Catena 115 (2014) 55–70

Contents lists available at ScienceDirect

Catena

j ourna l homepage: www.e lsev ie r .com/ locate /catena

Landslide susceptibility mapping at central Zab basin, Iran: A comparisonbetween analytical hierarchy process, frequency ratio and logisticregression models

Himan Shahabi a,⁎, Saeed Khezri b, Baharin Bin Ahmad a, Mazlan Hashim a

a Institute of Geospatial Science & Technology (INSTeG), Universiti Teknologi Malaysia (UTM), Skudai, 81310 Johor Bahru, Malaysiab Department of Physical Geography, Faculty of Natural Resources, University of Kurdistan, Iran

⁎ Corresponding author at: Department of Geoinformaand Real Estate, Universiti Teknologi Malaysia (UTM)Malaysia. Tel.: +60 166972788.

E-mail address: himanshahabi@gmail.com (H. Shahab

0341-8162/$ – see front matter © 2013 Elsevier B.V. All rihttp://dx.doi.org/10.1016/j.catena.2013.11.014

a b s t r a c t

a r t i c l e i n f o

Article history:Received 25 November 2012Received in revised form 3 November 2013Accepted 28 November 2013Available online xxxx

Keywords:LandslideGISCentral Zab basinRemote sensing

The purpose of this study is to compare the landslide susceptibility mapping models of logistic regression (LR),analytical hierarchy process (AHP) and frequency ratio (FR) applied in the central Zab basin(West Azerbaijan—Iran). Eight factors were used for landslide susceptibility mapping including slope, aspect,land cover, precipitation, lithology and the distance to roads, drainage, and faults that affect the occurrence oflandslides. To get more precision, speed and facility in our analysis, all descriptive and spatial information wasentered into GIS system. Satellite images (Landsat ETM+ and SPOT 5) are also used to prepare for land useand landslide-inventory mapping respectively. Landslide events as used as dependant variable and data layersas independent variable, making use of the correlation between these two variables in landslide susceptibility.The three models are validated using the relative landslide density index (R-index) and the receiver operatingcharacteristic (ROC) curves. The predictive capability of each model was determined from the area under therelative operating characteristic curve and the areas under the curves obtained using the LR, AHP, and FRmethods are 0.8941, 0.8115, and 0.8634, respectively. These results indicate that the LR and FR models arerelatively good estimators of landslide susceptibility in the study area. The interpretations of the susceptibilitymap reveal that precipitation, lithology and slope played major roles in landslide occurrence and distributionin the central Zab basin. In general, all three models were reasonably accurate. The resultant maps would beuseful for regional spatial planning as well as for land cover planning.

© 2013 Elsevier B.V. All rights reserved.

1. Introduction

As a natural hazard, landslides cause great damage to human livesand infrastructure (Nefeslioglu et al., 2008). Statistics from the Centerfor Research on the Epidemiology of Disasters (CRED) show that land-slides are responsible for at least 17% of all natural hazard fatalitiesworldwide (Pourghasemi et al., 2012). According to the IranianLandslide Working Party (2007), about 187 people have been killed inIran by landslides and losses resulting from mass movements to theend of September 2007 have been estimated at 126,893 billion IranianRials (almost $12,700 million dollars) using the 4900 landslide data-base (Iranian Landslide Working Party, 2007). Given the degree of lossin both human and economic terms, recognizing the identifying therisk factors inherent to landslide hazards makes zoning a necessity(Iranian Landslide Working Party, 2007).

Western Iran has some parameters which affect landslides suchas slope, precipitation, lithology, soil, earthquake and human activity.

tion, Faculty of Geoinformation, Skudai, 81310 Johor Bahru,

i).

ghts reserved.

The central Zab basin, in the south-west mountain sides of WestAzerbaijan province, is having landslide problems. The landslidesobserved and found in the area include old and new landslides.Although the area is not a complete agricultural zone, it has cultivatedland which is affected by landslides (Shahabi et al., 2012a).

Identifying the different types and processes of landslide formationcounts as major factors in preventing landslide induced damage. Thepreparation of a landslide susceptibility map is a major step in overalllandslide hazard management. Landslide susceptibility is defined asa quantitative or qualitative assessment of the classification, volume(or area) and spatial distribution of landslides which exist or potentiallymay occur in an area (Bednarik et al., 2010). Landslide susceptibilityassessment in a GIS environment is based upon a suitable selection offactors which play a dominant role in slope stability.

There are many qualitative and quantitative techniques available toanalyze the relationship between a landslide and conditioning parame-ters (Ayalew and Yamagishi, 2005). In recent years, many studies havebeen carried out on landslide hazard evaluation using GIS (Akgün andBulut, 2007; Akgun et al., 2008; Bednarik et al., 2012; Kayastha et al.,2013; Pradhan et al., 2012; Wan and Lei, 2009; Wan et al., 2010).

Many statistical models are used in landslide hazard analysis includ-ing logistic regression (Akgun et al., 2012; Bai et al., 2010; Bui et al.,

56 H. Shahabi et al. / Catena 115 (2014) 55–70

2011; Shahabi et al., 2012a;Wang et al., 2013), analytical hierarchy pro-cess and frequency ratio (Akgun et al., 2008; Shahabi et al., 2012b).

Previous investigations have not studied landslide hazard suscepti-bility nor utilized satellite images and statistical methods to study thelandslide susceptibility map in the study area. In previous studies inthe study area, researchers only used limited traditional methods.

Many previous studies using GIS technology have applied a singlemodel, discussed relevant factors, and validated landslide susceptibilitymaps to assess the performance of the applied model. The main differ-ence between previous studies and the present study is there has beenno comprehensive study to date involving the application, assessment,and validation of various statistical models in the same area oran accurate method for improving prediction accuracy. Therefore, thereis significant demand for landslide susceptibility mapping generated byLR, AHP and FRmodels in central Zab basinwith high predictive accuracy.

The purpose of this paper is to assess and compare the results oflandslide susceptibility maps using three statistics based approachesin the central Zab basin. This comparison involves four main steps. Thefirst step involves identifying the categories of the causative factorsresponsible for the occurrence of landslides based on the characteristicsof the landslide inventory. The second step estimates the relative contri-bution of these categories causing slope failure and establishes a rela-tion between the categories and the landslides. Then, the significant orthe most influential categories to assess the best statistical model aredetermined. Finally, the best indirect susceptibility map is comparedwith the best direct susceptibility map using two validation methods,R-index and ROC.

2. Study area

Zab basin occupies the southwestern section ofWest Azerbaijan andthenorthwestern part of Kurdistan. The area under present study covers

Fig. 1. General position of Zab basin regarding th

parts of mountains and slopes in southwestern West Azerbaijan inthe central portion of Zab basin between latitude (36° 8′ 25″) N and(36° 26′ 27″) N and longitude (45° 21′ 21″) E and (45° 40′ 44″) E.Central Zab basin has a north–south orientation and stretches almost30 km in an east–west direction. The study area covers some 520 km2

of its total area (Fig. 1). It is one of the most populated geographicalbasins and includes a city, three towns or small cities, and over 80villages (Khezri et al., 2013b). Most landslides are seen within a1000 m boundary of the main Sardasht–Piranshahr and Sardasht–Rabat roads. Here, a north-west extension branches off from theeast–west oriented ridges of Zab Valley, creating a different landscapefrom that of the internal sections of Azerbaijan and Kurdistan provinces.

