Cases and solutions

Assessment of landslide
susceptibility on the natural terrain
of Lantau Island, Hong Kong
F.C. Dai 7 C.F. Lee 7 J. Li 7 Z.W. Xu

Introduction
Abstract Steep terrain and the high frequency of
tropical rainstorms make landslide occurrence on Landslides in mountainous terrains often occur as a re-
natural terrain a common phenomenon in Hong sult of heavy rainfall, resulting in the loss of life and
Kong. For example, more than 800 slope failures damage to the natural environment. Potential sites that
were triggered by a rainstorm in November 1993 are landslide-prone should therefore be identified in ad-
on Lantau Island, Hong Kong. Maps of recent land- vance to reduce such damage. In this regard, landslide
slides interpreted from aerial photographs, in com- hazard assessment can provide much of the basic infor-
bination with a geographical information system, mation essential for hazard mitigation through proper
were used to evaluate the frequency and distribu- project planning and implementation. Landslide hazard
tion of landslides, with particular reference to such was defined by Varnes (1984) as the probability of occur-
physical parameters as lithology, slope gradient, rence of a potentially damaging landslide phenomenon
slope aspect, elevation, vegetation cover, and prox- within a specified period of time and within a given area.
imity to drainage line, all of which are considered The factors that determine the landslide hazard of an
to be influential in the occurrence of landslides. A area may be grouped into two categories: (1) the intrinsic
stepwise logistic regression model was obtained be- variables that contribute to landslide susceptibility, such
tween landslide susceptibility and the above men- as geology, slope gradient, slope aspect, elevation, soil
tioned physical parameters. The study area has geotechnical properties, vegetation cover, and long-term
been classified into five classes of relative landslide drainage patterns; and (2) the extrinsic variables that
susceptibility, namely, very low, low, moderate, tend to trigger landslides in an area of given susceptibili-
high, and very high, based on this methodology. ty, such as heavy rainfall, and earthquakes (Wu and Sidle
1995; Atkinson and Massari 1998). Obviously, the proba-
Keywords Geographical information systems bility of landslide occurrence depends on both the intrin-
(GIS) 7 Landslides 7 Logistic regression modeling sic and extrinsic variables. However, the extrinsic varia-
bles may change over a very short time span, and are
thus very difficult to estimate. If extrinsic variables are
not taken into account, the term “susceptibility” could be
employed to define the likelihood of occurrence of a
landslide event. The spatial distribution of the intrinsic
variables within a given area determines the spatial dis-
tribution of relative landslide susceptibility in that region
Received: 17 December 1999 7 Accepted: 21 March 2000 (Carrara and others 1995).
F.C. Dai A variety of techniques, such as heuristic, statistical, and
Institute of Geographical Science and Natural Resources, deterministic approaches, have been developed to assess
Chinese Academy of Sciences, Beijing 100101, P.R. China landslide susceptibility. In heuristic approaches, expert
C.F. Lee (Y) opinions are used to estimate landslide potential from
Department of Civil Engineering, data on intrinsic variables. They are based on the as-
The University of Hong Kong, Pokfulam, Hong Kong, P.R. China sumption that the relationship between landslide suscep-
e-mail: leecf6hkucc.hku.hk tibility and the intrinsic variables are known and are
Tel.: c86-852-28592645 specified in the models. A set of variables are then en-
Fax: c86-852–2858061126 tered into the model to estimate landslide susceptibility
J. Li (Niemann and Howes 1991; Anbalagan 1992; Pachauri
Institute of Remote Sensing Applications, and Pant 1992; Atkinson and Massari 1998). One problem
Chinese Academy of Sciences, Beijing 100101, P.R. China with the heuristic models is that they need long-term in-
Z.W. Xu formation on the landslides and their causal factors for a
Institute of Geographical Science and Resources, similar geo-environmental condition or for the same site,
Chinese Academy of Sciences, Beijing 100101, P.R. China and these are, in most cases, not available. Another limi-

Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag 381

 Cases and solutions

tation in this method is in the reproducibility of results tual landslide cases and is therefore representative of fail-
and in the subjectivity in assigning weightings and rat- ure conditions and relatively objective. This methodology
ings to the variables. is then applied to the study area, i.e., Lantau Island
Deterministic approaches are based on slope stability (143 km 2) off Hong Kong. A flow chart of the methodolo-
analyses, and are only applicable when the ground condi- gy is shown in Fig. 1.
tions are fairly uniform across the study area and the
landslide types are known and relatively easy to analyze.
The infinite slope stability model has been widely used to
assess landslide susceptibility in small areas (Van Westen Description of the study area
1993; Terlien and others 1995; Wu and Sidle 1995). The
advantage of the deterministic models is that they permit Lantau Island, the largest outlying island within the terri-
quantitative factors of safety to be calculated, whereas the tory of Hong Kong, has been selected as a pilot study
main problem is the high degree of simplification that is area. This selection is made based on the (1) presence of
usually necessary for the use of such models. Statistical numerous landslides with a high spatial concentration;
models involve the statistical determination of the combi- (2) availability of existing datasets such as topographical
nations of variables that have led to landslide occurrence maps, geological maps, vegetation cover, and spatial dis-
in the past. Quantitative or semi-quantitative estimates tribution of landslides; (3) existence of steep terrain; and
are then made for areas currently free of landslides, but (4) undeveloped terrain that is most suitable for mapping
where similar conditions exist. Both simple and multivar- of landslide susceptibility using the GIS.
iate statistical approaches have been used widely in such Lantau Island is located in the southwest part of the ter-
an indirect mapping of landslide susceptibility (Yin and ritory of Hong Kong (Fig. 2). Largely because of the steep
Yan 1988; Gupta and Joshi 1989; Carrara and others 1991, terrain, the island is virtually undeveloped and uninha-
1995; Siddle and others 1991; Wang and Unwin 1992; Na- bited with the exception of small coastal patches of flat
ranjo and others 1994; Atkinson and Massari 1998). land. The land area with slope gradients 1 257 accounts
Statistical techniques are generally considered the most for 44% of the total land. Elevation ranges from the sea
appropriate approach for landslide susceptibility mapping level to over 900 m above sea level and changes abruptly.
at scales of 1 : 20,000 to 1 : 50,000 because, on this scale, it The bedrock geology of the study area is dominated by
is possible to map out in detail the occurrence of past Mesozoic volcanic rocks and younger intrusive igneous
landslides, and to collect sufficient information on the rocks (Table 1). The volcanic rocks, which comprise tuffs
variables that are considered to be relevant to the occur- and lavas with intercalated sedimentary rocks, crop out
rence of landslides (Naranjo and others 1994). Tradition- in the west and north of the study area. Intrusive rocks
ally, susceptibility assessment and mapping were labo- consist mainly of granites, and dykes of various composi-
rious and time-consuming because of the time and effort tions. The Paleozoic sedimentary strata comprising meta-
required for the manual handling and processing of the morphosed siltstone, sandstone, and carbonaceous silt-
data. In this aspect, the geographical information systems stone occur as a small outcrop in the northwest coastal
(GIS) have become an important tool for landslide sus- areas of the study area. Superficial deposits of the Qua-
ceptibility mapping. GIS is a general-purpose computer- ternary age form large, flat-lying areas. In hilly terrain,
based technology designed to capture, store, manipulate, colluvium, including debris flow and other slope debris
analyze and display diverse sets of spatial or geo-refer- deposits, mostly of Pleistocene age, commonly mantles
enced data. Goodchild (1993) states that the GIS technol- side slopes and valleys as a result of numerous individual
ogy has the ability to perform a variety of tasks: (1) pre- episodes of mass wasting and erosion during the Quater-
processing of data from large stores into a form suitable nary period. Small alluvial deposits occur in hilly areas,
for analysis exemplified by reformatting, change of pro- but alluvium is generally restricted to fans developed
jection, resampling, and generalization; (2) supporting downslope of the colluvial deposits. Beach deposits of
analysis and modeling: forms of analysis, calibration of sand usually form in front of alluvial deposits, especially
models, forecasting, and prediction; and (3) post-proc- in coastal bays deposited under the combined influence
essing of results through reformatting, tabulation, report of higher sea levels and fluctuating climatic conditions in
generation, and mapping. The overlay operation com- recent times (Geotechnical Control Office 1988a, b). A re-
monly applied within the GIS is useful in both heuristic golith, or mantle of weathered rock, occurs over most of
and statistical approaches (Gupta and Joshi 1989; Carrara the study area. The effects of weathering vary with rock
and others 1991, 1995; Wang and Unwin 1992; Mark and types and this is reflected in the topographic relief. Intru-
Ellen 1995; Van Westen and others 1997; Fernandez and sive rocks and the Paleozoic sedimentary rocks are most
others 1999). deeply weathered and eroded rocks, and form the lower
The purpose of this study is to present a method that ground. The acidic volcanic rocks are resistant to deep
utilizes GIS and employs statistical analysis to define the weathering and erosion. The area is structurally affected
physical parameters contributing to the occurrence of by two sets of faults trending NE–NNE and NNW–NW,
landslides. This method allows a series of statistically respectively.
meaningful and independent variables to be included in The foot-slope terrain where elevation is low and slope
the assessment model. The procedure is based on the ac- gradient is gentle is generally covered by natural woody

