplic-
ity, assume that the prior knowledge on ␪ can be described by a
multivariate normal distribution with a mean of ␮␪ and a covari-
ance matrix of C␪. The objective of probabilistic back-analysis is

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010 / 99

J. Geotech. Geoenviron. Eng. 2010.136:99-109.

 then to improve the probability distribution of ␪ based on ob-
served slope failure information.
One common concern in back-analysis is that the prediction
from the theoretical slope stability model g共␪兲 may be subjected
to model error. In such a case, even if the calculated factor of
safety is unity, the actual factor of safety of the slope may not be
unity. To quantify the effect of model imperfection, the model
uncertainty can be modeled as a random variable, which is de-
fined as
␧ = y − g共␪兲
where y = actual factor of safety and ␧ = random variable charac-
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
terizing the modeling uncertainty. For simplicity, assume ␧ fol-
lows the normal distribution with a mean of ␮␧ and a standard
deviation of ␴␧.
With the above assumptions, the improved distribution of ␪
can also be approximated by a multivariate normal distribution
关see Appendix I based on Tarantola 共2005兲兴. Let ␮␪兩d and C␪兩d
denote the improved mean and covariance matrix of ␪, respec-
tively. As a multivariate normal distribution can be fully deter-
mined by its mean and covariance matrix, the task in the
probabilistic back-analysis is then reduced to determining ␮␪兩d
and C␪兩d.
For a general slope stability model g共␪兲, ␮␪兩d is a point that
maximizes the chance to observe the slope failure event, and it
denotes the most probable combination of parameters that had led
to the slope failure event. ␮␪兩d can be obtained by minimizing the
following misfit function 2S共␪兲 关see Appendix I based on Taran-
tola 共2005兲兴:
关g共␪兲 + ␮␧ − 1兴T关g共␪兲 + ␮␧ − 1兴
2S共␪兲 = + 共␪ − ␮␪兲TC␪−1共␪ − ␮␪兲
␴2␧
The improved covariance matrix of ␪, C␪兩d, which describes the
magnitude of uncertainty in each component of ␪ as well as the
dependence relationships among various components of ␪, can be
determined as follows 关see Appendix I based on Tarantola
共2005兲兴:
C␪兩d = 冉 冊
G TG
␴2␧
+ C␪−1
−1
冏 冏
⳵ g共␪兲
G=
⳵␪ ␪=␮␪兩d
where G = row vector representing the sensitivity of g共␪兲 with
respect to ␪ at ␮␪兩d.
The back-analysis based on the above equations is called
“probabilistic back-analysis method by optimization” in this
paper. At a first glance, the application of this method seems not
easy, as in Eq. 共2兲 the minimization is coupled with the slope
stability model g共␪兲. However, as will be shown later in this
paper, the minimization can be automated in a commonly avail-
able spreadsheet, thus avoiding the possible troublesome pro-
gramming work.
The expressions in Eqs. 共2兲–共4兲 are indeed quite general, and
are applicable no matter g共␪兲 is linear or not. However, when
g共␪兲 is approximately linear, ␮␪兩d and C␪兩d can be determined
analytically with the following equations without resorting to the
minimization procedure 关see Appendix II based on Tarantola
共2005兲兴:
100 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010
J. Geotech. Geoenviron. Eng. 2010.136:99-109.
␮␪兩d = ␮␪ + C␪HT共HC␪HT + ␴2␧兲−1关1 − g共␮␪兲 − ␮␧兴 共5兲
冉 冊
−1
H TH
C␪兩d = + C␪−1 共6兲
␴2␧
H= 冏 冏
⳵ g共␪兲
⳵␪ ␪=␮␪
共7兲
where g共␮␪兲 = predicted factor of safety calculated at point ␮␪ and
共1兲 H = row vector representing the sensitivity of g共␪兲 with respect to
␪ at point ␮␪.
The back-analysis based on Eqs. 共5兲–共7兲 is called “simplified
probabilistic back-analysis” in this paper. Compared with the
method by optimization, the simplified method is easier to apply,
as it does not involve any minimization procedure. The limitation
of the simplified method is that it is applicable only when g共␪兲 is
largely linear. Fortunately, based on the authors’ experience, many
stability models are in fact rather linear in spite of the fact that
numerical iterations are usually involved in slope stability mod-
els, which was also noticed by Mostyn and Li 共1993兲.
The following comments are made regarding the two methods
proposed in this paper for back-analysis of slope failure.
1. In the two methods described above, the emphasis is to im-
prove the knowledge on slope stability parameters ␪ through
back-analysis. The probability density function of the model
uncertainty variable ␧ is not updated in the back-analysis.
2. The probabilistic back-analysis procedures described above
are based on the assumption that the prior distribution of ␪
can be represented or approximated by a multivariate normal
distribution. This assumption simplifies the implementation
of the back-analysis methods. If this assumption is not satis-
fied, the prior distribution of ␪ can be transformed into a
共2兲 multivariate normal distribution using techniques like Nataf
transformation 共Der Kiureghian and Liu 1986兲, and the trans-
formed parameters can then be back-analyzed using the
above procedures. The transformation would introduce addi-
tional computational work into the back-analysis. In the ex-
amples presented in this paper, the prior distributions are
assumed to be multivariate normal.
共3兲
3. The credibility of the back-analysis results depends on the
accuracy of the identified soil stratigraphy 共including the ac-
curate location of the slip surface兲 as well as the understand-
ing of the failure mechanism. The back-analysis results may
be meaningless if the back-analysis is carried out based on an
共4兲
unrealistic stratigraphy or based on a wrong failure mecha-
nism. It is assumed in this study that both soil stratigraphy
relevant to the back-analysis and the location of the slip sur-
face are well defined through postfailure investigation. More-
over, it is assumed that the slope stability model g共␪兲 can
correctly reflect the failure mechanism. Back-analysis of
slope failures when the stratigraphy is uncertain is out of the
scope of this study.
4. Eqs. 共2兲–共7兲 show that the back-analyzed distribution of ␪
depends on ␮␪ and C␪, i.e., the prior knowledge about ␪, as
well as the slope failure information. Hence, the back-
analyzed distribution is a combination of the prior knowl-
edge about ␪ and the knowledge learned from the slope
failure event. As will be seen later in this paper, a change in
prior distribution will cause changes in the back-analyzed
