Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Engineering Geology 288 (2021) 106144

Contents lists available at ScienceDirect

Engineering Geology
journal homepage: www.elsevier.com/locate/enggeo

Reliability analysis of an existing slope at a specific site considering rainfall


triggering mechanism and its past performance records
Xin Liu , Yu Wang *
Department of Architecture and Civil Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong, PR China

A R T I C L E I N F O A B S T R A C T

Keywords: Accurate estimation of landslide probability or slope failure probability when there is a rainstorm event is crucial
Slope reliability for assessment and mitigation of rainfall-induced landslide risk. Slope reliability methods provide a rigorous way
Rainfall of estimating landslide probability based on slope failure mechanism and probability theory. However, it is well
Uncertainties
recognized in literature that the landslide probability estimated from existing slope reliability analysis methods
Bayesian updating
Digital twin
are often much larger than the observed landslide frequency. To improve accuracy of the slope failure proba­
bility, this study proposes a slope reliability analysis method for an existing slope at a specific site, which
considers both rainfall triggering mechanism and the slope’s performance records during previous rainfall
events. It was found that the fact that the slope survived from previous rainfall events could be utilized to
effectively reduce uncertainties in soil parameters and might reduce the estimated landslide probability by one to
two orders of magnitudes. In addition, the proposed method provides a real-time slope failure probability for a
given rainfall event, and the estimated slope failure probability varies as the considered rainfall event evolves
with time.

1. Introduction mechanism was often ignored in the previous reliability studies of slope
stability. Therefore, the estimated Pf often does not reflect magnitudes of
Rainfall-induced landslides are major geohazards in tropical and different rainfall events. Even when rainfall infiltration is considered,
subtropical zones (e.g., Tang et al., 2019; Gong et al., 2021). With the slope reliability analysis often produces an overestimated slope failure
rapid development of digital technologies, digital twin of slopes is probability. Christian and Baecher (2011) summarized ten unresolved
emerging as a promising tool for life-cycle risk management of landslide problems in geotechnical reliability analysis that hinders its wide ap­
hazards (e.g., Zhang et al., 2019; Cheverda et al., 2020). A digital twin of plications. The top challenge in the list is “why are failures less frequent
a slope is a virtual slope model that can be continuously improved by than our reliability studies predict?” The actual landslide probability is
slope performance records and monitoring data obtained from its often relatively small. For example, the Geotechnical Engineering Office
physical counterpart in reality (e.g., Grieves, 2015). It is valuable for (GEO) in Hong Kong reported that the annual failure rate in 2016 is
evaluating actual slope performance and providing key information for 0.07% and 0.527% for engineered and non-engineered slopes, respec­
decision-making and mitigation of landslide risk, such as landslide early tively (Wai et al., 2018). It was found that the estimated Pf from previous
warning (e.g., Baum and Godt, 2010; Chen and Zhang, 2014; Segoni reliability analysis are “at least an order of magnitude larger than the
et al., 2015; Chae et al., 2017; Kong et al., 2020). observed frequency”. One possible reason for such a difference might be
An important task for a digital twin of slope is estimation of slope attributed to huge uncertainties in soil parameters often observed from
failure probability (Pf) when there is a rainstorm scenario. Slope reli­ site investigation data, even when there are many measured data
ability analysis methods provide a rigorous way of estimating Pf for available (e.g., Tofani et al., 2017). To address this issue, existing in­
existing slopes based on slope failure mechanism and probability theory formation can be utilized to improve uncertainty quantification and
(e.g., Li and Lumb, 1987; Baecher and Christian, 2003). Various factors prediction of Pf by slope reliability updating (e.g., Zhang et al., 2011;
and uncertainties can be properly incorporated in slope reliability Papaioannou and Straub, 2012; Zhang et al., 2014; Schweckendiek
analysis in a quantitative manner. However, rainfall triggering et al., 2014; Li et al., 2015; Jiang et al., 2020). For example, Zhang et al.

* Corresponding author.
E-mail addresses: xliu268-c@my.cityu.edu.hk (X. Liu), yuwang@cityu.edu.hk (Y. Wang).

https://doi.org/10.1016/j.enggeo.2021.106144
Received 18 December 2020; Received in revised form 14 March 2021; Accepted 17 April 2021
Available online 21 April 2021
0013-7952/© 2021 Elsevier B.V. All rights reserved.
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

(2011) demonstrated that an existing slope in natural conditions can be The survival records contain at least two pieces of valuable infor­
deemed as a full-scale experiment under various triggering events, and mation: (1) A slope existing at a specific site is stable before a future
its performance that it failed or survived from a past triggering event can rainfall event occurs; and (2) the slope stays stable after multiple,
be used as valuable information for updating Pf and improving its ac­ possibly many, previous rainfall events with different rainfall intensities,
curacy. Such performance records are also readily available in durations, and patterns (e.g., hourly rainfall pattern). Although a few
geotechnical practice. However, such performance records have rarely studies have considered pervious performance records (e.g., Zhang
been utilized in estimating failure probabilities of rainfall-induced et al., 2014), only one artificial rainfall event was considered. In addi­
landslides, especially the survival records from multiple past rainfall tion, many reliability studies of slope stability ignore the first piece of
events with different rainfall intensities, durations, and patterns. information as triggering mechanisms are seldom considered in deter­
Furthermore, combinations of weak soil strength parameters may be ministic slope stability analysis. The proposed method incorporates both
generated when random variables or random fields are used to represent pieces of valuable information abovementioned for updating un­
soil parameters in slope reliability analysis, and they result in slope certainties in a slope existing at a specific site and performing slope
failure before imposing any triggering factor, especially when the vari­ reliability analysis to produce a real-time slope failure probability under
ability of soil strength parameters is large. This inevitably leads to a target rainfall scenario.
overestimation of slope failure probability. The estimation of slope Fig. 1 schematically illustrates a framework of the proposed method.
failure probability can be enhanced by utilizing the fact that an existing It starts with deterministic modeling and preliminary uncertainty
slope at a specific site is stable before imposing any triggering factor and modeling for an existing slope, which are similar to those modules in
after previous rainfall events. existing slope reliability analysis methods (e.g., Liu et al., 2019). Then,
This study aims to propose a slope reliability analysis method for the fact that the slope is stable before imposing any rainfall event is
estimating time-variant Pf of a slope existing at a specific site under a utilized to obtain posterior distributions of uncertain parameters, which
specific target rainfall scenario, such as a time-variant rainfall event. To is referred to as initial site-specific uncertainty updating in this study.
improve estimation of Pf, the proposed method explicitly considers the When there are available past rainfall records from which the slope has
fact that an existing slope at a specific site is stable before imposing any survived, details of the past rainfall events (e.g., rainfall intensity,
triggering factor, and it incorporates survival records from multiple past duration and pattern) are used for subsequent site-specific uncertainty
rainfall events for improving quantification of uncertainties in soils and updating. Based on the updated site-specific uncertainties, a scenario-
prediction of Pf. The proposed method will be introduced in the next based slope reliability analysis is performed to assess the slope failure
section, followed by its implementation procedure in Section 3. Then, it probability induced by a target rainfall scenario. The scenario-based
is illustrated using an infinite slope example under rainfall infiltration reliability analysis results offer key information required for landslide
considering its past performance records in Section 4. Section 5 sum­ risk assessment and mitigation. The following subsections will respec­
marizes major conclusions drawn from this study. tively introduce the modules of deterministic modeling, preliminary
uncertainty modeling, site-specific uncertainty updating, and scenario-
2. The proposed method based slope reliability analysis.

