Geomorphology 42 (2002) 213 228 www.elsevier.

com/locate/geomorph

Landslide characteristics and slope instability modeling using GIS, Lantau Island, Hong Kong
F.C. Dai a, C.F. Lee b,*
a

Institute of Geographical Sciences and Natural Resources, Chinese Academy of Sciences, Beijing 100101, Peoples Republic of China b Department of Civil Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong, Peoples Republic of China Received 22 September 2000; received in revised form 14 March 2001; accepted 20 March 2001

Abstract Steep terrain and high a frequency of tropical rainstorms make landslide occurrence on natural terrain a common phenomenon in Hong Kong. This paper reports on the use of a Geographical Information Systems (GIS) database, compiled primarily from existing digital maps and aerial photographs, to describe the physical characteristics of landslides and the statistical relations of landslide frequency with the physical parameters contributing to the initiation of landslides on Lantau Island in Hong Kong. The horizontal travel length and the angle of reach, defined as the angle of the line connecting the head of the landslide source to the distal margin of the displaced mass, are used to describe runout behavior of landslide mass. For all landslides studied, the horizontal travel length of landslide mass ranges from 5 to 785 m, with a mean value of 43 m, and the average angle of reach is 27.7. This GIS database is then used to obtain a logistic multiple regression model for predicting slope instability. It is indicated that slope gradient, lithology, elevation, slope aspect, and land-use are statistically significant in predicting slope instability, while slope morphology and proximity to drainage lines are not important and thus excluded from the model. This model is then imported back into the GIS to produce a map of predicted slope instability. The results of this study demonstrate that slope instability can be effectively modeled by using GIS technology and logistic multiple regression analysis. D 2002 Elsevier Science B.V. All rights reserved.
Keywords: Landslides; Runout; Logistic multiple regression; Geographical Information Systems (GIS)

1. Introduction Landslides in mountainous terrain often occur during or after heavy rainfall, resulting in the loss of life and damage to the natural and/or built environment. Mapping or delineating areas susceptible to landslides is essential for land-use activities and management decision-making in mountainous areas.
Corresponding author. Tel.: +852-28592645; fax: +85228580611. E-mail address: leecf@hkucc.hku.hk (C.F. Lee).
*

Sites that are prone to landslides can be identified by both analytical and empirical methods. A variety of approaches have been used in slope instability mapping and can be classified into qualitative factor overlay, statistical models, and geotechnical process models. In the qualitative approach, several maps representing the spatial distribution of those physical parameters which may have influence on the occurrence of landslides are combined into a hazard map using subjective decision rules, based on the experience of geoscientists involved (Anbalagan, 1992; Pachauri and Pant, 1992; Sarkar et al., 1995).

0169-555X/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 9 - 5 5 5 X ( 0 1 ) 0 0 0 8 7 - 3

214

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

The limitations in this approach are in the reproducibility of results and in the subjectivity in decision rules. Statistical models involve the statistical determination of the combinations of physical parameters that have led to past landslides. Quantitative or semiquantitative estimates are then made for areas currently free of landslides, but where similar conditions exist. Both multiple regression analysis and discriminant analysis have been used to explore relations between landslide occurrence and the terrain variables (e.g. Yin and Yan, 1988; Carrara et al., 1991, 1995; Brunori et al., 1996; Dhakal et al., 1999). A major deterrent to such techniques has undoubtedly been the logistics of collecting and calculating quantitative data (Rowbotham and Dudycha, 1998). Another problem is that the probability values computed from such techniques can often fall outside the 0 to 1 range of the probability values, which makes it difficult to relate the output to a systematic probability surface. Recently, logistic regression, one of a family of generalized linear models that are well suited to analyzing a presence absence dependent variable, has been used to predict slope instability (Carrara et al., 1991; Mark and Ellen, 1995; Rowbotham and Dudycha, 1998). Geotechnical process approaches are based on slope stability analyses, and are applicable only when the ground conditions are fairly uniform across the study area and the landslide types are known and relatively easy to analyze (e.g. Terlien et al., 1995; Wu and Sidle, 1995). The advantage of the geotechnical process models is that they permit quantitative factors of safety to be calculated, while the main problem is the high degree of simplification that is usually necessary for the use of such models. An assessment of landslide hazard requires knowledge of the landslide characteristics and runout behavior of landslide mass. This research was undertaken with a view to characterizing landslides on natural terrain of Lantau Island, Hong Kong, and then developing a Geographical Information Systems (GIS) approach to modeling slope instability. This study area is prone to landslides when subjected to heavy rainstorms. For example, widespread landslides occurred in Lantau Island, following heavy rainfall on 18 July 1992 and 5 November 1993 (Franks, 1999). There are four objectives in this research: (1) to characterize landslides on natural terrain of Lantau

Island in Hong Kong; (2) to determine the statistical correlations between landslide frequency and the physical parameters contributing to the initiation of landslides; (3) to develop a methodology for modeling slope instability using GIS; and (4) to characterize the runout behavior of landslide mass. One assumption of slope instability modeling is that the occurrence of landslides in the past is indicative of the potential for landslides to occur in the future. By identifying the physical parameters contributing to the initiation of landslides, and by incorporating them in a GIS-based logistic multiple regression model, regional slope instability on Lantau Island was modeled.