distribution of ␪. This implies that, in a probabilistic back-
analysis, information from other sources on the slope stabil-
ity parameters is also important.
 5. In minimizing Eq. 共2兲, it is possible that the global minimum
point cannot be found. To reduce the chance of missing the
global minimum point, the minimization of Eq. 共2兲 can be
solved starting from various initial points. If indeed several
points are found to satisfy the minimization criteria, the dif-
ferences between posterior means and posterior covariance
matrices evaluated at these points would increase when these
points are more separated. However, among various points,
the point that gives the minimal value of 2S共␪兲 is most plau-
sible because it has the maximum likelihood, and the poste-
rior mean and covariance matrix should be estimated at this
point.
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
6. The optimization method and simplified method for back-
analysis are quite general. Both methods are applicable to
updating multiple model input parameters, no matter the
slope failure occurs in a single soil layer or in layered soils.
Both methods are also applicable no matter ␪ is correlated or
not in its prior distribution.
7. It is known that soil property varies spatially. If spatial vari-
ability is to be considered, one should obtain the representa-
tive distributions of the spatial averages of the input
parameters along the slip surface separately prior to the back-
analysis. These prior distributions can then be used in the
back-analysis to obtain the updated posterior distributions of
the spatial averages of the input parameters.
Step-by-Step Implementation
In the two suggested methods, the improved joint probability dis-
tribution of the uncertain parameters is approximated by the mul-
tivariate normal distribution, and the task of back-analysis is then
reduced to the determination of the mean 共␮␪兩d兲 and covariance
matrix 共C␪兩d兲 of the multivariate normal distribution. The formu-
las to calculate ␮␪兩d and C␪兩d have been summarized in Eqs.
共2兲–共4兲 and Eqs. 共5兲–共7兲, respectively. The probabilistic back-
analysis methods described above can be implemented in three
steps:
1. Select a stability model, g共␪兲 and identify the uncertain pa-
rameters ␪;
2. Quantify the knowledge on ␪ prior to the back-analysis and
on the model uncertainty of g共␪兲. This step is in fact a pro-
cess of determining ␮␪, C␪, ␮␧, and ␴␧; and
3. Improve the probability distribution of ␪ considering the
slope failure event. In this step, ␮␪兩d and C␪兩d, which contain
the improved knowledge on ␪, are calculated using the for-
mulas presented in the previous section.
In the following, these three steps are illustrated in detail using
two case studies.
Example 1
Shek Kip Mei Slope Failure
The slope analyzed here is a cut slope located at Shek Kip Mei,
Hong Kong, which failed during a heavy rainstorm on August 25,
1999. After the slope failure, FMSW 共2000兲 carried out a detailed
postfailure investigation into this slope, and the findings reported
by FMSW 共2000兲 formed the basis of this case study. Fig. 1
shows a representative cross section of this slope as well as the
identified slip surface 共FMSW 2000兲. In this figure, Layers II, III,
IV, and V denote slightly decomposed granite 共SDG兲, moderately
decomposed granite 共MDG兲, highly decomposed granite 共HDG兲,
and completely decomposed granite 共CDG兲, respectively. Based
JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010 / 101
J. Geotech. Geoenviron. Eng. 2010.136:99-109.
50
40
V
Slip surface
Groundwater table
Elevation (m)
IV by FMSW (2000)
30
II
V
II V/IV II V
II
20 IV
III/II V
IV/III
II
II
10
0 10 20 30 40
Horizontal distance (m)
Fig. 1. A representative cross section of the Shek Kip Mei landslide
共II, III, IV, and V: slightly, moderately, highly, and completely de-
composed granite兲 关adapted from FMSW 共2000兲兴
on GEO 共2000兲, while SDG and MDG can be considered as
rocks, HDG and CDG are classified as soils. The groundwater
table at the moment of slope failure judged by FMSW 共2000兲 is
also plotted in Fig. 1. Based on this groundwater table, the pore
water pressure ratio along the slip surface is in the range of
0–0.30, and the average pore pressure ratio is about 0.13.
Slope Stability Analysis Model and Uncertain Parameters
The method of Spencer 共1967兲 is used to calculate the factor of
safety of this cut slope. Let c and ␾ denote the average cohesion
and average friction angle along the slip surface, respectively. As
c and ␾ are uncertain, they are modeled as random variables. In
addition to the above two shear strength parameters, the pore
water pressures at the moment of slope failure are also unknown.
Here the effect of pore water pressure is modeled by a homoge-
neous equivalent pore water pressure ratio along the slip surface,
ru, which is defined as the ratio of pore water pressure to vertical
total stress. Although ru is not an accurate representation of a pore
pressure distribution along the slip surface, it is an index with
effect equivalent to that of the complex and unknown pore pres-
sure 共El-Ramly et al. 2005兲. Based on the above discussion, the
parameters to be back-analyzed for this slope can be denoted by
the vector ␪ = 兵c , ␾ , ru其T.
Assessment of Uncertainties
Fig. 2 shows the saturated triaxial test results from CDG samples
retrieved using a Mazier sampler outside the distressed zone of
the slope. FMSW 共2000兲 judged that the average cohesion and
friction angle of the CDG are 8 kPa and 38°, respectively. The
lower bound cohesion and friction angle inferred from the triaxial
tests are 2 kPa and 36°, respectively. Assuming that the lower
bound implied by the triaxial tests is two standard deviations
away from the mean value, the standard deviations of cohesion
and friction angle are judged to be 3 kPa and 1°, respectively. The
above statistics for cohesion and friction angle do not consider the
effect of test uncertainty. The existence of test uncertainty would
introduce additional uncertainty into the soil strength parameters.
On the other hand, the effect of spatial variability of soil proper-
ties is also not considered in the previously assessed statistics for
 350
300
q' = 0.5(σ1'-σ3' ) (kPa)
250
200
150
100
50
0
0 50 100 150 200 250
p' = 0.5(σ 1 '+ σ 3 ' ) (kPa)
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