382 Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag

 Cases and solutions

Fig. 1
Flow chart of the methodology

forest, accounting for F27% of the total land, whereas occasionally intense during the rainstorms and typhoons.
the mid-slopes with moderate elevation and slope gra- Given the steep natural terrain mantled with a layer of
dient are covered by dense shrub and grass, occupying 25 superficial deposits and the frequent intense rainfall, it is
and 28%, respectively. Bedrock outcrops occur on the not surprising that landslides commonly occur on such
steep terrain with gradients exceeding 407, and on moun- natural terrain.
tain peaks. On 4–5 November 1993, the late-season passage of ty-
The climate is subtropical and monsoonal, with mild, dry phoon Ira dumped over 400 mm of rainfall in 24 h end-
winters and hot, humid summers. Rainfall is high, and ing at 10 : 00 a.m. on 5 November for most parts of Lan-
tau Island, with a maximum of over 700 mm in the Tung
Chung area where the rainfall was most intense (Wong
and others 1998). As a result of intense rainfall, over 800
landslides occurred on the natural terrain of Lantau Is-
land. The locations of all observable natural terrain land-
slides in the study area that took place during and prior
to the November 1993 rainstorm were identified with the
use of aerial photographs by the Geotechnical Engineer-
ing Office (GEO), Hong Kong (Wong and others 1998),
resulting in an average density of 6.1 landslides/km 2. The
locations of the landslide scars are shown in Fig. 3. Field
observations revealed that the landslides have the follow-
ing characteristics: (1) the volume of failure generally
ranged from tens of cubic meters to over a 1,000 m 3; (2)
the failures generally occurred along the colluvium–be-
drock contact, and the predominant failure mode is of
the translational type, a slipping of a thin layer of collu-
vium with a planar failure surface; and (3) most land-
slides started as slides and quickly converted to flows be-
cause of the water involved and the steep terrain below
Fig. 2 the debris sources (Wong and others 1998; Dai and oth-
Location of the study area ers 1999).

Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag 383

 Cases and solutions

Table 1
Brief description of solid rocks and superficial deposits of the study area

Age Lithology General lithological description

Quaternary Holocene Littoral deposits Essentially beach and dune sand with occasional
gravel horizons
Alluvium Generally loose brownish-gray silty sand with suban-
gular gravel. Occasionally contains cobble and
boulder horizons.
Pleistocene and Holocene Alluvial deposits Dense silty sand or clay with gravel
Colluvium Composed of sedimentary-derived, volcanic-derived,
granitic-derived and mixed colluvium. Generally
loose debris flow and other slope debris deposits
Pleistocene Colluvium Composed of a range of materials, which vary from
boulder colluvium, to gravelly colluvium with clay
and sand, to finer textured, gravelly sands and clay.
The deposits may be completely decomposed to a
mottled, colored sandy silt or clayey silt similar to
the in-situ residual deposits of their parent materi-
als.
Mesozoic Upper Jurassic Minor intrusive rocks Composed of quartz porphyry, feldspar porphyry,
quartz monzonite
Major intrusive rocks Composed of fine-grained, medium- to coarse-
grained, and coarse-grained granite and coarse-
grained porphyritic granitoid rock
Lower to Middle Jurassic Sedimentary, volcaniclastic Including sedimentary and water-laid volcaniclastic
and extrusive igneous rocks, acid lavas, coarse tuff, dominantly pyroclastic
rocks rocks with some lavas, and undifferentiated volcanic
rocks
Paleozoic Carboniferous Sedimentary rocks Composed of metamorphosed siltstone, sandstone
and carbonaceous siltstone