2. Description of the study area Lantau Island is located in the southwest part of the territory of Hong Kong and is the largest outlying island within the territory (Fig. 1). Primarily because of its steep terrain, the island is virtually undeveloped and uninhabited with the exception of small coastal patches of flat land. Land area with slope gradients greater than 25 accounts for 44% of the total land. Elevation ranges from sea level to over 900 m above sea level and changes abruptly. The bedrock geology of the study area is dominated by Mesozoic volcanic rocks and the younger

Fig. 1. Location of the study area.

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

215

intrusive igneous rocks (Fig. 2). The volcanic rocks, which comprise tuffs and lavas with intercalated sedimentary rocks, crop out in the west of the study area. Intrusive rocks consist mainly of granites, and dykes of various compositions. Paleozoic sedimentary strata comprising metamorphosed siltstone, sandstone and carbonaceous siltstone occur as a small outcrop in the northwest coastal areas of the study area. Superficial deposits of the Quaternary age form large, flatlying areas. In hilly terrain, colluvium, including debris-flow and other slope debris deposits, mostly of late-Pleistocene to Holocene age, commonly mantle side slopes and valleys as a result of numerous individual episodes of mass wasting and erosion during the Quaternary period. Colluvium occurs as relatively thin ribbon-like deposits filling drainage courses. However, there are deposits which are considerably thicker and of greater areal extent on some hillslopes in the study area. The colluvium derived from volcanics typically consists of subangular cob-

bles and boulders, of feldsparphyric rhyolite with some tuff, in a matrix of mottled, reddish brown and yellowish brown gravelly, sandy, slightly clayey silt. Small alluvial deposits occur in hilly areas, but alluvium is generally restricted to fans developed downslope of the colluvial deposits. Beach deposits of sand usually form in front of alluvial deposits, especially in coastal bays deposited under the combined influence of higher sea levels and fluctuating climatic conditions in recent times (Geotechnical Control Office, 1988a,b). A regolith, or mantle of weathered rock, occurs over most of the study area. The effects of weathering vary with rock types, being reflected in topographic relief. Intrusive rocks and the Paleozoic sedimentary rocks are most deeply weathered and eroded, forming the lower ground. The acidic volcanic rocks are more resistant to deep weathering and erosion. As indicated in Fig. 2, the area is structurally affected by two sets of faults trending NE NNE and NNW NW, respectively.

Fig. 2. Simplified geological map of the study area.

216

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

The climate is sub-tropical and monsoonal, with mild, dry winters and hot, humid summers. Rainfall is heavy, and occasionally intense during the rainstorms and typhoons. Mean annual rainfall over the period 1961 1991 is in the range of 2000 to 2400 mm. Recent major rainstorm events occurred on 5 November 1993 (24-h rolling rainfall of 745 mm, 1 in 796 year event) and on 18 July 1992 (24-h rolling rainfall of 454 mm, 1 in 28 year event). The hillslopes are drained by numerous small streams, most of which flow only during or after heavy or prolonged rainfall. The hillsides are often deeply incised as a result of erosion caused by ephemeral streams. In general, piezometric records from previous site investigations indicate that the regional groundwater table lies either just within the slightly to moderately weathered bedrock or within the overlying saprolite (Franks, 1999). The relatively high permeability of the colluvium deposits, when compared to the underlying saprolite or weathered bedrock, allows for the development of transient perched groundwater tables at the interface during or following periods of intense rainfall.

3. Data sources The study area was examined using the ArcView GIS software. The data available for this study include topography, land-use classification, a terrain morphological map, superficial and bedrock geology, and the locations and trails of landslides. All locational, geological, and geomorphological features provided by the different thematic maps mentioned above were imported into the ArcView GIS, or digitized using the GIS software PC Arc/Info, and then transferred to ArcView for subsequent analyses. Contour lines and drainage lines are obtained from the 1:20,000 scale topographic maps with a contour interval of 20 m. Elevation data were obtained from the digital elevation model (DEM) with a resolution of 20 20 m derived from the 1:20,000 scale digital contour lines of the area. Two data layers are derived from these elevation data, namely slope aspect and slope gradient. Proximity to drainage line is calculated using GIS functions. Superficial and bedrock geological data are obtained from 1:20,000 scale solid and superficial

geological maps developed by the Hong Kong Geological Survey of the Geotechnical Engineering Office (GEO), previously known as the Geotechnical Control Office (GCO). The maps covering the study area describe the geological groups, each comprising geological units of broadly similar lithology. For ease of analysis, the groups were further reclassified into nine categories: alluvial, terrace and beach deposits (ATB), debris flow deposits and talus (DF), sedimentary rock (SR), metasedimentary rock (MSR), intrusive rock (IR), minor intrusive rock (MIR), ash tuff, tuffite and tuff breccia (BCT), trachydacite, dacite and rhyolite lava (TDR), and volcaniclastic sedimentary rock (VSR), based on their stratigraphy and genesis. The 1:20,000-scale digital terrain classification maps covering the study area, developed by the GEO, were available to the authors. This dataset contains terrain classification information that includes erosion and stability, terrain component and morphology, and slope gradient, which was derived from Geotechnical Areas Studies Programme (GASP) primarily using aerial photography interpretation (API) technique (Brand, 1988; Geotechnical Control Office, 1988a,b). Based on the terrain classification information, terrain morphology which describes the physical appearance of the slope and the general shape of the slope profile (straight, concave or convex) is extracted and then reclassified into 10 categories for simplicity: hillcrest or ridge (A), straight sideslope (B), concave sideslope (C), convex sideslope (D), straight footslope (E), concave footslope (F), convex footslope (G), drainage plain (H), rock outcrop (M), and others, such as reclamation and coastal plain, (O). All the footslope and drainage plain terrain consists of colluvium, and all the sideslope terrain consists of insitu geological materials (Geotechnical Control Office, 1988a,b). The landslide database used was derived from the Geotechnical Engineering Office work in which landslide locations and trails were digitized from 23 temporal sets of 1:20,000 to 1:40,000 scale stereoscopic aerial photographs dating from 1945 to 1994 (Evans, 1998; King, 1999). The aerial photographs used thus cover a period of 50 years and recent landslides as old as about 10 years were visible before re-vegetation masked most scars. Recent landslides (Fig. 3) were observed on aerial photographs as a

