Landslides occur when unstable rock and soil masseson slopes are disturbed by earthquakes, intense storms,human activities such as road construction, or acombination of these factors (Keefer, 1984; Aleottiand Chowdhury, 1999; O'Hare and Rivas, 2005).Landslide is one of the most destructive natural hazards(Chung and Fabbri, 2005; Guzzetti et al., 2005) and amajor concern for watershed management (Sidle et al.,1985). In recent years, many researchers have used
logistic regression to predict probabilities of landslideoccurrence by analyzing the functional relationshipsbetween the instability factors and the past distributionof landslides (Guzzetti et al., 1999; Dai and Lee, 2003;Ohlmacher and Davis, 2003; Ayalew and Yamagishi,2005; Can et al., 2005;Wang et al., 2005; Yesilnacar andTopal, 2005). The assumption is that factors, whichcaused landslides in the past, are the same as those,which will trigger landslides in the future.
Most logistic regression models published so farhave been targeted at rainfall-induced landslides (e.g.,Dai and Lee, 2003). However, in areas such as Taiwan,where landslides can be triggered by earthquakes andintense storms brought on by typhoons (tropical cy-clones), it is important to model both types of landslides
336 K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
for the following two reasons. First, rainfall- andearthquake-induced landslides are likely to differ interms of related environmental factors (Aleotti andChowdhury, 1999). Unless these differences are consid-ered, we can only get a partial understanding of land-slide occurrence. Second, it has been suggested that, inareas such as Taiwan, rainfall-induced landslides tend toincrease in both number and magnitude and steeperslopes become more vulnerable after a strong earth-quake (Chang and Slaymaker, 2002; Lin et al., 2003;Dadson et al., 2004; Cheng et al., 2005). Therefore, anearthquake-induced model is likely to be able to augmenta rainfall-induced model in its predictive capability.
In this landslide study, we used landslides triggeredby a typhoon in 1996 and a major earthquake in 1999 tobuild a typhoon-induced (i.e., rainfall-induced) modeland an earthquake-induced model, respectively. We thenused landslides triggered by three typhoons after theearthquake to validate the typhoon-induced model andto test the usefulness of the earthquake-induced modelfor predicting typhoon-induced landslides. These eventsrepresent a natural experiment (Dadson et al., 2004),which allows us to investigate the difference as well asthe interaction between typhoon- and earthquake-
Fig. 1. Topographic map (left) and oblique aerial
induced models, topics that have not yet been coveredin the recent literature of landslide modeling. Theorganization of this paper is as follows. Section 2describes the study area and landslide data. Section 3explains logistic regression and the explanatory vari-ables used in this study. Section 4 presents the results,compares the topographic and locational differencesbetween typhoon- and earthquake-induced landslides,and shows how the earthquake-induced model can helpexplain typhoon-induced landslides. Section 5 discussesthe findings of this study and their implications,followed by a short conclusion in Section 6.
2. Study area and landslide data
2.1. Study area
Our study area is the Hoshe basin in central Taiwan(Fig. 1). This basin represents a typical mountainouswatershed in Taiwan that experiences landslides trig-gered by both storms and earthquakes. As part ofNational Taiwan University's experimental forest, theHoshe basin also has more data and references availablethan other mountainous watersheds in Taiwan.
photograph (bottom right) of the study area.
337K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
Draining an area of 92 km2, the Hoshe River is thesource of the Cheyulan River, a tributary of the ChoshuiRiver, the longest river in Taiwan. Three Miocenelithologic formations are distributed away from theterraces of the Hoshe River: the Hoshe (Hs) Formationwith shales and thin interbeds of sandstones, theNanchuang (Nc) Formation with layers of sandstonesand shales, and the Kueichulin (Kcl) Formation withsandstones (Fig. 2). Exposed rocks are heavily fracturedby joints from folding and faulting (Chen and Su, 2001).Two sets of strike–slip faults, striking WNW and NNE,respectively, dominate the basin.
The Hoshe basin ranges from 770 m to N2850 m inelevation, with a mean of 1760 m. Elevations generallydecrease from the divide toward the stream channel in thenorth central of the basin (Fig. 1). The highest elevation islocated along the southeastern border. Slopes range from0° to 77°, with a mean of 33°. The climate is subtropicalwith rainforest vegetation, with mean annual temperatureof 16 °C andmean annual precipitation of 2300mm.Over70% of rainfall occurs between July and September,brought on by typhoons and heavy summer convectionalthunderstorms.