Fig. 2. Results of triaxial tests on CDG 共␴⬘1 = major principal effec-
tive stress, ␴⬘3 = minor principal effective stress兲 关adapted from
FMSW 共2000兲兴
cohesion and friction angle. As soil property fluctuates spatially
along the slip surface, a high value at a certain point will be
balanced by lower values at other points. The existence of spatial
variability reduces the uncertainty of soil property along the slip
surface. Due to lack of information for satisfactory evaluation of
test uncertainty and spatial variability, the effects of test uncer-
tainty and spatial variability are assumed to be largely cancelled
out. With this assumption, the previously assessed statistics for
cohesion and friction angle are used to represent the statistics of
cohesion and friction averaged over the slip surface.
The coefficient of correlation between c and ␾ is not reported
in FMSW 共2000兲. Previous studies showed that c and ␾ may be
negatively correlated 共e.g., Cherubini 2000兲. In this study, the
back-analysis will be first carried out assuming c and ␾ are sta-
tistically independent. A sensitively analysis will be carried out
later to check the effect of the correlation between c and ␾ on the
back-analysis results.
There was no measurement on the pore pressures at the mo-
ment of slope failure. Therefore, the prior knowledge on average
pore pressure ratio is assessed mainly by judgment. In this study,
we assume that the average pore pressure ratio ru at the moment
of slope failure follows a normal distribution with a mean of 0.15
and a standard deviation of 0.1.
We further assume that c, ␾, and ru are statistically indepen-
dent. Therefore, our prior knowledge on c, ␾, and ru can be rep-
resented by a multivariate normal distribution with a mean of
␮␪ = 兵8 , 38, 0.15其T, and a covariance matrix as follows:
冤 冥
32 0 0
C␪ = 0 12 0
0 0 0.12
The units for c and ␾ in the above ␮␪ and C␪ are kPa and degree,
respectively. To be clear, the prior distribution used here is called
Prior 1.
In the method of Spencer 共1967兲, all the three equilibrium
conditions 共horizontal force equilibrium, vertical force equilib-
rium, and moment equilibrium兲 are satisfied. Christian et al.
共1994兲 suggested that the model uncertainty of the simplified
method of Bishop 共1955兲 has a mean of 0.05 and a standard
deviation of 0.07. As the results from Bishop’s simplified method
are usually close to rigorous methods that satisfy all three equi-
librium conditions, it is assumed that the model uncertainty of the
102 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010
method of Spencer 共1967兲 can also be modeled as a normally
distributed random variable with a mean of 0.05 and a standard
deviation of 0.07, i.e., ␮␧ = 0.05 and ␴␧ = 0.07.
Probabilistic Back-Analysis
As mentioned before, since the improved ␪ follows a multivariate
normal distribution, the task of probabilistic back-analysis is to
determine ␮␪兩d and C␪兩d. For a general slope stability model, ␮␪兩d
can be found through maximizing the chance to observe the slope
failure event, or equivalently, through the minimization of Eq. 共2兲.
300 350 400 450
The work involved in minimizing Eq. 共2兲, however, is unlikely to
be trivial because the slope stability model g共␪兲 is coupled with
the minimization problem. In this study, a spreadsheet template
has been developed to automate the probabilistic back-analysis
method by optimization. The layout of the spreadsheet is shown
in Fig. 3, which can be accomplished in three steps.
1. Construct the slope stability analysis model g共␪兲 to relate the
calculated factor of safety with c, ␾, and ru. The method of
Spencer 共1967兲 is constructed in the spreadsheet in this step.
In Fig. 3, the slope mass is divided into 20 slices. The factor
of safety of the slope can be obtained conveniently through
the Solver routine in a spreadsheet. In Fig. 3, the formula in
cell N27 is set to be “=1 * J27.” The settings in the Solver are
“Minimize N27, by changing cell I27 and J27, subjected to
M24= 0 and N24= 0.” The implementation of the method of
Spencer 共1967兲 in the spreadsheet in this step is similar to
that suggested by Low 共2003兲.
2. Find the improved mean ␮␪兩d by optimization. In the second
step, the prior mean 共␮␪兲, prior covariance matrix 共C␪兲,
model uncertainty 共␮␧ and ␴␧兲, and observed information
共factor of safety equal to unity兲 are typed into the spread-
sheet. Based on the factor of safety calculated from Step 1,
the formula for misfit function 2S共␪兲 in Eq. 共2兲 is typed into
cell N33. The posterior mean ␮␪兩d can be obtained by mini-
mizing cell N33 using the Solver. The setting in the Solver is
“Minimize N33, by changing cell I27, J27, G25, G26, and
G27, subjected to M24= 0 and N24= 0.”
3. Calculate the improved covariance matrix C␪兩d. In this step, a
sensitivity analysis at ␮␪兩d can be performed to obtain the
sensitivity vector G as defined in Eq. 共4兲. This sensitivity
vector is then substituted into Eq. 共3兲 to calculate C␪兩d.
To reduce the chance of missing the global minimum point,
the minimization of the misfit function is repeated starting from
several initial points selected within the possible ranges of the
parameters. It is found that all these minimization problems con-
verge to the same minimum point. Therefore, it seems that in this
problem, the point that minimizes the misfit function is unique,
and this point is then regarded as ␮␪兩d. Based on ␮␪兩d and C␪兩d
共8兲 calculated from the optimization method, the improved distribu-
tion for each parameter can be obtained. The prior and the im-
proved probability density functions for c, ␾, and ru are compared
in Fig. 4.
Although in this study the factor of safety is calculated using
the method of Spencer 共1967兲, other slope stability analysis meth-
ods can also be used in a similar fashion as long as they can be
implemented in a spreadsheet. In fact, many slope stability mod-
els have been implemented in spreadsheets, such as Bishop’s sim-
plified method 共Bishop 1955; El-Ramly et al. 2002兲, Janbu’s
generalized method 共Janbu 1987; Low and Tang 1997兲, and Chen-
Morgenstern’s method 共Chen and Morgenstern 1983, Low et al.
1998兲. Although in this paper it is shown that the probabilistic
back-analysis method by optimization can be implemented in a
J. Geotech. Geoenviron. Eng. 2010.136:99-109.
 A B C D E F G H I J K L M N O