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

217

Fig. 3. Shaded relief map of the study area showing locations of landslides (black dots).

distinctive light tone, which is generally bare of vegetation (King, 1999). This indicates that the aerial photographs record about a 60-year period of landslide activity. The location of each identified landslide crown was recorded on the 1:5000 scale base map, and the centerline of any landslide mass trail was marked with a line. In the interpretation of aerial photographs, the GEO classified the width of the landslide scars as greater or less than 20 m wide, and landslides with a width of greater than 20 m were referred to as wide. This may be attributed to the fact that landsliding with a width of greater than 20 m is not a common occurrence. The ground slope angle across the landslide head, calculated from the 1:5000 scale topographical maps, was noted. All these features have been digitized by the GEO, and are available to the authors. A 1:50,000 scale coverage of land-use types for the whole territory of Hong Kong, based on the interpretation of SPOT images with verification of

field checking by Chi (Unpublished data) in 1998, is used for the analysis. Although 35 land-use types were mapped, these were simplified into six categories for the purposes of this study: (1) developed land, such as cropland, roads, structures, reservoirs, and reclamation (DL); (2) forested land (FL); (3) shrub-forested land (SFL); (4) densely grassed land with a shrub coverage of less than 40% (DGL); (5) moderately grassed land with > 50% coverage (MGL); and (6) sparsely grassed land on rock outcrop-dominated areas (SGL). It should be noted that land-use cover is considered to be only estimates, because of increased development of coastal flat-lying lands with time and possible temporal change in land-use types over the past several decades. The above-mentioned vector datasets are then rasterized to the DEM resolution in ArcView for subsequent analyses. Each landslide was assumed to be within a single 20 m pixel.

218

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

4. Physical characteristics of landslides Physical characteristics of landslides, including landslide description, the physical parameters contributing to the initiation of landslides, and runout behavior of landslide mass are analyzed respectively as follows. 4.1. Landslide description Landslide classification systems are usually based on a combination of material and movement mechanism. Using the system proposed by Cruden and Varnes (1996), most of the landslides in the study area are probably debris slides, debris flows, complex debris slide-flows, or composite debris slide-flowfalls, all of which may be either open-slope or channelized (Evans et al., 1999). About 80% of the 2135 landslides recorded were less than 20-m in source width. Field checking indicates that the failures generally occurred along the colluvium bedrock contact, and that the predominant failure mode is of the translational type, involving a slipping of a thin layer of colluvium with a planar failure surface. Most landslides started as slides and quickly converted to flows because of the water involved and the steep terrain below the landslide sources (Dai et al., 1999). The vast majority of the landslides have the following common features: a source area, defined by a surface of rupture which comprises the main scarp and the scarp floor, and a landslide trail downslope of the source where landslide mass transport predominates, though erosion and deposition may also occur, and a

deposition fan where the majority of the landslide mass is deposited (Fig. 4). It should be noted that a deposition fan might not be well developed for many failures on open slopes because the landslide mass is completely deposited on the trail path. The GEO carried out a systematic study of the 56 natural terrain failures in three selected areas within the study area, and a factual and a diagnostic report on the investigations and observations of the landslides were given by Wong et al. (1997) and Wong et al. (1998), respectively. Field inspections of these landslides have also been carried out by the authors (Dai et al., 1999). The distributions of source length, source width, and failure depth are shown in Fig. 5. For the landslides examined, the source lengths vary between 6 and 40 m, with a mean value of about 15 m, and the source widths range from 3 to 20 m, with a mean value of about 10 m. The landslides generally have a failure depth varying between 0.5 and 2.0 m with a mean value of about 1.4 m. The vast majority of the landslides examined involved the failure of a thin surface layer of highly permeable bouldery colluvium. In slightly over 50% of the landslides examined on site, erosion pipe holes, usually near the interface of the colluvium and the underlying less permeable material, were observed in the loose colluvium exposed at the back scarps of the landslides. It seemed that these landslides were probably triggered by the development of a transient water table above the interface between the colluvium and the less permeable underlying material, resulting from direct surface infiltration and subsurface seepage (Wong et al., 1998; Dai et al., 1999). Given the heterogeneous nature of the colluvium layer and the likely presence of preferential flow paths in the layer, subsurface seepage flows leading to a build-up of seepage pressures acting within selected zones in the layer might also have contributed to triggering the landslides (Wong et al., 1998). 4.2. Physical parameters contributing to the initiation of landslides To examine the physical parameters contributing to the initiation of landslides, the landslides which occurred in the study area were correlated with those parameters considered to have influence on their occurrence. These physical parameters include lithol-

Fig. 4. Description of typical natural terrain landslides.