This study focuses on rainfall-induced landslides, a topic frequently


encountered in the field of engineering geology. For a slope existing at a 2.1. Deterministic modeling
specific site, it is stable before occurrence of any triggering factor. Such
an existing stable slope has experienced and survived from full-scale A slope stability analysis model is developed for an existing slope in
experiments under multiple previous rainfall events (e.g., Zhang et al., the deterministic modeling module. The slope geometry information can
2011, 2014). Such survival records provide valuable information for be obtained from site investigation and digital elevation maps. The latter
updating uncertainties of site-specific variables (e.g., Schweckendiek is especially useful for many inaccessible slopes in mountainous areas.
and Kanning, 2009; Zhang et al., 2011; Schweckendiek et al., 2014; This study adopts a commonly used two-stage approach for slope sta­
Chowdhury et al., 2015). The survival records are free and readily bility analysis under rainfall infiltration (e.g., Zhang et al., 2016). In the
accessible since rainfall data have been widely recorded in many first stage, transient seepage analysis is performed to obtain pore water
regions. pressure profiles under rainfall infiltration. The seepage analysis can be
solved analytically or numerically using commercial or free software.

Deterministic modeling of Available records of past


Preliminary uncertainty modeling
an existing slope rainfall events

Initial site-specific uncertainty


updating considering that the slope
stays stable initially before
imposing any rainfall event

Subsequent site-specific uncertainty


updating considering slope survival
records from past rainfall events

Scenario-based slope reliability Impose a target rainfall


analysis scenario

Prediction of slope failure


probability

Fig. 1. A framework of the proposed method.

2
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

For example, this study adopts a public domain software called [ ( )p ]2


Hydrus1D (Šimůnek et al., 2013) to perform seepage analysis under k = ks Se0.5 1 − 1 − Se1/p (4)
rainfall infiltration. In the second stage, the pore water pressure profile
is exported from the software and used for slope stability analysis to where k is the hydraulic conductivity of unsaturated soils; ks is the
determine whether or not slope failure occurs. saturated hydraulic conductivity.
Since rainfall-induced landslides are often shallow failures (e.g.,
Tang and Zhang, 2011; Zhang et al., 2014), an infinite slope model is 2.2. Preliminary uncertainty modeling
adopted for slope stability analysis for illustration purpose, and it might
be only suitable for distributed basin-scale analysis. If a specific slope is Uncertainties are unavoidable in slope engineering. In geotechnical
given, the actual slope configuration needs to be represented by an practice, site investigation data for a slope are often limited, and
advanced slope model with important elements, such as root cohesion, therefore knowledge about soil parameters (e.g., cohesion, friction
evapotranspiration, hysteresis cycles of unsaturated soils, development angle, and unsaturated soil properties) of the slope are often limited too.
of fracture, and two-dimensional or three-dimensional geometry and The uncertainty in soil properties has been deemed as a major uncer­
infiltration, etc. The proposed method is also applicable to complex tainty (e.g., Phoon and Kulhawy, 1999a&1999b), which significantly
slope models. Fig. 2 shows an infinite slope under rainfall infiltration. Its affects slope failure (e.g., Li et al., 2016; Liu et al., 2019). Various un­
factor of safety (i.e., FS) is calculated using limit equilibrium methods certainties can be reasonably considered in slope reliability analysis, and
(LEM) that have been widely adopted in geotechnical engineering they are often treated as uncertain variables and modelled using random
practice (e.g., Duncan and Wright, 2005). Since LEM consider only static variable or random field theory (e.g., Fenton and Vanmarcke, 1990). For
force and/or moment equilibrium, the slope deformation and displace­ example, the soil cohesion is often modelled as a lognormal random
ment constraints are not considered. According to LEM (e.g., Duncan variable with a prescribed mean and standard deviation (i.e., STD). The
and Wright, 2005; Lu and Godt, 2008), the factor of safety of a slip statistics can be determined by site investigation data, along with data
surface at depth z is calculated as from similar sites and engineering judgement (e.g., Wang et al., 2010;
′ ′ Wang and Cao, 2013). More site investigation data are preferred to
c + (Wcos2 β − σs )tanϕ
FSLEM = (1) reasonably characterize soil properties and enhance the prediction of
Wsinβcosβ
landslide failure probability (e.g., Tofani et al., 2017). The uncertainty
where c’ is the effective soil cohesion; ϕ’ is the effective soil friction modeling abovementioned is preliminary, and it will be further updated
angle; β is the slope angle; W is the total weight of soil at the depth z, i.e., with performance records of the slope, as described in the following

W = z0(γd + θ(z)γw)dz; γd is the dry unit weight of soil; θ(z) is the subsection.
volumetric water content along the depth; σ s is the suction stress as
calculated using Lu and Likos (2006), which is given by 2.3. Site-specific uncertainty updating
{
− (ua − uw ) ua − uw ≤ 0 2.3.1. Bayesian sequential updating scheme
σs = (2)
Se (ua − uw ) ua − uw > 0 As introduced before, an existing slope stays stable after many pre­
vious rainfall events, and it is stable before a future rainfall occurs. Such
where ua is the pore air pressure; uw is the pore water pressure; ua - uw is stable or survival records are used in this study to update probability
the matric suction, and it is negative in saturated soils and positive in distributions of uncertain parameters (e.g., soil parameters) obtained
unsaturated soils; Se is the effective degree of saturation in unsaturated from the preliminary uncertainty modeling. In geotechnical practice, FS
soils, and it can be obtained from soil water characteristic curve (SWCC). of a slope has been widely used as an index of stability. A slope is deemed
For example, the van Genuchten model (van Genuchten, 1980) is stable if its actual FS is large than 1, otherwise it fails. Nonetheless, the
expressed as calculated FS (e.g., FSLEM calculated using Eq. (1) based on limit equi­
θ − θr 1 librium methods) inevitably involves approximation and uncertainties,
Se = = p (3) which affect accuracy of the calculated FS and need to be considered in
θs − θr {1 + [α(ua − uw ) ]n }
evaluating slope performance. For example, Zhang et al. (2009) sepa­
where θr is the residual water content; θs is the saturated water content; rated the model uncertainty of limit equilibrium methods from param­
α, n, and p are SWCC parameters, p = 1–1/n. eter uncertainty based on a series of slope centrifuge test data. The
For unsaturated soils, the hydraulic conductivity varies with degree actual FS of a given slope considering an explicit model uncertainty may
of saturation, which can be represented by the Mualem’s model (Mua­ be defined as
lem, 1976) as FS = FSLEM + ε (5)

where FSLEM is the calculated FS using Eq. (1) based on limit equilibrium
Rainfall methods; ε is the model uncertainty or difference between the actual and
intensity q calculated FS, which may follow a normal distribution as reported in the
previous studies (e.g., Zhang et al., 2009; Zhang et al., 2011).
The fact that a slope existing at a specific site is stable before
imposing any rainfall is firstly used to update probability distributions of
depth z
uncertain parameters (e.g., soil parameters including cohesion, friction
angle, and saturated hydraulic conductivity) under a Bayesian updating
framework (e.g., Cao et al., 2016; Wang et al., 2016). Let fp(x) and f0(x)
W a slip surface respectively denote the prior and posterior probability density function
(PDF) of an uncertain parameter x, then f0(x) is given by
T
P(FS0 > 1|x )fp (x)
f0 (x) = f (x|FS0 > 1 ) = ∫ (6)
N P(FS0 > 1|x )fp (x)dx
slope angle β
where f0(x) is the posterior PDF of x given that the slope is initially stable
Fig. 2. Illustration of an infinite slope example under rainfall infiltration. before imposing any rainfall; fp(x) is the prior PDF of x obtained from