Fig. 2. Geological map of the study area. Mod
2.2. Landslide data
Taiwan is located at the collision zone between thePhilippine Sea Plate and the Eurasian Plate. Crustalshortening, estimated at 80 mm per year (Yu et al.,1997), is accompanied by frequent large earthquakes.Earthquake-triggered landslides for this study werecaused by the Chi-Chi earthquake, which struck centralTaiwan on September 21, 1999 with a momentmagnitude of 7.6. This earthquake occurred on theChelungpu thrust fault, and the epicenter was locatedabout 60 km to the northwest of the study area. Theimpact of the Chi-Chi earthquake, including landslides,has been reported in numerous publications (e.g., Liao,2000; Lin et al., 2003; Lin and Tung, 2003).
Taiwan has an average of four typhoons per year (Wuand Kuo, 1999). Typhoon-triggered landslides for thisstudy were caused by four major typhoons between1996 and 2005. Typhoon Herb in 1996 came before theChi-Chi earthquake, and Toraji in 2001, Mindulle in2004, and Haitang in 2005 came after the earthquake.Because the previous major earthquake (a momentmagnitude of 6.0) occurred 80 years before typhoon
ified after Chang and Slaymaker (2002).
Table 2Image data sources relevant to each event
Event Satellite image sets Aerial photograph sets
Typhoon Herb – 1995/01/08July 31, 1996 1996/08/21 a 1996/08/18Chi-Chi earthquake 1999/01/01 a 1999/03/06Sep. 21, 1999 1999/12/10 a 1999/10/31Typhoon Toraji – 2001/07/02July 28, 2001 2001/08/01 a 2001/08/22Typhoon Mindulle – 2004/04/21Jun. 30, 2004 2004/07/08 a 2004/07/12Typhoon Haitang 2005/07/06 b –July 16, 2005 2005/07/25 b –
– No images.a 12.5 m SPOT images.b 2 m FORMOSAT-2 images.
Table 1The duration, total rainfall, and maximum 1-h intensity for eachtyphoon
338 K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
Herb, landslides triggered by Herb are believed to bewithout the antecedent earthquake effect. Table 1 liststhe duration, total rainfall, and maximum 1-h intensityfor each typhoon from stations within the study area.The tracks of these typhoons were similar: theyapproached Taiwan from the Pacific in the northwesterlydirection, made landfall on the northeast or east coast ofTaiwan, and continued in the northwesterly directionbefore leaving the island.
We interpreted and delineated new landslidestriggered by the Chi-Chi earthquake and typhoonsHerb, Toraji, and Mindulle by comparing ortho-rectifiedaerial photographs taken before and after each event.These color orthophotographs were compiled by theAerial Survey Office of Taiwan's Forestry Bureau fromthe stereo pairs of 1:5000 aerial photographs. They havea pixel size of 0.35 m and an estimated horizontalaccuracy of 0.5 m. While delineating landslides on theorthophotographs, we also used SPOT images asreferences. We interpreted and mapped new landslidestriggered by typhoon Haitang on panchromatic FOR-MOSAT-2 images. These images have a spatialresolution of 2 m. Table 2 summarizes the data sourcesfor landslides and their dates.
Fig. 3 shows the distribution of landslides induced bythe earthquake and typhoons. Table 3 lists the number, thedescriptive statistics of sizes (range, mean, and standarddeviation), the total area, and the area percentage of thestudy area of these landslides. Landslides triggered by theChi-Chi earthquake had the largest mean (1.02 ha) andrange (b0.1–20.0 ha). However, the largest number(=1750) of landslides was triggered by typhoon Toraji,the first major typhoon following the earthquake. Some ofthese Toraji-triggered landslides were also large, up to18.5 ha in size. Most landslides, induced by either theearthquake or typhoons, were shallow landslides (Linet al., 2003; Cheng et al., 2005). For this study, we usedlandslides induced by the Chi-Chi earthquake andtyphoon Herb to develop the models, and landslidesinduced by typhoons Toraji, Mindulle, and Haitang formodel validation and prediction.
3. Methods
3.1. Logistic regression
Logistic regression is useful when the dependentvariable is categorical (e.g., presence or absence) and theexplanatory (independent) variables are categorical,numerical, or both (Menard, 2002). The logit modelfrom a logistic regression has the following form:
logitðyÞ ¼ aþ b1x1 þ b2x2 þ b3x3 þ : : : þ e ð1Þwhere y is the dependent variable, xi is the i-th explanatoryvariable, a is a constant, bi is the i-th regressioncoefficient, and e is the error term. The logit of y is thenatural logarithm of the odds:
logitðyÞ ¼ ln½ p=ð1−pÞ� ð2Þwhere p is the probability of the occurrence of y, andp/(1−p) is the odds. To convert logit( y) back to theprobability p, Eq. (2) can be rewritten as:
To develop our landslide models, we used adependent variable that separated landslide areas fromstable areas and explanatory variables that representedrelated environmental factors. Of the twelve explanatoryvariables, eight were numerical: elevation, slope, aspect,distance to fault line, distance to channel, distance toridge line, the NDVI (normalized difference vegetationindex), and the wetness index. Four variables werecategorical: surface shape, lithology, order of sub-basin,
Fig. 3. Distribution of landslides triggered by typhoon Herb in 1996, the Chi-Chi earthquake in 1999, and typhoons Toraji, Mindulle, and Haitang in2001, 2004, and 2005.
