You are on page 1of 12

Original Article

Landslides (2006) 3: 3950 DOI: 10.1007/s10346-005-0005-0 Received: 21 September 2004 Accepted: 24 May 2005 Published online: 17 January 2006 Springer-Verlag 2005

Hans-Balder Havenith Alexander Strom Fernando Caceres Eric Pirard

Analysis of landslide susceptibility in the Suusamyr region, Tien Shan: statistical and geotechnical approach

Abstract The Suusamyr region is located in the northern part of the Tien Shan Range in Central Asia. In 1992, this region was hit by the Ms = 7.3 Suusamyr earthquake triggering several large landslides along the Suusamyr Valley and on the southern slopes of the adjacent Suusamyr Range. One of these landslides had been investigated by geophysical and geotechnical methods in order to determine local trigger factors. The present paper focuses on the inuence of geological and morphological factors upon landslide occurrence on a regional scale. The analysis is based on a digital data set including landslides triggered in 1992 and several older landslides as well as various types of digital elevation models (DEMs), ASTER image data, and geological and active fault maps. These data were combined to compute landslide susceptibility (LS) maps using statistical methods, Landslide Factor and Conditional Analyses (LFA, CA), as well as a geotechnical one, the Newmarks Method (NM). The landslide data set was also analyzed with respect to the sizefrequency relationship. Keywords Landslides . Susceptibility . GIS . ASTER image . Suusamyr . Kyrgyzstan Abbreviations LS: Landslide Susceptibility . LFA, (M)CA: Landslide Factor Analysis, (Modied) Conditional Analysis . NM, ND: Newmarks Method, Newmark Displacement . PP: Predictive Power . Sc (Lb): map-scaled Scarp (Landslide body) area density . UCU: Unique Condition Unit . FS: Factor of Safety Introduction In the frame of the European project (19972000), Landslide Risk triggered by Earthquakes in Kyrgyzstan, Tien Shan, some landslides in northern Kyrgyzstan had been investigated by geophysical and geotechnical methods to determine local landslide trigger. The entire study also included dynamic modeling and slope stability computations to analyze the dependency of slope failure initiation on groundwater pressures, material shear strength, and seismic factors, such as ground motion amplication (Havenith et al. 2003). The present work was carried out to study the spatial inuence of geological and morphological factors upon landslide occurrence on a regional scale: the mapping of landslide susceptibility (or potential as in Gritzner et al. 2001). The basic principle of this approach is that under similar environmental conditions, the spatial distribution of past and recent slope-failures is the key for predicting slope movements in the future (Kuchai 1975; Carrara et al. 1995). LS mapping has become very popular since the development of efcient spatial analysis tools available with any GIS software. An important milestone in the application history of statistical LS analyses is the work published by Carrara et al. (1995). It provides a critical review of almost all relevant existing statistical methods. Another prominent study is the one by Guzzetti et al. (1999), which evaluates the efciency of various methods by comparing the results with each

other. Among the various available statistical methods, we have chosen one generally referred to as Conditional Analysis (CA). We also tested a modied version of this method based on a simple Landslide Factor Analysis, called here Modied Conditional analysis (MCA). An example of successful application of the CA is the assessment of LS in the Northern Apennines carried out by Clerici et al. (2002). Statistical-probabilistic methods are probably the most commonly used techniques, but other also proved to be valuable tools, such as the application of neural networks (Fernandez-Steeger 2002; Lee et al. 2004), fuzzy sets (Ercanoglu and Gokceoglu 2002), and physical or process-based models to LS mapping. The latter, also called geotechnical methods, are generally applied if a particular trigger factor is taken into consideration, such as soil wetness (Vanacker et al. 2003) and seismic effects (Khazai and Sitar 2000). In addition to the Conditional Analysis, we will apply a process-based technique, the Newmarks Method (NM), to LS mapping in the Suusamyr area and compare the produced maps with each other on the basis of their respective Predictive Power (PP). One goal of this study is to evaluate on the basis of this PP which method is the most efcient in providing predictions for large areas, such as the entire Tien Shan Range. Target area The Tien Shan is a Cenozoic orogenic belt in Central Asia with an eastwest extension of about 2500 km and a maximum width of more than 500 km. The structure of the Tien Shan is characterized by alternating, roughly eastwest trending mountain ranges and intermountain basins often bounded by neotectonic faults. This is also the case for the northern part of our study area, the Suusamyr Basin (20002500 m a.s.l.) lled with Neogene (siltstone and claystone) and Quaternary (alluvial and glacial deposits) sediments (Fig. 1). Active fault zones delimit the basin from the high-mountain ranges (3500 4800 m a.s.l.) made of Paleozoic granitic, sedimentary, or volcanic rocks (Fig. 1c). The entire target region also includes a part of the Naryn valley (Figs. 1c and 2a). In 1992, the Aramsu fault located in the west of the basin ruptured and produced a Ms = 7.3 event (main fault scarp location shown in Fig. 1c). Relatively few slope failures were triggered by this earthquake, such as the Belady rock avalanche and associated debris ow shown in Fig. 2a (location in Fig. 1c). During the same event, several landslides were triggered or re-activated on the southern slope of the ChetKorumdy ridge and partly destroyed the BishkekOsh highway (Fig. 2c). Input data and preprocessing The rst data set mainly included information that could be obtained free of charge, such as the NOAA (National Oceanic and Atmospheric Administration) and the SRTM (Shuttle Radar Topographic Mission) DEMs as well as LANDSAT 5 and 7 ETM+

Landslides 3 . 2006

39

Original Article
images. The low resolution of the NOAA DEM (about 1 km) clearly revealed to be insufcient to analyze surface morphology at regional scale. Hence, for this study we only used the 90 m SRTM DEM. The LANDSAT 5 and 7 ETM+ images were used as basis for georeferencing all collected data. Landslide bodies and scarps were outlined separately on KFA-1000 and KFA-3000 images, some location controls were made with CORONA images (example of zoomed CORONA image in Fig. 2b). Fault zones were mapped on the basis of the same images and eld observations. The recorded landslides include all kinds of coherent failures in rock (rock slides) and earth slopes (debris and earth slides, slumps) as well as several rock avalanches (e.g., Belady and Seit rock avalanches developed in granitic rocks, in Fig. 2a and 2b). Rock falls and debris ows, such as the one that developed from the Belady rock avalanche mass (Fig. 2a), were not mapped. Recently, two ASTER images (sequential images of the June 9, 2001; Fig. 3) were acquired in order to get a better topographic and image information. Indeed, ASTER images include 14 spectral bands, 3+1 (1,2,3N+3B) in the VNIR (visiblenear infrared, 15 m resolution) domain, 6 in the SWIR (short-wave infrared, 30 m), and 5 in the TIR (thermal infrared, 90 m) domains, while LANDSAT images only include 6 bands, at lower spatial resolution in the VNIR and SWIR domain. The 15 m 3N (nadir-looking) and 3B (backward looking) ASTER images in the VNIR domain provide a stereo-pair that can be used to construct reliable 30 m DEMs. The processing of the ASTER images for DEM extraction and orthorectication of the images has