As the study region is located in amajor Zagros thrust direction, tec-tonic faults are themain cause of slope failures. The regionsmorphologyis strongly affected by tectonic forces. The most important faults arelocated in western Iran, highlighting and giving importance to thiscause in the region of study. The target research zone, tectonically, is lo-cated in Sanandaj–Sirjan region, and its north-east region is located inMahabad–Khoy zone (Khezri et al., 2013a). Also, due to the steeprocks and several faults as well as their different functions, steepand layer lengths are different (Fig. 2). Stones, homogenous phyliteformations, marble, lime, green andesite, dolomite and sandstone arethe dominant lithology in these zones.

3. Material and methods

3.1. Landslide inventory

Landslide susceptibility mapping is based on a landslide inventoryincluding geology (lithology), geomorphology (slope, and aspect),land cover, precipitation and distance to roads, faults, land cover,and drainage network. Landslide inventory map of the study area was

e political provincial and national borders.

Fig. 2. Cross-section in 1:100,000 geological mapping across the Zab Valley in the northern part of the research area.

57H. Shahabi et al. / Catena 115 (2014) 55–70

identified by SPOT 5 satellite on 25 May 2008. Extensive field studieswere used to check the landslides size and shape, to identify thetype of movements and the materials involved, and to determine thelandslides state of activity (active, dormant, etc.).

The landslides identified from large-scale aerial photography werecarried out in 1996 and 2005 were used to train visual interpretationof SPOT imagery. Interpretation was conducted for the entire 2.5 mSPOT scene. Obvious landslides are visible in all four scenes. The easewith which landslides are detectable depends on characteristics ofboth the landslide (e.g. size and contrast with surrounding terrain)and the sensor used (e.g. spatial and spectral resolution) (Kääb, 2008).Size and contrast with surrounding terrain are the most importantlandslide characteristics, with large landslides and landslides surroundedby undisturbed vegetation (e.g. natural landslides) being the mostnoticeable. The minimum area covered by an inventoried and mappedlandslide is 1600 m2 extracted from SPOT 5 images (Shahabi et al.,2012a). Fig. 3 shows SPOT-5 satellite images and the location of twolandslides in the central Zab basin.

Geomorphologists have developed a landslide inventorywith a scaleof 1:10,000 to 1:25,000 for local zoning through interpretation of SPOT 5satellite images, field surveys and literature analysis from 2008 until2011. Fig. 4 shows the landslide inventory map of central Zab basin.

A total of 85 landslides were identified (Fig. 4), covering an area of21.86 km2, and accounting for 4.05% of the study area. The minimum,mean and the maximum landslide areas are 0.004, 0.027 and0.305 km2 respectively. The landslide properties were recorded ona standard landslide inventory data sheet. Smaller landslides arerepresented by a dot. Minor and lateral scarps may be distinguishedas well as up-slope deformations such as tension cracks or minorlandslides. According to the landslide classification by Cruden andVarnes (1996), the landslide inventory contains 41 shallow and 44deep-seated landslides, including 29 earth slides and 15 rock slides.Landslide properties were recorded on a standard landslide inventorydata sheet, but the main purpose herein was to map only thelandslide boundaries.

3.2. Spatial database construction

The first step in landslide susceptibility mapping is to acquire infor-mation about landslides that have occurred in the past (Akinci et al.,2011). Based on this idea, the first step in our study is to establish a spa-tial database for a spatially explicit analysis of the degree of susceptibil-ity. In the preparation phase, all necessary geometric thematic editingwas done on the original datasets and a topology was created. In thenext step, all vector layers were converted into raster format and thespatial datasets were processed in Arc GIS. The study area covers16,323.720 pixels. Data entry and production were the most cumber-some and time-consuming steps of this study. Since the study area ismountainous with no roads, accessibility to the landslides was difficultor even impossible.

A total of eight factors were used for landslide susceptibilitymapping including slope, aspect, land cover, precipitation, lithologyand the distance to roads, drainage, faultswhich are considered control-ling factors in the occurrence of landslides in the study area. A briefdescription of the preparation each controlling factor is given below.

The critical point was the selection of appropriate pixel size for posi-tional accuracy and precision of susceptibility levels in the resultantmap. The positional accuracy needed for 1:50,000 scale maps must be150 m. For this reason, a pixel size of 50 m was selected for our digitalelevation model (DEM). The DEM was generated from a triangulatedirregular network (TIN) model that was used to generate slope anglesand aspectmaps. Fault lines were derived from 1:100,000 scale geologymaps of the Iranian Geology Organization. A similar process was carriedout for road lines and drainage networks. In addition, the kilometersquare density of drainage networks, road and fault lines were alsoused to demonstrate the importance of the features in the wholestudy area. The 1-km2 search distance, with a 564-m search radius ofa 100 m offset, was used to store the information of these parametermaps in a uniform thematic database, and the size of each pixel for allthe products was 100 m × 100 m.

Another dataset used was land cover, which was interpreted fromLandsat ETM+ images taken on 21 April 2009 which were calibratedusing field observations. Due to significant cloud cover, the results ofthe classification were edited and simplified by manual digitization.The interpreted images were then digitally processed to furthermodify the boundaries by supervision classification with ERDAS(Earth Resource Data Analysis System) software to develop a statisticalcharacterization of the reflectance for each information class. Sevenmain land cover types were considered, namely grassland, pasture, set-tled, natural forest, man-made forest, dry farm land and barren land.Based on validation from the field observations, the land cover map isaccurate to the order of the Landsat image spatial resolution (~30 m).Despite image pre-processing by Landsat ETM+ imagerywhich includ-ed geo-referencing and ortho-rectification, geo-rectification accuracywas further improved by using 15 ground control points (GCPs) obtain-ed during field visits. Field surveys were undertaken in order to gain anunderstanding of the spatial and site-specific characteristics of therecent land cover types in the study area. After geo-referencing theresultant image, a combination of bands 1, 4 and 7 was used to makecomplex color pictures, and an operational information layer wascreated by the Categorization of Utmost Probability method (Dymondet al., 2006).

Considering that the best undisturbed morphological conditions(conditions before failure) and influencing factors can be extractedfrom the vicinity of the landslide itself, seed cells were selected usinga 25 m buffer zone around each landslide point. This buffer intervalwas chosen because the width and length of the landslide depletionzone rarely exceed 25 m (Che et al., 2011). As the factor raster mapswere produced at 20 m spatial resolution, 25 m buffer resulted in atleast 4 seed cells per landslide scar. Waterways of central Zab basin

Fig. 3. Different steps of preparing the map for landslide susceptibility mapping.