339K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
and road buffer. Table 4 shows the data source andresolution of these variables, and Table 5 shows theirdescriptive statistics including the range, mean, and
Table 3Descriptive statistics of landslide areas (ha) from the Chi-Chi earthquake an
Table 4Data, resolution/scale, data source, and procedure for the independentvariables
Data Resolution/Scale
Data source Procedure
DEM 40 m Aerial SurveyOffice ofTaiwan'sForestry Bureau
Elevation, slope,aspect, wetnessindex, surfaceshape, and orderof sub-basin arederived fromDEM data.
SPOT 12.5 m SPOT Image NDVI is derivedfrom SPOT imagery.
Geologicmap
1:50000 Taiwan'sCentralGeologicalSurvey
Lithology and faultlines are compiledfrom the geologicmap.
Topographicmap
1:5000 Aerial SurveyOffice ofTaiwan'sForestry Bureau
Roads, streamchannels, and ridgelines are compiledfrom the topographicmap.
340 K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
Topography is important for landslide studies. Thisstudy used common topographic variables of elevation,slope, and aspect. Landslides tend to occur on steeperslopes, especially where the slope is covered by a thincolluvium. Aspect can influence moisture retention and
Table 5Descriptive statistics of the numerical and categorical explanatory variables
Variable Numerical variables
Range
Elevation (m) 768–2858Slope (deg) 0–77Sin of aspect 1∼−1Cos of aspect 1∼−1Plane curvature (m−1) 16.50∼−12.38Profile curvature (m−1) 15.34∼−11.70Distance to fault line (m) 0–6130Distance to channel (m) 0–987Distance to ridge line (m) 0–686Wetness Index 1.99–17.14NDVI −0.64–0.71
Variable Class Categorical variables
% Study area
Lithology Terrace 4.39Hs 60.58Nc 31.28Kcl 3.75
Road buffer Yes 9.04No 90.96
vegetation, which in turn can affect soil strength andsusceptibility to landslide. For earthquake-triggered land-slides, aspect can also play a role in determining thedistribution of landslide density (Tibaldi et al., 1995).Because aspect is a circular measure, aspect measures fordata analysis are typically converted to their sine or cosinevalues, which range from −1 to 1 (Chang, 2005). A sinetransformation emphasizes the contrast between east andwest exposures, and a cosine transformation the contrastbetween north and south exposures.
Other topographic variables for this study includedslope curvature, the order of sub-basin, and the wetnessindex. A curvature can be concave or convex. Aconcave surface can lead to concentration of subsurfacedrainage, resulting in high pore water pressure and alikely condition for triggering landslides, especiallyshallow landslides (Pierson, 1980; Ayalew et al., 2004).In this study, we used both profile curvature and plancurvature. A negative profile curvature indicates anupwardly convex surface along the direction ofmaximum slope; a positive profile curvature, anupwardly concave surface; and a value of zero, a flatsurface. A positive plan curvature indicates an upwardlyconvex surface across the direction of maximum slope; anegative plan curvature, an upwardly concave surface;and a value of zero, a flat surface. The order of sub-basinrelates to the location of sub-basin (upstream or
S.D.
Mean
1760.23 451.8832.69 10.400.18 0.660.01 0.73
−0.06 1.070.08 0.96
2411.07 1435.40260.27 175.33188.31 140.41
5.37 1.360.55 0.11
Class %Studyarea
Variable
Order of sub-basin 1st 59.952nd 20.373rd 12.344th 3.215th 4.13
341K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
downstream) and the size of contributing area. Thewetness index combines local upslope contributing area(a) and slope (tanβ) in the formula, ln(a/tanβ), tomeasure topographic control on hydrologic processes(Beven and Kirkby, 1979; Quinn et al., 1995). The orderof sub-basin and the wetness index measure basin- andlocal-level pore water pressure, respectively.
We used three rock formations (Kueichulin, Nan-chuang, and Hoshe) for logit modeling, with alluvialterrace serving as the reference category in themodeling. These rock formations vary in the propertiesof the slope-forming materials such as strength andpermeability and can therefore affect slope failure.