3
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

preliminary uncertainty modeling based on site investigation data; x where the prior distribution is f0(x), instead of fm-1(x) as in Eq. (9).
denotes the concerned uncertain parameter, such as soil cohesion and
friction angle; FS0 is the actual FS before imposing any rainfall; P(FS0 > 2.3.2. Solution of Bayesian updating by Monte Carlo simulation and
1|x) is the conditional probability of {FS0 > 1} given a x value, and its sample reassembling
estimation will be discussed later. A key issue for solving posterior distributions f0(x) and fm(x) in Eqs.
When there are available records of past rainfall events from which (6–10) is to estimate P(FS0 > 1|x) and P(FSm > 1|x). They can be
the slope has survived, the survival records can be used to further update calculated using existing reliability methods, such as first-order reli­
probability distributions of uncertain parameters (e.g., soil parameters) ability method (FORM) (e.g., Low et al., 2011), response surface method
(e.g., Zhang et al., 2011; Papaioannou and Straub, 2012; Zhang et al., (RSM) (e.g., Li et al., 2011; Ji and Low, 2012; Jiang et al., 2015),
2014). Let fm(x) denote the posterior PDF of uncertain variable x importance sampling (IS) (e.g., Ching et al., 2009), direct Monte Carlo
considering survival records from m (m ≥ 1) past rainfall events and FSi simulation (MCS) (e.g., Liu et al., 2020), and subset simulation (SS) (e.g.,
denote the FS after the i-th past rainfall events, then f1(x), f2(x), and fm(x) Au and Beck, 2001; Wang et al., 2011; Au and Wang, 2014; Li et al.,
can be obtained using a Bayesian sequential updating approach (e.g., 2016). This study adopts direct MCS for estimating probabilities due to
Cao et al., 2016) as its conceptual simplicity. To implement a direct MCS, the first step is to
P(FS1 > 1|x)f0 (x) generate N random samples of uncertain vector x with all uncertain
f1 (x) = f (x|FS0 > 1 ∩ FS1 > 1 ) = ∫ (7) variables from their prior distributions fp(x), leading to an initial sample
P(FS1 > 1|x)f0 (x)dx
set S = {x1, x2, …, xN}. Then, these N random samples are used to
P(FS2 > 1|x )f1 (x) perform repetitively deterministic slope stability analysis (e.g., using the
f2 (x) = f (x|FS0 > 1 ∩ FS1 > 1 ∩ FS2 > 1 ) = ∫ (8) model described in subsection 2.2) and estimate P(FS0 > 1|x) in Eq. (6)
P(FS2 > 1|x )f1 (x)dx
as


N
I(FS0 (xi ) > 1 and x − Δx < x < x + Δx|xi ∈ S )
P(FS0 > 1|x ) ≈ P(FS0 > 1|x − Δx < x < x + Δx, xi ∈ S ) = i=1 (11)

N
I(x − Δx < x < x + Δx|xi ∈ S )
i=1


P(FSm > 1|x )fm− 1 (x)
fm (x) = f (x|FS0 > 1 ∩ FS1 > 1 ∩ … ∩ FSm > 1 ) = ∫ where ∆x is a small interval of x; xi is the i-th random sample of uncertain
P(FSm > 1|x )fm− 1 (x)dx
vector x in the sample set S; I(⋅) is an indicator function, whose value is
(9) equal to 1 when the expression in the bracket is true, otherwise it is 0.
Since the PDF in Eq. (7) is f0(x) instead of fp(x), the N random
where P(FSm > 1|x) is the conditional probability of {FSm > 1} for a
samples in S generated from fp(x) may not be directly used to estimate P
given x value.
(FS1 > 1|x) in Eq. (7). Conventionally, generation of another set of
Note that not all uncertain variables can be updated sequentially
random samples satisfying {FS0 > 1} from f0(x) using, e.g., Markov
using different past rainfall events. The uncertain variables can be
Chain Monte Carlo simulation (MCMCS) (e.g., Au and Wang, 2014), is
divided into two categories, i.e., time-invariant variables and time-
needed for estimating P(FS1 > 1|x) in MCS. Therefore, extensive
variant variables (e.g., Schweckendiek et al., 2014). Time-invariant
computational costs are required for re-generating random samples and
variables stay almost unchanged during the period of rainfall events
determining whether each generated random sample satisfies {FS0 > 1}.
and they can be updated sequentially. Soil properties, including cohe­
To improve the computational efficiency, a MCS sample reassembling
sion, friction angle, and saturated hydraulic conductivity, are often
method (Wang, 2012) is adopted in this study to make use of existing
considered as time-invariant variables (e.g., Zhang et al., 2011; Zhang
random samples in S, instead of repetitively generating new random
et al., 2013), when there are no large deformation and structure changes
samples from the updated PDF (i.e., f0(x), f1(x), …, and fm(x)).
of soils during the rainfall period. On the other hand, time-variant var­
Note that the N0 conditional random samples satisfying {FS0 > 1}
iables vary with time and may be different before each rainfall event,
constitute a sample set S0 = {xi | FS0(xi) > 1} that is a subset of the
such as soil moisture conditions and the initial matric suction that are
original random sample set S, and the remaining N–N0 samples
affected by vegetation with roots and canopy. The proposed method is
constitute a remaining sample set R0 = {xi | FS0(xi) < 1}. Since S0 and R0
only applicable when the contour conditions do not change. Some time-
are mutually complementary samples subsets of S, P(FS1 > 1|x) can be
variant properties may be modelled as random variables in the slope
estimated using the Total Probability Theorem (e.g., Ang and Tang,
model. For example, the initial matric suction can be deemed as a time-
2007; Wang, 2012)
variant variable because it is influenced by rainfall, evaporation, and
surface flow and may change significantly in a short period. When there P(FS1 > 1|x ) = P(FS1 > 1|x, xi ∈ S0 )Pf0 (S0 ) + P(FS1 > 1|x, xi ∈ R0 )Pf0 (R0 )
is a relatively long interval between two rainfall events, the initial (12)
matric suction before two rainfall events can be modelled as two inde­
pendent variables. It can only be updated with the survival record of one where Pf0(R0) is the probability of R0 given the PDF f0(x) = f(x|FS0 > 1),
past rainfall event considered, instead of multiple past rainfall events. and it is equal to Pf0(R0) = Pf0[xi|FS(xi) < 1] = 0 as random samples in R0
Thus, the posterior PDF of a time-variant variable x is given by satisfy {FS0 < 1} instead of {FS0 > 1}; and Pf0(S0) is the probability of S0
given f0(x), i.e., Pf0(S0) = 1 − Pf0(R0) = 1.
P(FSm > 1|x )f0 (x)
fm (x) = f (x|FS0 > 1 ∩ FSm > 1 ) = ∫ (10) Substituting Pf0(R0) = 0 and Pf0(S0) = 1 into Eq. (12) yields P(FS1 >
P(FSm > 1|x )f0 (x)dx
1|x) = P(FS1 > 1| x, xi ∈ S0), and it can be estimated using N0 random

4
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

samples in S0 as

N0

I(FSm (xi ) > 1 and x − Δx < x < x + Δx |xi ∈ S0 )
P(FS1 > 1|x ) ≈ P(FS1 > 1|x − Δx < x < x + Δx, xi ∈ S0 ) = i=1 (13)

N0
I(x − Δx < x < x + Δx|xi ∈ S0 )
i=1

After P(FS0 > 1|x) and P(FSm > 1|x) are estimated from the above­
Similar to S0, a sample subset Sm-1 (m ≥ 1) is defined as Sm-1 = {xi | mentioned MCS sample reassembling method, Eqs. (6–10) can be used
FS0(xi) > 1 ∩ … ∩ FSm-1(xi) > 1} with Nm-1 conditional random samples to update probability distributions of uncertain parameters accordingly.
satisfying {FS0 > 1 ∩ … ∩ FSm-1 > 1}. The remaining N – Nm-1 random By using the MCS sample reassembling method, it only requires gener­
samples constitute a remaining sample subset Rm-1. Sm-1 and Rm-1 are ating N random samples from fp(x), instead of repetitively generating
also mutually complementary sample subsets of S, and P(FSm > 1|x) is random samples from f0(x), f1(x), …, and fm(x), leading to a significant
given by saving in computational costs, especially when m is large.