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

219

4.2.1. Lithology and geological structure It may be reasonably expected that the properties of the slope-forming materials, such as strength and permeability that are involved in the failure, are related to the lithology, which therefore should affect the likelihood of failure. The correlation of landslide frequency with lithology is shown in Fig. 6a. It can be seen that there are three geological categories with relatively high landslide frequency, namely trachydacite, dacite and rhyolite lava (TDR), sedimentary rock (SR), and metasedimentary rock (MSR), with the former being the highest. As mentioned previously, the available evidence tends to suggest that surface thin colluvium may have played an important role in the majority of landslides. However, colluvial deposits that are less than approximately 2 m thick are not identified on the 1:20,000 scale geological maps (Evans et al., 1999). Hence, landslides in thin colluvium are recorded as occurring within the underlying geological group. This is not considered to be a serious problem as the properties of the thin colluvial layers will be very dependent on the bedrock geology from which they are derived. Immediately downslope from geological group boundaries, unmapped colluvial deposits may have been partly derived from the upper geological group rather than from the underlying unit. However, the proportion of landslides affected by this situation will be very small (Evans et al., 1999). Structural information is also available from the digital geological maps. However, visual examination of spatial distributions suggests that the correlation between landslides and mapped linear structural features at the 1:20,000-scale is not good, and the structural information is thus excluded in this study.
Fig. 5. Histograms showing characteristics of initial landslides: (a) source length, (b) source width, and (c) failure depth.

ogy and structure, slope gradient and slope morphology, slope aspect, elevation, proximity to drainage line, and land-use type. The digital map of landslide distribution was overlain on the raster data layers of physical parameters mentioned above using the GIS, and landslide frequency, which is the number of landslides per squared kilometer, was calculated for each category on the physical parameter maps.

4.2.2. Slope gradient and slope morphology Slope gradient has a great influence on the susceptibility of a slope to landsliding. On a slope of uniform, isotropic material increased slope gradient correlates with increased likelihood of failure. However, variations in soil thickness and strength are two factors which vary over a wide range for both failure and non-failure sites. To quantify the relative frequency of landslides on different slope gradients, it is necessary to consider the distribution of the slope gradient categories using the available digital eleva-

220

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

Fig. 6. Correlations between landslide frequency (landslides per squared kilometer) and the physical parameters (symbols refer to text).

tion model (DEM). Examination of landslide frequency with the corresponding slope gradient categories shows an increase with slope gradient until the maximum frequency is reached in the 35 40 category, followed by a decrease in the >40 category (Fig. 6b).

Slope morphology can probably affect the susceptibility of a slope to landslide in several ways. The shape of a slope influences the direction of and amount of surface runoff or subsurface drainage reaching a site. Concentration of subsurface drainage within a concave slope, resulting in higher pore water

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

221

pressures in the axial areas than on flanks, is one possible mechanism responsible for triggering landslides (Pierson, 1980). An analysis of correlation between landslide frequency and slope morphology is carried out with the use of the terrain morphological map, and the result is shown in Fig. 6c. It can be seen that the landslide frequency is generally higher for concave sideslopes (C), and for rock outcrop (M), followed by straight sideslopes (B). The landslide frequency on the sideslope terrain is much higher than that on the footslope terrain. It can also be noted from Fig. 6c that for both sideslopes and footslopes, the landslide frequency is highest for concave slopes. 4.2.3. Slope aspect The aspect of a slope can influence landslide initiation. Moisture retention and vegetation is reflected by slope aspect, which in turn may affect soil strength and susceptibility to landslides. If rainfall has a pronounced directional component by influence of a prevailing wind, the amount of rainfall falling on a slope may vary depending on its aspect (Wieczorek et al., 1997). To investigate the relative relationship between landslide frequency and slope aspect, the DEM was used to calculate the aspect of a slope within the study area. The distribution of aspect among the mapped landslides is shown in Fig. 6d. It can be seen that on north-facing slopes the landslide frequency is relatively low, and it increases with the orientation angle, reaching the maximum on southfacing slopes, and then declines. 4.2.4. Elevation The correlation of landslide frequency with elevation is shown in Fig. 6e. At very high elevations there are mountain summits that usually consist of weathered rocks, whose shear strength is much higher. At intermediate elevations, however, slopes tend to be covered by a thin colluvium, which is more prone to landslides. At very low elevations, the frequency of landslides is low because the terrain itself is gentle, and is covered with thick colluvium or/and residual soils, and a higher perched water table will be required to initiate slope failure. 4.2.5. Land-use type Extensive investigations have shown that land-use cover or vegetation cover, especially of a woody type

with strong and large root systems, helps to improve stability of slopes (Gray and Leiser, 1982; Greenway, 1987). Vegetation provides both hydrological and mechanical effects that generally are beneficial to the stability of slopes. Franks (1999) examined natural terrain landslides in the Tung Chung area, North Lantau Island, and concluded that a sparsely vegetated slope is most susceptible to failure. The correlation between land-use type and landslide frequency is shown in Fig. 6f. It can be seen that the landslide frequency on densely grassed land (DGL) is the highest, followed by moderately grassed land (MGL). 4.2.6. Proximity to drainage line An analysis has been carried out to assess the influence of drainage lines on landslide occurrence. For this purpose, proximity to drainage line is identified, and the results are divided into eight categories. It can be found that as the distance from drainage line increases, landslide frequency generally decreases (Fig. 6g). This can be attributed to the fact that terrain modification caused by gully erosion may influence the initiation of landslides.