Factors affecting landslide of vegetation cover interpreted from the Satellite Pour
l’Observation de la Terra (SPOT) image with the aid of
susceptibility field checking by Wang and Chi (unpublished data). For
the purpose of this study, these vegetation types were
In this study, a digital elevation model (DEM) was gener- then simplified into four types, namely, (1) bare land, (2)
ated from the triangulated irregular network (TIN) model grassland, (3) shrubland, and (4) woodland. Field obser-
that was derived from digitized contours (20-m interval) vations indicated that six different physical parameters
of the 1 : 20,000-scale topographic maps. Slope gradient, were expected to contribute to landslide susceptibility
slope aspect, and elevation layers were derived from the within the study area. These parameters include lithology,
DEM. For drainage lines digitized from the same topo- slope gradient, slope aspect, elevation, vegetation cover,
graphic maps, buffer zones of 50, 100, 150, and 200 m and proximity to drainage line. The digital map of land-
were generated. The Hong Kong Geological Survey slide distribution was then overlaid on the above men-
1 : 20,000-scale solid and superficial geological maps cov- tioned geographic and geological parameters or data
ering the study area were used to identify the geological layers using the GIS, and bivariate relations were then
groups, with each group comprising geological units of produced.
broadly similar lithology. A geological data set was de-
rived by digitizing the 1 : 20,000-scale geologic maps. For Lithology
ease of analysis, the groups were further reclassified into Lithology exerts a fundamental control on the geomor-
seven categories of geological materials with similar engi- phology of a landscape. The nature and rate of geomor-
neering properties. They are: loose littoral or alluvial de- phological processes, including the landslide process, is
posits of the Holocene (H), alluvial deposits and sedi- partially dependent on the lithology and weathering char-
mentary-, volcanic-, or granitic-derived colluvium of the acteristics of the underlying materials. The correlation of
Holocene to Pleistocene (HP), volcanic-, sedimentary- the frequency of landslides with lithology is shown in
and granitic-derived colluvium of the Pleistocene (P), Fig. 4a. It can be seen that there are two geological cate-
volcanic rocks (V), sedimentary rock (S), intrusive ig- gories with relatively high landslide densities, namely,
neous rocks (I), and minor intrusive igneous rocks (MI). volcanic rocks (V) and sedimentary rocks (S), although
The vegetation cover data set was digitized from the the frequency of landslides is lowest in the loose deposits
1 : 50,000-scale vegetation cover paper map with 139 types of the Holocene (H), followed by minor intrusive igneous

384 Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag

 Cases and solutions

Fig. 3 erage, the frequency of landslides was greater on steep

Landslide scar distribution map slopes than on gentle slopes (Fig. 4b). Examination of the
distribution of landslide densities with the corresponding
slope gradient classes shows an increase with slope gra-
dient until the maximum density is reached in the 30–407
rock (MI), sediments of the Pleistocene (P), and intrusive class, followed by a gradual decrease in the 40F507 and
igneous rocks (I). It should be noted that thin colluvium 1 507 classes. This is a reflection that steep natural slopes
in steep terrain, which is most susceptible to landsliding, with outcropping bedrock and hence much higher shear
is not fully reflected in the geological maps used in this strength may not be susceptible to shallow landsliding
study, and can only be represented by lithological charac- since little colluvium or residual soil is available. On the
teristics of underlying bedrock. other hand, moderately steep terrain is often covered by
Structural information is also available from the digital a thin layer of colluvium, which is more susceptible to
geological maps. However, qualitative examination of rainfall-induced failure (Wong and others 1998; Dai and
spatial distributions suggests that the correlation between others 1999).
landslides and mapped linear structural features at the
1 : 20,000 scale is not good, and the structural information Slope aspect
is, thus, excluded in this study. The aspect of a slope (i.e., the direction that it faces) has
the potential to influence its physical properties and its
Slope gradient susceptibility to failure. The processes that may be oper-
Slope gradient is an essential component of slope stabili- ating include exposure to sunlight, drying winds, and,
ty analysis. As slope gradient increases, the level of gravi- possibly, rainfall (Evans and others 1999). The distribu-
tation-induced shear stress in the colluvium or residual tion of aspect among the mapped landslides is shown in
soils increases as well. Gentle hill slopes are expected to Fig. 4c. It can be seen that the frequency of landslides
have a flow frequency of landslides because of generally was highest on south- and southwest-facing slopes, indi-
lower shear stresses associated with low gradients. On av- cating that natural terrain landsliding is more common

Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag 385

 Cases and solutions

Fig. 4a–f 500 m in elevation. At very high elevations there are

Correlations between the frequency of landslides (landslide mountain summits that are usually characterized by wea-
scars per km 2) and the physical parameters thered rocks, and the shear strength of these is much
higher. At intermediate elevations, however, slopes tend
to be covered by a thin colluvium, which is more prone
on these slopes. The frequency of landslides was lowest to landslides. At very low elevations, the frequency of
on those slopes facing north, northeast, and northwest, landslides is low because the terrain is gentle, and is cov-
while the frequency of landslides remained moderate on ered with thick colluvium or/and residual soils, and a
the east-, southeast-, west-, and southwest-facing slopes. higher perched water table will be required to initiate
Field inspection suggests that this may be attributed to slope failure.
fact that there is more vegetation cover on north slopes.
Vegetation cover
Elevation The presence or absence of thick vegetation may affect
The relation of the frequency of landslides with elevation landslide susceptibility, but there is much conflicting evi-
is shown in Fig. 4d. It can be seen that hill slopes be- dence in the literature concerning this issue (Collinson
tween 100 and 500 m in elevation had frequencies of and Anderson 1996). Franks (1999) examined natural ter-
landslide that were 2.7–6.4 times greater than those on rain landslides in north Lantau Island and concluded that
hill slopes that are less than 100 m and greater than sparsely vegetated slopes are most susceptible to failure.