Four distance variables were included in this study.Wecreated a 40-m road buffer, which divided the study areainto either within or outside of the buffer. The frequencyof landslides may increase along roads because ofundercuts (Ayalew and Yamagishi, 2005). We measureddistances to fault line, channel, and ridge line. As thedistance to fault line decreases, landslides may increase infrequency due to increases in rock fracture and weather-ing. Likewise, as the distance to channel decreases,landslides may increase in frequency due to channelerosion. Moreover, at least one study has indicated anegative relationship between distance to ridge line andearthquake-induced landslides (Chang and Hsu, 2004).
Finally, we calculated the NDVI and used it as anindependent variable. Defined as (near IR band− redband) / (near IR band+red band), the NDVI measuresthe density of surface vegetation. Dense vegetation,especially of a woody type with strong and large rootsystems, can improve slope stability (Wu and Swanston,1980).
3.3. Raster-based models
Data analysis for this study was raster-based using acell size of 40 m because half of the explanatory variables(elevation, slope, aspect, thewetness index, surface shape,and the order of sub-basin) for logit modeling werederived from 40mDEMs, standard DEMs in Taiwan.Weconverted (rasterized) each digitized landslide map into40 m cells and coded each cell 1 for landslide and 0 forstable. The numbers of landslide cells for each event are asfollows: 783 for typhoon Herb, 4213 for Chi-Chiearthquake, 6294 for typhoon Toraji, 2161 for typhoonMindulle, and 1146 for typhoon Haitang. We alsorasterized the lithology map, calculated the NDVI valuefor each cell, and measured straight-line distances in cellsto create the four distance variables. From the prepareddatabase, we took a random sample of 500 landslide cellsand 500 stable cells for developing each model. It is
generally recommended in logistic regression to use equalproportions of 1 (landslide) and 0 (stable) cells (e.g., Daiand Lee, 2002; Yesilnacar and Topal, 2005). By using arandom sample, we also avoided the problem of spatialautocorrelation. We ran logistic regression analysis in theStatistical Package of Social Sciences (SPSS) using astepwise (forwardWald) method to avoid the collinearityproblem between explanatory variables.
3.4. Assessment of model performance
To assess model performance, we first classified acell in the output as a landslide cell if its probability oflandslide occurrence was 0.5 or greater, and a stable cellif its probability was less than 0.5. Then we used the areaconcordance (Borghuis et al., in press) to measure howwell landslide cells from the model overlapped withlandslide areas manually digitized from orthophoto-graphs. Expressed in percentage, the area concordanceis calculated by:
½ðoverlapped landslide areaÞ=ðtotal landslide area on digital mapÞ� � 100
ð4Þ
The area concordance is a simple, straightforwardmeasure based on landslide areas only. It is differentfrom a measure such as a confusion matrix, which usesboth landslide and stable areas. Because landslide areasare much smaller than stable areas in our study area (amaximum ratio of 1-to-8.5 after typhoon Toraji), anyquantitative measures derived from a confusion matrixwould be strongly biased towards stable areas.
3.5. Model validation and prediction
Model validation is a necessary phase of landslidemodeling (Chung and Fabbri, 2003). To test the validityof the logit models, we used them to predict typhoon-induced landslides by Toraji, Mindulle, and Haitang.The testing was conducted in two parts. First, wecompared landslide areas predicted by the typhoon-induced model with landslides manually digitized fromorthophotographs (for Toraji and Mindulle) and FOR-MOSAT-2 images (for Haitang) and calculated the areaconcordance values for each event. Second, wecompared landslide areas predicted by the earthquake-induced model with digitized landslide areas that werenot accounted for by the typhoon-induced model. Thesecond part of the testing was designed to evaluate theusefulness of the earthquake-induced model in predict-ing typhoon-induced landslides after the earthquakeevent. The landslide data in Table 3 already suggest that
342 K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
the Chi-Chi earthquake was probably a major contrib-uting factor for the large number and sizes of landslidesassociated with typhoon Toraji, the first typhoon afterthe earthquake. We were also interested in finding out if,and to what extent, landslides triggered by typhoonsMindulle and Haitang were affected by the earthquake.
4. Results
4.1. Performance of the typhoon-induced model
The typhoon-induced model is significant at the 1%level (pb1.01), with Cox & Snell R2 =0.28, Nagelk-erke R2 =0.37, and ROC (relative operating charac-teristic) = 0.717. The ROC is a statistic that measuresthe ability of the model to correctly classify cases oflandslide and cases of stable area (Pontius and Batchu,2003). Among the explanatory variables, elevation,slope, sine of aspect, cosine of aspect, distance tochannel, the wetness index, and the NDVI are signif-icant at the 1% level, and distance to ridge line,Nanchuang Formation, and road buffer are significantat the 5% level. Compared to digitized landslidestriggered by typhoon Herb, the model has an areaconcordance value of 76.9% [(96.3 ha /125.2 ha)×100]. Fig. 4a shows the probability map derived fromthe model. Areas along the stream channels tend tohave higher probabilities (N0.5).