P(FSm > 1|x ) = P(FSm > 1|x, xi ∈ Sm− 1 )Pfm− 1 (Sm− 1 ) + P(FSm > 1|x, xi
∈ Rm− 1 )Pfm− 1 (Rm− 1 ) 2.4. Scenario-based slope reliability analysis
(14)
This study develops a scenario-based slope reliability analysis
where Pfm− 1(Rm− 1) is equal to zero given fm-1(x) as all random samples in method for assessing slope failure probability given a target rainfall (e.
Rm-1 do not satisfy {FS0 > 1 ∩ … ∩ FSm-1 > 1}; and Pfm− 1(Sm− 1) is equal to g., including a past 24 h-rainfall and an additional 6 h rainfall fore­
unity. casted). The target rainfall may have unique features including duration,
Substituting Pfm− 1(Rm− 1) = 0 and Pfm− 1(Sm− 1) = 1 into Eq. (14) intensity, and pattern (e.g., hourly rainfall pattern). Such rainfall fea­
yields P(FSm > 1|x) = P(FSm > 1| x, xi ∈ Sm− 1). Then, P(FSm > 1|x) in Eq. tures may significantly affect the slope failure probability (e.g., Zhang
(14) and Eqs. (8–9) (i.e., for m = 2 and m, respectively) can be estimated et al., 2014; Gao et al., 2017). However, previous statistical studies often
with Nm-1 random samples in the sample subset Sm-1 as consider only accumulative rainfall and ignore other rainfall features. In
contrast, a scenario-based slope reliability analysis is developed in this

N∑
m− 1
I(FSm (xi ) > 1 and x − Δx < x < x + Δx |xi ∈ Sm− 1 )
P(FSm > 1|x ) ≈ P(FSm > 1|x − Δx < x < x + Δx, xi ∈ Sm− 1 ) = i=1
N∑
(15)
m− 1
I(x − Δx < x < x + Δx|xi ∈ Sm− 1 )
i=1

study to model the whole target rainfall scenario and estimate the
induced real-time slope failure probability Pf that varies with time. Let F
Note that Eqs. (13) and (15) look similar to Eq. (11), except that their = {FSt < 1} denote the slope failure event caused by a target rainfall
sample subsets are different. This suggests that although random sam­ scenario where FSt denotes the actual FS after the target rainfall is
ples in Sm-1 were initially generated from fp(x), they can also be used as imposed and the subscript t means the “target rainfall”. The slope failure
random samples generated from fm-1(x) to estimate a probability using probability Pf is defined as
direct MCS. The only difference between S and Sm-1 is the occurrence ∫
possibility (or weight) of each random sample (e.g., Wang, 2012; Li Pf = f (x)dx (17)
et al., 2015). The possibility of a random sample is equal to 1/N in S, and FSt <0

it changes to 1/Nm-1 in Sm-1 for a given fm-1(x).


where f(x) denotes the joint PDF of uncertain parameters.
For time-variant variables, because the PDF in Eq. (10) is f0(x), P
Note that f(x) is updated as described in subsection 2.3 considering
(FSm > 1|x) in Eq. (10) should be estimated with random samples in the
survival records from past rainfall events. In the first analysis, f(x) is
sample subset S0 as
equivalent to the prior PDF, fp(x), of uncertain parameters obtained from

N0

I(FSm (xi ) > 1 and x − Δx < x < x + Δx |xi ∈ S0 )
P(FSm > 1|x ) ≈ P(FSm > 1|x − Δx < x < x + Δx, xi ∈ S0 ) = i=1
(16)

N0
I(x − Δx < x < x + Δx|xi ∈ S0 )
i=1

5
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

Fig. 3. Flowchart of the proposed method using direct MCS and sample reassembling.

preliminary uncertainty modeling using limited site investigation data. ∫


Then, site-specific uncertainties are updated considering that the slope Pf = Pmf = fm (x)dx (20)
stays stable before imposing any rainfall, the posterior PDF f0(x) is used
FSt <0

as f(x) in the second analysis. When survival records from m (m ≥ 1) past where Pt, P0f , and Pmf denote respectively the slope failure probability
rainfall events are considered, the posterior PDF fm(x) should be used to under a target rainfall scenario when ignoring any past performance
estimate slope failure probability in the (m + 2)-th analysis. Thus, by record, considering the fact that the existing slope is stable before any
replacing f(x) in Eq. (17) with fp(x), f0(x) and fm(x), slope failure prob­ rainfall, and considering survival records from m past rainfall events.
abilities of these three cases are respectively given by The Pf in Eqs. (18–20) can be estimated using direct MCS. To improve
∫ computational efficiency, the MCS sample reassembling method
Pf = P t = fp (x)dx (18) described in subsection 2.3.2 is used to avoid the repetitive generation of
FSt <0
random samples. The N random samples generated from fp(x) are reused

for the subsequent round of MCS. Estimating the probability of {FSt < 1}
Pf = P0f = f0 (x)dx (19) is similar to estimating {FSm > 1} in Eq. (16). The Nm conditional
FSt <0

6
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

random samples satisfying {FS0 > 1 ∩ … ∩ FSm > 1} constitute the simulated pore water pressure profile is exported and used to
sample subset Sm. They can be used as Nm random samples generated perform slope stability analysis for calculating FSj(xi).
from fm(x) (Wang, 2012) with an equal occurrence possibility of 1/Nm. (9) Repeat Step (8) above for all N random samples.
Thus, Pm
f in Eq. (20) can be estimated with random samples in Sm as (10) Perform subsequent site-specific Bayesian updating of uncer­
tainty with consideration of survival records from the j-th past
1 ∑
Nm
rainfall event. The Nj samples satisfying {FS0(xi) > 1 ∩ … ∩
Pmf = I[FSt (xi ) < 1 |xi ∈ Sm ] (21)
Nm i=1 FSj(xi) > 1} are used to constitute a sample subset Sj. Eqs. (7–10)
are used to obtain posterior distributions of uncertain variables.
where xi are the i-th random sample that follows the prior distribution (11) Repeat Steps (8)–(10) for all m past rainfall events.
fp(x) and satisfies {FS0 > 1 ∩ … ∩ FSm > 1}. (12) A target rainfall is imposed and used to create a new slope model
Eq. (21) is also applicable to estimating P0f when m = 0. It is able to with rainfall infiltration.
incorporate the observed slope performance records into estimation of (13) Use xi in the sample subset Sm as input of the slope model
slope failure probability, including facts that the slope stays stable considering the target rainfall to calculate FSt(xi).
before imposing any rainfall event and that the slope has survived from (14) Repeat Step (13) above for all Nm random samples in the sample
many past rainfall events. Using the survival records leads to a relatively subset Sm.
small estimate of slope failure probability. As more past rainfall events (15) Pf is estimated by Eq. (21) using the MCS sample reassembling
are considered, the estimated slope failure probability may gradually method.
decrease and approach to the observed slope failure frequency.
The abovementioned procedure has been coded into an in-house
3. Implementation procedure MATLAB package. It can be easily performed once the input parame­
ters (e.g., slope geometry, prior distribution of soil parameters, and
Fig. 3 shows a flowchart for implementing the proposed method rainfall records) are given. A software package with a user-friendly
using direct MCS. The details of each step are summarized as follows: graphical user interface will be released in future for facilitating its
application in geotechnical practice.
(1) Develop a deterministic slope stability analysis model without
rainfall for a slope existing at a specific site based on its geometry 4. Illustrative example
and soil parameters.
(2) Determine uncertain parameters of the slope model and their The proposed method is illustrated by a real slope located in the east
prior distributions from site-investigation data. of Tung Chung, Lantau Island in Hong Kong (Evans and Lam, 2003). The
(3) A total of N random samples of uncertain parameters are gener­ slope contains a soil layer of completely decomposed volcanic (CDV) at
ated from their prior distributions. These samples are denoted as shallow depth and slightly decomposed volcanic as bedrock. Monitoring
x1, x2, … xi, …, xN. instruments were installed at the site during 1999 to 2001 to obtain pore
(4) Use xi as input of the slope model without rainfall to perform water pressure and rainfall data. Since the shallow slope failure is the
slope stability analysis for calculating FSLEM. A random sample typical failure mode for landslides triggered by rainfall, the slope was
for model uncertainty ε is added to the computed FSLEM to obtain modelled by an infinite slope under rainfall infiltration. The slope has an
the actual FS, i.e., FS0(xi). angle of 35◦ and a vertical soil thickness of 2 m. The soil layer is initially
(5) Repeat Step (4) above for all N random samples. unsaturated with a uniform matric suction. The unit weight of dry soil
(6) Perform initial site-specific uncertainty updating considering the and water are taken as 16kN/m3 and 10kN/m3, respectively. Hydrus1D
fact that the slope is stable before imposing any rainfall. The N0 (Šimůnek et al., 2013) is adopted to perform seepage analysis under
samples satisfying {FS0(xi) > 1} are used to constitute a sample rainfall infiltration. Rainfall boundary is imposed on the top slope sur­
subset S0. Eq. (6) is used to obtain the posterior distribution of an face without ponding water. The slope bottom is modelled as an
uncertain variable. impermeable boundary. Controlling factors of landslides need to be
(7) Suppose that the slope has survived from m past rainfall events. considered in the slope model, and they could be determined using
Each past rainfall event is considered to develop a new slope multivariate statistical techniques including the Eigenvalue Variance
model with the observed rainfall infiltration during each past and the Canonical Correlation methods. For example, Chae and Seo
rainfall event. Thus, a total of m slope models are developed in (2010) found that slope angle, permeability, and elevation were three
this step.
(8) Use xi as input of the slope model with consideration of the j-th
past rainfall event to perform seepage analysis. Then, the