1
Step 1: Build the slope stability model
2 Slice# α b h u c tanφ s A B C D Q Q*cos( α - θ )
3 1 1.428 1.058 3.672 6.106 7.676 0.78 371.5 0.105 0.075 -0.03 9E-04 -48.8 -36.1
4 2 1.025 1.058 5.963 9.915 7.676 0.78 371.5 0.49 0.133 -0.13 0.044 -48.5 -45.7
21 19 0.315 1.058 2.091 3.477 7.676 0.78 371.5 0.885 -0.27 -0.03 0.08 33.05 30.76
22 20 0.098 1.058 0.717 1.193 7.676 0.78 371.5 0.826 -0.43 -0 0.044 42.37 35.18
23 Posterior mean, obtained by ∑Q ∑Q*cos(a- θ )
24 γ H θ,μθ|θ|d minimizing N33 1E-07 1E-07
25 17 20.66 c 7.676
26 φ 37.96 β Fs Fs
27 ru 0.171 0.689 0.956 0.956
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.

28

29
Step 2: Calculate the posterior mean Eq. (2)
30
31 μθ Cθ
32 8 9 0 0 με σε Observed data 2S( θ)
33 38 0 1 0 0.05 0.07 1 0.067
34 0.15 0 0 0.01
35

36 Posterior covariance,
Step 3: Calculate the posterior covariance matrix obtained by Eq. (3)
37
Sensitivity vector, Eq. (4)
38 Cθ|d
39 7.435 -0.17 0.103
40 G 0.027 0.027 -1.61 -0.17 0.981 0.011
41 0.103 0.011 0.003
42
43 Notes:
44 1. γ and H denote the unit weight of soil and the slope height, respectively;
45 2. β and F s denote the angle of interslice force and factor of safety of the slope, respectively;
46 3. α , b , h , and u are the base inclination angle, width, height and pore pressure of a slice, respectively;
47 4. s, A, B, C, D are five intermediate variables that facilitate coding in the spreadsheet;
48 5. When ∑Q = 0 and ∑Q*cos(a- θ ) = 0, force and moment equirlibriums are satisfied, respectively.
49

Fig. 3. Layout of the spreadsheet for probabilistic back-analysis of slope failure

spreadsheet, we do not mean that this method must be imple-
mented in the spreadsheet. Other tools can also be used.
Fig. 5 shows the relationships between the factor of safety and
c, ␾, and ru around ␮␪. These relationships are rather linear in the
likely ranges of these parameters. Therefore, the slope stability
parameters may also be back-analyzed using the simplified
method. To implement the simplified method, one can perform a
sensitivity analysis at point ␮␪ to get g共␮␪兲 and H first, and then
substitute them into Eqs. 共5兲 and 共6兲 to calculate ␮␪兩d and C␪兩d,
respectively. As g共␮␪兲 and H can be calculated using a conven-
tional slope stability program, the simplified method can be
implemented even without resorting to a spreadsheet. To check
the accuracy of the simplified method, the updated probability
density functions for c, ␾, and ru obtained from this simplified
method are also plotted in Fig. 4. The results obtained from the
simplified method are almost identical to those from the optimi-
zation procedure.
The probabilistic back-analysis also provides dependence in-
formation among the model parameters. Let ␳c,␾, ␳c,ru, and ␳␾,ru
denote the correlation coefficients between c and ␾, c and ru, and
␾ and ru, respectively. In Prior 1, c, ␾, and ru are assumed statis-
tically independent, so the values of ␳c,␾, ␳c,ru, and ␳␾,ru are all
zero in the prior distribution. After the back-analysis, the values
of ␳c,␾, ␳c,ru, and ␳␾,ru become ⫺0.06, 0.69, and 0.20, respec-
tively. The correlations among the variables in the posterior dis-
tribution reflect the slope failure event, which imposes a
constraint of factor of safety equal to unity at the moment of slope
failure. More specifically, in the posterior distribution a higher
pore water pressure should be accompanied by higher shear
strength parameters such that the calculated factor of safety is
close to unity. Since the variance of ␾ is very small, the variation
in shear strength parameters is dominated by the change in cohe-
sion. This makes the correlation coefficient between c and ru
larger than the correlation coefficients among other variables.
Information Allocation in Back-Analysis
In the probabilistic back-analysis methods, the distributions of all
uncertain parameters are modified by the slope failure event si-
multaneously. A logic question that one may raise is: would the
degrees of modification achieved by the back-analysis be the
same for all the parameters? This question is answered by Fig. 4,
which shows that the changes in the probability distributions for
the three parameters are in fact not the same. Among the three
parameters, while the probability distribution of ru is changed
significantly, the prior and modified distributions for ␾ are almost
identical. The degree of improvement for a parameter seems to be
related to its contribution to the uncertainty in factor of safety. To
see this, a simple uncertainty analysis can be performed to calcu-
late the variance of the predicted factor of safety based on the
prior knowledge 共Ang and Tang 2007兲

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010 / 103

J. Geotech. Geoenviron. Eng. 2010.136:99-109.

 0.3 1.5
(a) Prior (a)

Probability density
Optimization method

Factor of safety
0.2 Simplified method 1.0

0.1 0.5

0.0 0.0
0 5 10 15 20 0 3 6 9 12 15
c (kPa) c (kPa)

1.5
0.8 (b)
(b) Prior

Factor of safety
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.

Probability density

0.6 Optimization method 1.0

Simplified method
0.4
0.5
0.2
0.0
0.0 33 35 37 39 41 43
34 36 38 40 42 o
φ ()
φ (ο )

1.5
12 (c)
Prior

Factor of safety
(c)
Probability density

Optimization method 1.0

9
Simplified method
6 0.5

3
0.0
0 -0.10 0.00 0.10 0.20 0.30 0.40
-0.2 0.0 0.2 0.4 0.6 0.8 1.0 ru
ru
Fig. 5. 共a兲 Relationship between factor of safety and c 共␾ = 38°, ru
Fig. 4. Comparison of prior and back-analyzed distributions based = 0.15兲; 共b兲 relationship between factor of safety and ␾ 共c
on Prior 1 共Shek Kip Mei slope兲 = 8 kPa, ru = 0.15兲; and 共c兲 relationship between factor of safety and
ru 共c = 8 kPa, ␾ = 38°兲 共Shek Kip Mei slope兲

var共Fs兲 ⬇ 冏 冏 ⳵ g共␪兲
⳵c
2

␪=␮␪
var共c兲 + 冏 冏
⳵ g共␪兲
⳵␾
2

␪=␮␪
var共␾兲
Fig. 7. For comparison, the prior distributions of c, ␾, and ru are

冏 冏
⳵ g共␪兲 2
+ var共ru兲 + ␴2␧
⳵ ru ␪=␮␪
in which Fs = factor of safety. Through Eq. 共9兲, one can identify
the contribution of uncertainty from each parameter to the vari-
ance of Fs, and the results are presented in Fig. 6. The degrees of
changes in the probability distributions in Fig. 4 for the three
variables are consistent with their contributions to the uncertainty
in Fs as shown in Fig. 6: while the probability distribution of ru is
most updated, the probability distribution of ␾ is almost not
also plotted in this figure. Both the optimization method and sim-
plified method are used to calculate the posterior distributions.
共9兲
The two methods yield practically the same results, as shown in
Fig. 7. Among the three variables, while c and ru are more up-
dated, the distribution of ␾ is almost not changed. Fig. 8 shows
the contribution of uncertainty from each variable to the variance
80
Contribution to variance of Fs (%)

changed.
60
To further verify the above conclusion, the slope failure is
back-analyzed with a different prior distribution on ru, i.e., ru is
normally distributed with a mean of 0.20 and a standard deviation 40
of 0.07. The mean value of ␾ is also increased to 39°. The prior
mean for ␪ in this case is ␮␪ = 兵8 , 39, 0.20其T. The prior covariance
matrix in this case is 20