5. Slope instability modeling 5.1. Logistic multiple regression Logistic multiple regression is a multivariate technique which considers several physical parameters that may affect probability. It accepts both binary and scalar values as the independent variables, which allows for the use of variables that are not continuous or qualitatively derived. The advantage of logistic multiple regression modeling over other multivariate statistical techniques including multiple regression analysis and discriminant analysis is that the dependent variable can have only two valuesan event occurring or not occurring, and that predicted values can be interpreted as probability since they are constrained to fall in the interval between 0 and 1. In the present study, the dependent variable is a binary variable representing the presence or absence of landslides. The technique of logistic multiple regression yields coefficients for each variable based on data derived from samples taken across a study area. These coefficients serve as weights in an

222

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

algorithm which can be used in the GIS database to produce a map depicting the probability of landslide occurrence. Quantitatively, the relationship between the occurrence and its dependency on several variables can be expressed as: Prevent 1=1 e
Z

ceptibility values to be calculated within the GIS environment. 5.2. Variables selection and sampling All the physical parameters considered to be relevant to the occurrence of landslides, as noted previously, including lithology, slope gradient, slope aspect, slope morphology, elevation, land-use type, and proximity to drainage line, were selected as the initial independent variables in the present study. For each variable, the same categorization scheme as that used to study the relation of landslide frequency with variable categories previously is adopted for consistency. For the purpose of the statistical analysis, sample data representing both absence and presence of landslide must be provided to fit the logistic multiple regression model. The way in which these data are obtained will affect both the nature of the regression relation and the nature and accuracy of the resulting estimates (Atkinson and Massari, 1998). In this study, the data set of landslide inventory is an indispensable data source representative of samples of landslide presence. All locations of the 2135 landslides studied were thus used to extract automatically from the existing data layers the physical parameters that characterize landslide locations. To eliminate bias in the sampling process, an equal number of points were chosen from the not-yet-landslide area as samples representing the absence of landslide. These locations were obtained using a spatially uniform sampling scheme but excluding a 40-m buffer zone for all landslides so as to minimize the impact of the size of landslides. Each sample point has its respective binary value on the presence/absence of landslide, as well as information on independent variables. These sample data were then used to input to the logistic multiple regression algorithm within the SPSS (SPSS, 1997), a desktop statistical software package, to obtain the coefficients for the logistic multiple regression model. 5.3. Modeling results A logistic multiple regression model was constructed initially based on the physical parameters as defined above. Then, at each step, variables are

where Pr(event) is the probability of an event occurring. In the present situation, the Pr(event) is the estimated probability of landslide occurrence. As Z varies from 1 to + 1, the probability varies from 0 to 1 on an S-shaped curve. Z is the linear combination: Z B0 B1 X1 B2 X2 . . . Bn Xn where Bi (i = 0, 1,. . ., n) is the coefficient estimated from the sample data, n is the number of independent variables (i.e. landslide-related physical parameters), and Xi (i = 1, 2,. . ., n) is the independent variable. However, in a strict sense, it is not a probability since the dynamic variables triggering landslides, such as rainfall, are not accounted for. It may be more appropriate to term it hereafter slope instability or landslide susceptibility based on the quasi-static physical parameters. In logistic multiple regression, a coding scheme should be selected for the categorical variables by creating a new set of variables that correspond in some way to the original categories. The number of new variables required to represent a categorical variable is one less than that of the number of categories. The coefficients of the logistic multiple regression model are estimated using the maximum-likelihood method. In other words, the coefficients that make the observed results most likely are selected. Since the relationship between the independent variables and the probability is nonlinear in the logistic multiple regression model, an iterative algorithm is necessary for parameter estimation. Logistic multiple regression modeling is intended to describe the likelihood of landslide occurrence on a regional scale, and is very suitable for the assessment of slope instability, since the observed data consist of locations (points) or cells with a value of 0 (absence of landslide) or 1 (presence of landslide). This method allows a spatial distribution of probabilities or sus-

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

223

evaluated for removal one by one if they do not contribute sufficiently to the regression equation. The variables included in the model were slope gradient, lithology, elevation, slope aspect, and landuse cover. In the present analysis, the likelihood-ratio test was always used for determining whether variables should be added to the model. This involves estimating 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 remaining in the model (0.1 in this study), the variable is removed from the model and the model statistics are recalculated to see if any other variables are eligible for removal. Both proximity to drainage line and slope shape were not significant and were thus eliminated from the stepwise procedure. The coefficients for the final logistic multiple regression are shown in Table 1. Note that all the variables in the model are binary variables representing presences or absences of the corresponding variables. For each variable, the last category is used as

the default reference category, and the coefficient of that category is thus overridden. Fig. 7 is a histogram of the predicted landslide susceptibility for the training samples used in this analysis. Theoretically, if we have a model that successfully distinguishes the two groups based on a classification cutoff value of 0.5, the cases for which landslide has occurred should be to the right of 0.5, while the cases for which landslide has not occurred should be to the left of 0.5. The more the two groups cluster at their respective ends of the plot, the better it is. From Fig. 7, it can be shown that the model produced a concordance rate of 81.7% and that 85.2% of the actual landslides were correctly classified with the use of 0.5 as a classification cutoff value (default in SPSS). By examining this histogram of predicted susceptibilities, one can see what a different classification rule should be adopted when applying the model to each cell in the study area. To map future potential slope instability in the study area, the logistic multiple regression model was then transferred into the ArcView GIS, and applied to the independent variables representing the