Fig. 1 a Map of Central Asia. b NOAA DEM of the Northern and Central Tien Shan Mountains. c UTM zone 43N-projected SRTM (light shaded colors) and ASTER DEM (dark shaded colors, outlined) and digital geological map available for the target area, overlay of earthquakes (Ms

> 4, location of 1992 main shock is indicated), fault lines (brown), landslide body (white), and landslide scarps (black); locations of the Chet Korumdy ridge, the Belady and Seit rock avalanches, and the main 1992 fault rupture are shown

40

Landslides 3 . 2006

Fig. 2 a In 1992 earthquake-triggered Belady rock avalanche and associated debris ow (aerial photograph 1996). b Prehistoric Seit rock avalanche with scarp and accumulation outlined (CORONA image 1968). c In 1992 earthquake-triggered or re-activated landslides on the southern slope of the Chet Korumdy ridge (Photograph of 1998). Locations are indicated in Fig. 1c

Fig. 3 Mosaic of two ASTER images (limit marked by dashed line), 3-2-1 VNIR bands, with overlay of scarp (red) and landslide body (yellow) outlines as vertical view (a) and perspective views from the lower Suusamyr valley towards the NW (b) and from the western Suusamyr Range towards the E (c)

been carried out with the PCI Geomatica 8.0 Orthoengine software. A detailed description of the DEM generation within this PCI software is given by Al-Rousan et al. (1997). After extraction of the original DEMs, the latter had to be cleaned: rst, artifacts such as unrealistic peaks or holes (generally conic in shape) were manually removed; second, a median lter with 55 cell windows was applied.

Finally, the processed ASTER DEMs and the images were merged to produce a mosaic. The ASTER DEM mosaic and the 90 m SRTM DEM of the Suusamyr region are shown in Fig. 1, overlaid is also the geological map (with SRTM hill-shading) that had been digitized for a rectangular area including the Suusamyr Basin and adjacent ranges. Spectral information included in the various bands ASTER images that can be extracted by different processing techniques, e.g., by Landslides 3 . 2006 41

Original Article
Fig. 4 Frequency density function for landslide areas in the Suusamyr region

computation of ratios between the spectral bands. We principally applied the Principal Component (PC) transformation. For the LS analysis we used the three rst Principal Component images extracted from the combination of the 3 VNIR and 6 SWIR bands of the ASTER image mosaic. Landslide sizefrequency relationship Landslides were rst analyzed with regard to the size(cumulated) frequency distribution, considering the landslide body area as size unit since information about landslide volumes is generally not provided. It is certain that the use of areas instead of volumes will affect all the following analyses. In particular, the size of rock slides with a relatively large volume will be underestimated compared to the size of landslides in soft sediments characterized by volumes that may be signicantly smaller (for a similar area). This aspect will be considered in the nal discussion. The sizefrequency relationship of landslides (including also rockslides) were computed using the method suggested by Malamud et al. (2004) and analyzed in terms of the Frequency Density Function (f) of the landslide areas (AL Eq. (1)): f ( AL ) = NL AL ( 1)

relationship can be t by a power-law with an exponent of about 1.94. This value is signicant lower than those obtained for various types of landslide distributions Stark and Hovius (2001) and Malamud et al. (2004), which are larger than 2.2. Till now we cannot explain why our data set reveals a power law tail with such a low exponent. Creation of GIS platform As GIS tool we mainly used the Arcview 3.2 software and in particular its Spatial Analysis extension. The main part of the processing was carried out with data in grid (raster) format; hence, the mapping units are the pixels with a size depending on the involved data (90 m if SRTM DEM, 30 m if the ASTER DEM, 15 m if the PC images were used). Vectors, such as scarp, landslide body, and fault outlines were transformed into grids according to the following procedure. First, grids of distances to each feature were computed; in the case of the scarp outlines, the values of pixels within 100 m around the lines were set to 1, the others to 0 (i.e., 100 m buffer around lines); in the case of the landslide body polygons, a 50 m buffer (in addition to the landslide body area) was used. In the case of the fault zones, the computed Distance-to-Fault map was simply reclassied in 5 distance-units (see Tables 1 and 2). On the basis of the DEMs, three morphological factor maps were computed on the basis of a moving 33 cell window: the slope angle (SLOPE), slope aspect (ASPECT), and surface curvature (CURVATURE) maps. This processing combined with the previous median ltering introduces signicant smoothing of the DEM characteristics, which needs to be taken into consideration for the following analyses. While only one method is available for the computation of slope angles and aspects with the standard Spatial Analysis tool for
Table 1 Geological-tectonic factors and classications used for the LS analysis (GEOLOGY map extent)

where NL is the number of landslides with areas between AL and AL + AL . In a loglog graph (Fig. 4), this function shows the typical behavior of natural event records (earthquakes, landslides, etc.) characterized by a more or less linear tail (in loglog graph) for the large sizes which can be t by a power-law and a rollover for smaller events with decreasing frequency density for very small landslide areas. The rollover can partly be explained by undercounting of the small events (less accurate mapping) which can be taken into consideration by particular distribution ts proposed by Stark and Hovius (2001). On the other hand, Malamud et al. (2004) noticed that several data sets veried as complete exhibit a rollover with a maximum frequency density for an area of about 400500 m2 . Our data set reveals a maximum frequency density for about 7500 m2 , which may indicate that the inventory is incomplete for landslides smaller than 70008000 m2 . Beyond an area of 100,000 m2 the sizefrequency 42 Landslides 3 . 2006

Geology lithology Quaternary Neogene Pal. granite Pal. sediments Pal. volcanic

Sc 1.02 4.76 0.78 0.51 0.00

Lb 1.27 4.88 0.62 0.35 0.00

Fault distance (m) 0500 5002000 20005000 500010000 1000030000

Sc 4.37 2.16 1.15 0.56 0.17

Lb 4.20 2.57 0.99 0.46 0.27

Table 2 Morphological factors and spectral information used for the LS analysis (ASTER mosaic extent)

Slope ( ) 04 48 812 1216 1620 2024 2428 2832 3238 >38

Sc 0.38 1.11 1.54 1.55 1.40 1.10 0.73 0.53 0.46 0.32

Lb 0.54 1.62 1.91 1.63 1.23 0.82 0.50 0.24 0.15 0.13

Aspect ( ) 030 3060 6090 90120 120150 150180 180210 210240 240270 270300 300330 330360