58 H. Shahabi et al. / Catena 115 (2014) 55–70

were digitized and all the needed operations for using this informationin Ilwis environment were carried out. According to the results of30 year observations from 3 weather stations which include Sardasht,Mirabad and Piranshahr in West Azerbaijan province, February is thecoldest month with a temperature average of 7.7 °C, while the hottestmonth is August with a temperature average of 24.1 °C. In ourmethodology we used long-term precipitation for a 30 year period(1980–2010). The annual precipitation of central Zab basin is865 mm, and the precipitation disperses symmetrically everymonth. The mean annual rainfall is around 1310 mm, most ofwhich falls between the April and May. Measurements of daily rain-fall data were utilized to create average annual rainfall contourmaps,which were then used for a spatial interpolation by kriging withinArc GIS. Fig. 5 is a schematic illustration of the different map prepa-ration steps for landslide susceptibility mapping used in this study.

3.3. Statistical analysis

The preparation of a landslide susceptibilitymap in central Zab basininvolves manipulation, analysis using FR, AHP and LR model and thevalidation the results using R-index and ROC methods.

3.3.1. Analytical hierarchy process (AHP)The AHP developed by Saaty (1980) provides a flexible and easily

understood way to analyze complicated problems. The AHP is a multi-objective, multi-criteria decision-making approach that allows theactive participation of decision-makers in reaching agreement, andgives managers a rational basis on which to make decisions (Saaty andVargas, 1984). The AHP gained wide application in site selection,suitability analysis, regional planning, and landslide susceptibilityanalysis (Ayalew et al., 2004).

Fig. 4. SPOT-5 satellite images at the central Zab basin landslide.

59H. Shahabi et al. / Catena 115 (2014) 55–70

AHP is a problem-solving construction and a methodical process forrepresenting the elements of any problem (Yalcin et al., 2011). To applythis approach, it is necessary to break a complex unstructured problemdown into its component factors; to arrange these factors in a hierarchicorder; to assign numerical values to subjective judgments based on therelative importance of each factor; and to synthesize the judgments todetermine the priorities to be assigned to these factors (Sahnounet al., 2012). AHP requires the creation of a reciprocal pair-wise compar-isonmatrix. Entries into thematrix are found from comparison betweeneach layer based on a 9-point rating scale as developed by Saaty (1977)(see Table 1), where a value of 1 is given to imply the criteria undercomparison are of equal importance to the final solution and 9

Fig. 5. Landslide inventory m

expresses extreme importance of one criterion over another (Saatyand Vargas, 1991).

In these techniques, firstly, the effects of each parameter to thesusceptibility of landslides relative to each other were determined bydual evaluation to determine the preferences in the effects of theparameters on the landslide susceptibility maps (Yalcin, 2008).

Normally, the determination of the parameter values relative to eachother is a situation that is dependent on the choices of the decision-maker. However, in this study, both the comparison of the parametersrelative to each other and the determination of the decision alternatives,namely the effect values of the sub-criteria of the parameters (weight),were based on the comparison of landslide inventory maps whichwere

ap of central Zab basin.

Table 1Scales for pairwise comparisons (Saaty and Vargas, 1991).

Importance Definition Explanation

1 Equal importance Contribution to objectiveis equal

3 Moderate importance Attribute is slightly favoredover another

5 Strong importance Attribute is strongly favoredover another

7 Very strong importance Attribute is very stronglyfavored over another

9 Extreme importance Evidence favoring oneattribute is of the highestpossible order of affirmation

2, 4, 6, 8 Intermediate values When compromise is neededReciprocals Opposites Used for inverse comparison.

60 H. Shahabi et al. / Catena 115 (2014) 55–70

constructed using field surveys, previous inventory maps, and satelliteimages with the other data layers.

The weight of each factor obtained from thematrix weighting factorwasmultiplied by its weight class. The result of the susceptibility map isdetermined by factors with high local representation such as linea-ments. There are also studies that used natural (lithology, lineament,etc.) and artificial (roads and other engineering structures), or causal(slope, lithology, etc.) and triggering (rain, seismicity, etc.) factorstogether (Ayalew and Yamagishi, 2005; Feizizadeh et al., 2013; Yalcin,2008). The selection of the eight causal factors in this study is basedon these four criteria, and also considers general literature input anddata availability.

Pair-wise comparison, however, is subjective and the quality of theresults is highly dependent on the expert's judgment. In AHP, an indexof consistency, known as the consistency ratio (CR) Eq. (1), is used toindicate the probability that the judgment matrix was randomly gener-ated (Saaty, 1977).

CR ¼ CI=RI ð1Þ

where RI is the average of the resulting consistency index depending onthe order of the matrix given by Saaty (1977) and CI is the consistencyindex and can be expressed as Eq. (2)

CI ¼ λmax−nð Þ= n−1ð Þ ð2Þ

where λmax is the largest or principal eigenvalue of the matrix and canbe easily calculated from the matrix, and n is the order of the matrix.

CR was calculated for all the models where the AHP was used. If theCR values were greater than 0.1, the models were automaticallydiscarded. Using a weighted linear sum procedure (Voogd, 1983) theacquired weights were used to calculate the landslide susceptibilitymodels (Komac, 2006).

3.3.2. Frequency ratio (FR)The spatial relationship between all landslides and each related

factor were derived using the frequency ratio. The frequency ratiois the ratio between the landslides in the class as a percentage of alllandslides and the area of the class as a percentage of the entire map(Lee and Pradhan, 2007) (see Eq. (3)). When evaluating the probabilityof a landslidewithin a specific period of time andwithin a certain area, itis of major importance to recognize the conditions that can cause thelandslide and the process that could trigger the movement (Shahabiet al., 2012c; Yalcin et al., 2011). The correlation between landslideareas and associated factors that cause landslides can be allocatedfrom the connections between areas without past landslides and thelandslide-related parameters (Bednarik et al., 2012). In order to preparethe landslide susceptibility map quantitatively, the frequency ratiomethod was implemented using GIS techniques.

Therefore, the frequency ratios of each factor's type or range will becalculated from their relationship with landslide events. The frequencyratio was calculated for sub-criteria of parameter, and then the fre-quency ratios were summed to calculate the landslide susceptibilityindex (LSI) (Eq. (3)) (Lee and Sambath, 2006).

LSI ¼ Fr1 þ Fr2 þ Fr3 þ…þ Frn ð3Þ

where Fr is the rating of each factor's type or range, n is the numberof factor.

According to the frequency ratio method, the ratio is that of the areawhere the landslide occurred to the total area, so that a value of 1 is anaverage value. A value of N1 means the percentage of the landslideis higher than the area and refers to a higher correlation, whereasvalues lower than 1 mean a lower correlation (Akgün and Bulut,2007). In general, to predict landslides in central Zab basin, it wasnecessary to assume that landslide occurrence was determined bylandslide-related factors, and that future landslides will occur underthe same conditions as past landslides.