Table 1
Statistics of uncertain parameters used in the slope example.
Uncertain Probability Mean Standard Coefficient of
parameter distribution deviation variation
(STD) (COV)

Effective cohesion c, Lognormal 5 1.5 0.30


kPa
Effective friction Lognormal 32 9.6 0.30
angle ϕ, ◦
Saturated hydraulic Lognormal 7.2 3.6 0.50
conductivity ks,
mm/h
Initial matric Lognormal 10 3.0 0.30
suction ψ , kPa
Model error ε, Normal 0.02 0.07 3.50
Fig. 4. Simulated pore water pressure profiles of the slope after a 24 h-rainfall
dimensionless
with intensity q = 7.2 mm/h from 500 random samples.

7
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

(a) Cohesion (b) Friction angle

(c) Saturated hydraulic conductivity (d) Model error

(e) Initial matric suction


Fig. 5. Site-specific uncertainty updating results considering the survival record from an antecedent 24 h-rainfall with q = 7.2 mm/h.

the most important factors in rainfall-induced landslides. The uncertain 0.106, θs = 0.469, α = 1.06 m− 1, and n = 1.395). The hydraulic con­
parameters in this example include soil cohesion, friction angle, satu­ ductivity varies with degree of saturation based on a Mualem’s model (i.
rated hydraulic conductivity, model uncertainty, and the initial matric e., Eq. (4) with p = 0.283).
suction. Table 1 summarizes their statistics used in this example. The Direct MCS was adopted to generate 5000 samples of uncertain pa­
initial matric suction and the saturated hydraulic conductivity are rameters for estimating the slope failure probability caused by a target
modelled as random variables because they play an important role in rainfall scenario. Each random sample was used as input to perform
simulating hydraulic responses of soils under rainfall infiltration (e.g., seepage analysis of the slope under the rainfall using Hydrus1D. Then,
Zhang et al., 2013). As the CDV resembles silt loam or loam texture, the the simulated pore water profiles were exported and used for slope
SWCC parameters of a typical loam reported by Cai and Ugai (2004) are stability analysis. Eq. (1) was used to calculate FS at various depths and
used in this study due to a lack of site investigation data. The SWCC of the minimum FS was used as the FSLEM of the slope. For considering the
the soil is modelled by a van Genuchten model (i.e., Eq. (3) with θr = LEM modeling uncertainty, a random number was simulated for

8
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

consideration of the stable slope before a target rainfall and a survival


record from an antecedent 24 h-rainfall with q = 7.2 mm/h. Fig. 5(a)
shows prior and posterior PDF of cohesion. In preliminary uncertainty
modeling, the cohesion follows a prescribed lognormal distribution with
a mean of 5 kPa and a STD of 1.5 kPa (i.e., coefficient of variation (COV)
= 1.5/5 = 0.3). Its posterior PDF f0 is obtained by incorporating the fact
that the slope stays stable before imposing any rainfall. f0 is narrower
and has a higher peak than the prior PDF fp. The mean value of cohesion
decreases to 4.73 kPa and the STD of cohesion decreases to 0.99 kPa,
corresponding to a COV of 0.209 (i.e., 0.99/4.73 = 0.209). Then, the
posterior PDF f1 is obtained by considering the survival record from the
past rainfall. f1 becomes even narrower and higher than fp and f0. The
mean value of cohesion decreases to 4.67 kPa and the STD of cohesion
decreases to 0.81 kPa, corresponding to a COV of 0.173. It is obvious
that incorporating the survival record significantly reduces the uncer­
tainty of cohesion. In addition, posterior PDFs after Bayesian updating
have relatively thin tails on both sides, indicating a reduction in the
possibility of relatively small or large cohesion values. Similar trends are
also observed in other uncertain parameters, as shown in Figs. 5 (b-e).
After uncertainty updating with the survival record from the past rain­
Fig. 6. Results of slope reliability analysis considering the survival record from
fall, the COV of friction angle, saturated hydraulic conductivity, initial
one past rainfall event (the target rainfall duration = 24 h).
matric suction and model error decreases from 0.3, 0.5, 0.3 and 3.5 to
0.157, 0.250, 0.163 and 1.826, respectively. Note that uncertainties of
representing ε in Eq. (5) and calculating the final FS. The slope fails if the
soil parameters are likely to be overestimated when the prior distribu­
final FS is less than 1; otherwise, it is deemed as stable. tions are obtained from limited site investigation data due to significant
Fig. 4 shows the simulated typical pore water pressure profiles from
statistical uncertainty. On the other hand, incorporating the survival
500 random samples, when the prescribed rainfall duration is 24 h and
record from a past rainfall into site-specific uncertainty updating
the rainfall intensity q = 7.2 mm/h. The top of the slope becomes
effectively reduces uncertainties in soil parameters.
saturated after 24-h rainfall and it has a zero-pore water pressure
To explore the effect of past performance on estimating Pf, Fig. 6
because no ponding water is allowed on the slope surface. Rainfall
shows the estimated Pf induced by a 24 h-rainfall when the performance
infiltration reduces the negative-pore water pressure (or matric suction)
during the past rainfall event is ignored or considered. Fig. 6 shows that
from the slope surface to a deep depth, which compromises the slope
the estimated Pf increases with the rainfall intensity and that the esti­
stability, particularly at shallow depth. In most samples, the matric
mated Pf considering the fact that the slope is stable before rainfall is
suction at a deep depth stays unaffected. In comparison, positive pore
smaller than that when ignoring any past performance record, especially
water pressure develops in some samples with high hydraulic conduc­
when the rainfall intensity is less than 7.2 mm/h (i.e., rainfall intensity
tivity in soil and the soil becomes fully saturated. The result shows the
of the past rainfall). When the survival record from a 12 h-rainfall with q
importance of the initial matric suction and the hydraulic conductivity
= 7.2 mm/h is considered, the estimated Pf is smaller than that when
in seepage analysis under rainfall infiltration.
ignoring any past performance and considering the stable slope before
rainfall. The estimated Pf decreases significantly if duration of the past
4.1. Scenario-based slope reliability analysis considering the stable slope rainfall that the slope has survived from increases from 12 h to 24 h and
and the survival record from one past rainfall event 36 h. For example, the Pf caused by a rainfall with q = 18 mm/h is at the
order of magnitude of 10− 1 when ignoring any past performance, and it
Fig. 5 illustrates site-specific uncertainty updating results after decreases to the order of magnitude of 10− 2 and 10− 3 when the survival
record from a 24 h-rainfall and a 36 h-rainfall is considered. The results
indicate that slope reliability analysis with consideration of the slope
being stable before imposing any rainfall and the survival record of a
past rainfall event tends to produce a smaller Pf than that when ignoring
any past performance. The reduction of Pf is significant when the slope
has survived from an intense past rainfall event. The slope failure
probability becomes more dependent on future heavy rainstorms than
light rainfalls after the incorporation of past slope survival records.
Fig. 7 shows a time-variant Pf estimated under a target rainfall with a
constant rainfall intensity of 10.8 mm/h. It is shown that Pf increases
with duration of the target rainfall. Similar to Fig. 6, the estimated Pf
decreases significantly when considering the fact that the slope is stable
before rainfall and the survival record from one past rainfall event. The
results shown in Figs. 6&7 indicate that the Pf estimated from the pro­
posed method is able to reflect magnitudes of different rainfall scenarios.
When compared with previous studies in literature that ignored past
performance of the existing slope, incorporating the survival record
from a past rainfall may effectively decrease the estimated Pf.