Sc 1.42 1.33 1.13 1.12 1.20 1.10 0.88 0.82 0.80 0.56 0.65 1.02

Lb 1.52 1.25 0.95 1.06 1.01 1.09 0.91 0.74 0.83 0.67 0.83 1.15

CURVa (1/100m) <60 6050 5040 4030 3020 2010 1010 1020 2030 3040 4050 5060 >60

Sc 0.52 0.36 0.49 0.70 0.80 0.93 1.01 1.20 1.25 1.39 1.24 1.06 0.92

Lb 0.62 0.64 0.81 1.04 1.17 1.21 1.15 1.12 0.93 0.67 0.41 0.26 0.22

PC 1 1 2 3 4 5 6 7 8 9 10 11 12

Sc 0.53 1.09 1.51 1.24 1.07 0.82 0.64 0.50 0.41 0.24 0.12 0.08

Lb 0.45 1.06 1.61 1.45 1.01 0.60 0.39 0.29 0.13 0.08 0.06 0.19

PC2 1 2 3 4 5 6 7 8 9 10 11 12

Sc 1.37 1.47 1.31 1.25 1.10 0.58 0.33 0.22 0.22 0.26 0.19 0.08

Lb 1.06 1.31 1.44 1.33 1.15 0.60 0.26 0.10 0.06 0.07 0.07 0.19

PC3 1 2 3 4 5 6 7 8 9 10 11 12

Sc 0.20 0.05 0.05 0.11 0.49 1.18 1.59 1.05 0.84 0.61 0.37 0.29

Lb 0.38 0.45 0.30 0.22 0.76 1.44 1.54 0.96 0.88 0.66 0.20 0.14

Curv is the tangent curvature (1000 to avoid decimals), i.e., in inverse proportion to the radius of the curve adjusted to the earth surface at the pixel location ( is concave and + is convex), with values multiplied by 100 to reduce the number of decimals; Note, density values are enhanced by grayscale (light gray to black: <0.8, 0.81.2, 1.21.6, 1.62.4, >2.4), per unit the largest density is underlined; PC is Principal Component Sc (Lb) is the map-scaled scarp (landslide body) area density

Arcview, the CURVATURE can be calculated in terms of plan, prole, or tangent curvature. We considered the latter type as most useful for our study since it combines the characteristics of the two rst ones which are both relevant to landslide susceptibility analyses, i.e., predisposition to ow convergence or divergence and to erosion or deposition. We found only a few studies including the CURVATURE as potential landslide susceptibility factor, generally referring to the aforementioned hydrological, hydrogeological, or geomorphological characteristics of the curvature (Gritzner et al. 2001; Ayalew et al. 2004; Lee et al. 2004). In this paper, also other possible effects of this morphological factor (often neglected) will be discussed. The images of the three rst Principal Components computed from the merged ASTER data were added to the GIS platform as 15 m grids. All factor classes are summarized in Tables 1 and 2. Landslide factor analysis The Landslide Factor Analysis, the direct correlation between factor classes and landslide distribution, is commonly applied as rst approach to evaluate the environmental dependencies of landslide occurrence. For this study, landslide occurrence was divided into landslide scarp location and landslide body position in order to distinguish between the effects of the factors on the detachment failure and on the mass movement (and accumulation). It should be noticed, that most LS studies do not consider such a separation (Dai and Lee 2002; Ayalew et al. 2004; Lee et al. 2004 among others) or do not explain their use (Clerici et al. 2002). Another possible distinction can be made between the area within the landslide scar and the area surrounding it, which is more representative for prefailure conditions. This approach has been applied by S uzen and Doyuran (2004) to the LS mapping in Turkey by introducing the use of Seed-cells around the landslide. In this rst general study, it was decided that not only prelandslide conditions should be considered for the LS analysis but also those of developing and ancient landslides since they contribute to the overall future landslide potential (e.g., by re-activations). Thus, by buffering the features, both prelandslide (often more than 50% of landslide scarp area) and landslide conditions are taken into consideration by our analysis. A more detailed work is planned to evaluate LS only on the basis of prelandslide conditions.

The Landslide Factor Analysis is based on the computation of scarp (and landslide body) area densities within a certain factor class (or unit). The analyses were carried out over two different map extents (Geological map and ASTER mosaic, shown in Fig. 1c) and, in order to be comparable, the densities were scaled to the map extent according to the following Eq. (2): Sc(Lb) = buffer class counts map counts class counts buffer counts (2)

where Sc (Lb) is the map-scaled scarp (landslide body) area density, buffer-class-counts the number of pixels of the same class (= Unique Condition Unit, see next paragraph) within all scarp (landslide body) buffers, class-counts the number of all class (UCU) pixels, map-counts the number pixels within the map extent, and buffercounts the number of pixels within all scarp (landslide body) buffers. Thus, the rst factor is the simple density of class pixels within all landslide scarp (body) buffers and the second, the inversed total density of landslide scarp (body) buffers within the map. The density of all buffers of one type within a certain map extent is a constant; for the scarps it is 0.78 and 0.88% within the GEOLOGY map and ASTER mosaic extent, respectively; for the landslide body, it is 1.04 and 1.07% within the GEOLOGY map and ASTER mosaic extent, respectively. The results of the LFA are presented in Tables 1 and 2. Five factors have been included, one geological, one tectonic (Table 1), and three morphological (Table 2). Further, landslide distributions were spatially correlated with the PC images (Table 2). The latter are not considered as factors (direct inuence upon slope stability), but rather as additional information about particular earth surface characteristics. Indeed, it can be supposed that landslide prone areas are marked by a particular reectance related to the morphology and type of surface material (e.g., weathered rocks), to hydrogeological conditions (increased wetnesspresence of vegetation), etc. These conditions may present a spatial variation, which is not characterized by the former factor types (e.g., general geological map without any structural or hydrogeological or geotechnical information). From the Tables 1 and 2, it can be observed that landslide occurrence is most strongly depending upon geological and tectonic factors. Landslides 3 . 2006 43