冤 冥
32 0 0
C␪ = 0 1 2
0 共10兲 0
cc' f'φ ru
ru eε
0 0 0.072
The above prior distribution is called Prior 2. The posterior dis- Fig. 6. Contributions to the variance of factor of safety 共Fs兲 from
tributions of c, ␾, and ru calculated based on Prior 2 are shown in different parameters based on Prior 1 共Shek Kip Mei slope兲

104 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010

J. Geotech. Geoenviron. Eng. 2010.136:99-109.

 0.3 0.3
(a) Prior (a) Prior

Probability density
Probability density
Optimization method Optimization method
0.2 Simplified method 0.2 Simplified method
μc|d = 8.04 kPa μc|d = 7.71 kPa
σc|d = 2.57 kPa σc|d = 2.80 kPa
0.1 (Optimization method) 0.1 (Optimization method)

0.0 0.0
0 5 10 15 20 0 5 10 15 20
c (kPa) c (kPa)

0.8 0.8
(b) Prior (b) Prior
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.

Probability density

Probability density
0.6 μφ|d = 39.00 o Optimization method 0.6 μφ|d = 38.02 o Optimization method
σφ|d = 0.99 o Simplified method σφ|d = 1.00 o Simplified method
0.4 (Optimization method) 0.4 (Optimization method)

0.2 0.2

0.0 0.0
35 37 39 41 43 34 36 38 40 42
ο ο
φ () φ ()

12 12
(c) Prior (c) Prior

Probability density
Probability density

Optimization method 9 Optimization method

9
Simplified method Simplified method
6 μru|d = 0.20 6 μru|d = 0.17
σru|d = 0.05 σru|d = 0.05
3 (Optimization method) 3 (Optimization method)

0 0
-0.2 0.0 0.2 0.4 0.6 0.8 1.0 -0.2 0.0 0.2 0.4 0.6 0.8 1.0
ru ru

Fig. 7. Comparison of prior and back-analyzed distributions based Fig. 9. Comparison of prior and back-analyzed distributions based
on Prior 2 共Shek Kip Mei slope兲 共␮␪i兩d and ␴␪i兩d denote back- on Prior 3 共Shek Kip Mei slope兲
analyzed mean and standard deviation of ␪i, respectively兲

information allocation analysis for an uncorrelated prior distribu-

of Fs. The degree of updating for each variable is also consistent tion in this section still gives useful insights on how information
to its contribution to the variance of Fs. is allocated in the back-analysis.
In Eq. 共9兲 it is assumed that the model parameters are statisti-
cally independent. When model input parameters are correlated, Effect of Correlation between Model Input Parameters
the correlation among the input parameters could also affect the To examine the effect of the correlation between c and ␾ on the
contribution of a parameter to the variance of Fs 共e.g., Ang and back-analysis of a slope failure, Prior 1 is modified by assuming c
Tang 2007兲. In such a case, the individual contribution of a pa- and ␾ are negatively correlated with a coefficient of correlation of
rameter to variance of Fs is not directly obvious. Nevertheless, the ⫺0.50. The resultant prior distribution is called Prior 3 here.
Based on the above assumption, the mean of ␪ in Prior 3 is ␮␪
= 兵8 , 38, 0.15其T, and the covariance matrix of Prior 3 is
80

冤 冥
32 − 1.5 0
Contribution to variance of Fs (%)

60
C␪ = − 1.5 12 0 共11兲
0 0 0.12

40
The slope failure is then back-analyzed again using Prior 3, and a
comparison between prior and posterior distributions of c, ␾, and
ru are shown in Fig. 9. Both the optimization method and simpli-
20 fied method are used to calculate the posterior distribution, and
the two methods again yield practically identical results, as shown
in Fig. 9. Through the back-analysis, the values of ␳c,␾, ␳c,ru, and
0 ␳␾,ru are changed from ⫺0.50, 0, and 0 to ⫺0.51, 0.62, and
cc' f'φ ru
ru eε ⫺0.12, respectively. The negative correlation coefficient between
␾ and ru is caused by the combined effects of the positive corre-
Fig. 8. Contributions to the variance of factor of safety 共Fs兲 from lation between c and ru and the negative correlation between c
different parameters based on Prior 2 共Shek Kip Mei slope兲 and ␾. For comparison, the posterior values of ␳c,␾, ␳c,ru, and ␳␾,ru