Table 1 Regression coefficients estimated for the model Variable Constant term Slope gradient () Categories 0 10 10 15 15 20 20 25 25 30 30 35 35 40 R 40 0 50 50 100 100 150 150 200 200 250 250 300 300 350 350 400 400 450 450 500 500 550 550 600 600 650 650 700 >700 Coefficient 9.755 10.678 4.369 3.374 2.639 1.153 0.863 0.077 11.214 11.441 11.445 11.181 11.322 11.212 10.959 10.816 10.550 10.434 9.240 9.235 8.629 7.653 Variable Slope aspect Categories Flat N NE E SE S SW W NW developed land forested land shrub-forested land densely grassed land moderately grassed land sparsely grassed land alluvial, terrace and beach deposits debris flow deposits and talus sedimentary rock metasedimentary rock intrusive rock minor intrusive rock ash tuff, tuffite and tuff breccia trachydacite, dacite and rhyolite lava volcaniclastic sedimentary rock Coefficient 0.431 0.112 0.468 0.670 0.548 0.882 0.547 0.303 7.121 0.033 0.257 0.225 0.258 7.298 0.984 0.233 0.716 2.076 2.413 0.990 0.076

Elevation (m)

Land-use type

Lithology

224

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

Fig. 7. Histogram of predicted landslide susceptibility.

present conditions for each cell within the study area. For general purpose, the range of the susceptibility to landslides is classified into 4 categories: (a) Very low (0 0.2), (b) Low (0.2 0.35), (c) Moderate (0.35 0.5), and (d) High (>0.5). The ranges of the individual categories were derived based on the histogram of the estimated susceptibility to landslides shown in Fig. 7. The final product of the analysis is shown in Fig. 8. Zones classified as being of very low susceptibility are distributed in clusters on the coastal lowland and on the top of high mountains that are characterized by relatively gentle gradient, while zones of low susceptibility are sparsely distributed. In the zones of moderate susceptibility, the combination of physical parameters may adversely influence slope stability. When disturbed, the slopes are prone to landslides. The high susceptibility category exhibits a strongly clustered pattern of spatial distribution. This category bears a high potential for landslide occurrence, and is characterized by relatively high elevations and steep terrain. Most of the locations of the identified landslides actually fall within this category, and existing ground conditions are very likely to create serious landslide problems. Generally, the slope instability map reflects the potential for initiating a landslide on a slope, but does not indicate how far the landslide will travel. One of the possible solutions to this problem is that one may

use this slope instability map and runout behavior of landslide mass that will be discussed below to roughly estimate possible travel distance of potential landslides. Land use planners, developers and general public may use this map to determine areas where landslides may be a problem in site development. It should be noted that the complexity of the failure processes means that any evaluation of stability contains a considerable amount of uncertainty. The reliability of the assessment results depends on a multitude of factors ranging from the quality of the database, the introduction of potential errors associated with data entry, manipulation, and analysis within the GIS, to the limitations and assumptions inherent in the statistical techniques (Rowbotham and Dudycha, 1998). In addition, temporal and spatial distribution of rainfall, as a trigger for landslide occurrence in the study area, is not accounted for, though the landslide database used in this study including landslide incidence in about 60 years may even-out the spatiotemporal rainfall effects. It might be better to incorporate rainfall variables within the logistic multiple regression analysis. In this regard, a more detailed spatio-temporal approach to landslide hazard assessment is being carried out by the authors on a much larger scale, using a DEM with a resolution of 2 2 m and spatio-temporal landslide information derived from multi-temporal aerial photographs by using

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

225

Fig. 8. Map of relative landslide susceptibility.

aerial photogrammetric method, detailed 1:5000 scale superficial geological maps and rainfall records.

6. Runout behavior of landslide mass In landslide hazard assessment, the possible spatial impact of landslides needs to be estimated. A wellknown index expressing the runout behavior of landslide mass is the angle of the line connecting the head of the landslide source to the distal margin of the displaced mass. This angle has been designated as the angle of reach (Hsu, 1978; Corominas, 1996) or the angle of apparent friction (Wong and Ho, 1996). The angle of reach is considered the most suitable and practical parameter for use in assessing the mobility of landslide mass in view of its close modeling of the parameters for characterizing the rate of energy loss during mass movement and its consideration of the effect of downslope gradient (Wong and Ho, 1996).

Most studies (e.g. Hsu, 1978; Corominas, 1996; Wong and Ho, 1996) focus on the relationship between the angle of reach and the volume of failure. Generally, the angle of reach decreases (or mobility increases) with an increase in landslide volume. For natural terrain landslides in the study area, Wong et al. (1998) carried out a study on the relation between the angle of reach and landslide volume based on the assumption that the mobility of landslide mass can be significantly affected by the mechanism of mass movement. They classified the movement of landslide mass into three modes: (1) gravity (or sliding) mode without a significant influence from the action of surface water; (2) hydraulic mode that means landslide mass ran into stream courses and was subsequently subjected to significant action of surface running water; (3) mixed mode, intermediate between the above two modes. They concluded that the angle of reach is highest for landslides of the sliding mode and lowest for landslides of the hydraulic mode.

