Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Engineering Geology 161 (2013) 1–15

Contents lists available at SciVerse ScienceDirect

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

Assessment of rainfall-induced shallow landslide susceptibility using a


GIS-based probabilistic approach
Hyuck Jin Park a,⁎, Jung Hyun Lee a, Ik Woo b
a
Department of Geoinformation Engineering, Sejong University, Seoul, Republic of Korea
b
Department of Coastal Construction Engineering, Kunsan National University, Kunsan, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: This study proposes a probabilistic analysis method to assess shallow landslide susceptibility over an exten-
Received 28 September 2012 sive area by integrating an infinite slope model with GIS (Geographic Information System) and Monte Carlo
Received in revised form 17 April 2013 simulation, taking into consideration the inherent uncertainty and variability in input parameters. The me-
Accepted 22 April 2013
chanical parameters of soil materials (such as cohesion and friction angle) used in the infinite slope analysis
Available online 28 April 2013
have been identified as the major source of uncertainty because of their spatial variability; therefore, these
Keywords:
parameters were considered as random variables in this probabilistic landslide analysis. To properly account
Infinite slope model for the uncertainty in input parameters, the probabilistic analysis method used was Monte Carlo simulation.
Shallow landslide The process was carried out in a GIS-based environment because GIS has effective spatial data-processing
Hydrogeological model capacity over broad areas. In addition, the hydrogeological model was coupled with the infinite slope
Probability of failure model to evaluate increases in pore water pressure caused by rainfall.
Monte Carlo simulation The proposed approach was applied to a practical example to evaluate its feasibility. The landslide inventory
map and the spatial database for input parameters were constructed in a grid-based GIS environment and a
probabilistic analysis was implemented using Monte Carlo simulation. To evaluate the performance of the
model, the results of the probabilistic landslide susceptibility analysis were compared with the landslide in-
ventory. The probabilistic approach demonstrated good predictive performance when compared with the
landslide occurrence location. In addition, deterministic analysis was carried out using fixed single-input
data for comparison with the results from the proposed approach. In this comparison, the probabilistic anal-
ysis showed better performance than the deterministic analysis. In addition, the results showed that proper
consideration and understanding of uncertainties play an important role in accurately predicting shallow
landslide susceptibility.
© 2013 Elsevier B.V. All rights reserved.

1. Introduction difficult because an enormous amount of spatial data must be acquired


from the region and processed. Consequently, GIS (Geographic Infor-
Landslides are a common geological hazard during the rainy sea- mation System) has been widely used to analyze spatial information
son in Korea. On average, landslides cause the loss of 23 lives each relevant to landslides because of the strong capability of GIS for spatially
year, which accounts for nearly 25% of annual casualties due to natu- distributed data processing. In addition, complex techniques requiring a
ral disasters. In 2002, 75 people were killed in landslide-related inci- large number of map overlays and table calculations become feasible by
dents throughout the country, and 18 people were killed in 2011 by using GIS (Soeters and van Westen, 1996). This capability has facilitated
rainfall-induced shallow landslides in the central Seoul metropolitan many GIS-based studies on landslide susceptibility assessment.
area (Korean Geotechnical Society, 2011). However, the development Landslide susceptibility assessment methods can be classified as
and urbanization of mountainous areas continue; thus, slope instabil- qualitative or quantitative. Generally, qualitative approaches are based
ity and the potential for landslides have increased. entirely on the judgment of experts conducting the susceptibility as-
The occurrence of landslides is controlled by various spatial and cli- sessment (van Westen et al., 1999). Therefore, qualitative approaches
matic factors, such as geology, topography, hydrogeological conditions, are seldom used as susceptibility assessment methods for large-scale
vegetation and rainfall. Thus, prediction of landslide susceptibility is areas because of the lack of a concrete physical concept with slope fail-
ures (Xie et al., 2004). Quantitative landslide susceptibility assessment
methods can be divided into two categories: statistical methods and
⁎ Corresponding author at: Department of Geoinformation Engineering, Sejong
University, 98 Gunja-dong, Gwangjin-gu, Seoul 143-747, Republic of Korea. Tel.: +82 2
geotechnical approaches (Aloetti and Chowdhury, 1999). The statistical
3408 3965; fax: +82 2 3408 4341. methods, such as bivariate or multivariate statistical analyses, analyze
E-mail address: hjpark@sejong.ac.kr (H.J. Park). the relationship between landslide occurrences and related factors,

0013-7952/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.enggeo.2013.04.011
2 H.J. Park et al. / Engineering Geology 161 (2013) 1–15

such as soil type, land use, slope geometry, vegetation and parent mate- analysis coupled with the hydrogeological model, in this study the
rial. The most significant drawback of statistical methods is the collec- probabilistic analysis method was used considering the strength pa-
tion of data over large areas regarding landslide distribution and rameters of the soil as random variables. Monte Carlo simulation was
factor maps (van Westen, 2000). In addition, statistical methods only used for the probabilistic analysis with GIS integration. The proposed
consider relationships between landslide occurrences and related fac- approach was applied to assess the feasibility of the approach and to
tors, but not the failure mechanism. In contrast, the geotechnical build a landslide susceptibility map for the study site. Application of
approach analyzes the mechanical condition of slopes and evaluates the probabilistic analysis method in the infinite slope model coupled
their stability using mathematical calculations (Montgomery and with the hydrogeological model for landslide susceptibility assessment
Dietrich, 1994; Wu and Sidle, 1995; Luzi and Pergalani, 1996; requires spatial topographic, geologic, geotechnical and hydrogeo-
Gokceoglu et al., 2000; Gorsevski, 2002; Guimaraes et al., 2003; Zhou logical information. In addition, inventories of previous landslide occur-
et al., 2003; Frattini et al., 2004; Godt et al., 2008; Liu and Wu, 2008; rences are required to test the assessment results. Therefore, we
Li and Chi, 2011; Ho et al., 2012). In this approach, the physical proper- acquired information on landslide occurrence, precipitation, digital to-
ties of a particular slope are obtained from field investigations and lab- pography data, geology and the random properties of strength parame-
oratory tests, and a physically based model is used to evaluate landslide ters for slope materials using field investigation and laboratory tests,
susceptibility. The physically based model is a promising approach for and implemented the probabilistic analysis for the study area. Next,
the susceptibility analysis of shallow landslides because of its capacity the results of the susceptibility analysis were compared with the land-
to reproduce the physical processes governing landslide occurrence slide inventory for validation.
(Fell et al., 2008). Moreover, the general grid-based structure and
wide availability of GIS provide a convenient framework allowing anal- 2. Infinite slope model and steady-state hydrogeological model
ysis over wide areas (Sorbino et al., 2010). Most recent physically based
models are combined with a hydrogeological model to evaluate the ef- To use a geotechnical failure model and apply the mechanical
fect of pore water pressure. This is because pore water pressure, which properties of slope materials to the susceptibility analysis, the infinite
increases as a consequence of rainfall, triggers shallow landslides, and slope model is used as a physically based model for rainfall-induced
the hydrogeological model can evaluate variations in rainfall-driven shallow landslides. The slope is assumed to extend infinitely in its
pore water pressure. dip direction, and sliding is considered to occur along a plane parallel
However, in the physically based model, sufficient and accurate to the face of the slope. In previous works, the infinite slope model
information has to be obtained to construct an accurate landslide sus- was used in geotechnical analysis approaches because of its effective-
ceptibility map. That is, in landslide susceptibility analysis, the quality ness in landslide hazard analysis (Kamai, 1991; Terlien, 1996; Pack et
and quantity of information for the generation of landslide suscepti- al., 1998, 2001; Frattini et al., 2004; Huang et al., 2006; Rosso et al.,
bility zonation maps have been the main concerns, because such 2006; Godt et al., 2008; Picarelli et al., 2008; D'Amato Avanzi et al.,
data are often limited in extent and have imperfections or variable 2009; Harp et al., 2009; Apip et al., 2010; Griffiths et al., 2011;
quality (van Westen et al., 2006). In particular, the strength parame- Santoso et al., 2011; Ho et al., 2012). This model is also appropriate
ters of slope materials, such as cohesion and friction angle, are inher- for Korean landslides, where sliding surfaces are commonly located
ently spatially heterogeneous, as the slope material is produced by a at shallow depths (NIDP, 2000). The infinite slope model has been
natural process (Baecher and Christian, 2003; Carrara et al., 2008; used previously in site-specific studies, but the adoption of GIS allows
Chowdhury et al., 2010). In addition, the input parameters used in the analysis, modeling and spatialization of slope stability conditions
the geotechnical approach should be obtained from a very wide using this simple geotechnical model for a broad area.
study area, likely with limited sampling; thus, uncertainties are inev- As shown in Fig. 1, the factor of safety for an infinite slope can be
itably involved in the physically based model analysis. However, the calculated as:
deterministic approach, which is generally adopted by physically
based model analyses, is not appropriate, because it considers only a c′ þ ðγ·z−γw ·hw Þ cos2 α tan ϕ′
FS ¼ ð1Þ
single fixed value, while disregarding the variability and uncertainty γ·z· sin α· cos α
in input parameters. It is more appropriate to use the probabilistic
analysis method, because it properly considers uncertainties in the where α is the slope angle, γ is the unit weight of soil, γw is the unit
data. Nevertheless, most of the previous physically based model stud- weight of water, z is the soil depth from the ground surface, hw is the
ies that were coupled with the hydrogeological model used the deter- saturated soil thickness above the slip surface, c′ is the effective
ministic analysis approach to estimate the potential or relative
instability of slope over a large area without considering the uncer-
tainties in input parameters (Acharya et al., 2006; Rosso et al., 2006;
Salciarini et al., 2006, 2008; Godt et al., 2008; Harp et al., 2009; Lee
and Ho, 2009; Apip et al., 2010; Sorbino et al., 2010; Chiang and
Chang, 2011; Ho et al., 2012; Segoni et al., 2012). Baum et al. (2005)
employed the probabilistic analysis approach based on historical
landslide records but used the deterministic analysis approach in
the evaluation of the physically based model for susceptibility analy-
sis. Recently, Santoso et al. (2011) applied the probabilistic analysis
approach in the physically based model approach, but only the coeffi-
cient of permeability was considered as a random variable, and not
the strength parameters of the slope materials, which were known
as major sources of uncertainty in the geotechnical approach. Some
authors employed probabilistic analysis approaches in the physically
based model analysis, but the hydrogeological model was not adopted,
and subsequently a constant groundwater level was assumed for the
entire study area in their analyses (Zhou et al., 2003; Shou and Chen,
2005; Shou et al., 2009). Therefore, to deal properly with the uncertain-
ty and variability in input parameters for the physically based model Fig. 1. The infinite slope model.
H.J. Park et al. / Engineering Geology 161 (2013) 1–15 3