4.2. Scenario-based slope reliability analysis considering survival records


from multiple real rainfall events
Fig. 7. Time-variant slope failure probability considering the survival record
from one past rainfall event (the target rainfall intensity = 10.8 mm/h). This subsection explores effects of survival records from multiple

9
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

(a) The first rainfall event

(b) The second rainfall event

(c) The third rainfall event

Fig. 8. Three real rainfall events occurred in Hong Kong (Evans and Lam, 2003).

rainfall events on updating site-specific uncertainty and estimating Pf. event is uncertain, the initial matric suction is deemed as a time-variant
Fig. 8 shows three real rainfall records during June and July of 2001 variable. The initial matric suctions before three past rainfall events and
from a rain gauge at the slope site (Evans and Lam, 2003). Their dura­ the target rainfall event are modelled as independent random variables
tions and hourly rainfall patterns are quite different, though their cu­ following the same prior distribution shown in Table 1, and they are
mulative rainfalls are equal to about 300 mm. The whole hourly rainfall represented by ψ 1, ψ 2, ψ 3, and ψ 4. A total of 5000 random samples of
pattern of each rainfall event was modelled using Hydrus1D for seepage these uncertain variables were generated from their prior distributions.
analysis. Then, the simulated pore water pressure profiles were exported According to the implementation procedure in Section 3, the proposed
and used for slope stability analysis. method was adopted to update probability distributions of uncertain
It was reported that the slope has survived from these three rainfall parameters with survival records from three past rainfall events.
events (Evans and Lam, 2003). The survival records from them can be Fig. 9 shows prior and posterior distributions of uncertain parame­
used to update site-specific uncertainties using the proposed method. ters used in this example. In Fig. 9, the prior distribution is denoted as fp,
The uncertain parameters of this example include effective cohesion c’, while the posterior distributions updated after various previous perfor­
effective friction angle ϕ’, saturated hydraulic conductivity ks, model mance records are represented by f0, f1, f2, and f3, respectively. It is
uncertainty ε and the initial matric suction ψ before each rainfall, and shown in Fig. 9(a) that the posterior PDF of cohesion becomes narrow
their prior distributions are shown in Table 1. In this example, c’, ϕ’, ks, and has a high peak after site-specific uncertainty updating. Both the
and ε are deemed as time-invariant variables which can be updated mean and STD of cohesion decreases as more past rainfall events are
sequentially. Because the initial matric suction before each past rainfall considered. The STD of cohesion decreases significantly from 1.5 kPa to

10
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

(a) Cohesion (b) Friction angle

(c) Saturated hydraulic conductivity (d) Model error

(e) Initial matric suction


Fig. 9. Results of site-specific uncertainty updating with consideration of three past rainfall events.

only 0.63 kPa when survival records from three past rainfall events are posterior distribution of model uncertainty has an increased mean value
considered, equivalent to a reduction of COV from 0.3 to 0.136. Similar and a decreased STD, showing a reduction of COV from 3.5 to 0.968. It is
trends are also observed in Figs. 9(b-d) for other time-invariant vari­ also shown in Figs. 9(a-d) that posterior PDFs after updating often have
ables: friction angle, saturated hydraulic conductivity and model error. high peak and thin tails on both sides. Therefore, the posterior PDFs look
Their posterior PDFs also become narrow and have high peak after site- more similar to truncated normal or lognormal distributions. The trun­
specific uncertainty updating with survival records from past rainfall cation of the probability tails helps to narrow distributions of uncertain
events. In Fig. 9(b), the posterior distribution of friction angle has an parameters and reduce their uncertainties. These results show that
increased mean value and a decreased STD, showing a reduction of COV incorporating survival records from past rainfall events can bring a
from 0.3 to 0.123. Fig. 9(c) shows that the COV of saturated hydraulic significant reduction in variability of uncertain parameters after site-
conductivity decreases from 0.5 to only 0.209 due to the reduction in specific uncertainty updating. Survival records from multiple different
both mean and STD after uncertainty updating. Fig. 9(d) shows that the past rainfall events have a more significant effect on sequential

11
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

Fig. 10. A target rainfall event.

Fig. 9(e) shows posterior distributions of the initial matric suction


before each rainfall, which are obtained from uncertainty updating with
the survival record from only one past rainfall event. Its STD decreases to
1.58 kPa–1.78 kPa after updating, showing a reduction of COV from 0.3
to less than 0.2. The reduction of variability of initial matric suction is
less significant than time-invariant variables. This may be attributed to
the fact that the initial matric suction is time-variant and can only be
updated once with the survival record from only one past rainfall event,
instead of multiple rainfall events.
It is also of significance to explore the effect of survival records from
multiple rainfall events on estimating Pf. Fig. 10 shows a target rainfall
event lasting for two days with an accumulative rainfall of 247 mm. The
proposed method was used to estimate Pf caused by the rainfall
considering survival records from the past three rainfall events shown in
Fig. 8. Fig. 11 illustrates the relationship between the estimated Pf and
the number of considered past rainfall events in estimating Pf. Pf is
Fig. 11. Slope reliability updating results considering multiple past rain­ estimated as large as 0.185 when ignoring any past performance. If the
fall events. fact that the slope stays stable before imposing any rainfall is considered,
Pf is estimated as 0.074. Then, it decreases to 0.032 and 0.012 when
survival records of one (i.e., Fig. 8(a)) or two (i.e., Figs. 8(a) and (b))
past rainfall events are considered. If survival records of all three rainfall
events are considered, Pf decreases to 0.012, which is only 1/15 of the Pf
estimated when ignoring any past performance. The Pf gradually be­
comes stable and will not continue to decrease unless there is a new
record that the slope survived from a heavier rainfall than previous
rainfall events. Thus, the Pf is more dependent on future heavy rain­
storms than light rainfalls that the slope has experienced before. Since
the slope stability is analyzed with the same slope stability model by
limit equilibrium methods, its capability of detecting slope instability
remains the same. The results indicate that survival records from
different past rainfall events provide extremely valuable information for
effectively improving the estimate of Pf and tackling the problem of Pf
overestimation.
Furthermore, Fig. 12 illustrates the real-time Pf during the target
rainfall (see Fig. 10) when ignoring any past performance and consid­
ering survival records from past rainfall events (see Fig. 8). It is shown
that Pf increases with the elapsed time during the target rainfall, espe­
cially after 24 h when the rainfall becomes relatively intense. It seems
that the increase of Pf is not significant during the target rainfall, espe­
Fig. 12. Time-variant slope failure probability during the target rainfall.
cially when ignoring any past performance. This may be attributed to
the fact that the rainfall was poured with an intensity as large as 120
mm/h in a very short period (i.e., 0:00–6:00 of the second day in
uncertainty updating of time-invariant variables than that from a single
Fig. 10). In addition, the estimated Pf with consideration of survival
past rainfall event. Since real rainfall events often have different, and
records from past rainfall events is much smaller than that when
even unique, patterns, survival records of many previous rainfall events
ignoring any past performance. The Geotechnical Engineering Office
can be utilized for site-specific uncertainty updating using the proposed
(GEO) in Hong Kong reported that the annual failure rate of non-
method.
engineered slopes in Hong Kong in 2001 was 72/21550 = 0.33%