3.3.3. Logistic regression (LR)Logistic multiple regression allows one to evaluate amultivariate re-

gression relationship between a dependent and several independentvariables. The advantage of logistic regression is that through the addi-tion of an appropriate link function to the usual linear regressionmodel,the variables may be either continuous or discrete, or any combinationof both types, and they do not necessarily have normal distributions(Lee, 2005b). The logistic regression algorithm applies maximum likeli-hood estimation after transforming the dependent variable into a logitvariable (the natural log of the odds of the dependent occurring ornot). In this way, logistic regression estimates the probability of acertain event occurring (Dai and Lee, 2002).

In the case of landslide susceptibility mapping, the goal of LRwould be to find the best fitting (yet reasonable) model to describethe relationship between the presence or absence of landslides(dependent variable) and a set of independent parameters such asslope, aspect and lithology (Ayalew and Yamagishi, 2005). The in-dependent variables in this model are predictors of the dependentvariable and can be measured on a nominal, ordinal, interval orratio scale. The relationship between the dependent variable andindependent variables is non-linear (Ozdemir and Altural, 2012).

For this study, the dependent variable must be input as either 0 or 1,so the model applies well to landslide possibility analysis. Logisticregression coefficients can be used to estimate ratios for each of theindependent variables in the model (Pradhan and Lee, 2010).

In such cases, the probability of phenomenon (P) is estimated, andthe alternative model employed is logistic regression analysis, givenby Eqs. (4) and (5):

Y ¼ Logit pð Þ ¼ β0 þ β1X1 þ β2X2 þ…þ βnXn ð4Þ

Y ¼ Logit pð Þ ¼ lnp

1−p

� �ð5Þ

where p is the occurrence probability that the dependent variable (Y) is1, p / (1 − p) is the so-called odd or frequency ratio, β0 is the intercept,and β1 to βn are the coefficients which measure the contribution of theindependent factors (X1, X2,…, Xn) to the variations in Y (Lee, 2005b). Ineffect, Eq. (4) estimates the Logit (p) values and extracts the p fromthese values. The spatial databases of each factor were converted toASCII format files to be used in the statistical package SPSS. The correla-tions between landslide and each factor were calculated.

A positive sign indicates that the explanatory variable has increasedthe probability of change, and a negative sign implies the oppositeeffect. This statistical technique is a multivariate estimation methodthat examines the relative strength and significance of factors

Fig. 6. Landslide susceptibility pre-disposing factors: (a) slope; (b) aspect; (c) lithology; (d) land cover; (e) distance to fault; (f) distance to drainage; (g) distance to road; (h) precipitation.

61H. Shahabi et al. / Catena 115 (2014) 55–70

Table 2Pair-wise comparison matrix, factor weights and consistency ratio of the data layers.

Factors 1 2 3 4 5 6 7 8 9 10 Weights

Distance to drainage (m)(1) 0–1000 1 0.502(2) 1000–2000 1/3 1 0.256(3) 2000–3000 1/5 1/3 1 0.120(4) 3000–4000 1/6 1/4 1/2 1 0.074(5) N4000 1/7 1/5 1/3 1/2 1 0.050Consistency ratio: 0.031

Distance to road (m)(1) 0–1000 1 0.502(2) 1000–2000 1/3 1 0.256(3) 2000–3000 1/5 1/3 1 0.120(4) 3000–4000 1/6 1/4 1/2 1 0.074(5) 4000–5000 1/7 1/5 1/3 1/2 1 0.050(6) 5000–6000 1/7 1/8 1/9 1/8 1/5 1 0.023(7) 6000–7000 3 1/2 1/6 1/5 1/3 1/7 1 0.039(8) N7000 3 1/2 1/6 1/5 1/3 1/7 1 0.039Consistency ratio: 0.091

Lithology(1) Q2

t 1 0.124(2) ∈l 1/3 1 0.053(3) ∈m 1/3 1 1 0.053(4) ∈z 1/2 2 2 1 0.083(5) Kl 1/5 1/3 1/3 1/4 1 0.024(6) Kml 1/3 1 1 1/2 3 1 0.053(7) Kph 1/4 1/2 1/2 1/3 2 1/2 1 0.031(8) Gn 3 5 5 4 7 5 6 1 0.273(9) An 1 3 3 2 5 3 4 1/3 1 0.124(10) La 2 4 4 3 6 4 5 1/2 2 1 0.187Consistency ratio: 0.017

Slope (°)(1) 0–10 1 0.110(2) 10–20 3 1 0.173(3) 20–30 4 3 1 0.393(4) 30–40 3 3 1/3 1 0.239(5) 40–50 1/3 1/4 1/6 1/4 1 0.062(6) N50 1/7 1/8 1/9 1/8 1/5 1 0.023Consistency ratio: 0.083

Aspect(1) Flat 1 0.026(2) N 3 1 0.071(3) NE 5 3 1 0.189(4) E 3 1 1/3 1 0.071(5) SE 3 1 1/3 1 1 0.071(6) S 3 1 1/3 1 1 1 0.071(7) SW 7 5 3 5 5 5 1 0.354(8) W 3 1 1/3 1 1 1 1/5 1 0.071(9) NW 3 1 1/3 1 1 1 1/5 1 1 0.071Consistency ratio: 0.009

Precipitation (mm)(1) 0–600 1 0.078(2) 600–700 2 1 0.151(3) 700–800 3 2 1 0.263(4) N800 7 3 2 1 0.507Consistency ratio: 0.075

Distance to faults (m)(1) 0–1000 1 0.034(2) 1000–2000 9 1 0.423(3) 2000–3000 1 1/9 1 0.032(4) 3000–4000 1 1/9 1 1 0.032(5) 4000–5000 7 1/3 9 9 1 0.257(6) 5000–6000 5 1/5 6 6 1/2 1 0.171(7) N6000 2 1/8 2 2 1/8 1/7 1 0.051Consistency ratio: 0.043

Land cover(1) Natural forest 1 0.323(2) Dry farm land 1/4 1 0.080(3) Grass land 1/2 3 1 0.208(4) Pasture 1/3 2 1/3 1 0.114(5) Settled 1/4 1 1/2 1/2 1 0.084(6) Man-made forest 1/3 1 1/2 1 1 1 0.095(7) Barren land 1/3 1 1/2 1 1 1 1 0.095Consistency ratio: 0.016

62 H. Shahabi et al. / Catena 115 (2014) 55–70

Table 2 (continued)

Factors 1 2 3 4 5 6 7 8 9 10 Weights

(1) Distance to drainage 1 0.019(2) Distance to road 4 1 0.055(3) Lithology 8 5 1 0.238(4) Slope 7 4 1/2 1 0.169(5) Aspect 5 2 1/4 1/3 1 0.086(6) Precipitation 9 6 2 3 5 1 0.335(7) Distance to fault 3 1/2 1/6 1/5 1/3 1/7 1 0.039(8) Land cover 4 1 1/5 1/4 1/2 1/6 2 1 0.056Consistency ratio: 0.034

63H. Shahabi et al. / Catena 115 (2014) 55–70

(Ozdemir and Altural, 2012). In logistic regression modeling, maxi-mizing likelihood function is used to obtain the coefficients ofindependent variables (Kleinbaum and Klein, 2010). Likelihoodfunction is in fact a probability. In particular, it is the probabilitythat the observed values of dependant variable might be predictedby those of independent variables. Likelihood function is definedas Eq. (6).