cohesion and ϕ′ is the effective friction angle (Coduto et al., 2010). For slope model such as in Eq. (2). However, uncertainties are inevitably
simplicity, Eq. (1) assumes that the groundwater level in an infinite involved in the determination of the strength parameters because of
slope is parallel to the surface and located at hw. In addition, it is as- the very large size of the study area and the limited number of sam-
sumed that groundwater flow is parallel to the slip surface, and a ples. Therefore, applying the deterministic approach to an extensive
slice of infinite slope mass has a unit width. study area can be particularly difficult or impossible because of the
However, pore water pressure reduces the effective normal stress uncertainties and difficulties in obtaining, checking and processing
and the shear strength of soil, ultimately causing slope failure; thus, large spatial data sets (Zhou et al., 2003). For these reasons, the
knowledge of the groundwater level is important for predicting and parameters should be considered as random variables rather than
preventing slope instability. It is practically impossible, however, to having single, deterministic values. Consequently, in this study, cohe-
measure the groundwater levels over an extensive area. Therefore, sion and friction angle were considered as random variables, and
previous studies have used a constant or randomly selected value for probabilistic analysis was employed to properly account for uncer-
the groundwater level for the entire study area (van Westen and tainty and variability in input parameters. To conduct a probabilistic
Terlien, 1996; Zhou et al., 2003; Xie et al., 2004; Griffiths et al., analysis, the random properties of the input parameters, such as the
2011). However, the groundwater level is dependent on soil type, mean, standard deviation and probability density function, were
rainfall intensity and hydraulic conductivity of soil, so it is not appro- obtained from the available data or from subjective engineering judg-
priate to use a constant or randomly selected value for a large area. For ments, such as literature reviews or an engineer's decision (Baecher
analysis of the pore water pressure regime, the present study used a and Christian, 2003; Hoek, 2007). The probability of slope failure
hydrogeological model that can evaluate rainfall-driven pore water was then evaluated using the random properties of the input param-
pressure and can be coupled to the infinite slope model. The eters and their distribution functions.
hydrogeological model uses analytical solutions for the pore water The Monte Carlo simulation approach was employed to evaluate
pressure response to rainfall and can be used for shallow landslide the probability of slope failure using random properties of input pa-
hazard analysis. The hydrogeological models can be classified into rameters. The Monte Carlo simulation is the most complete probabi-
two categories on the basis of the simplifying assumption: steady listic analysis because all random variables and the probability of
and transient models (Montgomery and Dietrich, 1994; Terlien et al., failure as the result of reliability analysis are represented by their
1995; Wu and Sidle, 1995; Pack et al., 1998; Baum et al., 2002; probability density function. The Monte Carlo simulation is relatively
Crosta and Frattini, 2003; Savage et al., 2004; Godt et al., 2008). The easy to implement on a computer and can accommodate a wide range
steady model assumes that rainfall infiltration is at its steady state, of functions, including those that cannot be expressed conveniently in
and saturated water flows parallel to the slope surface (Montgomery an explicit form (Baecher and Christian, 2003). In the simulation pro-
and Dietrich, 1994; Wu and Sidle, 1995). This model allows the predic- cedure, discrete values of each random variable were randomly se-
tion of the groundwater table as a function of groundwater flow and lected from their probability density function. Using a group of
rainfall intensity. The transient model performs transient seepage these discrete values, a factor of safety value was calculated with
analysis using the linearized solution of Richards' equation (Iverson, the other deterministic input data. This process was repeated numer-
2000; Baum et al., 2002) and provides more realistic analysis results. ous times for a pixel. From the large number of factor of safety values
However, the transient model needs abundant and accurate spatial that were generated, a probability density function of the factor of
information. Moreover, it is sensitive to some input data, such as safety was obtained, and the probability of failure was evaluated.
hydraulic properties of the soil, initial steady-state groundwater con- This process was repeated for other pixels, and the slope failure prob-
dition and soil depth (Salciarini et al., 2006; Sorbino et al., 2007, ability map was then produced.
2010). Therefore, in this study the steady-state model was used, be-
cause the measurement of spatial parameters, such as initial steady- 4. Application of the proposed model to the study area
state groundwater condition over a large area, is difficult.
In the steady-state model, pore water pressure can be computed as- 4.1. Study area and landslide location detection
suming a hydrogeological steady state, with the depth of saturated soil
being sufficient to sustain a lateral discharge proportional to the specific Spatial information from the Pyeongchang district in eastern
contributing area (upslope area per unit contour length) (Pack et al., Korea was used for the combined probabilistic infinite slope model
2001). Subsequently, a steady-state shallow subsurface flow model, analysis and the hydrogeological model. The study site lies between
derived from TOPMODEL assumptions (Beven and Kirkby, 1979), was latitudes 37°33′20″N and 37°39′26″N and longitudes 128°29′49″E
adopted for the infinite slope stability model using the following and 128°36′36″E (Figure 2). The prevalent geological units exposed
equation: in the study area are Mesozoic Nokam formation and igneous rock
(Imgye granite) (Table 1). These are located in the Precambrian meta-
c′ þ ½γ·z−m·γw ·z cos2 α tan ϕ′ morphic rocks with uniformity. The metamorphic rocks consist large-
FS ¼ ly of banded gneiss, with smaller amounts of migmatitic gneiss, schist
γ·z· sin α· cos α
   
c′ þ γ·z− min RT · sina α ; 1:0 ·γ w ·z cos2 α tan ϕ′ and quartzite. Triassic Nokam formation is composed of fine sand-
¼ ð2Þ stone with gray sandy shale, originating from thick clastic successions
γ·z· sin α· cos α
of marginal marine to nonmarine environments, whereas the Jurassic
plutonic rock, Imgye granite, mainly occurs as a large batholith trend
where m is the ratio of the groundwater level to the soil depth
  NW–SE and as small stocks along the Ogcheon Belt consisting of gran-
¼ hzw , R is rainfall intensity, T is the soil transmissivity, which is ite with minor syenite and diorite. The Jeongseon limestone, which is
soil hydraulic conductivity times soil thickness, and α is the specific the third prevalent unit and Ordovician in age, is mainly shallow marine
contributing area (Beven and Kirkby, 1979; O'Loughlin, 1986; in origin and consists predominantly of limestone with smaller amounts
Montgomery and Dietrich, 1994; Hsu, 1998; Dietrich et al., 2001; of sandstone and shale. This area has experienced an extensive intrusion
Borga et al., 2002; Huang et al., 2006, 2007; Reid et al., 2007). of granitoids by the Daebo Orogeny that continued from the early Juras-
sic to early Cretaceous and was the most severe in intensity; thus, all of
3. Monte Carlo simulation the previous formations were intensely deformed (Geological Society of
Korea, 1962; Lee, 1988). The topography consists mostly of mountain-
The strength parameters of soil, such as cohesion and friction ous areas, which cover approximately 60% of the study area. Elevations
angle, are required to evaluate the factor of safety using an infinite range from 430 to 1390 m, with an average of 736.1 m, and the areas
4 H.J. Park et al. / Engineering Geology 161 (2013) 1–15

Fig. 2. The study area and location of the landslides that occurred in July 2006.