12
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

based on an observed landslide number of 72 and the registered slope Au, S.K., Wang, Y., 2014. Engineering Risk Assessment with Subset Simulation. John
Wiley & Sons Singapore Pte. Ltd., Singapore, Singapore.
number of 21,550 in Hong Kong (Lam et al., 2004). The slope failure
Baecher, G.B., Christian, J.T., 2003. Reliability and Statistics in Geotechnical
probability considering past slope performance records ranged from Engineering. John Wiley & Sons, Ltd., England, Chichester.
0.001 to 0.012, which is consistent with the actual annual slope failure Baum, R.L., Godt, J.W., 2010. Early warning of rainfall-induced shallow landslides and
rate. In contrast, the slope failure probability without considering past debris flows in the USA. Landslides 7 (3), 259–272. https://doi.org/10.1007/
s10346-009-0177-0.
slope performance records was estimated as large as 0.185, two orders of Cai, F., Ugai, K., 2004. Numerical analysis of rainfall effects on slope stability. Int. J.
magnitude larger than the slope failure rate observed in practice. This Geomech. 4, 69–78. https://doi.org/10.1061/(ASCE)1532-3641(2004)4:2(69).
shows that incorporating slope survival records from past rainfall events Cao, Z.J., Wang, Y., Li, D.Q., 2016. Site-specific characterization of soil properties using
multiple measurements from different test procedures at different locations – a
is able to enhance the estimation of slope failure probability and avoid Bayesian sequential updating approach. Eng. Geol. 211, 150–161. https://doi.org/
the problem of overestimating Pf. 10.1016/j.enggeo.2016.06.021.
Chae, B.G., Seo, Y.S., 2010. Suggestion of an evaluation chart for landslide susceptibility
using a quantification analysis based on canonical correlation. Korean Soc. Econ.
5. Conclusions Environ. Geol. 43, 381–391.
Chae, B.G., Park, H.J., Catani, F., Simoni, A., Berti, M., 2017. Landslide prediction,
monitoring and early warning: a concise review of state-of-the-art. Geosci. J. 21 (6),
This study proposed a scenario-based reliability analysis method for
1033–1070. https://doi.org/10.1007/s12303-017-0034-4.
an existing slope considering rainfall triggering mechanism and its Chen, H.X., Zhang, L.M., 2014. A physically-based distributed cell model for predicting
survival records from past rainfall events. The survival records were regional rainfall-induced shallow slope failures. Eng. Geol. 176, 79–92. https://doi.
org/10.1016/j.enggeo.2014.04.011.
used to reduce site-specific uncertainties in soil parameters and other
Cheverda, V., Reshetova, G., Lisitsa, V., Protasov, M., 2020. Digital twin of multiscale
variables and improve accuracy of the estimated Pf. Effects of one and geological media: Faults, fracture corridors, caves. Seismic simulation and imaging.
multiple past rainfall events on site-specific uncertainty updating and Pf In: 4th EAGE Workshop on High Performance Computing for Upstream in the Middle
estimation were explored using an infinite slope example under rainfall East: HPC through the 4th Industrial Revolution. European Association of
Geoscientists and Engineers, EAGE, pp. 1–5. https://doi.org/10.3997/2214-
infiltration. A few key conclusions can be summarized as: 4609.201903294.
Ching, J., Phoon, K.K., Hu, Y.G., 2009. Efficient evaluation of reliability for slopes with
(1) For a given slope, its Pf is dependent on a target rainfall scenario circular slip surfaces using importance sampling. J. Geotech. Geoenviron. 135 (6),
768–777. https://doi.org/10.1061/(ASCE)GT.1943-5606.0000035.
that may have different characteristics (e.g., duration, intensity, Chowdhury, R., Zhang, S., Flentje, P., 2015. Reliability updating and geotechnical back-
and hourly rainfall pattern). Different rainfall characteristics (e. analysis. Adv. Geotech. Eng. https://doi.org/10.1680/AIGEV2.32644.0012 (The
g., hourly rainfall pattern) might lead to different Pf. The illus­ Skempton Conference).
Christian, J.T., Baecher, G.B., 2011. Unresolved problems in geotechnical risk and
trative example showed that, after Bayesian updating using the reliability. GeoRisk 2011, 50–63. https://doi.org/10.1061/41183(418)3.
past performance records, the estimate of Pf for an existing slope Duncan, J.M., Wright, S.G., 2005. Soil Strength and Slope Stability. John Wiley & Sons
may be reduced by one to two orders of magnitudes, making the Inc.
Evans, N.C., Lam, J.S., 2003. Tung Chung East Natural Terrain Study Area Ground
obtained Pf consistent with the observed slope failure frequency.
Movement and Groundwater Monitoring Equipment and Preliminary Results (GEO
The challenge of overestimating Pf from traditional slope reli­ Report No. 142). Geotechnical Engineering Office (GEO), Hong Kong, China.
ability analysis is tackled using past performance records of the Fenton, G.A., Vanmarcke, E.H., 1990. Simulation of random fields via local average
subdivision. J. Eng. Mech. 116 (8), 1733–1749. https://doi.org/10.1061/(ASCE)
existing slope. The propose method can be used to estimate real-
0733-9399(1990)116:8(1733).
time slope failure probability Pf that varies as the considered Gao, L., Zhang, L.M., Chen, H.X., 2017. Likely scenarios of natural terrain shallow slope
rainfall evolves with time and provide key information for failures on Hong Kong Island under extreme storms. Nat. Hazards Rev. 18 (1),
decision-making and mitigation of landslide risk. B4015001 https://doi.org/10.1061/(ASCE)NH.1527-6996.0000207.
van Genuchten, M.T., 1980. A closed-form equation for predicting the hydraulic
(2) Incorporating the fact that the slope stays stable before imposing conductivity of unsaturated soils. Soil Sci. Soc. Am. J. 44 (5), 892. https://doi.org/
any rainfall and the slope’s survival record from past rainfall 10.2136/sssaj1980.03615995004400050002x.
events leads to a significant reduction in variability of uncertain Gong, W., Juang, C.H., Wasowski, J., 2021. Geohazards and human settlements: Lessons
learned from multiple relocation events in Badong, China – Engineering geologist’s
parameters. The effect of Bayesian updating of uncertainties is perspective. Eng. Geol. 285, 106051. https://doi.org/10.1016/j.
more significant for time-invariant variables (e.g., soil properties) enggeo.2021.106051.
when using survival records from multiple past rainfall events Grieves, M., 2015. Digital Twin: Manufacturing Excellence through Virtual Factory
Replication.
than that when using a single past rainfall event. Ji, J., Low, B.K., 2012. Stratified response surfaces for system probabilistic evaluation of
slopes. J. Geotech. Geoenviron. 138 (11), 1398–1406. https://doi.org/10.1061/
(ASCE)GT.1943-5606.0000711.
Declaration of Competing Interest Jiang, S.H., Li, D.Q., Cao, Z., Zhou, C.B., Phoon, K.K., 2015. Efficient system reliability
analysis of slope stability in spatially variable soils using Monte Carlo simulation.
J. Geotech. Geoenviron. 141 (2), 04014096 https://doi.org/10.1061/(ASCE)
The authors declare that they have no known competing financial GT.1943-5606.0001227.
interests or personal relationships that could have appeared to influence Jiang, S.H., Huang, J., Qi, X.H., Zhou, C.B., 2020. Efficient probabilistic back analysis of
spatially varying soil parameters for slope reliability assessment. Eng. Geol. 271,
the work reported in this paper. 105597. https://doi.org/10.1016/j.enggeo.2020.105597.
Kong, V.W.W., Kwan, J.S.H., Pun, W.K., 2020. Hong Kong’s landslip warning system—40
years of progress. Landslides 17 (6), 1453–1463. https://doi.org/10.1007/s10346-
Acknowledgements
020-01379-6.
Lam, T.T.M., Sun, H.W., Ho, K.K.S., 2004. Review of Landslides in 2001. GEO Report No.
The work described in this paper was supported by a grant from the 155. Geotechnical Engineering Office, Hong Kong, China.
Research Grants Council of the Hong Kong Special Administrative Re­ Li, D.Q., Chen, Y., Lu, W., Zhou, C., 2011. Stochastic response surface method for
reliability analysis of rock slopes involving correlated non-normal variables.
gion, China (Project No. T22-603/15N) and a Strategic Research Grant Comput. Geotech. 38 (1), 58–68. https://doi.org/10.1016/j.compgeo.2010.10.006.
from City University of Hong Kong (Project No. 7005551). The financial Li, D.Q., Zhang, F.P., Cao, Z.J., Zhou, W., Phoon, K.K., Zhou, C.B., 2015. Efficient
supports are gratefully acknowledged. reliability updating of slope stability by reweighting failure samples generated by
Monte Carlo simulation. Comput. Geotech. 69, 588–600. https://doi.org/10.1016/j.
compgeo.2015.06.017.
References Li, D.Q., Xiao, T., Cao, Z.J., Zhou, C.B., Zhang, L.M., 2016. Enhancement of random finite
element method in reliability analysis and risk assessment of soil slopes using Subset
simulation. Landslides 1–11.
Ang, A.H.S., Tang, W.H., 2007. Probability Concepts in Engineering: Emphasis on
Li, K.S., Lumb, P., 1987. Probabilistic design of slopes. Can. Geotech. J. 24 (4), 520–535.
Applications to Civil and Environmental Engineering, 2nd Ed. John Wiley & Sons,
https://doi.org/10.1139/t87-068.
Inc.
Liu, X., Wang, Y., Li, D.Q., 2019. Investigation of slope failure mode evolution during
Au, S.K., Beck, J.L., 2001. Estimation of small failure probabilities in high dimensions by
large deformation in spatially variable soils by random limit equilibrium and
subset simulation. Probabil. Eng. Mech. 16, 263–277. https://doi.org/10.1016/
S0266-8920(01)00019-4.