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010 / 105

J. Geotech. Geoenviron. Eng. 2010.136:99-109.


20 Silty Water table
sand
Clay 1
Elevation (m)
Slip surface
10
Clay 2
Clay 3
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
0 Based on Christian et al. 共1994兲, the model uncertainty of the
0 10 20
Horizontal distance (m)
Fig. 10. A representative cross section of Congress Street Cut
关adapted from Ireland 共1954兲兴
are ⫺0.06, 0.69, and 0.20, respectively, if the back-analysis is
started with Prior 1. This shows that prior distribution has an
important effect on the back-analyzed distribution. The back-
analyzed distribution is a combination of prior knowledge and
observed slope failure information. It is important to obtain high-
quality prior knowledge on model input parameters before the
back-analysis is carried out.
Example 2
In many cases the purpose of back-analysis is to improve one’s
knowledge on the shear strength parameters. The Shek Kip Mei
landslide example shows that the knowledge on pore water pres-
sure has important effects on the updating of shear strength pa-
rameters. When the knowledge on pore water pressure is poor, the
uncertainty caused by pore pressure may dominate the uncertainty
in Fs. Therefore, the information collected from the back-analysis
is mainly allocated to reduce the uncertainty in pore water pres-
sure rather than to reduce the uncertainty in the shear strength
parameters. However, if pore water pressure uncertainty is small
or pore water pressure is not involved in the back-analysis, the
probability distributions of shear strength parameters can be im-
proved more significantly. To illustrate this, the failure of Con-
gress Street Cut 共Ireland 1954兲 is analyzed as follows.
Congress Street Cut 共Fig. 10兲 was located mainly in saturated
clays and failed in an undrained manner in 1952 during the con-
struction of Congress Street in Chicago 共Ireland 1954兲. The slip
surface passed through four different layers of soil, i.e., a sandy
fill overlaying three saturated clay layers. The cohesion of the
sandy fill was zero 共Oka and Wu 1990兲. In the back-analysis, the
factor of safety of this slope is calculated using Bishop’s simpli-
fied method 共1955兲. Let su1, su2, and su3 denote the average und-
rained shear strengths of the three clay layers, and ␾ denote the
average friction angle of the sandy fill. The parameters to be
back-analyzed in this case are ␪ = 兵su1 , su2 , su3 , ␾其T.
For this slope, sampling was carried out using 51 mm diameter
Shelby tubes with samples taken at least 1 m apart 共Tang et al.
1976兲. Based on reported triaxial tests, the mean values of su1, su2,
and su3 are 43.14, 32.65, and 37.35 kPa, respectively, and the
standard deviations for these variables are 10.22, 5.47, and 6.77
kPa, respectively 共Yucemen et al. 1973; Tang et al. 1976兲. Vari-
ance reduction due to spatial variability was considered in deriv-
106 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010
J. Geotech. Geoenviron. Eng. 2010.136:99-109.
ing the statistics for su1, su2, and su3 共Yucemen et al. 1973; Tang et
al. 1976兲. The mean of ␾ is assumed to be 30° based on Oka and
Wu 共1990兲, and the coefficient of variation of ␾ is assumed to be
0.2. Assuming further su1, su2, su3, and ␾ are statistically indepen-
dent, the prior knowledge on ␪ can be described by a multivariate
normal distribution with a mean of ␮␪ = 兵43.14, 32.65,
37.35, 30其T and a covariance matrix of
冤 冥
10.222 0 0 0
0 5.472 0 0
C␪ = 2 共12兲
0 0 6.77 0
0 0 0 6.002
30 method of Bishop 共1955兲 is modeled as a normal random variable
with a mean of 0.05 and a standard deviation of 0.07, i.e., ␮␧
= 0.05 and ␴␧ = 0.07.
Sensitivity analysis shows that the slope stability model in this
case is approximately linear, so the simplified method is appli-
cable. Fig. 11 compares the prior and back-analyzed distributions
for su1, su2, su3, and ␾. The distribution of su3 is modified most
significantly by the observed slope failure. This example illus-
trates that when pore water pressure uncertainty is not involved in
the back-analysis, the probability distributions of shear strength
parameters may be improved more significantly. The distributions
of other three parameters have also been improved by the back-
analysis, but the degrees of improvement for them are not as
significant as that for su3. This is because the uncertainty in su3
has contributed most to the uncertainty in the factor of safety, as
shown in Fig. 12.
Fig. 11 shows the marginal posterior distributions of the
strength parameters. The posterior correlation matrix ␳␪兩d, which
reflects the dependence relationships among the four strength pa-
rameters, is
冤 冥
1 − 0.10 − 0.39 − 0.00
− 0.10 1 − 0.37 − 0.00
␳␪兩d = 共13兲
− 0.39 − 0.37 1 − 0.01
− 0.00 − 0.00 − 0.01 1
It shows that the strength parameters are negatively correlated as
expected to satisfy the constraint that the calculated factor of
safety based on the combination of posterior parameters should be
close to unity as a reflection of the slope failure event. Among the
four uncertain model input parameters, the correlation between
su1 and su3 as well as the correlation between su2 and su3 are
relatively stronger. The marginal distributions of su1, su2, and su3
are more updated so these variables have more freedom to adjust
their correlation coefficients.
Summary and Conclusions
The research reported in this paper and findings from the paper
are summarized as follows:
1. Back-analysis of slope failure is often carried out to improve
one’s knowledge on the parameters of a slope stability
model. To back-analyze multiple sets of slope stability pa-
rameters simultaneously under uncertainty, the back-analysis
can be implemented in a probabilistic way, in which uncer-
tain parameters are modeled as random variables, and their
distributions are improved using observed slope failure infor-
mation.
 0.5 80
(a) Prior

Contribution to variance of Fs (%)

Probability density
0.4
Back analysis
μsu1|d = 40.38 kPa 60
0.3
σsu1|d = 9.74 kPa
0.2
40
0.1

0 20
0 20 40 60 80
s u1 (kPa)
0
c1
s u1 sc2u2 sc3
u3 φf εm

0.5
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.

(b) Fig. 12. Contributions to the variance of factor of safety 共Fs兲 from
Prior
different parameters 共Congress Street Cut兲
Probability density

0.4
Back analysis
0.3 μsu2|d = 31.26 kPa
0.2 σsu2|d = 5.24 kPa failure caused by rainfall infiltration. The stability model of
this case is largely linear, and the back-analysis results from
0.1 the two methods are indeed similar.
0
4. Case studies in this paper show that the degree of change in
10 30 50 70 the probability distribution of a parameter achieved by the
s u2 (kPa) back-analysis depends on its contribution to the uncertainty
in the factor of safety: the probability distribution of the pa-
rameter that contributes the greatest uncertainty to the factor
0.5 of safety will be changed most. This conclusion is valid if the
(c)
Prior parameters to be updated are not correlated in the prior dis-
Probability density

0.4
Back analysis
tribution. For a slope failure in which the knowledge in pore
0.3
μsu3|d = 32.67 kPa water pressure is poor, the uncertainty in pore water pressure
0.2 σsu3|d = 4.27 kPa may dominate the back-analysis. In such a case, the knowl-
edge on shear strength parameters may not be effectively
0.1 improved by probabilistic back-analysis.
0
10 30 50 70
s u3 (kPa) Acknowledgements

This research was substantially supported by the Research Grants

0.5 Council 共RGC兲 of the Hong Kong SAR 共Project Nos. 620206 and
(d) Prior
6294/04E兲.
Probability density

0.4
Back analysis
0.3 μφ|d = 29.95 o
σφ|d = 6.00 o Appendix I. Theory of Probabilistic Back-Analysis
0.2
for a General Slope Stability Model
0.1