with slope greater than 30° account for 41.2% of the total area. On 15–16 landslides, occupying an area of about 771,100 m 2, were mapped
July 2006, this area experienced heavy rainfall of 429 mm, and over (Figure 2).
1400 landslides were reported (KIGAM, 2009). Fig. 3 illustrates the typ-
ical landslide types that occurred in the study area. 4.2. Spatial database construction
The accurate detection of landslide locations is important for land-
slide susceptibility analysis. Field surveys are widely considered to be The proposed infinite slope model requires knowledge of several
the most accurate detection method in the landslide inventory pro- input parameters, some of which can be derived from digital
cess. However, using a field survey as an incipient step in the topographic data using GIS, while others have to be obtained from
data-collection process is difficult, time consuming and costly, in par- geotechnical field work and laboratory tests, or derived from precip-
ticular in mountainous areas, where access is limited or impossible. itation records. Digitized terrain information is used to acquire the
Aerial photography and satellite imagery have the advantage of slope angle as a geomorphological attribute. To achieve this, contour
detecting landslides rapidly and accurately, and were therefore lines and survey base points with an elevation value were extracted
employed to overcome the lack of field data. from 1:5000 scale topographic maps, and a digital elevation model
Digital aerial photographs with a ground resolution of 50 cm were (DEM) with a 10 m resolution was constructed. Using the DEM, the
obtained from Samah Aerial Survey Co. Ltd and the National Geographic slope angle (Figure 5) and specific contributing areas for groundwater
Information Institute (NGII) of Korea. Aerial photographs taken before flow were calculated. Further, the soil thickness in the study area was
and after the landslides were used to detect event locations. The analog acquired from 1:25,000 scale digital soil maps produced by the
‘before’ photos were taken on 4 April 2005 and the ‘after’ photographs National Institute of Agricultural Science (http://soil.rda.go.kr) on
were taken on 27 May 2008 using the UltraCam-X sensor (Figure 4). the basis of field observations. The soil thickness was evaluated
In this study, complementary field surveys were used to verify from the depth to bedrock and used as a soil depth (z) in the infinite
landslide locations indicated by the aerial photographs. A total of 1480 slope model in Eq. (2) (Figure 6). In addition, a geological map

Table 1
Geological description of lithology in the study area.

Era Period System Formation Description

Cenozoic Quaternary Quaternary Sand, gravel, clay


Mesozoic Jurassic Felsite –
Imgye granite Biotite granite (porphyritic and schistose in parts)
Triassic Pyeongan Nokam Grayish-green fine sandstone, gray sandy shale
Gobangsan Gray, milky-white coarse sandstone and sandy shale
Unknown Sambangsan Light-gray quartzite, dark-gray medium to coarse sandstone
Paleozoic Ordovician Joseon Jeongseon limestone Gray, dark-gray, milky-white limestone intercalated with thin
layers of gray shale and sandstone
Precambrian Biotite gneiss Banded biotite gneiss, migmatitic gneiss, schist, quartzite
H.J. Park et al. / Engineering Geology 161 (2013) 1–15 5

Fig. 3. Landslides triggered by heavy rainfall on July 15–16, 2006.

(Geological Society of Korea, 1962) was used to obtain the contribut- for cohesion and friction angle were scattered (see Figure 9). Thus, in
ing lithology (Figure 7). the probabilistic analysis cohesion and friction angle were considered
In this study, the requisite geotechnical and hydrogeological soil pa- as random variables. To quantify their random properties, the means
rameters were obtained from laboratory tests for the study area. Soil and standard deviations for cohesion and friction angle were evaluated
samples were collected from the landslide occurrence locations in from the results of laboratory testing (Table 2). However, the probabil-
each geological unit, as shown in Fig. 7, and a 75-mm diameter alumi- ity density functions of strength parameters for each soil type could not
num tube sampler was used for sampling (Figure 8). For each sampling be determined from the test results because the data were insufficient
location, six to nine soil samples were obtained for the laboratory tests. for determining the distribution function. Therefore, the probability
A total of 36 direct shear tests were performed to obtain shear strength density functions were estimated based on judgment, experience, or re-
parameters, such as cohesion and internal friction angle, for each soil sults published by others (Hoek, 2007). Previous studies have noted
type. Laboratory permeability tests were also performed for hydraulic that cohesion and friction angle were usually normally distributed
conductivity of the ground materials. Twelve constant head tests and (Vanmarcke, 1977; Mostyn and Li, 1993; Lacasse and Nadim, 1996;
32 falling head tests were performed to obtain the hydraulic conductiv- Terlien, 1996; Nilsen, 2000; Zhou et al., 2003; Pathak and Nilsen,
ity of each soil type. The unit weights of each soil type were also 2004; Park et al., 2005; Liu and Wu, 2008; Zolfaghari and Heath, 2008;
obtained from laboratory tests. However, as mentioned in previous Wang et al., 2010; Melchiorre and Frattini, 2011; Park et al., 2012).
works (Chowdhury and Flentje, 2003; Zhou et al., 2003; Xie et al., Thus, a normal distribution was used as the probability density function
2004; Shou and Chen, 2005; Huang et al., 2006; Zolfaghari and Heath, for cohesion and friction angle in this study.
2008; Griffiths et al., 2011), cohesion and friction angle of slope mate- The build-up of soil pore water pressure is the major triggering
rials were considered to be the major sources of uncertainty because mechanism for slope failure, so groundwater level data are required
of spatial variability and limited sampling, and subsequently the values for the infinite slope model. The groundwater level is calculated from
6 H.J. Park et al. / Engineering Geology 161 (2013) 1–15

Fig. 5. Map showing the distribution of slope angles.

4.3. Simulation procedure

Fig. 11 presents a flowchart showing the sequential steps in the


probabilistic analysis procedure. The study area was divided into
10 m × 10 m pixels, and the input data were converted to form a
10 m × 10 m grid using Arc/GIS software. These data sets and their ran-
dom properties (Table 2) were used, along with Monte Carlo simula-
tion, to evaluate the probability of failure. This simulation procedure
was coded and implemented in MATLAB software. The sequence is
listed below:
1. The groundwater level (m) for each pixel was calculated using the
hydrogeological model, with rainfall intensity (10 mm/h) and
hydraulic conductivity of the soil along with slope geometry.
2. Using cumulative distribution for each random variable and a ran-
dom number generator, cohesion and friction angle were randomly
generated. This step could be accomplished systematically for each
variable by first generating a random number between 0 and 1,
and then obtaining the corresponding random number with the

Fig. 4. Identification of landslide occurrence based on aerial photographs (A) before


and (B) after the landslide and (C) photographs from the field survey.

the hydrogeological model using rainfall intensity and hydraulic con-


ductivity data. Therefore, the precipitation input for the hydrogeological
model, such as rainfall intensity, is one of the most important parame-
ters to be obtained. Rainfall intensity values for the hydrogeological
model were obtained from Jinbu AWS (automatic weather system),
which is located at latitude 37°10′52″N and longitude 128°27′26″E,
because this is the closest and most representative rain gage in the
study area. The average rainfall intensity for the study area was obtained
from the hourly rainfall records in Fig. 10 by taking rainfall records on
average during the rainstorm on 15–16 July 2006, and 10 mm/h was
evaluated. Fig. 6. Map showing the distribution of soil thickness.
H.J. Park et al. / Engineering Geology 161 (2013) 1–15 7

achieve an acceptable level of accuracy in simulation results was


determined to be 5000.
5. The probability of failure was evaluated. After performing a suffi-
cient number of iterations for steps 2 and 3, the probability of fail-
ure was determined. The probability of failure is the ratio of the
number of the iterations where the factor of safety is less than 1
to the total number of iterations evaluated.
6. The above calculation was repeated for all other pixels and the spa-
tial distribution map for the probability of failure was produced.