L ¼ ΠN

i¼1pyii � 1−pið Þ 1−yið Þ ð6Þ

where pi is the probability of slide occurrence defined by Eqs. (4) and(5), and yi represents the presence (1) or absence (0) of the observedmass movement. Likelihood ratio index is used if evaluation of rela-tive significance of any individual independent variable is requiredin the process of predicting the dependant variables through logisticregression model. The index is equal to −2 Log (L), and will have avalue ranging between 0 to +∞, defined by the L range.

Significance of an independent variable will be higher than thatof another independent variable when its exclusion results in a widervariation in the likelihood ratio. The linear model provided by logisticregression analysis rests on slides occurredwith respect to independentvariables or earlier situations that led to landslide (Lee and Sambath,2006). Therefore, to obtain the optimized regression coefficientandslide layer was defined as the response variable (Y), and thepresence of mass movement was designated as 1 and its absence as 0.

4. Appraisal of landslide individual factors

4.1. Pre-disposing factors of landslide susceptibility in central Zab basin

There are no universal guidelines regarding the selection of factorsin landslide susceptibility mapping (Ayalew et al., 2005). The selectionof causal factors therefore needs to take the nature of the study areaand data availability into account. A total of 8 potential pre-disposingfactors were considered in this study, namely slope, aspect, land cover,precipitation and the distance to roads, drainage, faults,, and geologicalfactors. These factors are divided into several classes. Extent area andlandslide frequency determines the proportion and effect of each class.The list of all the data layers is illustrated in Fig. 6.

4.1.1. SlopeSlope is viewed as the major controlling factor in landslide forma-

tion. It is frequently used in the calculation of landslide susceptibilitymaps (Clerici et al., 2002; Lee, 2005a; Nefeslioglu et al., 2008; Yalcin,2008; Yalcin et al., 2011). The slope map of the study area was dividedinto six slope categories. ArcGIS softwarewas used for this classificationand for the calculation of the relationships to landslide susceptibility.The landslide percentage in each slope group class is determinedas a percentage of slopes. The result indicates that most landslides(38) occur when the percentage of the slope is greater than 30°(Table 3 and Fig. 6a).

4.1.2. AspectAspect has an accelerated role in a landslide event. On many occa-

sions, slope aspect has controlled the amount of water in the slopes andhillsides (Shahabi et al., 2012a). Aspect regions are classified in nine cat-egories according to the aspect class as: ″ at (−1°), north (0°–25.5°;315°–360°), north-east (22.5°–59.5°), east (59.5°–135.5°), south-east(135.5°–153.5°), south (153.5°–215.5°), south-west (215.5°–251.5°),west (251.5°–298.5°), and north-west (298.5°–341.5°). Table 3 showsthat half of the basin's landslides occurred in the north and north-eastand a lower proportion of landslides occurred in other classes (Fig. 6b).

4.1.3. Geology and lithologyGeological factor is considered as an independent variable in land-

slide formation in that lithology and its varied structures tend to leadto a variation in stone stability and strength and also to a varied soiltexture (Ayalew and Yamagishi, 2005). It is important to note that thegeology of our study area is very complex and the lithology units com-prise several formations. The formations were therefore classified inten categories with respect to landslide susceptibility. Table 3 revealedthat most slide events in the study area occurred in loose formations in-cluding fine-grained sediments, particularly in alluvial terraces. Stones,homogenous phylite formations, marble, lime, green andesite, dolomiteand sandstone have the widest distribution in the region (Fig. 6c).

4.1.4. Land coverThe lack of appropriate cover such as vegetation is related to an

increase in landslide frequency (Shahabi et al., 2012b). In thisstudy seven categories of land cover were determined and comparedto landslide susceptibility (Fig. 6d). Man-made forests are moreprone to landslides compared to other categories. Although dryfarm land has the most landslides among land cover classes, the ex-tent of the area is bigger than other classes, particularly man-madeforest class (Table 3).

4.1.5. Distance to faultFaults are an important susceptibility factor. Highly faulted zones are

areas of particularly high incidence of unstable slopes and the degree offracturing and shearing plays an important role in determining slopestability (Varnes et al., 1984). It has generally been observed that theprobability of landslide occurrence increases at sites close to lineaments.In this study seven different buffer zones to existing faults were created(Fig. 6e). Most of the landslides have occurred in class 0–1000 m,whichnot only affect the surface material structures but also contributes toterrain permeability causing slope instability (Table 3).

4.1.6. Distance to drainageDrainage density is another controlling factor of landslide formation.

Drainage layer was selected so as to simultaneously consider the rolesof inappropriate drainage and the undercutting of the hydrographic sys-tem (Shahabi et al., 2012b). The proximity of the slope to drainagestructures is another important factor in terms of stability. Streamsmay adversely affect stability by eroding the slopes or by saturatingthe lower part of the material, resulting in water level increases

Fig. 7. Landslide susceptibility map generated using (a) analytical hierarchy process, (b) frequency ratio and (c) logistic regression models in central Zab basin.

64 H. Shahabi et al. / Catena 115 (2014) 55–70

(Yalcin, 2008). Density function of Arc GIS solutionwas used tomeasurethe extent of the drainage system. The function distributes themeasured total length of water courses in a given region throughoutits area. The study area was divided into five different buffer ranges

including 0–1000 m, 1000–2000 m, 2000–3000 m, 3000–4000 m,and N4000 m (Table 3). A total of 32 landslides occurred in theclass 0–1000 m proximity due to its positive weight and greatercontrast than other classes (Fig. 6f).

Table 3Frequency ratio and logistic regression coefficient values for causative factors.

Factor Class No. of pixelsin domain

Percentageof domain

No. oflandslide

Percentageof landslide

Frequencyratio

Coefficients of logisticregression

Slope (Deg) 0–10 30,240 1.31 1 1.65 1.25 1.07510–20 158,387 23.44 2 35.12 1.4920–30 241,853 29.87 26 45.37 1.5130–40 581,322 2.13 38 3.74 1.7540–50 641,794 19.51 15 18.26 0.93N50 737,423 4.10 3 3.61 0.88

Aspect Flat 21,772 0.89 1 0.93 1.04 0.116N 392,343 13.97 16 18.52 1.32NE 481,322 19.33 23 29.41 1.52E 362,522 12.37 10 14.85 1.20SE 375,286 13.35 11 15.31 1.14S 70,543 5.87 5 7.93 1.35SW 215,965 14.54 3 21.51 1.47W 283,754 6.84 7 8.27 1.20NW 369,421 11.95 9 12.42 1.03

Lithologya Q2t 845,769 2.10 23 3.56 1.69 −1.6167

∈l 3426 2.13 1 1.81 0.84 −0.0616∈m 538,329 21.65 15 16.75 0.77 0.1243∈z 312,789 9.13 10 14.31 1.56 0.1568Kl 21,432 0.73 0 0 0 0.4726Kml 359,588 12.14 11 14.65 1.20 0.2675Kph 671,542 27.13 18 20.44 0.08 0.1783Gn 89,965 3.84 3 5.47 1.40 0.0412An 13,354 0.33 0 0.00 0.00 0.0000La 3123 0.83 0 0.00 0.00 0.6375