226

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

However, for landslide hazard assessment over a large area, the relationship between the angle of reach and the volume of landslide mass may not be a practical method because it is very difficult to predict the volume of a potential landslide and estimate the possible mode of mass movement. A possible, more practical approach may be used to estimate the variations in the angle of reach and/or travel distance of landslide mass based on historic records. In the present study, no attempt has been made to obtain the relation of the angle of reach with landslide volume because data on landslide volume are not available. In the dataset of landslide distribution, the elevations at the head and the distal margin of landslides, and the horizontal length of landslides are noted. This permits us to carry out a statistical analysis of the runout distance of landslide mass and the relationship between the horizontal length and change in elevation. Fig. 9 shows the histogram of horizontal length of landslides. Of all landslides studied, about 67% are less than 40 m and about 9% are greater than 100 m in horizontal length. For landslides with width of < 20 m, 74% are less than 40 m and 5% are greater than 100 m in horizontal length. However, for landslides with a width of >20 m, about 41% is less than 40 m and 27% are greater than 100 m in horizontal length. The average horizontal lengths are 35.3 and 72.6 m for landslides with a width of < 20 m and landslides with a width of >20 m, respectively. This indicates that the horizontal length of landslide mass may increase with the width of landslides, or landslide volume. A linear regression analysis is performed to obtain the best relation between the horizontal length and change in elevation of landslides. Outliers are defined as being significantly different from points with more than three standard deviations from the mean. These significant outliers are then excluded from the analysis and the regression is refitted so as to obtain an equation of general applicability. Of the 2135 landslides studied, 32 outliers are determined and then excluded. For landslides with a width of < 20 m and landslides with a width of >20 m, 28 and 9 outliers are defined and excluded from the total 1691 and 444, respectively. This exclusion of the outliers is considered to have little influence on the statistical results used on regional scale, primarily

Fig. 9. Histogram showing the distribution of horizontal length of landslides.

because the percentage of outliers is quite small. The results are shown in Fig. 10. It can be seen that the average angle of reach is 27.7 for all landslides studied, and that a slight difference in the average angle of reach exists between the landslides with a width of < 20 m (29.0) and those with a width of >20 m (26.7). This indicates that a dependency of the angle of reach on landslide width or thus landslide volume may exist and that this dependency is

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228

227

runout distance of landslide mass when incorporated into a map showing slope instability and the digital elevation model (DEM) within GIS. 7. Conclusions With Lantau Island of Hong Kong as a study area, the pertinent landslide characteristics are described, and the relations of landslide frequency with the physical parameters contributing to the initiation of landslides are presented. The runout distance and the angle of reach of landslides are analyzed. GIS tools have made possible the production of innovative slope instability maps. In particular, they have facilitated the application of the logistic multiple regression analysis technique. Logistic multiple regression applied to training samples collected from existing data layers considered to be relevant to landslide occurrence was able to predict slope instability at a rate of about 85% concordance. The predicted susceptibilities generated from the model within the GIS environment were in turn used to produce a map of relative landslide susceptibility. The results of this study indicate that the model is useful and suitable for the scale adopted in this study. Acknowledgements Funding for this research was provided by the Research Grants Council of Hong Kong and the Hong Kong Jockey Club Research and Information Centre for Landslip Prevention and Land Development, at the University of Hong Kong. The authors wish to express their sincere appreciation for the generous support received from these two organizations.
Fig. 10. Change in elevation plotted as a function of horizontal length of landslide mass.

References
Anbalagan, D., 1992. Landslide hazard evaluation and zonation mapping in mountainous terrain. Engineering Geology 32, 269 277. Atkinson, P.M., Massari, R., 1998. Generalized linear modelling of landslide susceptibility in the Central Apennines, Italy. Computers and Geosciences 24, 373 385. Brand, E.W., 1988. Special lecture: landslide risk assessment in Hong Kong. In: Bonnard, C. (Ed.), Proceedings of 5th International Symposium on Landslides, Lausanne, Switzerland. Balkema, Rotterdam, pp. 1059 1073. Brunori, F., Casagli, N., Fischi, S., Garzonio, C.A., Moretti, S., 1996. Landslide hazard mapping in Tuscany, Italy: an example

not important because the percentage of landslides involved is quite small. There are a lot of uncertainties not considered in the present study. These uncertainties underlying the model may include the type of material, mechanism of failure, groundwater, the volume of failure, and geology, etc. The parameters obtained are applicable to predict the travel distance on regional scales, and provide an effective means for the assessment of