4.4. Evaluation of model performance

In landslide susceptibility analysis, model validation is a fundamental


step in determining model performance. Validation refers to comparing
the model predictions with the real-world data set to assess the accura-
cy or predictive power (Begueria, 2006). For model validation, previous
landslide susceptibility studies used the success rate, defined as the ratio
Fig. 7. Geological map and sampling location.
of successfully predicted landslides to the total number of actual land-
slides (Montgomery and Dietrich, 1994; Dietrich et al., 1995; Borga et
al., 1998, 2002; Duan and Grant, 2000; Lee, 2005). That is, the success
cumulative distribution function for each random variable through
rate focuses on the model's ability to correctly identify landslide occur-
appropriate transformations.
rence. However, an ideal landslide susceptibility map simultaneously
3. The random values for cohesion and friction angle obtained in the
maximizes the agreement between known and predicted landslide loca-
previous step were used as input data for the calculation of the fac-
tions and minimizes the area outside the known landslide locations that
tor of safety. Eq. (2) was used to evaluate the factor of safety with
is predicted to be unstable (Godt et al., 2008). Therefore, in this study a
randomly generated strength parameters and deterministic input
receiver operating characteristic (ROC) graph was used to assess the
parameters (e.g., soil thickness, slope angle and unit weight of
performance of models, because the ROC graph is a very useful tool for
soil).
visualizing and evaluating model performance and provides a diagnostic
4. The operations described in steps 2 and 3 were repeated until a
that may be used to distinguish between two classes (true class and
sufficiently large number could be obtained for the factor of safety.
modeled class) of events (Swets, 1988; Fawcett, 2006). In ROC analysis,
Based on the error analysis, the number of iterations required to
true class instances (observation or landslide occurrence) are compared
with modeled class instances (landslide prediction) using a classifica-
tion model. The observations are classified as positive or negative
depending on whether they indicate occurrence or nonoccurrence of a
landslide. Predictions are classified as yes or no depending on whether
they predicted ‘unstable’ or ‘stable’ conditions (Cepeda et al., 2010).
Fig. 12 presents the outcomes of a classification model using a confusion
matrix. Therefore, the grid-based prediction results of landslide suscep-
tibility obtained from the proposed analysis could be compared with the
landslide inventory, and one of four outcomes was assigned to each grid
cell. If the grid cell was predicted as unstable and was coincident with a
landslide occurrence in the inventory, it was counted as a true positive.
However, if the grid cell was predicted as unstable and it did not coin-
cide with a landslide location, it was counted as a false positive. If the
cell was predicted to be stable and was coincident with the area outside
the known landslide location, it was counted as a true negative, and if it
fell on the landslide occurrence grid, it was counted as a false negative
(Fawcett, 2006; Godt et al., 2008; Cepeda et al., 2010). Based on the con-
fusion matrix, two important measures of model performance
(true-positive rate and false-positive rate) are commonly used for
model validation. The true-positive rate (TPR) is the ratio of the number
of true positives to the total number of positives, and the false-positive
rate (FPR) is the ratio of the number of false positives to the total number
of negatives. These two metrics are used to plot the ROC graph, which is
a two-dimensional graph in which the TPR is plotted on the y-axis and
the FPR on the x-axis. Model results that plot toward the upper left of
the graph are generally considered superior on the ROC graph (Godt et
al., 2008). The origin of the graph (0, 0) represents a model result with
maximum underestimate. In contrast, the upper right of the graph (1,
1) represents the maximum overestimate, which means that the entire
area is predicted to be unstable. A perfect prediction would be located at
the upper left (0, 1). Since an ROC curve is a two-dimensional depiction
of model performance, a single scalar value representing the expected
performance is needed to compare model performance. A common
method is to calculate the area under the curve (AUC) (Hanley and
Fig. 8. Sampling procedure using aluminum tube samplers. McNeil, 1982; Bradley, 1997). Since the AUC is a portion of the area of
8 H.J. Park et al. / Engineering Geology 161 (2013) 1–15

Fig. 9. Examples of distribution graphs of strength parameters in two major geological units: (A) friction angles for Imgye granite, (B) friction angles for sandstone in Nokam
formation, (C) cohesion for Imgye granite and (D) cohesion for sandstone in Nokam formation.

the unit square, its value will always be between 0 and 1, and the larger factor of safety distribution. Fig. 13 shows the spatial distribution map
the area, the better the model's performance. Another discrete classifier for the probability of failure calculated using the coupled infinite slope
is the distance to perfect classification (Cepeda et al., 2010): model with the hydrogeological model and Monte Carlo simulation.
The area ratios for probabilities of failure of less than 10%, 10%–50%
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
and greater than 50% were 79.4%, 6.4% and 14.2%, respectively.
r¼ ð1−TPRÞ2 þ FPR2 : ð3Þ
When considering slope stability problems in engineering geology,
correct determination of the level of safety for a slope is needed. In
The smaller the r values, the better the model's performance. deterministic analysis, a value of FS = 1.0 is considered to be the cri-
terion that the safety of the slope can be evaluated. That is, a computed
5. Results factor of safety of less than 1.0 implies an unstable slope. However,
since there are many uncertainties that are involved in and affect the
The probabilistic analysis was conducted following the procedure analysis procedure, we need to estimate the probability of whether
described in Fig. 11, using Monte Carlo simulation with the input pa- the slope will be unstable in order to determine the level of slope
rameters listed in Table 2. First, the groundwater levels for each pixel safety in the probabilistic analysis (Silva et al., 2008). Several different
were calculated using the hydrogeological model and hydraulic input criterion values for natural slopes and landslides have been suggested
parameters. Then, the factor of safety was evaluated for each pixel for the probabilistic analysis. Ko Ko et al. (2004) suggested a failure
using the randomly generated cohesion and friction angle, slope geom- probability of 20% as the threshold for unstable areas, and Chowdhury
etry and groundwater levels. Based on the repeated factor of safety cal- and Flentje (2003) suggested 15% for moderate to steeply inclined nat-
culations, the probability of failure for each pixel was obtained from the ural slopes. However, a probability of failure of 10% has been used as the

Table 2
Characteristics of input parameters in the study area.

Geological formation Friction angle (°) Cohesion (kPa) Unit weight (kN/m3) Hydraulic conductivity (m/h) Soil classification

Mean COV (%) Mean COV (%)

Felsite 20.8 9.9 17.5 8.6 17.8 0.171 SP


Quartzite in Sambangsan 40.6 1.5 8.0 25.0 19.3 0.096 SW
Imgye granite 35.2 4.5 3.8 28.9 23.2 0.089 SP
Jeongseon limestone 28.4 8.0 4.4 14.5 17.9 0.019 SP
Sandstone in Nokam 40.2 2.4 10.4 13.5 18.4 0.090 SP
Sandstone in Gobangsan 37.1 1.6 7.8 13.8 18.7 0.084 SW
H.J. Park et al. / Engineering Geology 161 (2013) 1–15 9

of correctly predicted landslide grids (true positives) to the total


number of landslide occurrence grids (positives), was calculated as
0.734. In addition, 20.6% of non-landslide grids were predicted as un-
stable, which means that the FPR was 0.206 (Table 3). Next, this result
was plotted on the ROC graph (point D in Figure 14) to depict the rel-
ative trade-off between true positives and false positives. The AUC
and distance to perfect classification were evaluated from the ROC
graph, and their values were 0.764 and 0.336, respectively.
However, the coefficients of variation (COVs) for the friction an-
gles obtained from laboratory tests and used for the previous analysis
ranged from 1.5% to 9.9%, and these values were lower than those
(10%–15%) reported in previous studies (Schultze, 1975; Lee et al.,
1983; Harr, 1987). Moreover, the number of laboratory tests carried
out to determine the friction angle in this study was small relative
to the size of the study area, so it was expected that a higher uncer-
Fig. 10. Records of hourly and total rainfall in the study area.
tainty would be involved in this probabilistic analysis procedure.
Therefore, higher COVs than those used for previous probabilistic
analyses needed to be considered, and the probabilities of slope fail-
ure were evaluated with an increased COV of the friction angle
criterion for unstable slopes in probabilistic design and stability analysis (15%) to account for the higher uncertainties. Fig. 15 shows the spa-
procedures in several studies (Priest and Brown, 1983; AGS, 2000; Fell tial distribution map for the probability of failure using an increased
et al., 2008; Liu and Wu, 2008; Silva et al., 2008). Consequently, a prob- COV for the friction angle. Of the mapped landslides, 79.9% of the inven-
ability of slope failure of 10% was used as the criterion for unstable tory was evaluated as unstable, and 24.4% of the non-landslide grids
slopes in this study, and a grid cell with a probability of failure greater were predicted to be unstable. That is, the evaluated TPR and FPR
than 10% was predicted to be unstable. were 0.799 and 0.244, respectively (Table 3). Consequently, both TPR
In the results of the probabilistic infinite slope model analysis, and FPR values obtained from the analysis with increased COVs were
1086 of the 1480 grids where actual landslides occurred had proba- higher than those obtained using laboratory tests, meaning that the
bilities of failure greater than 10%. Thus, 73.4% of actual landslides probabilistic analysis with increased COV captured more landslide oc-
(or the mapped landslides in the inventory map) were evaluated currence and predicted a larger portion of the study area to be unstable
as unstable, meaning that the TPR, which is the ratio of the number than the laboratory test results. The evaluated TPR and FPR were plotted

Start

Geometric and geotechnical parameters


of pixel

Geometry of Evaluate groundwater level using Rainfall


pixel hydrogeological model

Randomly generate geotechnical


parameters

Calculate factor of safety of pixel

No. of No
iterations =
Increase # of iterations

Yes
Calculate probability of failure for pixel
Increase # of pixels

Collect probability of failure of each pixel


to generate Pf map

Stop

Fig. 11. Flow chart for evaluating slope failure probability.