13
X. Liu and Y. Wang Engineering Geology 288 (2021) 106144

material point methods. Comput. Geotech. 111, 301–312. https://doi.org/10.1016/ Tofani, V., Bicocchi, G., Rossi, G., Segoni, S., D’Ambrosio, M., Casagli, N., Catani, F.,
j.compgeo.2019.03.022. 2017. Soil characterization for shallow landslides modeling: a case study in the
Liu, X., Li, D.Q., Cao, Z.J., Wang, Y., 2020. Adaptive Monte Carlo simulation method for Northern Apennines (Central Italy). Landslides 14, 755–770. https://doi.org/
system reliability analysis of slope stability based on limit equilibrium methods. Eng. 10.1007/s10346-017-0809-8.
Geol. 264, 105384. https://doi.org/10.1016/j.enggeo.2019.105384. Wai, R.C.T., Lee, R.W.H., Law, R.H.C., 2018. Review of Landslides in 2016 (GEO Report
Low, B.K., Zhang, J., Tang, W.H., 2011. Efficient system reliability analysis illustrated for no. 341. Hong Kong, China).
a retaining wall and a soil slope. Comput. Geotech. 38 (2), 196–204. Wang, Y., 2012. Uncertain parameter sensitivity in Monte Carlo simulation by sample
Lu, N., Godt, J., 2008. Infinite slope stability under steady unsaturated seepage reassembling. Comput. Geotech. 46, 39–47. https://doi.org/10.1016/j.
conditions. Water Resour. Res. 44 (11), W11404 https://doi.org/10.1029/ compgeo.2012.05.014.
2008WR006976. Wang, Y., Cao, Z., 2013. Probabilistic characterization of Young’s modulus of soil using
Lu, N., Likos, W.J., 2006. Suction stress characteristic curve for unsaturated soil. equivalent samples. Eng. Geol. 159, 106–118. https://doi.org/10.1016/j.
J. Geotech. Geoenviron. 132 (2), 131–142. https://doi.org/10.1061/(ASCE)1090- enggeo.2013.03.017.
0241(2006)132:2(131). Wang, Y., Au, S.K., Cao, Z., 2010. Bayesian approach for probabilistic characterization of
Mualem, Y., 1976. A new model for predicting the hydraulic conductivity of unsaturated sand friction angles. Eng. Geol. 114 (3–4), 354–363. https://doi.org/10.1016/j.
porous media. Water Resour. Res. 12 (3), 513–522. https://doi.org/10.1029/ enggeo.2010.05.013.
WR012i003p00513. Wang, Y., Cao, Z., Au, S.K., 2011. Practical reliability analysis of slope stability by
Papaioannou, I., Straub, D., 2012. Reliability updating in geotechnical engineering advanced Monte Carlo simulations in a spreadsheet. Can. Geotech. J. 48 (1),
including spatial variability of soil. Comput. Geotech. 42, 44–51. https://doi.org/ 162–172. https://doi.org/10.1139/T10-044.
10.1016/j.compgeo.2011.12.004. Wang, Y., Cao, Z., Li, D., 2016. Bayesian perspective on geotechnical variability and site
Phoon, K.K., Kulhawy, F.H., 1999a. Characterization of geotechnical variability. Can. characterization. Eng. Geol. 203, 117–125. https://doi.org/10.1016/j.
Geotech. J. 36 (4), 612–624. https://doi.org/10.1139/t99-038. enggeo.2015.08.017.
Phoon, K.K., Kulhawy, F.H., 1999b. Evaluation of geotechnical property variability. Can. Zhang, H., Wang, R., Wang, C., 2019. Monitoring and Warning for Digital Twin-driven
Geotech. J. 36 (4), 625–639. Mountain Geological disaster. In: Proceedings of 2019 IEEE International Conference
Schweckendiek, T., Kanning, W., 2009. Updating piping probabilities with survived on Mechatronics and Automation, ICMA 2019. Institute of Electrical and Electronics
loads. In: 7th International Probabilistic Workshop, Delft, The Netherlands, Engineers Inc, pp. 502–507. https://doi.org/10.1109/ICMA.2019.8816292.
pp. 1–13. Zhang, J., Zhang, L.M., Tang, W.H., 2009. Bayesian framework for characterizing
Schweckendiek, T., Vrouwenvelder, A.C.W.M., Calle, E.O.F., 2014. Updating piping geotechnical model uncertainty. J. Geotech. Geoenviron. 135 (7), 932–940. https://
reliability with field performance observations. Struct. Saf. 47, 13–23. doi.org/10.1061/(ASCE)GT.1943-5606.0000018.
Segoni, S., Battistini, A., Rossi, G., Rosi, A., Lagomarsino, D., Catani, F., Moretti, S., Zhang, J., Zhang, L.M., Tang, W.H., 2011. Slope reliability analysis considering site-
Casagli, N., 2015. Technical note: an operational landslide early warning system at specific performance information. J. Geotech. Geoenviron. 137 (3), 227–238.
regional scale based on space–time-variable rainfall thresholds. Nat. Hazards Earth https://doi.org/10.1061/(ASCE)GT.1943-5606.0000422.
Syst. Sci. 15 (4), 853–861. https://doi.org/10.5194/nhess-15-853-2015. Zhang, J., Huang, H.W., Zhang, L.M., Zhu, H.H., Shi, B., 2014. Probabilistic prediction of
Šimůnek, J., Šejna, M., Saito, H., Sakai, M., van Genuchten, M.T., 2013. The Hydrus-1D rainfall-induced slope failure using a mechanics-based model. Eng. Geol. 168,
Software Package for Simulating the Movement of Water, Heat, and Multiple Solutes 129–140. https://doi.org/10.1016/J.ENGGEO.2013.11.005.
in Variably Saturated Media, Version 4.17, HYDRUS Software Series 3. Department Zhang, L.L., Li, J., Li, X., Zhang, J., Zhu, H., 2016. Rainfall-Induced Soil Slope Failure:
of Environmental Sciences, University of California Riverside, Riverside, California, Stability Analysis and Probabilistic Assessment. CRC Press, Boca Raton, USA.
USA, p. 342. Zhang, L.L., Zuo, Z.B., Ye, G.L., Jeng, D.S., Wang, J.H., 2013. Probabilistic parameter
Tang, H., Wasowski, J., Juang, C.H., 2019. Geohazards in the three Gorges Reservoir estimation and predictive uncertainty based on field measurements for unsaturated
Area, China – Lessons learned from decades of research. Eng. Geol. 261, 105267. soil slope. Comput. Geotech. 48, 72–81. https://doi.org/10.1016/J.
https://doi.org/10.1016/j.enggeo.2019.105267. COMPGEO.2012.09.011.
Tang, W.H., Zhang, L.M., 2011. Development of a risk-based landslide warning system.
GeoRisk 224 GSP, 25–49. https://doi.org/10.1061/41183(418)2.

14

You might also like