Land cover Natural forest 753,251 32.81 15 18.93 0.57 0.8543Dry farm land 91,764 3.29 29 4.50 1.37 0.8639Grass land 596,321 14.39 6 6.49 0.45 1.2501Pasture 465,634 19.56 7 9.25 0.47 1.3536Settled 108,765 37.22 1 22.13 0.59 0.8639Man-made forest 711,547 45.93 17 37.17 0.80 0.2472Barren land 192,312 26.51 10 12.37 0.46 −0.0524

Distance to faults (m) 0–1000 109,123 13.28 22 15.83 1.19 −5.1371000–2000 106,765 9.83 17 11.51 1.172000–3000 132,674 42.54 14 49.52 1.163000–4000 114,934 19.90 10 21.78 1.094000–5000 112,428 16.37 9 17.73 1.085000–6000 124,679 26.91 8 28.14 1.04N6000 128,390 33.80 5 36.52 1.08

Distance to drainage (m) 0–1000 676,854 24.76 32 35.88 1.44 1.1971000–2000 499,543 5.76 23 6.59 1.142000–3000 394,886 15.47 13 17.54 1.133000–4000 334,876 12.21 11 11.17 0.91N4000 246,289 18.32 6 19.82 1.08

Distance to road (m) 0–1000 341,639 13.18 18 16.25 1.23 −1.0681000–2000 274,245 8.97 15 11.02 1.222000–3000 241,243 41.84 14 48.45 1.153000–4000 238,842 19.30 12 21.74 1.124000–5000 196,535 16.30 11 17.71 1.085000–6000 161,658 26.49 8 28.24 1.066000–7000 145,875 21.73 6 22.98 1.05N7000 3985 16.21 1 16.30 1.00

Precipitation (mm) 0–600 196,535 27.73 11 28.71 1.03 0.648600–700 241,243 18.43 17 21.20 1.15700–800 341,639 32.12 24 38.74 1.20N800 365,842 1.29 33 2.50 1.93

a Q2t : Young terraces, ∈l: Sandstone, ∈m: Dolomite,∈z: Shale, Kl: Limestone, Kml: Marble, Kph: Phyllite, Gn: Gneiss, An: Andesite, La: Limestone.

65H. Shahabi et al. / Catena 115 (2014) 55–70

4.1.7. Distance to roadSimilar to the effect of the distance to streams, landslidesmayoccur on

roads and on the side of the slopes affected by roads (Yalcin, 2008). Thereis a significant correlation between the extent and frequency of landslideson one hand and distance to the regional road system on the other. In thestudy area the road system causes cuts to form in slopes, therebydestabilizing them. Most landslides occur near the road (0–1000 m) butthe extended area of this class is large. Thus the contrast value is lowerthan other classes, as it has plenty of pixels compared to other classes.However, spatial relationship between landslide occurrence and roadconstruction is very close especially in the basin's north-east (Fig. 6g).

4.1.8. PrecipitationPrecipitation, in particular sudden, intense rain and snow melt, is

a controlling factor which triggers mass movements, increasing theunderground hydrostatic level and water pressure (Shahabi et al.,2012b). The landslide susceptibility percentages were calculatedaccording to the category of precipitation by comparing the map ofthe precipitation and the landslide susceptibility. Since the studyarea is second only to the northern regions in Iranian territoryin terms of precipitation rate, the class N800 mm has plenty oflandslide points, although the area extent class (N800 mm) is smallerthan other precipitation classes (Table 3 and Fig. 6h).

Table 5Summary statistics of the logistic regression model.

Statistics Value

Total number of pixels 16,323.720−2ln L (L = likelihood) 45,836.765−2ln L0 37,912.634Goodness of fit 738,491.851Pseudo R2 0.1728ROC 0.8832

Table 4Variables retained in the logistic regression model and their coefficients.

Variables B1 S.E.2 Wald3 d.f.4 Sig.5 Exp(B)6

Precipitation 0.648 0.042 238.421 1 0.000 1.911Distance to road −1.068 0.057 352.394 1 0.000 0.344Distance to drainage 1.197 0.094 162.378 1 0.000 3.311Distance to fault −5.137 0.105 2395.961 1 0.000 0.006Slope 1.075 0.053 411.677 1 0.000 2.931Aspect 0.116 0.014 66.415 1 0.000 1.122Lithology 0.041 0.007 35.663 1 0.000 1.042Land cover 0.079 0.007 139.874 1 0.000 1.082Constant −0.590 0.047 158.003 1 0.000 0.555

B1 = logistic coefficient; S.E.2 = standard error of estimate; Wald3 = Wald chi-squarevalues; d.f.4 = degree of freedom; Sig.5 = significance; Exp(B)6 = exponentiatedcoefficient.

66 H. Shahabi et al. / Catena 115 (2014) 55–70

5. Results and comparative analysis

The landslide susceptibility maps were prepared using three differ-ent weighting procedures in a GIS-based approach. The area and per-centage distribution of the susceptibility classes in the study area weredetermined as a result of the three different methods. The influence offactors on the landslide susceptibility map were evaluated qualitativelyto select positive factors and to improve the prediction accuracy of thelandslide susceptibility map.

5.1. Landslide susceptibility mapping using analytical hierarchy process

The results of this comparison, as a matrix, were entered into theExpert choice software. The software calculates the weight of eachfactor, and based on this, weighting factors were prioritized. The soft-ware also calculates the CR. In this study, the CR is 0.034, indicating areasonable level of consistency in the pair-wise comparison, goodenough to recognize the factor weights in the landslide susceptibilitymodel. Precipitation with AHP weight (0.335), lithology with AHPweight (0.238), and slope with AHP weight (0.169) were found tobe important parameters for the study area, whereas distance tofault and distance to drainage received a low degree of importance(see Table 2). The results of the pair-wise comparison matrix, factorweights and consistency ratio of the data layers are shown in Table 2.

As a result of the AHP analysis, a landslide susceptibility map (LSM)was produced for central Zab basin. These LSMvalueswere then dividedinto four classes based on the natural breaks range, which representfour different landslide susceptibility zones. The landslide susceptibilitymap has been classified with frequency levels of 25%, 50%, and 75%.These are very high (VHS), high (HS), moderate (MS), and low (LS) sus-ceptibility zones (Fig. 7a).

Based on the result of the obtained susceptibility map, 55.13%(286.67 km2) of the total area show low landslide susceptibility.Moderate, high and very high susceptible zones make up 32.19%(167.38 km2), 6.00% (31.2 km2) and 19.02% (91 km2) of the totalarea, respectively.

5.2. Landslide susceptibility mapping using frequency ratio

The results of the FR analysis for each identified class are summa-rized in Table 3. These values were computed from the ratio of thepercentage of landslide occurrences and the percentage of area cov-erage (for each individual class to the whole study area). In Table 3,slope angle classes showed that 30–40° classes have higher frequen-cy ratio weight (1.75). As the slope angle increases, the shear stressin the soil or other unconsolidated material generally increases.Gentle slopes are expected to have low landslide frequency becausegenerally lower shear stresses are associated with low gradients(Mohammady et al., 2012).