Original Article
Indeed, the densities of scarps and landslide bodies are particularly large within the Neogene unit and in close vicinity to the faults. The Paleozoic rocks are clearly less prone to landslide occurrence than both the Neogene and Quaternary sediments. Factor-to-factor correlations and the map-fault zone overlay in Fig. 1c show that the presence of Neogene sediments at the surface and the location of faults are spatially connected. Thus, at this stage of the analysis, it cannot be dened, which of the two factors has the strongest inuence on landslide occurrence. On one hand, the clayey material with lower shear strength may explain the larger LS within the Neogene; on the other hand, reduced shear strength and enhanced seismic hazard (rupture, seismic source areas, seismic wave trapping) close to the faults also favor landslide occurrence. The morphological factors generally have a minor effect on the landslide and body location. Among them, the SLOPE is generally considered as the most important factor since larger slopes induce lower slope stability. However, here, even though the SLOPE appears as the most important morphological factor (largest density values), landslide body and scarp occurrence is not correlated with large but with relatively small slope angles. A similar paradox correlation has already been pointed out by Ayalew et al. (2004) and Clerici et al. (2002). As those authors, we think that the preferential location of landslides and their scarps within zones marked by small slope angles is likely to be related to a factor-to-factor spatial relationship, the interdependency between slope angle and geologicaltectonic factors. Indeed, numerous landslides and scarps formed within Neogene sediments (or close to faults), which are marked by milder slopes than areas within Paleozoic rocks (less prone to slope failure). Thus, the milder slopes can certainly be not considered as cause for enhanced landslide potential but result themselves from the presence of weaker geological materials prone to failure. ASPECT and CURVATURE are (almost) not spatially correlated with any of the two geological-tectonic factors. Their respective inuence has to be explained by other interactions with environmental conditions. The inuence of the ASPECT upon LS is generally explained by structural (rock foliation, orientation of bedding) and climatic (common wind direction) factors. For our region, such information is not provided. Thus, it cannot be quantitatively explained why landslides are preferentially located on slopes oriented to the NE. Considering the climatic factor, it can be supposed, however, that the dependency may be related to increased wetness on NE slopes due to larger snow accumulations in winter (preferred westerly winds) and slower melting in the shadows in spring time. The CURVATURE has apparently the weakest effect on slope stability but it is interesting to notice that, contrary to the former cases, preferred scarp and landslide body locations do not spatially correlate with the same class of this factor. Landslide bodies are preferentially located within slightly concave and scarps within slightly convex areas. Since the distinction between scarps and accumulations is generally not made, this difference has not yet been outlined according to our knowledge (Clerici et al. 2002 did not include the CURVATURE in their analysis). The discrepancy needs to be explained by taking into account various environmental factors. Those generally put forward and mentioned above are related to geomorphological and hydro(geo)logical aspects (Ayalew et al. 2004; Lee et al. 2004). These can explain why landslide bodies accumulate in concave zones classied as depositional areas (for sediments and convergence of water) but not why scarps form on convex slopes characterized by ow 44 Landslides 3 . 2006 divergence and deeper ground water table. We think that there are various physical reasons for this observation based on geomechanic and seismic aspects. First, under similar hydrogeological conditions, convex slopes are less stable (lower Factor of Safety) since a larger body (larger driving force) acts on the same sliding surface (equal resistant forces). Second, convex morphologies may indicate the presence of accumulation material (colluvium) characterized by lower shear resistance. Third, convex morphologies as ridge crests and hilltops are affected by ground motion amplication effects, which may have an impact on slope stability (Havenith et al. 2002). The morphology has a direct effect on the amplication (surface convexity induces seismic wave convergence) or indirect effects due to the presence of weathered material on hilltops. Hence, it is not surprising that landslides are commonly observed on convex slopes within areas affected by seismic shocks (Harp et al. 1981; Durville and M eneroud 1982 among others). Now, it still needs to be explained why zones marked by very strong convexity (very narrow ridge crests) are not prone to slope failure. One possible reason (which, however, cannot be proved) is that these morphologies mark the presence of very strong material among zones characterized by weaker materials. A similar correlation was put forward by Ayalew et al. (2004). This study and the one by Gritzner et al. (2001) reveal a bimodal relationship between landslide occurrence and CURVATURE. The doubled maximum, for slightly concave and for slightly convex zones, could be explained by the combined effect of the concave and convex morphology on the landslide body deposition and detachment not distinguished in the cited works. The interpretation of the previous statistical analysis cannot be generalized for all types of landslides, such as rock falls, which are not included in our study. Further, it should be reemphasized that the slope angle and curvature maps are affected by signicant smoothing; hence the slope angles and curvature prone to failure or deposition are likely to be somewhat larger in reality. From eld surveys we know for example that the upper part of the southern Chet Korumdy ridge slope has a 3034 slope angle where the ASTER and SRTM DEMs present a 2028 slope. Conditional analysis Original conditional analysis (CA) The principles of the Conditional Analysis (CA) are well outlined in Carrara et al. (1995) and graphically represented in Clerici et al. (2002). Basically, it consists in subdividing the entire area in units characterized by a specic combination of environmental conditionsthe Unique Condition Units (UCUs). In practice, we applied the combine function to several factor-grids to create one single grid including all the information of the former grids. Thereby, each pixel of the resulting grid belongs to a specic combination of the involved factorclasses (units), a UCU referred to by a new index. In order to evaluate the landslide susceptibility of the UCU, we computed the map-scaled scarp (landslide body) area densities for each UCU according to Eq. (1). Since the combination of even a few factor-grids may result in a grid with thousands of UCUs, some studies (Carrara et al. 1995) suggest applying neighborhood majority ltering to exclude small meaningless UCUs. This approach was also tested, but it was found that ltering computations strongly reduce the efciency of this method by increased processing-time and by requiring additional classication (introducing subjectivity). Hence, the results shown in the following were obtained without intermediate ltering. The disadvantage of not including ltering is that the number of UCU produced

by combinations of many factor classes is limited by computational constraints. Modied conditional analysis (MCA) In addition to the original CA, we also applied a modied type of the CA (called here MCA). It is based on the results of previous landslide factor analysis. The map-scaled densities obtained for the involved factor-classes are summed up and averaged for each pixel. The resulting map is reclassied (using Natural Breaks or Standard Deviation options) and per class the map-scaled density of scarps (landslide bodies) is computed to obtain the nal LS map. With each method, more than 10 grid-combinations were computed, at least ve for each map extent (GEOLOGY map and ASTER mosaic). The most relevant results are in Table 4, summarized after the next paragraph, in terms of maximum map-scaled densities and the respective count (number of pixels of the UCU) obtained for each map, and in terms of predictive power of the LS map. We dened the Predictive Power (PP) by the following Eq. (3): PP = averageofthe10%ofthemostsusceptibleUCU averageofthe90%oftheleastsusceptibleUCU (3)

with a probability of nonexceedance of 90% was considered). For the computation of the second type of map, we apply the (relatively free) assumption that all faults within the GEOLOGY map extent were at least once seismically activated (but not necessarily ruptured). Arias Intensity values produced over the area by the shocks and the attenuation of the seismic ground motion were computed using the empirical Arias Intensity attenuation relationship (Eq. (5)) proposed by Wilson and Keefer (1985) and valid for earthquake magnitudes less than 7: log Ia = 4.1 + M 2log R + 0.5 P (5)

Note that this comparative study was only carried out for the scarps; hence the susceptibility of a UCU is here dened as the map-scaled density of UCU within scarps (Sc).