Fig. 4. Probability map derived from the typhoon-induce
4.2. Performance of the earthquake-induced model
The earthquake-induced model is significant at the 1%level, with Cox & Snell R2=0.38, Nagelkerke R2=0.51,and ROC=0.758. Among the explanatory variables,elevation, slope, sine of aspect, cosine of aspect, distanceto fault line, distance to channel, distance to ridge line, thewetness index, the NDVI, the order of sub-basin, and roadbuffer are significant at the 1% level, and profile curvatureand Kueichulin Formation are significant at the 5% level.Compared to digitized landslides triggered by the Chi-Chiearthquake, the model has an area concordance value of74.5% [(502.0 ha/6674.1 ha)×100]. Fig. 4b shows theprobability map derived from the model. Areas withhigher probabilities (N0.7) are distributed mainly in thewestern part of the watershed.
4.3. Model comparison
Table 6 shows the regression coefficients and thesignificance level of the explanatory variables for thetyphoon- and earthquake-induced models. Distance tochannel and distance to ridge line are significant in bothmodels, but with opposite signs. Typhoon-triggeredlandslides tend to occur near stream channels and fartheraway from ridge lines. In contrast, earthquake-triggeredlandslides are more likely to be located near ridge linesand farther away from stream channels.
d model (a) and the earthquake-induced model (b).
Table 6Logistic regression coefficients and the significance level of explanatory variables for the typhoon- and earthquake-induced models
Numeric explanatoryvariables
Coefficient of logit model P-value
Typhoon Earthquake Typhoon Earthquake
Elevation (m) −0.001 −0.001 b0.01 b0.01Slope (deg) 0.065 0.090 b0.01 b0.01Sin of aspect 0.640 0.691 b0.01 b0.01Cos of aspect −0.526 −1.518 b0.01 b0.01Plane curvature (m−1) # # # #Profile curvature (m−1) # −0.067 # 0.037Distance to fault line (m) # 0.001 # b0.01Distance to channel (m) −0.004 0.002 b0.01 b0.01Distance to ridge line (m) 0.001 −0.002 0.011 b0.01Wetness index 0.219 0.158 b0.01 b0.01NDVI −0.455 −0.414 b0.01 b0.01
Categorical explanatory variables Class Coefficient of logit model P-value
–: Reference class of categorical independent variables. #: Variable not selected into the logit model.
Table 7Percent landslides triggered by typhoons Toraji, Mindulle, and Haitangthat were predicted by the typhoon model, the earthquake model, andboth models
% Landslidespredicted
Toraji-triggered
Mindulle-triggered
Haitang-triggered
By the typhoon model 44.4 48.1 50.8By the earthquake model 32.9 25.9 27.6By both models 77.3 74.0 78.4
343K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
The earthquake-induced model includes several sig-nificant explanatory variables not in the typhoon-inducedmodel. The inclusion of profile curvature suggests thatearthquake-induced landslides are more likely to occur onupwardly convex surfaces. The order of sub-basin variableindicates that more earthquake-induced landslides arelocated in lower-order (upstream) sub-basins. Further-more, the distance to fault variable suggests that moreearthquake-induced landslides are found away from faultlines within the study area, an unexpected result that mightbe due to the effect of multicollinearity (Hair et al., 1998).
Lithology is significant for both models. Thetyphoon-induced model selects the Nanchuang Forma-tion with alternating layers of shales and sandstones.The earthquake-induced model, on the other hand,selects the Kueichulin Formation made primarily offractured sandstones.
4.4. Model validation and prediction
The typhoon-induced model correctly predicts44.4%, 48.1%, and 50.8% of landslides triggered bytyphoons Toraji, Mindulle, and Haitang, respectively
(Table 7). With the additional contribution of theearthquake-induced model, the area concordance valuesare raised to 77.3%, 74.0%, and 78.4%, respectively.Fig. 5 shows landslide areas predicted by the typhoon-induced model, additional areas predicted by theearthquake-induced model, and unpredicted landslides.
5. Discussion
5.1. Location of landslide
Fig. 3 shows that the density of landslides triggeredby the Chi-Chi earthquakes generally decreases in a
Fig. 5. Landslides triggered by typhoons Toraji, Mindulle, and Haitang that were predicted by the typhoon-induced model, predicted by theearthquake-induced model, and unpredicted.
Fig. 6. Relative frequencies and landslide location (LL) values forlandslides triggered by typhoon Herb in 1996, the Chi-Chi earthquakein 1999, and typhoons Toraji, Mindulle, and Haitang in 2001, 2004,and 2005.