386 Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag

 Cases and solutions

To assess the effect of vegetation cover on the occurrence to ce, the probability varies from 0 to 1 on a S-shaped
of landslides, the correlation between vegetation type and curve. However, in a strict sense, it is not a probability
landslides is shown in Fig. 4e. It can be seen that the because the extrinsic variable triggering landslides, such
landslide density on bare land is relatively low, and is as rainfall, is not accounted for. It may be more appro-
highest on grassland. This is in agreement with the fact priate to term it hereafter as landslide susceptibility
that vegetation cover, especially of a woody type with based on the intrinsic physical parameters. Z is the linear
strong and big root systems, helps to improve the stabili- combination
ty of slopes (Gray and Leiser 1982; Greenway 1987).
ZpB0cB1X1cB2X2c 777 cBnXn
Proximity to drainage lines where B0 is the intercept of the model, n is the number
Intense gully erosion occurs in the study area and may of independent variables (i.e., landslide-related physical
be a factor in the occurrence of landslides. Some studies parameters), Bi (ip1,2... n) is the slope coefficient of the
(VanDine 1985; Reneau and Dietrich 1987) showed that model, and Xi (ip1,2... n) is the independent variable. In
the drainage lines (in the form of a natural gully channel) logistic regression, a coding scheme should be selected
are an important factor controlling debris flow suscepti- for the categorical variables by creating a new set of var-
bility. For the study area, the distance to a drainage line iables that correspond in some way to the original cate-
was calculated, and five classes were identified. It can be gories. The number of new variables required to repre-
found that as the distance from the drainage line in- sent a categorical variable is one less than that of the
creases, the frequency of landslide occurrence decreases number of categories. For example, in the present study
(Fig. 4f). This can be attributed to the fact that terrain we have four types of vegetation cover, and three new
modification caused by gully erosion may influence the variables would have to be created to represent the effect
initiation of landslides. Field checking also revealed that of vegetation cover. In this case, one variable would rep-
most landslides occurred within or adjacent to drainage resent the woodland, coded 1 if the vegetation cover is
lines or hollows. woodland and 0 otherwise; the second variable would
represent the shrubland, coded 1 and 0 otherwise; the
third variable would represent the grassland; and the
bare land would be represented by codes of 0 for three of
Mapping of landslide these variables. The parameters of the logistic regression
susceptibility model are estimated using the maximum-likelihood
method. In other words, the coefficients that make the
Logistic regression model observed results most “likely” are selected. Because the
Logistic regression modeling is a multivariate technique relationship between the independent variables and the
that considers several physical parameters that may affect probability is nonlinear in the logistic regression model,
probability. The advantage of logistic regression modeling an iterative algorithm is necessary for parameter estima-
over other multivariate statistical techniques, including tion.
multiple regression analysis and discriminant analysis, is The probability of the event not occurring is estimated as
that the dependent variable can have only two values –
an event occurring or not occurring, and that predicted Pr(no event)p1PPr(event)
values can be interpreted as probability because they are Logistic regression modeling is intended to describe the
constrained to fall in the interval between 0 and 1 (Klein- likelihood of landslide occurrence on a regional scale,
baum 1994). In addition, an appealing S-shaped descrip- and is very suitable for the assessment of landslide sus-
tion of the combined effect of several independent varia- ceptibility because the observed data consist of locations
bles on the dependent variable is also an important fac- (points) or cells with a value of 0 (absence of landslide)
tor for the popularity of logistic regression modeling. For or 1 (presence of landslide). This method allows a spatial
example, Bernknopf and others (1988) applied logistic re- distribution of probabilities or susceptibility values to be
gression to the economic evaluation of damage resulting calculated within the GIS environment.
from landslides in Cincinnati, Ohio. Mark and Ellen
(1995) used logistic regression to predict sites of rainfall- Independent variables and sampling
induced shallow landslides that initiate debris flows in Several different geological and geographical parameters
San Mateo County, California. In the present study, the considered to be relevant to the occurrence of landslides
dependent variable is a binary variable representing the were selected as the independent variables. Lithology,
presence or absence of landslides. The logistic model can slope aspect, and vegetation cover were treated as catego-
be written as rical independent variables, whereas slope gradient, eleva-
1 tion, and proximity to drainage line were continuous in-
Pr(event) p dependent variables (Table 2). In addition, those interac-
1ce Pz
tion terms between elevation and proximity to drainage
where Pr(event) is the probability of an event occurring. line, slope gradient and proximity to drainage line, and
In the present situation, the Pr(event) is the estimated elevation and slope gradient were also accounted for ini-
probability of landslide occurrence. As Z varies from –e tially in the model.

Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag 387

 Cases and solutions

Table 2
Independent variables initially included in the model

Index Variable Code

1 Lithology Lith
Holocene H
Holocene to Pleistocene HP
Pleistocene p
Volcanic rocks V
Sedimentary rock S
Intrusive igneous rock I
Minor intrusive igneous rock MI
2 Slope angle Slop
3 Slope angle squared Slop2
4 Slope aspect Asp
Flat F
N N
NE NE
E E
SE SE
S S
SW SW
W W
NW NW
5 Elevation Ele
6 Elevation squared Ele2 Fig. 5
7 Vegetation cover Veg Histogram of predicted landslide susceptibility
Bare land BL
Grassland GL
Shrubland SL
Woodland WL test is always used for determining whether variables
8 Proximity to drainage line Dis should be added to the model. This involves estimating
9 Proximity to drainage line squared Dis2 the model with each variable eliminated in turn and
looking at the change in the logarithm of likelihood when
each variable is deleted. If the observed significance level
is greater than the probability for stepwise (0.1 in this
For the purpose of the statistical analysis, sample data study) for remaining in the model, the variable is re-
representing both absence and presence of landslide must moved from the model and the model statistics are recal-
be provided to fit the logistic regression model. The way culated to see if any other variables are eligible for re-
in which these data are obtained will affect both the na- moval. The independent variables included in the model
ture of the regression relation and the nature and accura- were slope gradient, slope aspect, lithology, proximity to
cy of the resulting estimates (Atkinson and Massari drainage line, elevation and its value squared, slope gra-
1998). In this study, the data set of landslide inventory is dient squared, and the interaction term between elevation
an indispensable data source representative of samples of and slope gradient.
landslide presence. All locations of the landslide scars Figure 5 is a histogram of the predicted landslide suscep-
were thus used to extract the physical parameters (inde- tibility for the training samples used in this analysis.
pendent variables) automatically from the existing data Theoretically, if we have a model that successfully distin-
layers. Altogether, 1087 locations were chosen from the guishes the two groups based on a classification cutoff
not-yet-landslided area as samples representing the ab- value of 0.5, the cases for which landslide has occurred
sence of landslide. These locations were obtained using a should be to the right of 0.5, whereas the cases for which
random sampling scheme. The training data were then landslide has not occurred should be to the left of 0.5.
used to input to the logistic regression algorithm within The more the two groups cluster at their respective ends
the Statistical Package for Social Science (SPSS), a desk- of the plot, the better it is. From Fig. 5, it can be shown
top statistical software, to obtain the coefficients for the that the model produced a concordance rate of 82.8%
logistic regression model. with the use of 0.5 as a classification cutoff value (default
in SPSS). This result is in agreement with the work in
Modeling results northern Italy by Carrara and others (1991). By examin-
A logistic regression model was constructed initially ing this histogram of predicted susceptibilities, one can
based on the physical parameters as defined above. Then, see what a different classification rule should be adopted
at each step, variables are evaluated for removal one by when applying the model to each cell in the study area.
one if they do not contribute sufficiently to the regres- To map landslide susceptibility in the future in the study
sion equation. In the present analysis, the likelihood-ratio area, the logistic regression model was then transferred

388 Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag

 Cases and solutions

Table 3 highly stable and are not favorable to the development of

Classification of predicted susceptibility from the model landslides. Zone of “low” susceptibility, covering 18.0% of
the total land, is relatively dispersed in its spatial distri-
Estimated susceptibility Relative susceptibility class
bution, and hence the chance for landslide to develop
0.75F1.0 Very high within this class is small. The zone of “moderate” suscep-
0.55F0.75 High tibility covers 15.4% of the total area, and is featured by
0.3F0.55 Moderate lower sections of slopes and ridges. The “high” suscepti-
0.1F0.3 Low bility class, occupying 11.6% of the area, is mainly distri-
0.0F0.1 Very low buted in the middle section of slopes and bears a high
potential for landslide occurrence. The “very high” sus-
ceptibility class occupies 13.1% of the study area and ex-
hibits a strongly clustered pattern of spatial distribution.
into the ArcView GIS, and applied to the independent This category is distinguished from the “high” category
variables representing the present conditions for each cell by relatively high elevations and steeper terrain. Most of
within the study area. A fivefold classification scheme, the locations of the identified landslides actually fall
ranging from very high susceptibility to landsliding, to within this class.
very low, was employed for the predicted probabilities. It should be noted that the complexity of the failure
The ranges of the individual classes presented in Table 3 processes means that any evaluation of stability contains
were derived based on the histogram of the estimated a considerable amount of uncertainty. The use of the pre-
landslide susceptibility shown in Fig. 5. The final product dicted susceptibility map produced in this study is lim-
of the analysis is shown in Fig. 6. Zones classified as be-
ing of “very low” susceptibility, accounting for 41.8% of
the study area, are distributed in clusters on the coastal
lowland and on the top of high mountains that are char- Fig. 6
acterized by relatively gentle gradient. All these sites are Map of relative landslide susceptibility

Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag 389

 Cases and solutions

ited to regional-scale assessments and is not suitable for References

site-specific evaluations. The reliability of the assessment
results depends on a multitude of factors ranging from Anbalagan D (1992) Landslide hazard evaluation and zonation
the quality of the data base, the introduction of potential mapping in mountainous terrain. Eng Geol 32 : 269–277
errors associated with data entry, manipulation, and anal- Atkinson PM, Massari R (1998) Generalized linear modelling
ysis within the GIS, to the limitations and assumptions of landslide susceptibility in the Central Apennines, Italy.
Comput Geosci 24(4):373–385
inherent in the statistical techniques (Rowbotham and Bernknopf RL, Campbell RH, Brookshire DS, Shapiro CD
Dudycha 1998). In this study, a particular problem with (1988) A probabilistic approach to landslide hazard mapping
uncertainty is that the 1 : 20,000-scale topographic map in Cincinnati, Ohio, with applications for economic evalua-
cannot fully reflect the micro-topographic conditions tion. Bull Assoc Eng Geol 25(1):39–56
prerequisite for the occurrence of landslides because Carrara A, Cardinali M, Detti R, Guzzetti F, Pasqui V,
landslide occurrence in the study area is characterized by Reichenbach P (1991) GIS techniques and statistical models
small failure volumes, and that a slight change in micro- in evaluating landslide hazard. Earth Surf Proc Landforms
16 : 427–445
scale landform may have a strong influence on the occur- Carrara A, Cardinali M, Guzzetti F, Reichenbach P
rence of landslides. This, however, has not been reflected (1995) GIS-based techniques for mapping landslide hazard.
in the topographic maps. Another problem is that the In: Carrara A, Guzzetti F (eds) Geographical information sys-
1 : 20,000-scale geological maps used in this study cannot tems in assessing natural hazards. Kluwer, Dordrecht,
fully reflect the distribution of colluvium or residual soils pp 135–176
that are of critical significance to the occurrence of land- Collinson AJC, Anderson MG (1996) Using a combined slope
slides. In addition, the run-out distance of landslide de- hydrology/stability model to identify suitable conditions for
landslide prevention by vegetation in the humid tropies.
bris is not reflected in the predicted susceptibility map Earth Surf Proc Landforms 21 : 737–747
produced in this study. In this regard, a more detailed Dai FC, Lee CF, Wang SJ (1999) Analysis of rainstorm-induced
spatiotemporal approach to landslide risk assessment is slide-debris flows on natural terrain of Lantau Island, Hong
being carried out by the authors on a much larger scale, Kong. Eng Geol 51 : 279–290
using 1 : 1,000-scale topographic maps, detailed 1 : 5,000- Evans NC, Huang SW, King JP (1999) The natural terrain
scale superficial geological maps and multi-temporal aer- landslide study, phases I and II. GEO Report no 73. The Gov-
ial photographs as well as rainfall records. ernment of the Hong Kong Special Administrative Region
Fernandez CI, Castillo TFD, Handouni RE, Montero JC
(1999) Verification of landslide susceptibility mapping: a case
study. Earth Surf Proc Landforms 24 : 537–544
Franks CAM (1999) Characteristics of some rainfall-induced
Conclusions landslides on natural slopes, Lantau Island, Hong Kong. Q J
Eng Geol 32 : 247–259
GIS tools have made possible the production of innova- Geotechnical Control Office (1988a) Geotechnical area
tive landslide susceptibility maps. In particular, they have studies programme – North Lantau. GASP VI, Hong Kong
Government Press
facilitated the application of the logistic regression analy- Geotechnical Control Office (1988b) Geotechnical area
sis technique. Logistic regression applied to training sam- studies programme – South Lantau. GASP XI, Hong Kong
ples collected from existing data layers considered to be Government Press
relevant to landslide occurrence was able to predict land- Goodchild MF (1993) The state of GIS for environmental
slide susceptibility at a rate of about 82% concordance. problem-solving. In: Goodchild MF, Parks BO, Steyaert LT
The predicted susceptibilities generated from the model (eds) Environmental modeling with GIS. Oxford University
within the GIS environment were in turn used to produce Press, New York, pp 8–15
Gray DH, Leiser AT (1982) Biotechnical slope protection and
a map of relative landslide susceptibility. The results of erosion control. Van Nostrand Reinhold, New York
this study indicate that the model is useful and suitable Greenway DR (1987) Vegetation and slope stability. In: Ander-
for the scale adopted in this study. son MG, Richards KS (eds) Slope stability. Wiley, New York,
pp 187–230
Acknowledgements Funding for this research was provided by Gupta RP, Joshi BC (1989) Landslide hazard zoning using the
the Research Grants Council of Hong Kong and the Hong Kong GIS approach – a case study from the Ramganga catchment,
Jockey Club Research and Information Center for Landslip Pre- Himalayas. Eng Geol 28 : 119–131
vention and Land Development, at the University of Hong Kleinbaum DG (1994) Logistic regression-a self-learning text.
Kong. The authors wish to express their sincere appreciation Springer, Berlin Heidelberg New York
for the generous support received from these two organizations. Mark RK, Ellen SD (1995) Statistical and simulation models
for mapping debris flow hazard. In: Carrara A, Guzzetti F
(eds) Geographical information systems in assessing natural
hazards. Kluwer, Dordrecht, pp 93–106
Naranjo JL, Van Westen CJ, Soeters R (1994) Evaluating
the use of training areas in bivariate statistical landslide haz-
ard analysis: a case study in Colombia. Int Inst Aerospace
Surv Earth Sci J 3 : 292–300
Niemann KO, Howes DE (1991) Applicability of digital terrain
models for slope stability assessment. Int Inst Aerospace Surv
Earth Sci J 1991–3 : 127–137

