Assessment of the Incidence of Landslides Using Numerical Information

PAPER Assessment of the Incidence of Landslides Using Numerical Information Atsushi HASEGAWA Takehiro OHTA, Dr. Sci. Assistant Senior Researcher, Senior Researcher, Laboratory Head, Geology Laboratory, Disaster Prevention Technology Division In recent years, railways have suffered frequent damage from landslides due to natural slopes adjoining tracks. Checking the stability of all trackside slopes however would be immensely time consuming and expensive. Developing a practicable and viable method to manage such natural slopes is therefore important. This paper quantified the factors relating to landslides such as geomorphologic features and vegetation, using a digital elevation model and a digital surface model. These evaluation factors were then applied to identify potential landslide locations. Keywords: landslide, gradient, slope type, vegetation, digital elevation model, digital surface model 1. Introduction Certain factors such as topography, geology, groundwater, and artificial change of landform are assumed to affect landslide occurrence. Until now, the occurrence possibility of landslides was estimated on the basis of aerial photographs and field surveys. These surveying methods rely mainly on visual inspection, making it possible to investigate a large number of slopes in relatively less time than required for other equipment based surveying methods such as drilling. The results of visual surveys however are not only solely qualitative but may also vary depending on the survey engineer performing the study. Furthermore, using these methods to investigate all of slopes adjoining railways requires a great deal of time and money. Therefore, it is necessary to develop a quick and inexpensive method to assess the stability of natural trackside slopes. Recent developments however with the computer and Geographical Information Systems (GIS), mean that national land numerical information is now being utilized. The airborne Light Detection and Ranging Technology (airborne LiDAR) in particular has undergone improvements and Digital Elevation Models (DEMs) and Digital Surface Models (DSMs) with a mesh size of less than 1m are now available. Consequently many studies on natural disasters have been carried out using the abovementioned DEMs. A number of these provided insights into the topographical changes caused by the disaster by comparing DEMs from before and after the events [1], []. Ohta et al. [3] therefore tried to extract landslide hazard locations using satellite imagery and DEM. This study aims to clarify the prime factors underlying landslide occurrences by interpreting aerial photographs and field surveys, and quantifying these factors through 1m-mesh DEM and DSM. A method was developed to extract locations where landslides would occur based on these quantified factors.. Extraction of landslide factors from several surveys This study examines four selected fields which share relatively homogeneous geology. A summary of each field is shown in Table 1..1 Method First, aerial photographs taken four times of each of all the fields were interpreted. Items which were interpreted are listed in Table. Each field was then surveyed to verify and complement the information obtained from the Field Area (km ) D 5 N 5 T 34 S 8 Table 1 Overview of the study field Geology Landform Dominant vegetation Igneous rock (Tuffaceous rock) Igneous rock (Granite) Sedimentary rock (Sandstone, Shale) Metamorphic rock (Schist) Gradual mountain U-shaped valley Riverbed gradient:gentle Steep mountain V-shaped valley Riverbed gradient:steep Gradual mountai U-shaped valley Riverbed gradient:gentle Steep mountain V-shaped valley Riverbed gradient:steep Conifer forest (matured) Mixed forest Conifer forest (matured) Broad-leaved forest Conifer forest (matured) Broad-leaved forest QR of RTRI, Vol. 54, No., May 13 79

Table Interpretation items of aerial photographs Category Landform Vegetation Modification Interpretation items Convex break of slope,concave break of slope,landslides,gully, Condition of riverbed,catchment slope,talus,outcrop etc. Classification(Broad-leaved forest,mixed forest, Conifer forest (matured),conifer forest (young), Deforestation,Secondary forest (young),bamboo Construction of roads and residential land,deforestation and Afforestation,Bank protection and Erosion control works etc. interpretation of the aerial photographs. Next, a 1m-mesh DEM and DSM by airborne LiDAR was generated for each field. By mapping the results of aerial photograph interpretation and the field survey onto the contour map made from the DEMs, the prime factors of landslide occurrence for each field were