344 K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
NW–SE trend. Because the Chelungpu fault is locat-ed to the northwest of the study area, the above obser-vation means that the density of landslides decreasesaway from the epicenter, a finding similar to thatreported by Dadson et al. (2004). Besides this generalspatial trend, we can study the topographic character-istics of earthquake-triggered landslides.
Earthquake-induced landslides tend to be located onhillslopes near ridge lines (Okunishi et al., 1999), whiletyphoon-induced landslides tend to be near stream chan-nels. To further analyze this difference, we used a simplelandslide location (LL) measure:
LL ¼ ðdistance to channelÞ=ðslope lengthÞ ð5Þ
LL ranges from 0 to 1. A landslide cell near a divide willhave a value close to 1, and a landslide cell near a channelwill have a value close to 0. Fig. 6 shows the relativefrequencies and the LL values for landslides triggered bythe Chi-Chi earthquake and four typhoons. We can firstcompare landslides triggered by typhoon Herb and theChi-Chi earthquake. The relative frequency of storm-triggered landslides decreases steadily with increasing LLfor LL values between 0.2 and 1.0 (i.e., from near achannel to a divide). In contrast, the relative frequency of
earthquake-triggered landslides is highest at the LL valueof 0.8 and remains high at 0.9 and 1.0 (i.e., near a ridge).A likely explanation for this topographic characteristic ofearthquake-triggered landslides is the “amplificationeffect of topography”, meaning that seismic motion isamplified at mountain tops (Geli et al., 1988). This effectcan lead to the weakening of the cohesion and strength ofrock and soil mass near hillcrests, which in turn can leadto more earthquake-triggered landslides (Havenith et al.,2006).
Fig. 7. Rose diagram showing the distribution of earthquake-triggered landslide cells by aspect.
Fig. 8. The mean and standard deviation of LL values for landslidestriggered by the earthquake and typhoon events.
345K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
Tibaldi et al. (1995) reported a correlation betweenearthquake-induced landslides and the orientation ofmountain slopes in the Ecuadorian Andes, withmaximum density on slopes facing towards ESE (i.e.,perpendicular or sub-perpendicular to the seismogeneticfault plane). This distribution bias with respect tohillslope orientation was also reported by Okunishi et al.(1999) on landslides induced by the Kobe earthquake in1995. In this study, we found high frequencies ofearthquake-triggered landslides on slopes facing south-east and south (Fig. 7). These landslides are thereforelocated on slopes opposite to approaching seismicwaves.
5.2. Effect of earthquake on typhoon-triggeredlandslides
We can divide Fig. 6 into two halves at the midpoint(LL=0.5). For the left half (LLb0.5), the relativefrequencies of landslides triggered by three typhoonsafter the Chi-Chi earthquake are all lower than those ofHerb-triggered landslides. The pattern is exactly theopposite in the right half (LLN0.5), except for a slightdeviation for Toraji-induced landslides at 1.0. Together,these two patterns suggest that typhoon-triggered land-slides after the Chi-Chi earthquake occurred in areas
closer to the divide than before the earthquake. Thisspatial distribution of typhoon-triggered landslidestherefore indicates the continued impact of the Chi-Chi earthquake during the 6 years following theearthquake event (1999 to 2005), presumably in areaswhere slope materials were already weakened by theearthquake.
We can also compare the chronological effect of earth-quake on typhoon-triggered landslides. Fig. 8 shows themean and standard deviation of the LL values of land-slides triggered by the earthquake and typhoon events.Two trends appear in Fig. 8. First, the mean LL values of
346 K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
landslides triggered by typhoons after the Chi-Chiearthquake in 1999 are all higher than before theearthquake. We ran a t-test comparing the means of land-slides triggered by typhoon Herb (mean=0.353, S.D.=0.064) and landslides triggered by typhoon Haitang(mean=0.449, S.D.=0.078). The result showed a signif-icant difference (t=−19.033, pb0.001). Second, themean LL values of landslides triggered by typhoonsafter the earthquake decrease over time. The two trendstogether suggest that, although the effect of the earthquakehas decreased over time, the effect still remains 6 yearsafter the earthquake event.
The decreased effect of the earthquake can also bemeasured by the percentage of earthquake-triggeredlandslides that were reactivated by each typhoon. Toderive the percentages, we overlaid typhoon-triggeredlandslides from each event with earthquake-triggeredlandslides and derived the areal percentage of overlap.The results show 40.3% for Toraji, 35.8% for Mindulle,and 30.1% for Haitang. Thus the percentage ofreactivation decreases as the time lapse between atyphoon and the Chi-Chi earthquake increases.