390 Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag

 Cases and solutions

Pachauri AK, Pant M (1992) Landslide hazard mapping Van Westen CJ (1993) Application of geographic information
based on geological attributes. Eng Geol 32 : 81–100 systems to landslide hazard zonation. Int Inst Aerospace Surv
Reneau SL, Dietrich WE (1987) The importance of hollows in Earth Sci Publ no 15, Enschede
debris flow studies: examples from Marin County, California. Van Westen CJ, Rengers N, Terlien MTJ, Soeters R (1997)
In: Costa JE, Wieczorek GF (eds) Debris flows/avalanches: Prediction of the occurrence of slope instability phenomena
process, recognition and mitigation. Reviews in engineering through GIS-based hazard zonation. Geol Rundsch
geology, vol 7. Geological Society of America, Boulder, 86 : 404–414
pp 165–180 Varnes DJ (1984) Landslide hazard zonation: a review of prin-
Rowbotham DN, Dudycha D (1998) GIS modelling of slope ciples and practice. Natural hazards no 3. UNESCO, Paris
stability in Phewa Tal watershed, Nepal. Geomorphology Wang SQ, Unwin DJ (1992) Modelling landslide distribution
26 : 157–170 on loess soils in China: an investigation. Int J Geograph In-
Siddle HJ, Jones DB, Payne HR (1991) Development of a form Systems 6(5):391–405
methodology for landslip potential mapping in the Rhondda Wong HN, Lam KC, Ho KKS (1998) Diagnostic report on the
Valley. In: Chandler RJ (ed) Slope stability engineering. Tho- November 1993 natural terrain landslides on Lantau Island.
mas Telford, London, pp 137–142. GEO report no 69. The Government of the Hong Kong Spe-
Terlien MTJ, Van Asch TWJ, Van Westen CJ (1995) Deter- cial Administrative Region
ministic modelling in GIS-based landslide hazard assessment. Wu W, Sidle RC (1995) A distributed slope stability model for
In: Carrara A, Guzzetti F (eds) Geographical information sys- steep forested basins. Water Resour Res 31(8):2097–2110
tems in assessing natural hazards. Kluwer, Dordrecht, Yin KL, Yan TZ (1988) Statistical prediction models for slope
pp 57–77 instability of metamorphosed rocks. In: Bonnard C (ed) Pro-
VanDine DF (1985) Debris flows and debris torrents in the ceedings 5th International Symposium on Landslides, vol 2.
southern Canadian Cordillera. Can Geotech J 22 : 44–68 Lausanne, Switzerland, pp 1269–1272

Environmental Geology 40 (3) January 2001 7 Q Springer-Verlag 391