10 H.J. Park et al. / Engineering Geology 161 (2013) 1–15

True class (observations)

Positives Negatives

Yes True Positives False Positives


Modeled class
(predictions)
No False Negatives True Negatives

P N

Fig. 12. Confusion matrix showing outcomes of a classification model.

together with the previous results at point E in Fig. 14 for comparison level was increased to the full thickness of the soil, the TPR and FPR
with the results from the probabilistic analysis using random properties values were evaluated and plotted on the ROC graph at point B in
of laboratory test results. The AUC (0.780) and distance to perfect clas- Fig. 14. The evaluated TPR and FPR were 0.601 and 0.169, respectively.
sification (0.316) for the analysis with increased COVs were evaluated Then, the AUC and distance to perfect classification for the fully saturated
(Table 3). These results indicated that the probabilistic analysis with in- slope model were evaluated as 0.716 and 0.433, respectively. As can be
creased COV of the friction angle showed better performance. This observed in Fig. 14, the results of the probabilistic analysis were located
shows that proper consideration and understanding of uncertainties more toward the upper left corner of the ROC graph. Moreover, the AUC
play an important role in accurately predicting shallow landslide values of the probabilistic analysis were larger than those of the deter-
susceptibility. ministic analysis, and the distances to perfect classification of the proba-
To compare the results of the probabilistic and deterministic analy- bilistic analysis were smaller than those of the deterministic analysis.
ses, a deterministic analysis was also conducted on the basis of the fac- Consequently, the probabilistic analysis showed better performance
tor of safety concept. In the deterministic analysis, the infinite slope than the deterministic analysis.
model without the coupled hydrogeological model was used, and the To check the effect of the hydrogeological model, an additional
same deterministic input parameters (e.g., slope angle and soil thick- deterministic analysis was conducted with consideration of the
ness) were used to evaluate the factor of safety. In addition, mean hydrogeological model. In this case, the groundwater level was eval-
values of the random variable (e.g., cohesion and friction angle) were uated from the hydrogeological model using rainfall intensity and
used as the representative single values for the deterministic analysis. hydraulic conductivity of the ground for each grid cell, and the factors
Moreover, since the hydrogeological model was not employed in the of safety were calculated deterministically for each cell using the
deterministic analysis, the groundwater level was considered as a calculated groundwater level. Fig. 17 shows the spatial distribution
fixed single value without considering rainfall intensity and hydraulic map for the factor of safety evaluated with the hydrogeological
conductivity of the ground. That is, the groundwater level value (m) model. The TPR and FPR were evaluated as 0.603 and 0.169, respec-
was set to constant values for half-saturated (m = 0.5) and fully satu- tively (point C in Figure 14). The AUC and distance to perfect classifi-
rated soil (m = 1.0). Fig. 16 shows the spatial distribution maps for cation were 0.717 and 0.431, respectively. As can be seen in Figs 16
the factor of safety in the deterministic analysis results when m values and 17 and Table 3, the results of the deterministic analysis with
were 0.5 and 1.0. For half-saturated soil, only 403 of 1480 actual land- the hydrogeological model were almost identical to the results of
slide grids were predicted as unstable (i.e., factor of safety less than the deterministic analysis with m = 1.0. This is because the value of
1.0), which means that TPR was calculated as 0.273 (Table 3). The FPR 10 mm/h rainfall intensity that was used in the hydrogeological
was also evaluated as 0.082, as listed in Table 3, and the obtained TPR model causes almost full saturation of the study area (Figure 18).
and FPR were plotted on an ROC graph (point A in Figure 14) for com- However, if the performance of the probabilistic analysis coupled
parison. The AUC and distance to perfect classification for the determin- with the hydrogeological model was compared with the performance
istic analysis of the half-saturated slope model were evaluated as 0.595 of the deterministic analysis coupled with the hydrogeological model,
and 0.732, respectively. In addition, in the case where the groundwater the point of the probabilistic analysis in the ROC graph would be
located closer to the (0, 1) position than that of the deterministic
analysis. In addition, the AUC value of the probabilistic analysis was
higher and the distance to perfect classification of the probabilistic

Table 3
Comparison of model performance for the deterministic analyses with different
groundwater levels and the probabilistic analyses with different COVs of fiction angle.

Model True-positive False-positive Area Distance to


rate (TPR) rate (FPR) under the perfect
curve classification
(AUC) (r)

Deterministic analysis 0.273 0.082 0.595 0.732


with m = 0.5
Deterministic analysis 0.601 0.169 0.716 0.433
with m = 1.0
Deterministic analysis 0.603 0.169 0.717 0.431
with hydrogeological
model
Probabilistic analysis 0.734 0.206 0.764 0.336
with COV of friction
angle = 1.5–9.9%
Probabilistic analysis 0.799 0.244 0.780 0.316
with COV of friction
Fig. 13. Map showing the slope failure probability predicted using the COV of friction
angle = 15%
angle obtained from field data.
H.J. Park et al. / Engineering Geology 161 (2013) 1–15 11

Fig. 14. ROC graph comparing the analysis results. A: deterministic analysis with m =
0.5, B: deterministic analysis with m = 1.0, C: deterministic analysis with the
hydrogeological model, D: probabilistic analysis with COV of friction angle =
1.5–9.9%, E: probabilistic analysis with COV of friction angle = 15%.

analysis was lower. Consequently, when the results of the probabilistic


analysis coupled with the hydrogeological model were compared with
those of the deterministic analysis coupled with the hydrogeological
model, the performance of the probabilistic analysis was superior.

6. Discussion

To evaluate the susceptibility of rainfall-induced shallow landslides,


this study adopted the probabilistic approach in a physically based
model analysis. The probabilistic approach demonstrated good predic-
tive performance regarding landslide occurrence locations in the land-
slide inventory over the study area. Moreover, the results of the
probabilistic analysis showed higher AUC and lower distance to perfect
classification values compared with the deterministic analysis, meaning
that the probabilistic analysis showed better performance than the de-
terministic analysis. However, there are some limitations in performing
Fig. 16. Map showing the factor of safety predicted for (A) m = 0.5 and (B) m = 1.0.
the probabilistic infinite slope model analysis and obtaining the input
parameters for the physically based model. Therefore, the shortcomings
need to be noted for future application because they may limit model
accuracy and should be improved.

6.1. Appropriateness of the hydrogeological model

In this study, the steady-state hydrogeological model was used to


evaluate the groundwater level (or pore water pressure) in the infinite
slope model. This is because we could obtain only a limited amount of
hydraulic information for the study area, and consequently could not
utilize the transient hydrogeological model. As can be observed in
Fig. 18, the analysis results showed that almost the entire study area
(92.7%) was evaluated as fully saturated (that is, m = 1.0) when the
hydrogeological model was used to calculate the groundwater level
with 10 mm/h rainfall intensity. Subsequently, as mentioned previ-
ously, the results of the deterministic analysis with the hydrogeological
model were almost identical to those of the deterministic analysis with
m = 1.0. We can consider two possibilities that caused the fully satu-
rated study area. First, 429 mm of rainfall over two days can be con-
sidered as an extreme event, taking into account an average annual
rainfall of 1475 mm in the study area. This means that almost 30% of
Fig. 15. Map showing the slope failure probability predicted using 15% COV for the fric- the average annual rainfall fell in two days. Consequently, we could
tion angle. say that the extreme case of heavy rainfall caused the full saturation
12 H.J. Park et al. / Engineering Geology 161 (2013) 1–15

of soil characteristics is required to perform the slope stability analysis.