5.3. DEM resolution and landslide prediction
This study used standard 40 m DEMs in Taiwan.Without finer-resolution DEMs, it is impossible toevaluate the effects of DEM resolution by doing asensitivity analysis like that performed by Claessenset al. (2005). Also, when evaluating the performance ofthe earthquake- and typhoon-induced models, we mustconsider the limitation posed by the cell size of 1600 m2
(0.16 ha). The cell size excludes landslides that aresmaller than 0.16 ha. Unpredicted landslides triggeredby typhoons Toraji, Mindulle, and Haitang had anaverage area of 0.2, 0.23, and 0.15 ha, respectively. Wecan therefore state that the cell size was a likely reasonfor the typhoon-induced model's failing to predict someof these unpredicted landslides.
6. Conclusion
In this study we used logistic regression and landslidestriggered by a major earthquake and a typhoon whichcame before the earthquake, to develop both earthquake-and typhoon-induced models. We then validated themodels by using landslides triggered by three typhoons,which came after the earthquake. Both models weresignificant at the 1% level. Results of model validationshowed the area concordance values around 75% inpredicting landslides triggered by three subsequenttyphoon events. By comparing the two models, this
study found that typhoon-triggered landslides tended tobe near stream channels and earthquake-triggered land-slides were more likely to be near ridge lines. This studyalso found that a major earthquake such as the Chi-Chiearthquake could still affect the spatial location oftyphoon-triggered landslides 6 years after the event.Taiwan as well as other mountainous areas subject totyphoons and earthquakes experience landslides that canbe triggered by both storms and earthquakes. By coveringboth types of landslides in a natural experiment, this studytakes a step closer to understanding one of the mostdestructive natural hazards.
Acknowledgements
This work was supported by the National TaiwanUniversity (95R0034-02) and the National ScienceCouncil (NSC 95-2415-H-002-031). We thank MikeFullen, an anonymous referee, and the editor for theirhelpful comments.
References
Aleotti, P., Chowdhury, R., 1999. Landslide hazard assessment:summary review and new perspectives. Bulletin of EngineeringGeology and the Environment 58, 21–44.
Ayalew, L., Yamagishi, H., Ugawa, N., 2004. Landslide susceptibilitymapping using GIS-based weighed linear combination, the case inTsugawa area of Agano River, Niigata Prefecture, Japan. Land-slides 1, 73–81.
Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logisticregression for landslide susceptibilitymapping in theKakuda-YahikoMountains, Central Japan. Geomorphology 65, 15–31.
Beven, K.J., Kirkby, M.J., 1979. A physically based, variablecontributing area model of basin hydrology. Hydrological SciencesBulletin 24, 43–69.
Borghuis, A.M., Chang, K., Lee, H.Y., in press. Comparison betweenautomated and manual mapping of typhoon-triggered landslidesfrom SPOT-5 imagery. International Journal of Remote Sensing.
Can, T., Nefeslioglu, H.A., Gokceoglu, C., Sonmez, H., Duman, T.Y.,2005. Susceptibility assessments of shallow earthflows triggeredby heavy rainfall at three catchments by logistic regressionanalyses. Geomorphology 72, 250–271.
Chang, K., 2005. Introduction to Geographic Information Systems, 3rded. McGraw-Hill, New York.
Chang, J.C., Slaymaker, O., 2002. Frequency and spatial distributionof landslides in a mountainous drainage basin: Western Foothills,Taiwan. Catena 46, 285.307.
Chang, T., Hsu, M., 2004. A comparison of spatial distribution ofstorm-triggered and earthquake-triggered landslides: the case ofthe Chenyulan Drainage Basin. Journal of Geographical Science35, 1–16 (in Chinese).
Chen, H., Su, D.I., 2001. Geological factors for hazardous debris inHoser, Central Taiwan. Environmental Geology 40, 1114–1124.
Cheng, J.D., Huang, Y.C., Wu, H.L., Yeh, J.L., Chang, C.H., 2005.Hydrometeorological and landuse attributes of debris flows anddebris floods during typhoon Toraji, July 29–30, 2001 in centralTaiwan. Journal of Hydrology 306, 161–173.
347K.-T. Chang et al. / Geomorphology 89 (2007) 335–347
Chung, C.F., Fabbri, A.G., 2003. Validation of spatial predictionmodels for landslide hazard mapping. Natural Hazards 30,451–472.
Chung, C.F., Fabbri, A.G., 2005. Systematic procedures of landslidehazard mapping for risk assessment using spatial predictionmodels. In: Glade, T., Anderson, M., Crozier, M.J. (Eds.),Landslide Hazard and Risk. Wiley, New York, pp. 139–177.