Geotechncial approach Newmarks method As geotechnical or process-based model we applied the simplied Newmark approach to the Suusamyr region within the GEOLOGY map extent. This method is commonly used to evaluate LS in seismically active regions since, in some cases, it was shown that such geotechnical models can successfully predict failure potential (Jibson et al. 1998; Miles and Ho 1999); however, in others, a clear mismatch between computation results and landslide occurrence was observed (Khazai and Sitar 2000). The original Newmarks Method (Newmark 1965) is based on a simple model of a block sliding on an inclined plane. It aims at computing the distance of sliding during the seismic shaking (i.e., coseismic sliding), the Newmark Displacement (ND). The Newmark Displacement is computed from an acceleration time history by integrating twice the values larger than the critical acceleration (the threshold acceleration required to initiate sliding). For a detailed description of the procedure applied to GIS we refer to the works of Jibson et al. (1998) and Miles and Ho (1999). Here, the method will only roughly be outlined. It consists of a simplied computation scheme based on Eq. (4) (Miles and Ho 1999) by replacing the laborious integration over an accelerogram. By this equation, Newmark Displacements are directly calculated from Arias Intensities (Ia, Arias 1970) and critical accelerations (ac ). log(ND) = 1.46 log( Ia ) 6.642 ac + 1.546 (4)

with ND in (cm), Ia in (m/s), and ac in (m/s2 ). Arias Intensity maps were created in several ways. In most cases we used the map of probabilistic seismic hazard (see Abdrakhmatov et al. 2003) computed for this region both in terms of peak-ground acceleration and Arias Intensity values (here, a period of 50 years

where M is the earthquake magnitude (here Ms), R the hypocentral distance (here, the distance to the mean fault rupture depth), P is a probability term including possible variations and uncertainty, here set to 0. The computations were carried out for various magnitudes (5.47) and rupture depths (515). The critical acceleration value (ac ) in Eq. (4) is evaluated on the basis of the factor of safety (FS), both computed by means of simplied equations suggested by Jibson et al. (1998) and Miles and Ho (1999). This calculation involves morphological (slope angle) and geotechnical input data (shown in Table 3) and was only carried out for slopes larger than 4 and if the obtained FS value is larger than 1 (to exclude static instability leading to negative critical acceleration values). The geotechnical input includes shear strength parameters (cohesion, c, and internal friction angle, ), the specic gravity of the material ()) and water, the thickness of the potential sliding mass (t) and the proportion of the thickness, which is saturated (m). The respective values are assigned to each unit-pixel of the GEOLOGY map. First estimates of the shear strength parameter and specic gravity values of the Paleozoic rocks are based on data presented from Hoek and Bray (1981), which can be considered as representative for such materials. The parameter values assigned to the Quaternary and Neogene sediments have been directly measured on samples taken from trenches on the Chet Korumdy ridge (see Havenith et al. 2000). For the whole map, the thickness of the potential sliding layer (t) has been xed at 10 m; the proportion of thickness that is saturated (m) was either set constant for the whole map or varied over some units (see Table 3). The values marked by(1) are the rst estimates, those marked by(2) were determined after the Landslide Factor Analysis: principally, the shear strength of the Neogene sediments which proved to be prone to slope failure and the water table depth in these materials were both reduced (lower c, and larger m-value); conversely, the shear strength values of the obviously less susceptible Paleozoic rocks (especially the volcanic rocks where no landslides occurred) and the water table depth were increased (larger c, and smaller m-value). Unlike the maps produced by CA and like the maps produced by the MCA, the ND maps (Fig. 5f and g) directly provide an indication of LS. Ideally, larger NDs should indicate a higher LS (Jibson et al. 1998). A correlation between ND classes (using the probabilistic Ia map) and the respective map-scaled densities of scarps (Sc) inside these classes is shown in Fig. 6. This graph reveals that larger NDs calculated on the basis of the probabilistic Ia map and rst estimates of input data (curve ScND1) do not predict larger Sc-values, while estimates taking into consideration the statistical results provide a better (yet not perfect) correlation between increasing NDs and S-values (ScND2). The same graph also presents the correlation curve between Sc and ND calculated by using the map of Ia attenuation from (all) faults for a Ms = 6.6 seismic event with a hypocenter depth Landslides 3 . 2006 45

Original Article
Table 3 Geotechnical parameters used for FS and ND computations

Geology lithology Quaternary Neogene Pal. granite Pal. sedim Pal. volcanic
a

c1 (Mpa) 0.02 0.07 0.10 0.05 0.05

c2 (Mpa) 0.02 0.04 0.10 0.15 0.15

1 ( ) 26 24 34 32 27

2 ( ) 24 18 42 40 50

(kg/m2 ) 2100 2500 2600 2600 2800

t (m) 10 10 10 10 10

m1 0.5 0.5 0.5 0.5 0.5

m2 0.5 0.9 0.2 0.2 0.2

c, are the cohesion and internal friction angle values; is the specic gravity of the material, t is the thickness of the potential sliding mass, and m the proportion of the thickness which is saturated denote, respectively, initial parameter values and values modied after Landslide Factor Analysis

b1,2

curves (maximum of 80 cm), which are clearly marked by increasing map-scaled densities of scarps and hence correlate with higher LS. The comparison between results obtained for different magnitudes and hypocenter depths reveals that the NDs are strongly inuenced by the magnitude (as could be expected). For an equal magnitude, LS values Fig. 6 corresponding to large ND classes increase with decreasing hypocenter depth. This result clearly conrms the previous observation of the strong impact of the faults on LS. Newmark-conditional analysis method The Newmark method was combined with the Conditional Analysis to involve factors, which are not included in the simplied Newmark approach: the ASPECT, the CURVATURE, and the FAULT DISTANCE. Principally, the original version of the CA was applied by including the ND map as an additional factor map and combining it with the maps of the cited factors (see Table 4). As shown above, the Distance-to-Faults factor can also be taken into consideration by means of Ia attenuation from faults. Below it will be discussed which approach (combination with the Distance-to-Faults map or use of Ia attenuation from faults map) is more efcient and valid. Discussion Comparison between analyses In the following, the outcomes of the LS analyses will be discussed on the basis of Table 4 and detailed views of a few LS maps shown in

Fig. 5 Comparison between eight different maps of LS within the northern Suusamyr Region in terms of map-scaled scarp densities (yellow-orange-red-violet Sc-scale) and Newmark Displacements (yellow-green-blue-violet ND-scale in cm, pink areas represent zones with FS < 1). Overlaid are scarps (blue), landslide bodies (white), and faults (brown). Abbreviations used in the map titles are explained below Table 4. Note rectangles outlining two landslide-prone zones in 5ch

of 5 km. The depth is relatively shallow to include the effects of possible surface activation. The related curve ScNDFi presents much larger ND values (up to 1000 cm10 m) than the former
Fig. 6 Graph showing the correlation between Newmark Displacement (ND) classes and map-scaled scarp densities; for initial (curve Sc-ND1) and modied (curve Sc-ND2) estimates of parameter values, using the probabilistic Ia map; correlations using ND calculated for modied parameter values and the map of Ia attenuation from the faults is shown by curve Sc-ND2-Fi