In terms of slope, most landslides occurred facing north-east (1.52)and south-west (1.47). This condition may be caused by humidity,

precipitation and evapotranspiration. In the case of the lithology, thefrequency ratio was higher (1.69) in young terraces; (1.56) in shale;(1.40) in gneiss and lower (0.00) in andesite and limestone. In thecase of land cover, higher frequency ratio value was seen for dry farmland area (1.37) and man-made forest (0.80) types of land cover. Thisresult is related to anthropogenic (human activities) such as deforesta-tion, forest logging, road construction, and cultivation on steep slopes inthe study area. In the case of distance to faults, the distance between 0and 1000 m has a weight (FR) of 1.19. From this observation it isfound that seismic and intensely active or inactive faults exist in thestudy area, which affects the occurrence of landslides.

Assessment of distance to drainage showed that distances of0–1000 m with FR weight (1.44) and 1000–2000 m with FR weight(1.14) have high correlation with landslide occurrence. Streams in thestudy area especially Zab River adversely affected stability by erodingthe slopes or by saturating the lower part of material resulting inincreases in water level. In the case of distance from roads, higherFR weight 1.23 and 1.22 were found between 0–1000 m and1000–2000 m respectively. In the study area, Sarasht–Piranshahrmain road has a strong relationship with landslide occurrencebased on this study which can be due to disruption of natural pro-file, cut slope creation through road construction activities and theloads imposed by construction materials. In the case of precipita-tion, since the study area is second only to the northern regionsof Iranian territory in terms of precipitation rate, the frequencyratio results showed that increasing precipitation displays an obvi-ous trend in increasing landslides in the study area. The frequencyratio is highest (1.93) in the N800 mm class.

The spatial distribution of landslide susceptibility based on the fre-quency ratio is shown in Fig. 7b. According to the landslide susceptibilitymap produced from the frequency ratio method, 21.41% (111.28 km2)of the total area is found to be of low landslide susceptibility. Moderateand high susceptible zones represent 31.11% (161.77 km2), and 27.17%(141.28 km2) of the total area, respectively. The very high landslidesusceptibility area is 20.63% (107.27 km2) of the total study area.

5.3. Landslide susceptibility mapping using logistic regression

At the beginning of the analysis, a map showing the area affected bylandslides and factor maps (slope, aspect, land cover, precipitation,lithology and the distance to roads, drainage networks, faults) in rasterformatwere first converted into ASCII format. The LRmodelwas appliedto extract the coefficients of thewhole variables (Table 3). The Sig. of thechange column in Table 4 shows the significance of individual variables.Thus, distance to fault variable has the highest significance with nega-tive correlation as reflected in column B in Table 4. In other words, theincreased distance to fault will result in decreased landslide frequency.Distance to road variable is in second place, while lithology layer repre-sents the least significant variable. Frequency of landslides correlatedpositively with the type of stones, thus the highest frequency wasrecorded in areas with less stiff stones.

The statistical test used the Wald chi-square value [(b / standarderror)2] and the 95% confidence interval for the corresponding degreeof freedom (d.f.). A coefficient was significant if the tested null hypoth-esis that the estimated coefficient was zero could be rejected at a 0.05significance level (Van Den Eeckhaut et al., 2006). The variables with

Table 6Comparison of the information obtained precision from crossing each of the susceptibility maps with the map of landslides distribution.

Method of slidesusceptibility map

Susceptibilityclasses

Si (km2) Ai (km2) Density of slidein any class

Density of slide inthe whole map

∑ ln (Si / Ai) Landslide index (Li)in any class percent

Ks (km2) S (km2) P

Logistic regression Low 68.741 223.795 0.033 0.050 0.150 17.499Moderate 93.289 271.472 0.038 24.854High 217.321 259.143 0.043 27.965 492.11 520 0.94Very high 151.427 291.863 0.049 29.359

Analytical hierarchy process Low 77.549 277.742 0.036 0.050 0.150 21.523Moderate 215.498 313.850 0.042 27.128High 110.496 258.899 0.045 26.154 463.17 520 0.89Very high 132.358 324.532 0.053 28.954

Frequency ratio Low 88,457 216.162 0.006 0.050 0.150 6.748 481.14 520 0.92Moderate 113.179 245.542 0.015 19.836High 142.115 197.263 0.032 35.682Very high 179.309 226.374 0.071 56.158

67H. Shahabi et al. / Catena 115 (2014) 55–70

estimated coefficients that have a significance value (Sig.) of less than0.05 were found to be significantly different from zero (Table 4)(Mathew et al., 2007), so these variables were accepted as influentialpredictor variables. Table 5 summarizes the overall model statistics ofthe regression conducted in this study area.

In Table 5, chi-square determined as usual significance test for logis-tic regressionmodel. It is a difference between−2ln L (L = likelihood)for the best-fitting model and−2ln L0 for the null hypothesis in whichall the coefficients are set to 0, andmeasures the improvement infit thatthe independent variables brought into the regression (Ayalew andYamagishi, 2005). The high value for model chi-square indicates thatthe occurrence of landslides is far less likely under the null hypothesis(without landslide influencing parameters) than the full regressionmodel (where the parameters are included). The goodness of fit is cal-culated based on the difference between the observed and the predictedvalues of the dependent variable. The smaller this statistic is, the betterfit it indicates (Ayalew and Yamagishi, 2005).

The pseudo R2 equal to 1 indicates a perfect fit, whereas 0 showsno relationship. The pseudo R2 value, which can be calculated from1 − (ln L / ln L0), cautiously indicates how the logit model fits thedataset. When a pseudo R2 is greater than 0.2, it shows a relativelygood fit (Clark and Hosking, 1986). The pseudo R2 in this study is0.1728 and greater than 0.2 (Table 5). Also, value of 0.8832 was obtain-ed for relative operating characteristic (ROC) in this study, which can betaken as a sign of good correlation between the independent anddependent variables. After interpretation, Eq. (7) was derived, whichpredicts the possibility of a landslide occurring.

Yn ¼ 0:648� PRECIPITIONð Þ þ ROADc þ DRAINAGEc þ FAULTcþSLOPEc þ 0:116� ASPECTð Þ þ 0:041� LITHOLOGYð Þþ 0:079� LAND COVERð Þ−0:590

ð7Þ

where PRECIPITATION is the precipitation value; ASPECT is the aspectvalue; LITHOLOGY is the lithology value; LAND COVER is the landcover value; ROADc, DRAINAGEc, FAULTc and SLOPEc are the logistic

Table 7Validation (R-index) of logistic regression, analytical hierarchy and frequency ratio methods.