0 Basic Theory
0 20
φ (ο)
Fig. 11. Comparison of prior and back-analyzed distributions
共Congress Street Cut兲
2. Two methods are presented for probabilistic back-analysis of
slope failure, i.e., a method based on optimization for a gen-
eral slope stability model, and a simplified method based on
sensitivity analysis. While the simplified method is easier to
apply in practice, it is only applicable when the slope stabil-
ity model is approximately linear.
3. For efficient analysis, a spreadsheet template has been devel-
oped to implement the back-analysis method based on opti-
mization. Also, the simplified method involves simply a
sensitivity analysis together with a conventional slope stabil-
40 60
The general probabilistic back-analysis theory is first introduced
here. Let g共␪ , r兲 be a general model to predict the response of a
system, where ␪ and r are vectors denoting unknown and known
parameters, respectively. In the following text, r is dropped from
the prediction model for simplicity. Let d be a vector representing
the actual system response. The objective of the probabilistic
back-analysis is to improve the probability distribution of ␪ based
on some observed system response dobs through the prediction
model g共␪兲.
Before the probabilistic back-analysis, one may already have
some prior knowledge on ␪ from various sources. Assume the
prior knowledge on ␪ can be denoted by a multivariate normal
distribution
冋
1
f共␪兲 = k1 exp − 共␪ − ␮␪兲TC␪−1共␪ − ␮␪兲
2
册 共14兲

ity program. The results from the optimization method and where k1 = normalization constant to make the probability density
the simplified method are compared in a case study of slope function valid 关i.e., making the integration of probability density

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010 / 107

J. Geotech. Geoenviron. Eng. 2010.136:99-109.

 function in Eq. 共14兲 over the entire domain of ␪ equal to unity兴. Extension to a Biased Prediction Model
Due to the existence of observational uncertainty, the observed
In the literature, prediction models are often assumed unbiased. In
data dobs may not be the same as the actual system response d.
geotechnical engineering, however, prediction models could be
Assuming that the observational uncertainty can be described by a
biased. For example, a two-dimensional slope stability model
multivariate normal distribution with a mean vector of zero and a
might be biased towards the conservative side as it neglects three-
covariance matrix of Cd, the probability density function of d
dimensional effects 共Christian et al. 1994兲. Assume that the model
given dobs is
uncertainty of a biased model follows a multivariate normal dis-

冋 册
tribution with a mean of ␮m and a covariance matrix of Cm. The
1 magnitude of ␮m measures the degree of bias in the model pre-
f共d兩dobs兲 = k2 exp − 共d − dobs兲TCd−1共d − dobs兲 共15兲
2 diction. For example, ␮m = 0 means the model is unbiased, and in
this case the model uncertainty is described by Eq. 共16兲. In the
where k2 is another normalization constant to make the probabil-
case where ␮m ⫽ 0, although the prediction from g共␪兲 is biased,
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.

ity density function valid.

the prediction from g⬘共␪兲 = g共␪兲 + ␮m is unbiased. Therefore, when
As any prediction model is only an abstraction of the real
g共␪兲 is biased, the above formulations also apply if g共␪兲 is sub-
world, model uncertainty always exists. The existence of model
stituted with g⬘共␪兲, i.e., g共␪兲 + ␮m. For example, the misfit func-
uncertainty makes the predicted response g共␪兲 different from the
tion for a biased model is
actual response d even when ␪ is exactly known. Assuming the
prediction model is unbiased and that the model uncertainty can
be described by a multivariate normal distribution with a mean 2S共␪兲 = 关g共␪兲 + ␮m − dobs兴TCD−1关g共␪兲 + ␮m − dobs兴
vector of zero and a covariance matrix of Cm, the probability + 共␪ − ␮␪兲TC␪−1共␪ − ␮␪兲 共22兲
density function of d given ␪ is

再 1
f共d兩␪兲 = k3 exp − 关d − g共␪兲兴TCm−1关d − g共␪兲兴
2
冎 共16兲 Application to Back-Analysis of Slope Failure
In back-analysis of slope failure, the observation data is that the
where k3 is again a normalization constant to make the probability factor of safety at the moment of slope failure is unity, which is a
density function valid. scalar. In such a case, the model uncertainty, the observation data,
With the above assumptions, the improved knowledge of ␪ and the observational uncertainty are all scalars. Let ␴d2 , ␮␧, and
considering both f共␪兲 and dobs can be described by its posterior ␴2␧ denote Cd, ␮m, and Cm in the back-analysis of slope failure,
probability distribution 共Tarantola 2005兲 respectively, where ␴d and ␴␧ are the standard deviations of ob-
servation and model uncertainty variables, respectively. Note that
f共␪兩dobs兲 = k4 exp兵关g共␪兲 − dobs兴TCD−1关g共␪兲 − dobs兴 in back-analysis of slope failure dobs = 1. Assume the observa-
+ 共␪ − ␮␪兲TC␪−1共␪ − ␮␪兲其 共17兲 tional uncertainty is negligible, ␴d = 0. In such a case, Eqs. 共22兲,
共19兲, and 共20兲 become Eqs. 共2兲–共4兲, respectively.
where k4 = normalization constant to make the probability density
function valid; f共␪ 兩 dobs兲 = posterior distribution of ␪ given dobs;
␮␪ = prior means vector of ␪; C␪ = prior covariance matrix of ␪
Appendix II. Probabilistic Back-Analysis Theory for
and CD = Cd + Cm.
an Approximately Linear Slope Stability Model
The posterior density function of ␪ may not have a closed-
form expression except when g共␪兲 is linear. Let ␪ⴱ be the point
For a general linear prediction model g共␪兲 = H␪, the posterior
where f共␪ 兩 dobs兲 is maximized. If the prediction model g共␪兲 is
mean and covariance of ␪ can be determined using the following
nonlinear but is linearizable in the neighborhood of ␪ⴱ, f共␪ 兩 dobs兲
relationships without resorting to the optimization procedure
can be approximated by a multivariate normal distribution with a
共Tarantola 2005兲
mean of ␮␪兩d and a covariance matrix of C␪兩d, where ␮␪兩d and C␪兩d
are 共Tarantola 2005兲
␮␪兩d = ␮␪ + C␪HT共HC␪HT + CD兲−1共dobs − H␮␪兲 共23兲
ⴱ
␮␪兩d = ␪ 共18兲

C␪兩d = 共HTCD−1H + C␪−1兲−1 共24兲

C␪兩d = 共GTCD−1G + C␪−1兲−1 共19兲
If g共␪兲 = H␪ is biased with the mean model uncertainty of ␮m, one
G=
⳵␪
冏 冏
⳵ g共␪兲
␪=␪*
共20兲
can view H␪ + ␮m as a unbiased prediction model, and the poste-
rior mean ␮␪兩d can then be determined as follows:

␪ⴱ can be found by maximizing Eq. 共17兲, or equivalently, mini-
mizing the misfit function 2S共␪兲, which is defined as follows
共Tarantola 2005兲:
2S共␪兲 = 关g共␪兲 − dobs兴TCD−1关g共␪兲 − dobs兴 + 共␪ − ␮␪兲TC␪−1共␪ − ␮␪兲
共21兲
The general probabilistic back-analysis method described above
is also known as theory of system identification 共e.g., Schweppe
1973兲.
␮␪兩d = ␮␪ + C␪HT共HC␪HT + CD兲−1共dobs − H␮␪ − ␮m兲 共25兲
If the slope stability model is largely linear, Eqs. 共25兲 and 共24兲
may be used to determine the improved mean and covariance
matrix of ␪ approximately, respectively. As mentioned before, in
the back-analysis of slope failure, the observed information is a
scalar. The scalar form of Eqs. 共25兲 and 共24兲 are Eqs. 共5兲 and 共6兲,
respectively. The accuracy of this simplified method has been
checked in the case study.