46

Landslides 3 . 2006

Table 4 Summary of the results obtained by the various types of analyses

Methoda -Extentb FA-AST CA-AST CA-GEO CA-AST CA-GEO CA-GEO CA-GEO CA-AST CA-AST CA-AST CA-GEO CA-GEO MCA-GEO CA-GEO NM-GEO NM-GEO NM-GEO NM-CA-GEO NM-CA-GEO NM-CA-GEO

Combination S S-A S-A S-C S-C A-C S-A-C S-A-C PC123 PC123-S-C S-A-C-F S-A-C-G S-A-C-G S-A-C-G-F ND ND2 ND2-Fi ND2-F ND2-A-C ND2-A-C-F

Max Sc 1.55 2.69 8.06 2.18 4.43 3.97 46.67 113.86 5.27 113.86 122.24 122.24 6.27 122.24 0.59 2.70 6.98 88.28 88.28 88.28

Counts 653507 64114 430 84762 7784 327 55 7 12083 1 9 9 49775 3 7332 102 262 1 1 1

PP 1.65 2.32 2.75 2.23 2.42 1.92 3.86 2.99 3.88 5.75 13.86 10.24 5.24 44.90 0.83 3.50 5.08 7.59 6.00 9.00

PP values are enhanced by grayscale (light gray to black: <2, 23, 35, 510, >10), largest values are underlined
a

Methods:LFA,LandslideFactorAnalysis;(M)CA,(Modied)ConditionalAnalysis;NM,Newmark Method Extent AST, ASTER mosaic; GEO, GEOLOGY map Combination: S, SLOPE; A, ASPECT; C, CURVATURE; G, GEOLOGY; F, FAULT DISTANCE; PCxyz, Principal Components x,y,z; ND, Newmark Displacement; (2), modied input; Fi, Ia attenuation from faults

b c

the Figs. 5 and 7. In Table 4, results of 20 different analyses (different methods, map extends, combinations between factors) are presented in terms of maximum map-scaled density of scarps (Max Sc), respective counts (of pixels with maximum Sc), and the Predictive Power (PP dened by Eq. (2)) of the LS maps. As could be expected, the LS mapping based on one single factor (here the SLOPE angle) is characterized by the lowest PP. The comparison between the results obtained for various combinations between the morphological factors conrms the outcomes of the Landslide Factor Analysis: the slope has the strongest inuence, followed by the ASPECT, and then by the CURVATURE. Note that the Predictive Powers of maps combining the morphological factors over the GEOLOGY map extent using the SRTM DEM are systematically larger than those based on the ASTER DEM mosaic. Within this ASTER mosaic extent the PP of the map combining all three morphological factors is about 3. The combination of the three rst Principal Component images produces a map with a larger predictive power (3.88). In order to take into consideration both the morphological and spectral information, Principal Component images were combined with two of the morphological factors, SLOPE and CURVATURE (the combination with the third factor was not possible due to the computational limits). This allowed us to construct a LS map with a predictive power of up to 5.75 on the basis of the ASTER image mosaic alone. Combining the morphological factor maps with the GEOLOGY or the FAULT DISTANCE map induces a signicant increase of the PP, up to 10.24 for the SLOPE-ASPECT-CURVATURE-GEOLOGY and up to 13.86 for the SLOPE-ASPECT-CURVATURE-FAULT DISTANCE landslide susceptibility map. These results conrm the strong inu-

ence of the geology and tectonics upon LS (larger than the effect of one single morphological factor). From the comparison between the respective PP values, we can further infer that the Distance-to-Faults is the most important LS factor considered here. Finally, the combination between all available factors (within the GEOLOGY map extent) produces a map with very large PP of almost 45 (to compare with the 1.65 for the single SLOPE map). The PP of the ND maps were dened in the same way as for the previous maps, but considering that here the susceptibility of UCU is indicated by the ND itself (not the map-scaled density of scarps, Sc). This means that the 10% of the (presumed) most susceptible UCU do not necessarily include the largest Sc-value (c.f. Fig. 6), which may have a negative effect on the PP. Indeed, the PP of the ND map based on the initial estimates of input parameter values is less than 1, i.e., larger ND predict lower LS, which is the opposite of what is expected. The ND computations carried out for the modied inputs produce a much better LS map (PP = 3.5) similar, in terms of PP, to the map obtained by combination of all morphological factors (PP = 3.86). If in addition, the FAULT DISTANCE factor is introduced by considering the Ia attenuation from faults (Fi), the predictive power of the NDFi map is improved (PP = 5.08). However, the best results in terms of PP are obtained if the ND map is combined with other factor maps (PP up to 9). By comparing the PP of the combination ND2F (PP = 7.59) with the PP of the NDFi map (PP = 5.08), it appears that the introduction of the FAULT DISTANCE factor within a Conditional Analysis is more efcient than within the geotechnical analysis. It is also less affected by ad-hoc hypotheses (denition of Magnitude, hypocenter depth, etc.) For a few cases, it will be analyzed on the basis of detailed map views (in Figs. 5 and 7), which features contribute to the Predictive Power of the LS maps. First, the common, then the distinctive characteristics will be outlined comparing the maps in Fig. 5: all maps characterized by a PP larger than 1 present a higher LS along the edges of the basin and of the at hills within the basin. The difference between low LS within the ranges and high LS on the slopes surrounding the basin becomes more important with increasing number of factor combinations. Indeed, the S-A-C map (based on ASTER DEM, in Fig. 5a) shows only a faint difference, while the S-A-C-G-F map (Fig. 5d) reveals a very strong LS along the edges and a LS of almost 0 within the ranges (exceptions are sediment-lled mountain valleys). If Principal Component images are added to the combination of the morphological factors (Fig. 5b), the difference between LS inside and along the border of the ranges is enhanced as well. However, it is likely that this distinction is not entirely based on the physical properties of the ground marked by a certain reectance, but on particular surface conditions within the higher parts of the ranges: the snow-cover. Actually, the snow-covered areas marked by a strong reectance are deselected due to the low LS in these zones. Since the snow-cover changes over the year(s) (against a roughly constant LS), the contribution of the PC images needs to be considered with caution. Note that for future analyses we intend to apply masks to deselect the snow areas before analysis. By comparing the results obtained by the CA and MCA using the S-A-C-G combination (Fig. 5c and e), it can be observed that the CA produces maps with more differentiated LS zones. In particular, the CA allows to better distinguish LS within the NEOGENE unit and clearly outlines very large LS zones within the landslide-prone areas, such as the Chet Korumdy ridge (left rectangle shown in Fig. 5ch) where several landslides were triggered or re-activated by the M=7.3 event in 1992. The greater efciency of the CA with respect to the Landslides 3 . 2006 47

Original Article
Fig. 7 Comparison between four different LS maps in the eastern area (LS is given by Sc-scale). Overlaid are scarps (blue), landslide bodies (white), and faults (brown). Used abbreviations are explained below Table 4. Rectangles outline scarps of several rock failures