However, sufficient soil information from landslide-prone areas has
rarely been obtained (Okimura, 1989; Derose et al., 1991), and soil
thickness is one characteristic for which information is difficult to ob-
tain. Therefore, many authors rely on simple solutions, such as consid-
ering a constant value for soil depth over the whole study area
(Montgomery and Dietrich, 1994; Khazai and Sitar, 2000) or assigning
a constant value for each geological formation (Savage et al., 2004).
Recently, more complex methods, such as multivariate statistical analy-
ses (Gessler et al., 2000; Tesfa et al., 2009), a process-based model
(Casadei et al., 2003; Pelletier and Rasmussen, 2009), a geomorphology-
based model (Catani et al., 2010; Segoni et al., 2012), or adoption of the
wetness index (Lee and Ho, 2009; Ho et al., 2012) have been proposed.
However, these complex methods are less frequently used in large-
scale stability analyses because they are very site specific and require
significant effort to be correctly applied and calibrated over large
areas (Segoni et al., 2012).
In this study, soil depth was obtained from digital soil maps at the
1:25,000 scale, and the average soil thickness for each soil type was
assigned to the study area. According to Melchiorre and Frattini (2011),
Fig. 17. Map showing the factor of safety predicted using a hydrogeological model. soil depth is the parameter that has the greatest influence on model un-
certainties, and is one of the variables that has the greatest effect on the
of the study area. As can be seen in Fig. 19, if groundwater levels were results of the analysis. Therefore, to improve the performance of the pro-
evaluated using 1, 2 and 5 mm/h rainfall intensities with the posed probabilistic model, more effort is required to obtain reliable soil
hydrogeological model, 35.0%, 48.3% and 78.6% of the study area, re- thickness data. Moreover, the soil thickness could be considered as a ran-
spectively, were evaluated as fully saturated compared with 92.7% of dom variable in the probabilistic analysis if we can obtain the random
the study area that was saturated when 10 mm/h rainfall intensity properties of soil depth distribution.
was used. In addition, if only steep slope areas in which landslides
could occur (slope angle greater than 15°) were considered, 14.1% 6.3. Dependency between strength parameters
(1 mm/h), 26.2% (2 mm/h) and 56.2% (5 mm/h) of the study area
were evaluated as fully saturated. Therefore, 10 mm/h rainfall could When dealing with more than one random variable in the probabilis-
be the major cause of full saturation, and this was an uncommon and tic analysis, uncertainty in one variable may be associated with uncertain-
extreme weather event in the study area. Another possibility is the lim- ty in another (Baecher and Christian, 2003). That is, the uncertainties may
itation of the steady-state hydrogeological model. The steady-state not be independent. Dependence among the uncertainties affects the
model assumed that the recharge was constant and equal to the rainfall results of probabilistic analyses, but these dependencies can be difficult
intensity, with the value calculated as the mean intensity during the to identify and to estimate. Therefore, many probabilistic analyses for
event (Frattini et al., 2004). Therefore, this model could not take into ac- slope stability have assumed that the random variables are independent
count transient variations in rainfall amounts during the rainstorm, and (Alonso, 1976; Chowdhury, 1984; Li and Lumb, 1987; Zhou et al., 2003;
consequently this has caused overestimation of the groundwater level. Park et al., 2005; Shou and Chen, 2005; Cassidy et al., 2008; Liu and Wu,
2008; Shou et al., 2009; Wang et al., 2010). This assumption is often
6.2. Soil depth made to simplify the analysis. Consequently, the random variables used
in the proposed probabilistic analysis in this study were also assumed
In this study, the infinite slope model was adopted as the physically to be independent. However, recently some researchers have shown
based model, and consequently information on the spatial distribution the effect of dependence between cohesion and internal friction angle
in probabilistic slope stability analyses (Yucemen et al., 1973; Low and
Tang, 1997; Liang et al., 1999; Cherubini, 2000; Hong and Roh, 2008;
Wu, 2008; Babu and Singh, 2009; Cho, 2010; Cho and Park, 2010). There-
fore, the assumption of independence needs critical evaluation, and con-
sequently the effect of dependence between strength parameters in
landslide susceptibility analysis was evaluated in this discussion. The cor-
relation coefficient between cohesion and friction angle is known to be −
0.5 from references (Mostyn and Li, 1993; Low and Tang, 1997;
Cherubini, 2000; Wu, 2008; Babu and Singh, 2009; Cho, 2010), and there-
fore a value of −0.5 was used as the correlation coefficient between cohe-
sion and friction angle in this study. The Monte Carlo simulation shown in
Fig. 11 was carried out with two correlated input variables. A negative
correlation implies that low cohesion values are associated with high
friction angle values and vice versa (Cho, 2010). According to Fenton
and Griffith (2003), the variance of shear strength is reduced if there
is negative correlation between cohesion and friction angle. Fig. 20
shows the spatial distribution map for the probability of failure using
correlated random properties of cohesion and friction angle obtained
from laboratory tests. The area ratios for probability of failure of less
than 10%, 10%–50% and greater than 50% were 79.1%, 6.5% and 14.4%, re-
Fig. 18. Ratio of groundwater level to soil depth calculated with the hydrogeological spectively, and these values were similar to the results of the analysis
model and 10 mm/h rainfall intensity. with uncorrelated variables. The TPR and FPR were evaluated to validate
H.J. Park et al. / Engineering Geology 161 (2013) 1–15 13

Fig. 19. Detailed map of groundwater level ratio evaluated with the hydrogeological model for rainfall intensities of (A) 1 mm/h, (B) 2 mm/h, (C) 5 mm/h and (D) 10 mm/h.

model performance, and their values were 0.739 and 0.208, respective- to evaluate the probability of failure using the infinite slope model
ly. The AUC and distance to perfect classification were 0.765 and 0.334, coupled with GIS and Monte Carlo simulation.
respectively, showing that the performance of probabilistic analysis • The proposed approach was applied to a practical example to evalu-
with correlated variables was similar to that with uncorrelated vari- ate its feasibility. To achieve this, landslide occurrence locations were
ables listed in Table 3. Therefore, the dependence between two random obtained from aerial photographs to construct a landslide inventory,
variables did not affect the results of the probabilistic infinite slope sta- and the strength parameters and hydraulic properties of slope mate-
bility analysis in this study area. rials were acquired from laboratory tests. The spatial database for
input parameters was constructed in a grid-based GIS environment.
7. Conclusions In addition, the average rainfall intensity was evaluated from hourly
rainfall records obtained from the nearby rain gage.
• This study adopted the probabilistic approach in a physically based • To evaluate the performance of the proposed probabilistic infinite
model analysis to deal properly with uncertainties caused by limited slope analysis, grid-based prediction results of the analysis were
information and inherent variability of material properties. In this compared with the landslide inventory using the ROC graph. Two
probabilistic analysis, the uncertainties in strength parameters of performance measures, TPR and FPR, were evaluated on the ROC
slope materials were considered and the analysis was implemented graph using a confusion matrix, and the scalar values representing
14 H.J. Park et al. / Engineering Geology 161 (2013) 1–15