228

F.C. Dai, C.F. Lee / Geomorphology 42 (2002) 213228 King, J.P., 1999. Natural Terrain Landslide Study: Natural Terrain Landslide Inventory. GEO Report No. 74, Geotechnical Engineering Office, Hong Kong, 127 pp. Mark, R.K., Ellen, S.D., 1995. Statistical and simulation models for mapping debris flow hazard. In: Carrara, A., Guzzetti, F. (Eds.), Geographical Information Systems in Assessing Natural Hazards. Kluwer Academic Publishing, The Netherlands, pp. 93 106. Pachauri, A.K., Pant, M., 1992. Landslide hazard mapping based on geological attributes. Engineering Geology 32, 81 100. Pierson, T.C., 1980. Piezometric response to rainstorms in forested hillslope drainage depressions. Journal of Hydrology (New Zealand) 19, 1 10. Rowbotham, D.N., Dudycha, D., 1998. GIS modelling of slope stability in Phewa Tal watershed, Nepal. Geomorphology 26, 151 170. Sarkar, S., Kanungo, D.P., Mehrotra, G.S., 1995. Landslide hazard zonation: a case study in Garhwal Himalaya, India. Mountain Research and Development 15, 301 309. SPSS, 1997. SPSS advanced statistics 7.5. Chicago, 578 pp. Terlien, M.T.J., Van Asch, T.W.J., Van Westen, C.J., 1995. Deterministic modelling in GIS-based landslide hazard assessment. In: Carrara, A., Guzzetti, F. (Eds.), Geographical Information Systems in Assessing Natural Hazards. Kluwer Academic Publishing, The Netherlands, pp. 57 77. Wieczorek, G.F., Mandrone, G., DeCola, L., 1997. The influence of hillslope shape on debris-flow initiation. In: Chen, C.L. (Ed.), Debris-flow hazards mitigation: mechanics, prediction, and assessment. American Society of Civil Engineers, New York, pp. 21 31. Wong, H.N., Ho, K.K.S., 1996. Travel distance of landslide debris. In: Senneset, K. (Ed.), Landslides, vol. 1, Balkema, Rotterdam, pp. 417 422. Wong, H.N., Chen, Y.M., Lam, K.C., 1997. Factual Report on the November 1993 Natural Terrain Landslides in Three Study Areas on Lantau Island. GEO Report No. 61, Geotechnical Engineering Office, Hong Kong, 42 pp. Wong, H.N., Lam, K.C., Ho, K.K.S., 1998. Diagnostic Report on the November 1993 Natural Terrain Landslides on Lantau Island. GEO Report No. 69, Geotechnical Engineering Office, Hong Kong, 98 pp. Wu, W., Sidle, R.C., 1995. A distributed slope stability model for steep forested basins. Water Resources Research 31, 2097 2110. Yin, K.L., Yan, T.Z., 1988. Statistical prediction models for slope instability of metamorphosed rocks. In: Bonnard, C. (Ed.), Proceedings 5th International Symposium on Landslides, Lausanne, Switzerland. Balkema, Rotterdam, pp. 1269 1272.

of automatic evaluation. In: Slaymaker, O. (Ed.), Geomorphologic Hazards. Wiley, Chichester, pp. 55 67. Carrara, A., Cardinali, M., Detti, R., Guzzetti, F., Pasqui, V., Reichenbach, P., 1991. GIS techniques and statistical models in evaluating landslide hazard. Earth Surface Processes and Landforms 16, 427 445. Carrara, A., Cardinali, M., Guzzetti, F., Reichenbach, P., 1995. GISbased techniques for mapping landslide hazard. In: Carrara, A., Guzzetti, F. (Eds.), Geographical Information Systems in Assessing Natural Hazards. Kluwer Academic Publishing, The Netherlands, pp. 135 176. Corominas, J., 1996. The angle of reach as a mobility index for small and large landslides. Canadian Geotechnical Journal 33, 260 271. Cruden, D.M., Varnes, D.J., 1996. Landslide types and processes. In: Turner, A.K., Schuster, R.L. (Eds.), Landslides Investigation and Mitigation. Special Report 247, Transportation Research Board, National Research Council. National Academy Press, Washington, DC, pp. 36 75. Dai, F.C., Lee, C.F., Wang, S.J., 1999. Analysis of rainstorm-induced slide-debris flows on natural terrain of Lantau Island, Hong Kong. Engineering Geology 51, 279 290. Dhakal, A.S., Amada, T., Aniya, M., 1999. Landslide hazard mapping and the application of GIS in the Kulekhani watershed, Nepal. Mountain Research and Development 19, 3 16. Evans, N.C., 1998. The natural terrain landslide study. In: Li, K.S., Kay, J.N., Ho, K.K.S. (Eds.), Slope Engineering in Hong Kong. Balkema, Rotterdam, pp. 137 144. Evans, N.C., Huang, S.W., King, J.P., 1999. The natural terrain landslide studyPhases I and II. GEO Report No. 73, Geotechnical Engineering Office, Hong Kong, 128 pp. Franks, C.A.M., 1999. Characteristics of some rainfall-induced landslides on natural slopes, Lantau Island, Hong Kong. Quarterly Journal of Engineering Geology 32, 247 259. Geotechnical Control Office (GCO), 1988. Geotechnical Area Studies Programme-North Lantau. GASP VI, Hong Kong Government, 124 pp. Geotechnical Control Office (GCO), 1988. Geotechnical Area Studies Programme-South Lantau. GASP XI, Hong Kong Government, 148 pp. Gray, D.H., Leiser, A.T., 1982. Biotechnical Slope Protection and Erosion Control. Van Nostrand-Reinhold, New York, 271 pp. Greenway, D.R., 1987. Vegetation and slope stability. In: Anderson, M.G., Richards, K.S. (Eds.), Slope Stability. Wiley, New York, pp. 187 230. Hsu, K.J., 1978. Albert Heim: observations on landslides and relevance to modern interpretations. In: Voight, B. (Ed.), Rockslides and Avalanches, vol. 1, Elsevier, Amsterdam, pp. 71 93.