investigated.. Landslide occurrence factors There are many prime landslide occurrences factors such as geology, landform, groundwater, artificial modification and so on. However, it is difficult to obtain information about underground space such as geology and groundwater from remote sensing data. As such, landform and vegetation conditions which may be related to landslides occurrences were investigated instead. The numbers of landslide locations examined in each region were as follows; D-region: 1; N-region: 15; T-region: 41; S-region: 68. All of these landslides were caused by heavy rainfall...1 Landform Results from the interpretation of aerial photographs and field surveys, indicated that landslides tend to occur on catchment slopes and/or on steep slopes. Consequently, studies were made into the influence of gradient and slope types on landslide occurrence by using contour maps and slope gradient maps which were modeled by DEM. In the DEM, the gradient at the target grid point (x i, y i ) was given by 1 u( x S x y i, y i ) u( x i, y i ) u( x i, y i ) u( x (, ) + 1 1 i i i x + + 1, y ) i 1 y where u(x i, y i ) is elevation at the target grid point (x i, y i ) and u(x i-1, y i ), u(x i+1, y i ), u(x i, y i-1 ), u(x i, y i+1 ) are the elevation at the four grid points adjacent to the target point. Examples of gradient histograms of landslide slopes are shown in Fig. 1. In these histograms, the average value of the gradient at the landslide point was around 4 degrees, and the standard deviation of the gradient was 7 to 9. In addition a tendency was found that frequency of the gradient increased sharply at 3 degrees. In these histograms, the ratio of the sum of the frequency of gradients of more than 3 was 88% in region D, 96% in region N, 73% in region T, and 95% in region S. Given the shallowness of the examined landslides it can be considered that the landslide slope gradients were equal to those before the landslide actually occurred. It was therefore concluded that the landslides were likely to occur on slopes with a gradient of 3 degrees or more. The frequency of landslides per slope type, defined as a combination of horizontal and vertical cross-sectional slope shape, was investigated [4] (Fig. ). The majority of landslides were with slopes of type Vv, Rv, Vs and Rs in any field. In particular, landslides were more likely to occur type Rv and Vv slopes called catchment slopes. It is considered therefore that one of the causes of landslides is a specific shape of slope, such as a catchment slope. (1) 14 Frequency (pixel) 1 8 6 4 Av. : 4. S.D. : 9.4 Frequency (pixel) 1 1 8 6 4 Av. : 43.4 S.D. : 8. 1 3 4 5 6 7 8 9 1 3 4 5 6 7 8 9 Gradient (degree) Gradient (degree) (a) D-field (b) N-field Fig. 1 Histograms of the gradient at landslide points 8 QR of RTRI, Vol. 54, No., May 13

Vertical cross-sectional shape of slope Convex (X) Rectilinear (R) Concave (V) Horizontal cross-sectional shape of slope Ridge (r) Straight (s) Valley (v) Xr Xs Xv Rr Rs Rv Vr Vs Vv Incidence of landslides 6 5 4 3 1 Broad-leaf D-field N-field T-field S-field Mixed forest Conifer (matured) Conifer (young) Deforestation Secondary forest Bamboo Fig. 3 Incidence of landslides for each vegetation D-field N-field T-field Fig. S-field 4 6 Occurrence of landslides (%) 1 Definition of Slope types (above) and occurrence percentage of landslides (below).. Vegetation Vegetation types of each field were classified into broad-leaved forest, mixed forest, conifer forest (matured), conifer forest (young), deforestation area, secondary forest (young), and bamboo forest. Vegetation type at landslide locations was then examined. Judging from the number of landslides per 1km for each type of vegetation, there is a tendency which appears showing that landslides are likely to occur in deforested areas in each field (Fig. 3). In some fields, a large number of landslides also occurred in conifer-forests (young) and secondary forests (young). One of the causes of landslides therefore seems to be connected to specific vegetation types such as deforested areas, coniferforests (young) and secondary forests (young). 3. Quantifying landform and vegetation conditions by DEM and DSM According to the results from the previous chapter, slope gradient and type and vegetation can be selected 8 Digital Surface Model (DSM) Digital Elevation Model (DEM) Ground surface Fig. 4 Difference between the DSM and DEM as factors associated with landslides. A method was then studied to quantify these factors, using the DSMs and DEMs. DSMs and DEMs used in this study is a 1m data mesh; the DEMs are ground elevation grid data, and the DSMs are land cover elevation grid data (Fig. 4). 