Bradley, A.P., 1997. The use of the area under the ROC curve in the evaluation of
machine learning algorithm. Pattern Recognition 30, 1145–1159.
Carrara, A., Crosta, G.B., Frattini, P., 2008. Comparing models of debris-flow susceptibil-
ity in the alpine environment. Geomorphology 94 (3–4), 353–378.
Casadei, M., Dietrich, W.E., Miller, N.L., 2003. Testing a model for predicting the timing
and location of shallow landslide initiation in soil mantled landscapes. Earth
Surface Processes and Landforms 28, 925–950.
Cassidy, M.J., Uzielli, M., Lacasse, S., 2008. Probability risk assessment of landslides: a
case study at Finneidfjord. Canadian Geotechnical Journal 45, 1250–1267.
Catani, F., Segoni, S., Falorni, G., 2010. An empirical geomorphology based approach to
the spatial prediction of soil thickness at catchment scale. Water Resources
Research 46, W05508. http://dx.doi.org/10.1029/2008WR007450.
Cepeda, J., Chavez, J.A., Martinez, C.C., 2010. Procedure for the selection of runout
model parameters from landslide back analyses: application to the Metropolitan
area of San Salvador, El Salvador. Landslides 7, 105–116.
Cherubini, C., 2000. Reliability evaluation of shallow foundation bearing capacity on c,
ϕ soils. Canadian Geotechnical Journal 37, 264–269.
Chiang, S.H., Chang, K.T., 2011. The potential impact of climate change on typhoon-
triggered landslides in Taiwan, 2010–2099. Geomorphology 133, 143–151.
Cho, S.E., 2010. Probabilistic assessment of slope stability that considers the spatial var-
iability of soil properties. Journal of Geotechnical and Geoenvironmental Engineer-
ing 136, 975–984.
Cho, S.E., Park, H.C., 2010. Effect of spatial variability of crosscorrelated soil properties
on bearing capacity of strip footing. International Journal for Numerical and Ana-
lytical Methods in Geomechanics 34, 1–26.
Fig. 20. Map showing the slope failure probability predicted using correlated random Chowdhury, R., 1984. Recent developments in landslide studies: probabilistic methods,
variables. state-of-art report. International Symposium on Landslides, pp. 209–228.
Chowdhury, R., Flentje, P., 2003. Role of slope reliability analysis in landslide risk man-
agement. Bulletin of Engineering Geology and the Environment 62, 41–46.
expected performance (AUC and distance to perfect performance) Chowdhury, R., Flentje, P., Bhattacharya, G., 2010. Geotechnical Slope Analysis. CRC
were evaluated for model validation, because the ROC graph is a Press, p. 737.
two-dimensional depiction of performance in which it is not easy Coduto, D.P., Yeung, M.R., Kitch, W.A., 2010. Geotechnical Engineering: Principles and
Practices, 2nd ed. Pearson (794 pp.).
to compare the model's performance. Crosta, G.B., Frattini, R., 2003. Distributed modeling of shallow landslides triggered by
• Based on the ROC graph, the model's performance in probabilistic intense rainfall. Natural Hazards and Earth System Sciences 3, 81–93.
and deterministic analyses was evaluated and compared. The proba- D'Amato Avanzi, G., Falaschi, F., Giannecchini, R., Puccinelli, A., 2009. Soil slip suscepti-
bility assessment using mechanical–hydrological approach and GIS techniques: an
bilistic approach demonstrated good predictive performance com- application in the Apuan Alps. Natural Hazards 50, 591–603.
pared with the landslide inventory. Moreover, the results of the DeRose, R.C., Trustrum, N.A., Blaschke, P.M., 1991. Geomorphic change implied by reg-
probabilistic analysis showed higher AUC and lower distance to per- olith slope relationships on steepland hillslopes, Taranaki, New Zealand. Catena 18,
489–514.
fect classification than the deterministic analysis. Dietrich, W.E., Reiss, R., Hsu, M.L., Montgomery, D.B., 1995. A process based model for
colluvial soil depth and shallow landsliding using digital elevation data. Hydrolog-
ical Processes 9, 383–400.
Acknowledgements Dietrich, W.E., Bellugi, D., Real De Asua, R., 2001. Validation of the shallow landslide
model, SHALSTAB, for forest management. Water Science and Application 2,
195–227.
This work was supported by the National Research Foundation Duan, J., Grant, G.E., 2000. Shallow landslide delineation for steep forest watersheds based
of Korea (NRF) grants funded by Korea government (MEST) (No. on topographic attributes and probability analysis. In: Wilson, J.P., Gallant, J.C. (Eds.),
2010-0021314 and No. 201M3A2A1050984). Terrain Analysis — Principles and Applications. John Wiley & Sons, pp. 311–329.
Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recognition Letters 27,
861–874.
References Fell, R., Corominas, J., Bonnard, C., Cascini, L., Leroi, E., Savage, W.Z., 2008. Guidelines for
landslide susceptibility, hazard and risk zoning for land use planning. Engineering
Acharya, G., De Smedt, F., Long, N.T., 2006. Assessing landslide hazard in GIS: a case Geology 102, 85–98.
study from Rasuwa, Nepal. Bulletin of Engineering Geology and the Environments Fenton, G.A., Griffith, D.V., 2003. Bearing capacity prediction of spatially random c-ϕ
65, 99–107. soils. Canadian Geotechnical Journal 40, 54–65.
AGS, 2000. Landslide risk management concepts and guidelines. Australian Frattini, P., Crosta, G.B., Fusi, N., Negro, P.D., 2004. Shallow landslides in pyroclastic
Geomechanics 35, 49–92. soils: a distributed modeling approach for hazard assessment. Engineering Geology
Aloetti, P., Chowdhury, R., 1999. Landslide hazard assessment: summary review and 73, 277–295.
new perspectives. Bulletin of Engineering Geology and the Environments 58, Geological Society of Korea, 1962. Changdong–Hajinburi Geological Map Sheet. KIGAM.
21–44. Gessler, P.E., Chadwick, O.A., Charman, F., Althouse, L., Holmes, K., 2000. Modeling soil
Alonso, E.E., 1976. Risk analysis of slopes and its application to slopes in Canadian sen- landscape and ecosystem properties using terrain attributes. Soil Science Society of
sitive clay. Geotechnique 26, 453–472. America Journal 64, 2046–2056.
Apip, K., Takara, K., Yamashiki, Y., Sassa, K., Ibrahim, A.B., Fukuoka, H., 2010. A distrib- Godt, J.W., Baum, R.L., Savage, W.Z., Salciarini, D., Schulz, W.H., Harp, E.L., 2008.
uted hydrological–geotechnical model using satellite-derived rainfall estimates for Transient deterministic shallow landslide modeling: requirements for suscepti-
shallow landslide prediction system at a catchment scale. Landslides 7, 237–258. bility and hazard assessments in a GIS framework. Engineering Geology 102,
Babu, G.L.S., Singh, V., 2009. Reliability analysis of soil nail wall. Georisk 3, 44–54. 214–226.
Baecher, G.B., Christian, J.T., 2003. Reliability and Statistics in Geotechnical Engineering. Gokceoglu, C., Sonmez, H., Ercanoglu, M., 2000. Discontinuity controlled probabilistic
Wiley (605 pp.). slope failure risk maps of the Altindag (settlement) region in Turkey. Engineering
Baum, R.L., Savage, W.Z., Godt, J.W., 2002. TRIGRS: a Fortran program for transient rain- Geology 55, 277–296.
fall infiltration and grid based regional slope stability analysis. Open File Report, Gorsevski, P.V., 2002. Landslide Hazard Modeling Using GIS. (Ph.D. thesis) University of
02-424. USGS, Colorado (38 pp.). Idaho (240 pp.).
Baum, R.L., McKenna, J.P., Godt, J.W., Harp, E.L., McMullen, S.R., 2005. Hydrologic mon- Griffiths, D.V., Huang, J., Fenton, G.A., 2011. Probabilistic infinite slope analysis. Com-
itoring of landslide-prone coastal bluffs near Edmonds and Everett, Washington. puters and Geotechnics 38, 577–584.
U.S. Geological Survey Open-File Report. (42 pp.). Guimaraes, R.F., Montgomery, D.R., Greenberg, H.M., Fernandes, N.F., Gomes, R.A.T.,
Begueria, S., 2006. Validation and evaluation of predictive models in hazard assessment Carvalho Jr., O.A., 2003. Parameterization of soil properties for a model of topo-
and risk management. Natural Hazards 37, 315–329. graphic controls on shallow landsliding: application to Rio de Janeiro. Engineering
Beven, K.J., Kirkby, M.J., 1979. A physically based, variable contributing area model of Geology 69, 99–108.
basin hydrology. Hydrological Sciences Journal 24, 43–69. Hanley, J.A., McNeil, B.J., 1982. The meaning and use of the area under a receiver oper-
Borga, M., Fontana, G.D., Marchi, L., 1998. Shallow landslide hazard assessment using a ating characteristic (ROC) curve. Radiology 143, 29–36.
physically based model and digital elevation data. Environmental Geology 32, Harp, E.L., Reid, M.E., McKenna, J.P., Michael, J.A., 2009. Mapping of hazard from
81–88. rainfall-triggered landslides in developing countries: examples from Honduras
Borga, M., Dalla Fontana, G., Gregoretti, C., Marchi, L., 2002. Assessment of shallow and Micronesia. Engineering Geology 104, 295–311.
landsliding by using a physically based model of hillslope stability. Hydrological Harr, M.E., 1987. Reliability Based on Design in Civil Engineering. McGraw-Hill, New
Processes 16, 2833–2851. York, p. 290.
H.J. Park et al. / Engineering Geology 161 (2013) 1–15 15