MCA is also marked by the larger PP (CA: PP = 10.24; MCA: PP = 5.24, see Table 4). The Chet Korumdy ridge is also well outlined on the ND maps (Fig. 6f and g) as zone with high LS (largest ND values in both maps). Actually, the upper part of this ridge is even represented as statically unstable (FS<1). In this regard we should mention that the geological map is not completely correct within the area of the Chet Korumdy ridge. Indeed, eld investigations (Havenith et al. 2003) revealed that this ridge is cored by Neogene sediments (as on the map) but covered by glacial Quaternary sediments with varying thickness. These deposits have a larger shear strength than the Neogene sediments (clays, silts) and hence are more stable. Besides the Chet Korumdy ridge, there is another landslide-prone area along the northern border of the Suusamyr Basin (right rectangle in Fig. 5ch). It is recognized as high LS zone only by maps based on a Conditional Analysis involving the FAULT DISTANCE factor (Fig. 5d and h). This may indicate that especially the clustering of landslides in this area is connected with the presence of the faults. Due to the high LS of the Neogene sediments, many analyses do not differentiate LS within the Paleozoic rocks. In this regard, we need to emphasize that the low LS predicted for the Paleozoic rocks is probably underestimated since correlations are made with failure areas. These do not represent the larger the volume of the detachment of rock slides compared to the one of earth slides in the soft Neogene and Quaternary sediments. In particular, the map based on the Modied Conditional Analysis including the SLOPE, ASPECT, CURVATURE, and GEOLOGY factors does not detect increased susceptibility close to rock slide scarps. An example is shown in Fig. 7c where rock slide (avalanche) scarps are outlined inside two black rectangles. This map (PP = 5.24) does not show any increased LS around these scarps while the map produced by the original CA applied to the same factors seems to better reect rock failure susceptibility (Fig. 7b). The best t between observed rock and landslide scarp location and computed large LS is provided by the combination of this latter map with the FAULT DISTANCE factor map Fig. 7d). This high ability to detect 48 Landslides 3 . 2006

unstable zones is quantitatively represented by the large predictive power of this map (PP45). The PC123-S-C map in Fig. 7a has a much lower PP (5.75). It is similar to the map in Fig. 7c, but clearly produces higher LS around rock failures than the latter (see Seit rock avalanche scarp within rectangle). This map, in particular, presents higher LS along ridge crests, marked by convex morphology, of representing the upper limit of rock slide scarps (see Seit rock avalanche scarp in Fig. 7a). There are still some examples where the generally more efcient statistical methods fail to predict failure, e.g., the Belady rock avalanche. The related scarp and surrounding area does not exhibit any particular feature that can be distinguished by the here involved factor combinations: it is located in the less landslide-prone granitic rocks, no faults were mapped in the vicinity, the slope is steep (statistically less susceptible in this region) oriented toward the south, and the scarp is not directly adjacent to a ridge crest. It is likely that other factors, such as the rock fabric and weathering need to be included to better predict such failures. Note that the ND1 map based on the initial input data estimates and characterized by a very low PP (0.83 in Table 4) detects a potentially unstable area in the upper part of the rock avalanche scarp, probably due to the large slope angle and the lower shear strength used for ND1. Postprocessing As mentioned above, no ltering was applied during the analyses in order to keep the approach as general as possible (ltering and reclassication introduces subjectivity). By there, thousands of UCU can be produced and even neighboring pixels may belong to UCUs with completely different Sc-values. This result can be considered as scientically relevant since landslides do not necessarily develop where the whole but only where a part of the slope is unstable. The propagation of instability from unstable to stable zones can be explained by failure dynamics. On the other hand, the strong spatial variability of the Sc-value makes it difcult to discern zones with a higher LS. In particular, risk assessment and management requires the distinction

of wider potentially unstable areas. Therefore, it might reveal to be useful to apply neighborhood ltering as postprocessing, with median lters applied to a specic engineering purpose (e.g., ner lters for assessment at the scale of private construction, smoother ltering for large engineering projects like dams). Since landslide risk is not the topic of this paper, maps obtained after ltering will not be presented. We will just outline the most susceptible zones in the Suusamyr Naryn Region, marked also by the presence of landslide clusters: the Chet Korumdy ridge (1) in Neogene and Quaternary sediments and several zones close the contact between soft sediments and grantitic rocks, most of them being crossed by active faults. It should also be noticed that combinations of many factors produce numerous singular UCU (count of 1, see Table 4) with very large Sc-value (above 10), which are only representative for particular features within some scarp zones (additional effect of no ltering). Thus, in general, the very large Sc-values (10150 class in violet in Figs. 5 and 7) mainly characterize existing sliding areas where re-activation may occur in future (e.g., the ve cited zones). Only very few small zones outside (but still close to) these areas belong to the highest LS class whereas medium to high LS is also predicted in areas where no landslides developed till now. These zones present a certain potential for new failures and hence need to be taken into consideration by risk assessment. Conclusion Several methods, statistical and geotechnical, were applied to analyze and map landslide susceptibility over the Suusamyr region. The map is based on various data inputs: digitized landslide scarp and accumulation as well as fault outlines, SRTM DEM, mosaic of two ASTER images (and associated DEM), and the digitized geological map. The DEMs were processed to extract the morphological information, SLOPE, ASPECT, and CURVATURE. The spectral information of the ASTER images was used in terms of Principal Component images. On the basis of the digitized fault lines a Distance-to-Fault map was computed. All applied methods, Landslide Factor Analysis, Conditional Analysis, and Newmark Displacement computation revealed the important inuence of the GEOLOGY and the FAULT DISTANCE on the landslide susceptibility compared to minor effects of the morphological factors (which may have been reduced by the ltering). Combining all factors, the most susceptible failure zones are located in Neogene sediments crossed by faults, on relatively mild slopes (<20 ) dipping towards the NE (or SE) and presenting a slight convex morphology. In general terms, almost all LS maps clearly delimit landslide-prone zones along the edges of the intramontane basins lled by soft sediments from wide stable areas within the Suusamyr Basin and the higher parts of the mountain ranges. Among the various LS mapping methods, the Conditional Analysis proved to be more efcient than the geotechnical methods to predict both rock and earth slope failures. Since the computational efforts are reasonable (if not too many classes are used), the CA is well suited to map LS over large areas, such as the entire Tien Shan. SRTM and ASTER DEMs (+ spectral information) are cost-effective input data enabling rapid but quite rough predictions of LS, while additional geological and tectonic information allows us to map LS with much more detail. A major advantage of the classic geotechnical methods with regard to statistical techniques is that they allow us to produce LS maps for regions without a priori knowledge of existing landslide