3.1 Quantifying landform conditions by DEM The slope gradient as one of the factors associated with landslides can be easily quantified by the gradients calculated from the DEMs (Equation 1). Investigations where made into how to classify slope types using DEMs. The slope types shown in Fig. are classified by a combination of the horizontal and vertical cross-sectional shape of the slope. Among the geomorphological quantities that can be calculated by the DEMs, Profile Curvature and Plan Curvature are available [5] to represent the crosssectional shape of the slope. In Fig. 5, the DEM has a grid point spacing d and Z1~Z9 are the elevation values of each grid point; then Profile Curvature (PrC) and Plan Curvature (PlC) are given by the following equations [5]: PrC = ( AD + BE + CDE) /( D + E ) () QR of RTRI, Vol. 54, No., May 13 81

Z1 z y x Z7 Z4 PlC = ( AE + BD + CDE) /( D + E ) (3) where A~E are given as follows: { } A = ( Z 4 + Z6) / Z5 / d { } Z8 B = ( Z + Z8) / Z5 / d d Fig. 5 Examples of DEM Z5 Z9 Z Z6 (4) (5) C = ( Z1+ Z3 + Z7 Z9) / 4d (6) D = ( Z 4 + Z6) / d (7) E = ( Z Z8) / d (8) Z3 Where PrC is negative, the vertical cross-sectional shape of the slope is convex, and where PrC is positive, the vertical cross-sectional shape of the slope is concave. Also if PlC is negative, the horizontal cross-sectional shape of the slope is a valley, and a positive PlC shows the horizontal crosssectional shape of slope is a ridge. In any curvature, shows that the slope shape is linear. It is possible to classify slope types by combining the Profile Curvature and Plan Curvature as shown in Fig. 6. However, because these curvature values are calculated from the elevation values of the eight neighboring grid points and target grid point, they will only represent the shape of a very narrow area on the ground surface. This cannot therefore be used to clas- Fig. 6 Profile curvature value Negative Positive Plan curvature value Positive Negative Xr Rr Vr Xs Rs Vs Xv Rv Vv Classification of slope types by profile curvature and plan curvature Fig. 7 Example of classification of slope types Xv Vv Xs Rs Vs Xr Vr 5 1 3 4 m sify the slope type by using a simple combination of these curvatures because these values do not show shape of the whole slope. Spatial filtering was then applied. The spatial filtering involves passing a square window over a surface, such as a DEM, and computing a statistical value of the covered window, returning the value to the grid point. The outcome of spatial filtering application (window of size 9 9, statistical value is mode) to the DEM is illustrated in Fig. 7. This method was used to classify valleys, ridges and taluses as Vv, Xr, Vr respectively, and it can be able to properly classify the slope types. 3. Quantifying vegetation by DEMs and DSMs Landslide occurrence is influenced by poor vegetation such as deforestation, where height of vegetation is lower than in conifer forests (matured) and broad-leaved forests. The DSMs therefore concentrated on these factors in the data showing land cover elevation. The DSMs are grid data of the land cover elevation, so vegetation height can be obtained by subtracting DEMs from DSMs. These sets of data were termed as Digital Canopy Models (DCMs). In order to classify vegetation using DCMs, each vegetation histogram from the DCMs was examined. Results showed the following tendency: (1) In broad-leaved forests, mixed forests, conifer forests (matured), bamboo forest areas. The average DCM values range from 15 to m, and DCM values are widely distributed between to 3m. () In conifer forests (young), secondary forest (young) areas. The values of DCMs are less than about 1m. (3) In deforested areas. DCM values are below about m, and DSM histograms peak at m. It is difficult to make a distinction among broadleaved, mixed, conifer (matured), and bamboo forests, because 8 QR of RTRI, Vol. 54, No., May 13

DCM value distributions and peak values are too similar. Since there is no significant difference in the frequency of landslide occurrence in these vegetation areas however (Fig. 3), it is reasonable to classify these types of forest together in a same group of low landslide incidence and taller vegetation. Following on, secondary (young) and conifer forests (young) can be grouped together given the relatively higher incidence of landslides and lower vegetation in these areas. Separating deforested areas from other types of land cover is fairly simple because the DSM histo- Deforestation area Secondary forest (young), Conifer (young) Broadleaf, Mixed forest, Conifer (matured), Bamboo (a) Vegetation map (by aerial photograph and field survey) gram for this type of area is significantly different from the histograms of other areas of