Ho, J.Y., Lee, K.T., Chang, T.C., Wang, Z.Y., Liao, Y.H., 2012. Influence of spatial distribu- Priest, S.D., Brown, E.T., 1983. Probabilistic stability analysis of variable rock slopes.
tion of soil thickness on shallow landslide prediction. Engineering Geology 124, Transactions of the Institution of Mining and Metallurgy 92, A1–A2.
38–46. Reid, M.E., Ellen, S.D., Brien, D.L., Fuente, J., Falls, J.N., Hicks, B.G., Johnson, E.C., 2007.
Hoek, E., 2007. Rock engineering; course note by Evert Hoek [online]. http://www. Predicting debris slide locations in northwestern California. USDA Forest Service
rockeng.utoronto.ca/Hoekcorner.htm. Gen. Tech. Rep. PSW-GTR-194.
Hong, H.P., Roh, G., 2008. Reliability evaluation of earth slope. Journal of Geotechnical Rosso, R., Rulli, M.C., Vannucchi, G., 2006. A physically based model for the hydrologic
and Geoenvironmental Engineering 134, 1700–1705. control on shallow landsliding. Water Resources Research 42, W06410. http://
Hsu, M.L., 1998. A grid based model for predicting shallow landslides: a case study in dx.doi.org/10.1029/2005WR004369.
Linkou, Taipei. EOS. Transactions of the American Geophysical Union 79, W 25. Salciarini, D., Godt, J.W., Savage, W.Z., Conversini, P., Baum, R.L., Michael, J.A., 2006.
Huang, J.C., Kao, S.J., Hsu, M.L., Lin, J.C., 2006. Stochastic procedure to extract and to in- Modeling regional initiation of rainfall induced shallow landslides in the eastern
tegrate landslide susceptibility maps: an example of mountainous watershed in Umbria Regional of Central Italy. Landslides 3, 181–194.
Taiwan. Natural Hazards and Earth System Sciences 6, 803–815. Salciarini, D., Godt, J.W., Savage, W.Z., Baum, R.L., Conversini, P., 2008. Modeling land-
Huang, J.C., Kao, S.J., Hsu, M.L., Lin, J.C., 2007. Influence of specific contributing area al- slide recurrence in Seattle, Washington, USA. Engineering Geology 102, 227–237.
gorithm on slope failure prediction in landslide modeling. Natural Hazards and Santoso, A.M., Phoon, K.K., Quek, S.T., 2011. Effects of soil spatial variability on rainfall-
Earth System Sciences 7, 781–792. induced landslides. Computers and Structures 89, 893–900.
Iverson, R.M., 2000. Landslide triggering by rain infiltration. Water Resources Research Savage, W.Z., Godt, J.W., Baum, R.L., 2004. Modeling time dependent areal slope stabil-
36, 1897–1910. ity. In: Lacera, W.A., Erlich, M., Fontoura, S.A.B., Sayao, A.S.F. (Eds.), Landslide
Kamai, T., 1991. Slope stability assessment by using GIS. Science and Technology Agency Evaluation and Stabilization. Proceedings of 9th International Symposium on Land-
of Japan, in Japanese. slides. Balkema, Rotterdam, pp. 23–36.
Khazai, B., Sitar, N., 2000. Assessment of seismic slope stability using GIS modeling. Schultze, E., 1975. The general significance of statistics for the civil engineer. Proceed-
Geographic Information Sciences 6, 121–128. ings of the 2nd International Conference on Application of Statistics and Probabil-
KIGAM (Korea Institute of Geoscience and Mineral Resources), 2009. Development of ity in Soil and Structural Engineering, Aachen.
Landslide Prediction Technology and Damage Mitigation Countermeasures. (203 pp.). Segoni, S., Rossi, G., Catani, F., 2012. Improving basin scale shallow modelling using re-
Ko Ko, C., Flentje, P., Chowdhury, R., 2004. Landslide qualitative hazard and risk assess- liable soil thickness maps. Natural Hazards 61, 85–101.
ment method and its reliability. Bulletin of Engineering Geology and the Environ- Shou, K., Chen, Y., 2005. Spatial risk analysis of Li-shan landslide in Taiwan. Engineering
ment 63, 149–165. Geology 80, 199–213.
Korean Geotechnical Society, 2011. Final Report on Mt. Woomyun Landslide: Causes Shou, K., Chen, Y., Liu, H., 2009. Hazard analysis of Li-shan landslide in Taiwan.
and Recommendation of Countermeasure (262 pp., in Korean). Geomorphology 103, 143–153.
Lacasse, S., Nadim, F., 1996. Uncertainties in characterizing soil properties. Geotechni- Silva, F., Lambe, T.W., Marr, W.A., 2008. Probability and risk of slope failure. Journal of
cal Special Publication 58, 49–75. Geotechnical and Geoenvironmental Engineering 134, 1691–1699.
Lee, D.S., 1988. Geology of Korea (Kyohak-Sa, Seoul). Soeters, R., van Westen, C.J., 1996. Slope instability recognition, analysis and zonation.
Lee, S., 2005. Application and cross validation of spatial logistic multiple regression for In: Turner, A.K., Schuster, R.L. (Eds.), Landslides Investigation and Mitigation, TRB,
landslide susceptibility analysis. Geosciences Journal 9, 63–71. Special Report, 247. National Academy Press, pp. 129–177.
Lee, K.T., Ho, J.Y., 2009. Prediction of landslide occurrence based on slope instability Sorbino, G., Sica, C., Cascini, L., Cuomo, S., 2007. On the forecasting of flowslide trigger-
analysis and hydrological model simulation. Journal of Hydrology 375, 489–497. ing areas using physically based models. Proc. of 1st North American Landslides
Lee, I.K., White, W., Ingles, O.G., 1983. Geotechnical Engineering. Pitman, Boston. Conference: AEG Publication, 23, pp. 305–315.
Li, Y.F., Chi, Y.Y., 2011. Rainfall induced landslide risk at Lushan, Taiwan. Engineering Sorbino, G., Sica, C., Cascini, L., 2010. Susceptibility analysis of shallow landslides
Geology 123, 113–121. sources area using physically based models. Natural Hazards 53, 313–332.
Li, K.S., Lumb, P., 1987. Probabilistic design of slopes. Canadian Geotechnical Journal 24, Swets, J., 1988. Measuring the accuracy of diagnostic systems. Science 240, 1285–1293.
520–535. Terlien, M.T.J., 1996. Modeling spatial and temporal variations in rainfall triggered
Liang, R.Y., Nusier, O.K., Malkawi, A.H., 1999. A reliability based approach for evaluating landslide. International Institute for Aerospace Survey and Earth Science, Publica-
the slope stability of embankment dams. Engineering Geology 54, 271–285. tion No. 32.
Liu, C., Wu, C., 2008. Mapping susceptibility of rainfall-triggered shallow landslides Terlien, M.T.J., van Westen, C.J., van Asch, T.W.J., 1995. Deterministic modeling in GIS
using a probabilistic approach. Environmental Geology 55, 907–915. based landslide hazard assessment. In: Carrara, A., Guzzetti, F. (Eds.), Geographical
Low, B.K., Tang, W.H., 1997. Reliability analysis of reinforced embankments on soft Information Systems in Assessing Natural Hazards. Kluwer Academic Publisher,
ground. Canadian Geotechnical Journal 34, 672–685. Dordrecht, pp. 57–77.
Luzi, I., Pergalani, F., 1996. Application of statistical and GIS techniques to slope insta- Tesfa, T.K., Tarboton, D.G., Chandler, D.G., McNamara, J.P., 2009. Modeling soil depth
bility zonation. Soil Dynamic and Earthquake Engineering 15, 83–94. from topographic and land cover attributes. Water Resources Research 10,
Melchiorre, C., Frattini, P., 2011. Modeling probability of rainfall-induced shallow land- W10438. http://dx.doi.org/10.1029/2008WR007474.
slides in a changing climate, Otta, Central Norway. Climate Change. http:// van Westen, C.J., 2000. The modeling of landslide hazard using GIS. Surveys in
dx.doi.org/10.1007/s10584-011-0325-0. Geophysics 21, 241–255.
Montgomery, D.R., Dietrich, W.E., 1994. A physically based model for the topographic van Westen, C.J., Terlien, M.T.J., 1996. An approach towards deterministic landslide
control on shallow landsliding. Water Resources Research 30, 1153–1171. hazard analysis in GIS; a case study from Manizales. Earth Surface Processes and
Mostyn, G.R., Li, K.S., 1993. Probabilistic slope analysis — state of play. Proceeding of Landform 21, 853–868.
Conference on Probabilistic Method in Geotechnical Engineering, pp. 89–109. van Westen, C.J., Seijmonsbergen, A.C., Mantovani, F., 1999. Comparing landslide
NIDP (National Institute for Disaster Prevention), 2000. Fundamental issues for land- hazard maps. Natural Hazards 20, 137–158.
slide hazard avoidance or mitigation plans. Research Report. (276 pp., in Korean). van Westen, C.J., Asch, T.W.J., Soeters, R., 2006. Landslide hazard and risk zonation —
Nilsen, B., 2000. New trend in rock slope stability analysis. Bulletin of Engineering why is it still so difficult? Bulletin of Engineering Geology and the Environment
Geology and the Environment 58, 173–178. 65, 67–184.
O'Loughlin, E.M., 1986. Prediction of surface saturation zones in natural catchments by Vanmarcke, E.H., 1977. Probabilistic modeling of soil profiles. Journal of Geotechnical
topographic analysis. Water Resources Research 22, 794–804. Engineering Division ASCE 1227–1246.
Okimura, T., 1989. Prediction of slope failure using the estimated depth of the potential Wang, Y., Cao, Z., Au, S., 2010. Efficient Monte Carlo Simulation of parameter sensitivity
failure layer. Journal of Natural Disaster Sciences 11, 67–79. in probabilistic slope stability analysis. Computers and Geotechnics 37, 1015–1022.
Pack, T.T., Tarboton, D.G., Goodwin, C.N., 1998. The SINMAP approach to terrain stabil- Wu, T.H., 2008. Reliability analysis of slopes. In: Phoon, K.K. (Ed.), Reliability Based
ity mapping. 8th Congress of the International Association of Engineering Geology, Design in Geotechnical Engineering. Taylor & Francis, pp. 413–447.
Vancouver, B.C., Canada, pp. 21–25. Wu, W., Sidle, R.C., 1995. A distributed slope stability model for steep forested basins.
Pack, T.T., Tarboton, D.G., Goodwin, C.N., 2001. Assessing terrain stability in a GIS using Water Resources Research 31, 2097–2110.
SINMAP. 15th Annual GIS Conference, GIS. Xie, M., Esaki, T., Zhou, G., 2004. GIS based probabilistic mapping of landslide hazard
Park, H.J., West, T.R., Woo, I., 2005. Probabilistic analysis of rock slope stability and ran- using a three dimensional deterministic model. Natural Hazards 33, 265–282.
dom properties of discontinuity parameters, Interstate Highway 40. Engineering Yucemen, M.S., Tang, W.H., Ang, A.H.S., 1973. A probabilistic study of safety and design
Geology 79, 230–250. of earth slopes. Civil Engineering Studies, Structural Research Series, 402 (Univer-
Park, H.J., Um, J.G., Woo, I., Kim, J.W., 2012. Application of fuzzy set theory to evaluate sity of Illinois, Urbana).
the probability of failure in rock slopes. Engineering Geology 125, 92–101. Zhou, G., Esaki, T., Mitani, Y., Xie, M., Mori, J., 2003. Spatial probabilistic modeling of
Pathak, S., Nilsen, B., 2004. Probabilistic rock slope stability analysis for Himalayan con- slope failure using an integrated GIS Monte Carlo simulation approach. Engineering
dition. Bulletin of Engineering Geology and the Environment 63, 25–32. Geology 68, 373–386.
Pelletier, J.D., Rasmussen, C., 2009. Geomorphically based predictive mapping of soil Zolfaghari, A., Heath, A.C., 2008. A GIS application for assessing landslide hazard over a
thickness in upland watersheds. Water Resources Research 45, W09417. http:// large area. Computers and Geotechnics 35, 278–285.
dx.doi.org/10.1029/2008WR007319.
Picarelli, L., Olivares, L., Avolio, B., 2008. Zoning for flowslide and debris flow in pyro-
clastic soils of Campania Region based on infinite slope analysis. Engineering
Geology 102, 132–141.

You might also like