Claessens, L., Heuvelink, G.B.M., Schoorl, J.M., Veldkamp, A., 2005.DEM resolution effects on shallow landslide hazard and soilredistribution modeling. Earth Surface Processes and Landforms30, 461–477.
Dadson, S.J., Hovius, N., Chen, H., Dade, W.B., Lin, J., Hsu, M., Lin,C., Horng, M., Chen, T., Milliman, J., Stark, C.P., 2004.Earthquake-triggered increase in sediment delivery from an activemountain belt. Geology 32, 733–736.
Dai, F.C., Lee, C.F., 2002. Landslide characteristics and slopeinstability modeling using GIS, Lantau Island, Hong Kong.Geomorphology 42, 213–228.
Dai, F.C., Lee, C.F., 2003. A spatiotemporal probabilistic modeling ofstorm-induced shallow landsliding using aerial photographs andlogistic regression. Earth Surface Processes and Landform 28,527–545.
Geli, L., Bard, P.Y., Jullien, B., 1988. The effect of topography onearthquake ground motion: a review and new results. Bulletin ofthe Seismological Society of America 78, 42–63.
Guzzetti, F., Carrara, A., Cardinali, M., Reichenbach, P., 1999.Landslide hazard evaluation: a review of current techniques andtheir application in a multi-scale study, Central Italy. Geomor-phology 31, 181–216.
Guzzetti, F., Stark, C.P., Salvati, P., 2005. Evaluation of flood andlandslide risk to the population of Italy. Environmental Manage-ment 36, 15–36.
Havenith, H., Strom, A., Caceres, F., Pirard, E., 2006. Analysis oflandslide susceptibility in the Suusamyr region, Tien Shan:statistical 1 and geotechnical approach. Landslides 3, 39–50.
Keefer, D.K., 1984. Landslides caused by earthquakes. GeologicalSociety of America Bulletin 95, 406–421.
Liao, H., 2000. Landslides triggered by Chi-Chi earthquake. MastersThesis, National Central University. Jhongli, Taiwan. (in Chinese).
Lin, M., Tung, C., 2003. A GIS-based potential analysis of thelandslides induced by the Chi-Chi earthquake. EngineeringGeology 71, 63–77.
Lin, C., Shieh, C., Yuan, B., Shieh, Y., Liu, S., Lee, S., 2003. Impact ofChi-Chi earthquake on the occurrence of landslides and debrisflows: example from the Chenyulan River watershed, Nantou,Taiwan. Engineering Geology 71, 49–61.
O'Hare, G., Rivas, S., 2005. The landslide hazard and humanvulnerability in La Paz City, Bolivia. Geographical Journal 171,239–258.
Ohlmacher, G.C., Davis, J.C., 2003. Using multiple logistic regressionand GIS technology to predict landslide hazard in northeastKansas, USA. Engineering Geology 69, 331–343.
Okunishi, K., Sonoda, M., Yokoyama, K., 1999. Geomorphic andenvironmental controls of earthquake-induced landslides. Transac-tions Japanese Geomorphological Union 20, 351–368.
Pierson, T.C., 1980. Piezometric response to rainstorms in forestedhillslope drainage depression. Journal of Hydrology (NewZealand) 19, 1–10.
Pontius Jr., R.G., Batchu, K., 2003. Using the relative operatingcharacteristic to quantify certainty in prediction of location of landcover change in India. Transactions in GIS 7, 467–484.
Quinn, P.F., Beven, K.J., Lamb, R., 1995. The ln(a/tanβ) index: how tocalculate it and how to use it within the TOPMODEL framework.Hydrological Processes 9, 161–182.
Tibaldi, A., Ferrari, L., Pasquare, G., 1995. Landslides triggered byearthquakes and their relations with faults and mountain slope ge-ometry: an example from Ecuador. Geomorphology 11, 215–226.
Wang, H., Liu, G., Xu, W., Wang, G., 2005. GIS-based landslidehazard assessment: an overview. Progress in Physical Geography29, 548–567.
Wu, Y.H., Swanston, D.N., 1980. Risk of landslides in shallow soilsand its relations to clearcutting in southeastern Alaska. ForestScience 26, 495–510.
Wu, C.C., Kuo, Y.H., 1999. Typhoons affecting Taiwan: currentunderstanding and future challenges. American MeteorologicalSociety Bulletin 80, 67–80.
Yesilnacar, E., Topal, T., 2005. Landslide susceptibility mapping: acomparison of logistic regression and neural networks methods ina medium scale study, Hendek region (Turkey). EngineeringGeology 79, 251–266.
Yu, S.B., Chen, H.Y., Kuo, L.C., 1997. Velocity field of GPS stationsin the Taiwan area. Tectonophysics 274, 41–59.