vegetation. Vegetation areas were thus classified according to the following DCM histogram value-based criteria. (1) ~m: deforested area () ~7m: secondary forest (young), conifer forest (young) (3) 7m~: broad-leaved forest, mixed forest, conifer forest (matured), bamboo forest Comparison of the vegetation map based on aerial photograph interpretation and field surveys (Fig. 8 (a)) and the DCM vegetation map (Fig. 8 (b)), demonstrated that the distribution and shape of the deforested area and the secondary forest (young) and/or conifer forest (young) area in the Fig. 8 (b) correlate well with what is shown in Fig. 8 (a). However, there are many DCM low value plots in the distribution area of broad-leaved and conifer forests. This is due to the influence of the canopy gap. In order to eliminate this influence spatial filtering was applied in the same way as for slope classification. The result of applying spatial filtering (the window was of size 9 9, the statistical value was the mean) to the DCM is illustrated in Fig. 8 (c). This result shows good reproduction of the distribution and shape of each vegetation group. However, there is a tendency to misclassify the secondary forest (young) and/or the conifer forest (young) as the deforested areas in zones where there is poor vegetation such as the bottom of valleys. 4. Assessment of the incidence of landslides (b) Classification result of DCM ~m ~7m 7m~ The factors closely related to landslide occurrence which are classified by DEM and DSM can be treated as numerical data. The value of each prime factor was extracted for each of the landslides and sound slopes. The contribution of each prime factor to the occurrence of landslides by logistic regression analysis was then assessed [6]. A method was established for extracting the landslide hazard areas by weighted overlay according to the extent of the contribution of each factor. The extraction flowchart is shown in Fig. 9, and the result of the extraction is illustrated in Fig. 1. It should be noted that the validity of this extraction has been confirmed by field survey. Airborne LiDAR Digital Surface Model (DSM) Digital Elevation Model (DEM) Digital Canopy Model (DCM) Gradient Profile curvature Plan curvature Fig. 8 ~1.5m 1.5~6m 6m~ (c) Classification result of DCM after spatial filtering Comparison of vegetation map and classification result of DCM Vegetation map Gradient map Slope type map Extraction map of shallow landslide hazard Fig. 9 Extraction flowchart QR of RTRI, Vol. 54, No., May 13 83

Acknowledgment This work was financially supported in part by the Japanese Ministry of Land, Infrastructure, Transport and Tourism. References Fig. 1 Result of extraction map of landslide hazard 5. Conclusion Landslide hazard area Natural slopes adjoining the railway This report proposes a method for extracting landslide hazard areas using DEMs and DSMs. Landslide occurrence factors do not only include landform and vegetation but also lithologic features, geological structure, weathering of rock, state of fracture, thickness of overburden, groundwater, etc. At present it is difficult to obtain numerical information about these factors by remote sensing. As mentioned above, it is possible to extract landslide hazard areas using only landform and vegetation conditions. It is believed that this method can be used as a primary screening method to schematically investigate slopes and streams that are distributed over wide area. Also, if the DEMs and DSMs are obtained periodically, it is possible to grasp changes in the environment, such as the expansion of deforested areas. [1] Ishimaru, S. et al., Topographic and site characteristics of slope failures by Typhoon1, 3, in Ribira, Hidaka district, Hokkaido: analysis using airborne Li- DAR data, Journal of the Japanese Landslide Society, Vol. 45, No., pp. 137-146, 8 (in Japanese). [] Nakamura, Y. et al., River bed variation in the area of the sediment yield and transportation by LIDAR, Journal of the Japan Society of Erosion Control Engineering, Vol.59, No.4, pp. 54-57, 6 (in Japanese). [3] Ohta, T. et al., Evaluation of Natural Slope Stability by Digital Elevation Models and Satellite Imagery, RTRI Report, Vol. 19, No. 1, pp. 7-3, 5 (in Japanese). [4] Suzuki, T., Introduction to Map Reading for Civil Engineers Volume1. Kokon Shoin co. ltd., Tokyo, Japan, pp. 1-16, 1997 (in Japanese). [5] Perer A. B. and Rachael A. M., Principles of Geographical Information Systems. Oxford University Press Inc., New York, U. S. A., pp.19-193, 1998. [6] Annette J. Dobson, An Introduction to Generalized Linear Models, Second Edition. Kyoritsu shuppan co. ltd., Tokyo, Japan, pp. 139-179, 8 (in Japanese). 84 QR of RTRI, Vol. 54, No., May 13