data, provided that digital geological data are available. The reliability of the predictions may, however, only be checked by spatial correlation with landslide distribution; it was shown that it strongly depends on the complexity of the simulated processes and on the quality of the collected shear strength, hydrogeologic, and structural information. Acknowledgements Some data used for our research were produced by a COPERNICUS project funded by the European Community (EC, DG XII contract no. IC15-CT97-0202). HBH wishes to thank the Fonds National de la Recherche Scientique of Belgium for the 5 years nancial support of his research at the University of Liege. References
Abdrakhmatov K, Havenith HB, Delvaux D, Jongmans D, Trefois P (2003) Probabilistic PGA and Arias Intensity maps of Kyrgyzstan (Central Asia). J Seism 7:203220 Al-Rousan N, Cheng P, Petrie G, Toutin T, Valadan Zoej MJ (1997) Automated DEM Extraction and Orthoimage Generation from SPOT 1B Imagery. Photogrammetric Eng Remote Sensing 63(8):965974 Arias A (1970) A measure of earthquake intensity. In Hansen RJ (ed) Seismic design for nuclear power plants. MIT Press, Cambridge, MA, pp 438483 Ayalew L, Yamagashi H, Ugawa N (2004) Landslide susceptibility mapping using GIS-based weighted linear combination, the case in Tsugawa area of Agano River, Niigata Prefecture. Jpn Landslides 1:7381 Carrara A, Cardinali M, Guzzetti F, Reichenbach P (1995) GIS technology in mapping landslide hazard. In: Carrara A, Guzzetti F (eds) Geographical Information Systems in assessing natural hazards. Kluwer Academic, Dordrecht, The Netherlands, pp 135175 Clerici A, Perego S, Tellini C, Vescovi P (2002) A procedure for landslide susceptibility zonation by the conditional analysis method. Geomorphology 48:349364 Durville JL, M eneroud JP (1982) Ph enom` enes g eomorphologiques induits par le s eisme dEl Asnam, Alg erie. Comparaison avec le s eisme de Campanie, Italie. Bull Liaison Labo P Ch 120:1323 Ercanoglu M, Gokceoglu C (2002) Assessment of landslide susceptibility for landslide-prone area (north of Yenice, NW Turkey) by fuzzy approach. Environ Geol 41:720730 Fern andez-Steeger TM (2002) Erkennung von Hangrutschungssystemen mit Neuralen Netzen als Grundlage f ur Georisikoanalysen. PhD Thesis, Universit at Karlsruhe Ghose S, Mellors RJ, Korjenkov AM, Hamburger MW, Pavlis TL, Pavlis GL, Mamyrov E, Muraliev AR (1997) The Ms = 7.3 1992 Suusamyr, Kyrgyzstan earthquake: 2. Aftershock focal mechanisms and surface deformation. Bull Seism Soc Am 87:2338 Gritzner ML, Marcus WA, Aspinall R, Custer SG (2001) Assessing landslide potential using GIS, soil wetness modeling, and topographic attributes, Payette River, Idaho. Geomorphology 37:149 165 Guzzetti F, Carrara A, Cardinali M, Reichenbach P (1999) Landslide hazard evaluation : a review of current techniques and their application in a multiscale study, Central Italy. Geomorphology 31:181 216 Harp EL, Wilson RC, Wieczorec GF (1981) Landslides from the February 4, 1976, Guatemala earthquake. The Guatemala earthquake of February 4, 1976. Geol Surv Prof Pap 1204-A:135 Havenith HB, Jongmans D, Abdrakhmatov K, Trefois P, Delvaux D, Torgoev IA (2000) Geophysical investigation of seismically induced surface effects: case study of a landslide in the Suusamyr valley, Kyrgyzstan. Surv Geophys 21:349369 Havenith HB, Jongmans D, Faccioli E, Abdrakhmatov K, Bard PY (2002) Site effects analysis around the seismically induced Ananevo rockslide, Kyrgyzstan. Bull Seism Soc Am 92:31903209 Havenith HB, Strom A, Jongmans D, Abdrakhmatov K, Delvaux D, Trefois P (2003) Seismic triggering of landslides, Part A: eld evidence from the Northern Tien Shan. Nat Hazard Earth Syst Sci 3:135 149 Hoek E, Bray J (1981) Rock slope engineering. Institution of Mining and metallurgy Jibson RW, Harp EL, Michael JA (1998) A method for producing digital probabilistic seismic landslide hazard maps: an example from Los Angeles, California Area. U.S. Geological Survey, Open-le Report 98-113 Khazai B, Sitar N (2000) Assessment of seismic slope stability using GIS modeling. Geogr Inf Sci 6(2):121128 KuchaiVK(1975)Estimationoftherecurrenceandactivityoflandslides.SovGeol5:142149(inRussian) Lee S, Ryu JH, Won JS, Park HJ (2004) Determination and application of the weights for landslide susceptibility mapping using an articial neural network. Eng Geol 71:289302 Malamud BD, Turcotte DL, Guzzetti F, Reichenbach P (2004) Landslide inventories and their statistical properties. Earth Surf Process Landforms 29:687711 Miles SB, Ho CL (1999) Rigorous landslide hazard zonation using Newmarks method and stochastic ground motion simulation. Soil Dyn Earth Eng 18:305323

Landslides 3 . 2006

49

Original Article
Newmark N (1965) Effects of earthquakes on dams and embankments. Geotechnique 15:137160 Stark CP, Hovius N (2001) The characterization of landslide size distributions. Geophs Res Lett 28(6):10911094 S uzen ML, Doyuran V (2004) Data driven bivariate landslide susceptibility assessment using geographical information systems: a method and application to Asarsuyu catchment, Turkey. Eng Geol 71:303321 Vanacker V, Vanderschaeghe M, Govers G, Willems E, Poesen J, Deckers J, De Bievre B (2003) Linking hydrological, innite slope stability, and land-use change models through GIS for assessing the impactofdeforestationonslopestabilityinhighAndeanwatersheds.Geomorphology52:299315 Wilson RC, Keefer DK (1985) Predicting the areal limits of earthquake-induced landsliding. In: Ziony (ed) Evaluating earthquake hazards in the Los Angeles regionAn Earth Science perspective, U.S. Geological Survey Professional Paper 1360, pp 316345 H.-B. Havenith F. Caceres E. Pirard ( Department GeomaC, University of Liege, B52/3, Sart Tilman B52, 4000 Liege, Belgium e-mail: eric.pirard@ulg.ac.be Tel.: +32-4-3669528 Fax: +32-4-3669520 )

A. Strom Institute of Geospheres Dynamics, Russian Academy of Sciences, Leninsky Avenue, 119334 Moscow, Russia Present address: H.-B. Havenith Institute of Geophysics, Swiss Seismological Service, ETH-H onggerberg, CH-8093 Zurich, Switzerland e-mail: havenith@sed.ethz.ch

50

Landslides 3 . 2006

You might also like