108 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010

J. Geotech. Geoenviron. Eng. 2010.136:99-109.

 References
Ang, A. H.-S., and Tang, W. H. 共2007兲. Probability concepts in engineer-
ing: emphasis on applications to civil and environmental engineering,
2nd Ed., Vol. 1, Wiley, New York.
Bishop, A. W. 共1955兲. “The use of the slip circle in the stability analysis
of slopes.” Geotechnique, 5共1兲, 7–17.
Chen, Z., and Morgenstern, N. R. 共1983兲. “Extensions to the generalized
method of slices for stability analysis.” Can. Geotech. J., 20共1兲, 104–
119.
Cherubini, C. 共2000兲. “Reliability evaluation of shallow foundation bear-
ing capacity on c⬘, ␾⬘ soils.” Can. Geotech. J., 37共1兲, 264–269.
Chowdhury, R., Zhang, S., and Flentje, P. 共2004兲. “Reliability updating
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
and geotechnical back-analysis.” Proc., Advances in Geotechnical En-
gineering: The Skempton Conf., R. J. Jardine, D. M. Potts, and K. G.
Higgins, eds., Thomas Telford, London.
Christian, J., Ladd, C., and Baecher, G. 共1994兲. “Reliability applied to
slope stability analysis.” J. Geotech. Engrg., 120共12兲, 2180–2207.
Der Kiureghian, A., and Liu, P. L. 共1986兲. “Structural reliability under
incomplete probability information.” J. Engrg. Mech., 112共1兲, 85–
104.
El-Ramly, H., Morgenstern, N. R., and Cruden, D. M. 共2002兲. “Probabi-
listic slope stability analysis for practice.” Can. Geotech. J., 39共3兲,
665–683.
El-Ramly, H., Morgenstern, N. R., and Cruden, D. M. 共2005兲. “Probabi-
listic assessment of stability of a cut slope in residual soil.” Geotech-
nique, 55共1兲, 77–84.
FMSW. 共2000兲. “Report on the Shek Kip Mei landslide of 25 August
1999.” Rep. Prepared by Fugro Maunsell Scott Wilson Joint Venture
for the Geotechnical Engineering Office, Government of Hong Kong,
Hong Kong.
Geotechnical Engineering Office 共GEO兲. 共2000兲. Guide to rock and soil
description, Government of Hong Kong, Hong Kong.
Gilbert, R. B., Wright, S. G., and Liedtke, E. 共1998兲. “Uncertainty in
back-analysis of slopes: Kettleman Hills case history.” J. Geotech.
Geoenviron. Eng., 124共12兲, 1167–1176.
Ireland, H. O. 共1954兲. “Stability analysis of Congress Street open cut in
Chicago.” Geotechnique, 4共4兲, 163–168.
Janbu, N. 共1987兲. “Slope stability computation.” Embankment-dam engi-
neering: Casagrande volume, R. C. Hirschfeld and S. J. Poulos, eds.,
Kriege, New York, 47–86.
JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JANUARY 2010 / 109
J. Geotech. Geoenviron. Eng. 2010.136:99-109.
Low, B. K. 共2003兲. “Practical probabilistic slope stability analysis.”
Proc., 12th Pan-American Conf. on Soil Mechanics and Geotechnical
Engineering and 39th U.S. Rock Mechanics Symp., MIT, Cambridge,
Mass., 2, 2777–2784.
Low, B. K., Gilbert, R. B., and Wright, S. G. 共1998兲. “Slope reliability
analysis using generalized method of slices.” J. Geotech. Geoenviron.
Eng., 124共4兲, 350–362.
Low, B. K., and Tang, W. H. 共1997兲. “Probabilistic slope analysis using
Janbu’s generalized procedure of slices.” Comput. Geotech., 21共2兲,
121–142.
Luckman, P. G., Der Kiureghian, A., and Sitar, N. 共1987兲. “Use of sto-
chastic stability analysis for Bayesian back calculation of pore pres-
sures acting in a cut at failure.” Proc., 5th Int. Conf. on Application of
Statistics and Probability in Soil and Struct. Engr., University of Brit-
ish Columbia, Vancouver, Canada.
Mostyn, G. R., and Li, K. S. 共1993兲. “Probabilistic slope stability
analysis—State-of-play.” Probabilistic methods in geotechnical engi-
neering, K. S. Li and S.-C. R. Lo, eds., Balkema, Rotterdam, The
Netherlands.
Oka, Y., and Wu, T. H. 共1990兲. “System reliability of slope stability.” J.
Geotech. Engrg., 116共8兲, 1185–1189.
Schweppe, F. C. 共1973兲. Uncertain dynamic systems, Prentice-Hall,
Englewood Cliffs, N.J.
Spencer, E. 共1967兲. “A method of analysis of embankments assuming
parallel inter-slice forces.” Geotechnique, 17共1兲, 11–26.
Tang, W. H., Yucemen, M. S., and Ang, A. H.-S. 共1976兲. “Probability
based short-term design of soil slopes.” Can. Geotech. J., 13共3兲, 201–
215.
Tarantola, A. 共2005兲. Inverse problem theory and methods for model pa-
rameter estimation, 2nd Ed., Elsevier Science, New York.
Tiwari, B., Brandon, T. L., Marui, H., and Tuladhar, G. R. 共2005兲. “Com-
parison of residual shear strengths from back-analysis and ring shear
tests on undisturbed and remolded specimens.” J. Geotech. Geoenvi-
ron. Eng., 131共9兲, 1071–1079.
Wesley, L. D., and Leelaratnam, V. 共2001兲. “Shear strength parameters
from back-analysis of single slips.” Geotechnique, 51共4兲, 373–374.
Yucemen, M. S., Tang, W. H., and Ang, A. H.-S. 共1973兲. “A probabilistic
study of safety and design of earth slopes.” Structural Research Series
No. 402, Dept. of Civil Engineering, University of Illinois at Urbana,
Ill.