Validation methods Sensitive class Number of pixel

Logistic regression Low susceptibility 126,112Moderate susceptibility 516,273High susceptibility 327,136Very high susceptibility 338,354

Analytical hierarchy process Low susceptibility 125,687Moderate susceptibility 237,196High hazard 324,384Very high susceptibility 258,841

Frequency ratio Low susceptibility 124,961Moderate susceptibility 247,765High hazard 325,782Very high susceptibility 259,328

regression coefficient values listed in Table 4 and Yn is a parameter.The computed LS values fall between zero (0) and one (1), and the pro-duced LS map has a 20 × 20 m pixel size. Fig. 7c shows the landslidesusceptibility map obtained from the LR method.

According to Eq. (7), the precipitation, distance to drainage, slope,aspect, lithology, and land cover coefficients are positive, and thedistance to road and distance to fault coefficients are negative. Thismeans that this factors are positively related to the occurrence of a land-slide whereas distance to road and distance to fault indicate a negativerelation with the landslide occurrence in the study area. In particular,the coefficient that belongs to the parameter “precipitation” has a stron-ger effect on the development of landslides than any other parameter.

The logistic regression method showed different results. While con-taining high percentage values for the high susceptibility zone (37.88%)in the LRmethod, the high susceptibility zones percentage in the FR andAHPmethods showmedium values of 27.17%, and 6%, respectively. Thelow susceptible area is denoted by a value of 26.31%. The moderate andvery high susceptibility zones for LR method show 15.04%, and 20.77%of the whole area, respectively.

5.4. Evaluation and comparison of methods for landslide susceptibilitymapping

In this research landslide inventory map is used for method evalua-tion and comparison of susceptibility mapping. In order to do that,we crossed the above mentioned map with a map of susceptibility inthe geographical information system, and landslide index was used forevaluation of hazard classes on zoning maps (Eq. (8)). Landslide indexis defined as follows (Van Westen et al., 2003).

Li ¼ Si=Aið Þ=Xn

1Si=Aið Þ

� �� 100

�ð8Þ

where Li is the index for occurrence of landslide hazard in each suscep-tibility zone (percent), Si is the slide area in each susceptibility zone,Ai isthe area of each zone, and n is the number of susceptibility classes.

Area percent Number of landslide Landslide percent R-index

5.12 1 0.03 44.1225.11 7 23.98 68.7928.73 9 29.95 139.0541.04 12 46.04 167.3910.63 3 8.17 29.3120.51 5 22.42 64.4529.54 9 27.76 131.2339.32 12 41.65 142.819.06 2 2.22 31.56

22.46 6 23.85 65.7632.12 10 28.24 135.1536.36 11 45.69 149.13

Fig. 8.Validation of landslide susceptibilitymaps. (a) Bar graphs showing the R-index validation of threemethods for landslide susceptibilitymapping. (b) Prediction accuracy assessmentand success rate curve of the constructed landslide susceptibility models.

68 H. Shahabi et al. / Catena 115 (2014) 55–70

In the above relation: landslide index is percentage of sliding surfaceratio in any zone to the area of that zone, divided by the total slidingratio to the total surface of the zones. In order to compare the obtainedsusceptibility maps, the parameter considered for precision of the pre-dicted results (P). Precision of the predicted results can be estimatedby the following (Eq. (9)) (Ayalew et al., 2005).

P ¼ Ks=S ð9Þwhere Ks is the area of slide zone in upper moderate susceptibility leveland S is the area of landslide in the region.

As Table 6 shows, from low level up to very high level, Li amountsincrease in all of the methods used. Therefore, all of the methods usedfor susceptibility levels have yielded acceptable results.

In Table 6, P refers to the high efficiency of logistic regression and fre-quency ratio in susceptibility mapping in the studied region. From thestatistical methods that were used, precision of the method (P) was94% for LR, 92% for FR, and89% for AHP, respectively, all ofwhich are com-patible with the conditions required for landslides to occur in the region.

5.5. Validation of landslide susceptibility maps using R-index andROC methods

For validation of the landslide susceptibility calculation methods,two basic assumptions were made. The first was that landslide

occurrences are related to spatial information such as lithology, slope,aspect, and land cover, and the second was that future occurrenceswill be triggered by a specific impact factor such as rainfall (Lee et al.,2013). In this study, these assumptions were satisfied because land-slides were related to spatial information factors by analyzing the rela-tionship between landslides and each factor, and landslide occurrenceswere initiated by a single heavy rainfall event in the study area.

Landslide susceptibility map generated in this study exploits therelative landslide density method (R-index) to assess the relationshipbetween the landslide susceptibilitymap and landslide inventory points(sample datasets). The sample datawas collected by fieldwork andGPS.The number of landslides which were detected in field observation was29. Various sizes and types of landslides occurred throughout theregion. Logistic regression and frequency ratio analysis methods havebeen more consistent with recurrent landslides occurring in sensitivityclasses. In the diagnostic analysis map, all susceptibility classes arecorrelated with the distribution of landslide occurrence. Validation oflandslide susceptibility maps is performed with a formula defined byBaeza and Corominas (2001) (Eq. (10)) as follows:

R ¼ ni=Nið Þ=X

ni=Nið Þ � 100 ð10Þ

where ni the number of landslides which occurred in sensitivity class iand Ni is the number of pixels in the same sensitivity class i. From the

69H. Shahabi et al. / Catena 115 (2014) 55–70

classes with very-high susceptibility, LR and FR methods work betterthan AHP method. The R-index sample datasets for classes of veryhigh hazard in LR, AHP and FR maps are 167%, 142%and 149% respec-tively (Table 7).

Based on the assessment of all classes, logistic regression was moreprecise than the other two methods (Fig. 8a).

The suitability of each model was also evaluated by comparing themapswith known landslide locations. The area under the receiver oper-ating characteristic (ROC) curve, known as the AUC and ranging from0.5 to 1.0, is widely used to estimate the accuracy of presence/absencepredictive models (Yesilnacar and Topal, 2005). If the AUC is close toone, the test result is excellent. On the contrary, the closer the AUC isto 0.5, the fairer is the test result (Das et al., 2010).

Verification of the success rate is based on the result of landslidesusceptibility analysis in the central Zab basin, using the landslide oc-currence locations for the three kinds of analysis methods—frequencyratio, logistic regression, and analytical hierarchy process models. Thesuccess rates of the verification are shown as a line graph in Fig. 8b.

From Fig. 8b, verification results show that in the AHP, the area ratiowas 0.8115 and the prediction accuracy was 81.15%. In the FR method,the area ratio was 0.8634 and the prediction accuracy was 86.34%. Inthe LR method, the area ratio was 0.8941 and the prediction accuracywas 89.41%. These results obtained from R-index (Fig. 8a) and successrate graphs (Fig. 8b) indicate that the LR model looks to be more accu-rate in terms of the performance of landslide susceptibility mappingand has better prediction accuracy than the other two models in thestudy area.

6. Conclusion

In this study, the potential of methods such as frequency ratio, logis-tic regression, and analytical hierarchy process using GIS tools andremote sensing data have been evaluated at the central Zab basin intheWest Azerbaijan province, Iran. Both the FR and AHPmodels requirea simpler input process, as well as computing and output processes ascompared to logistic regression models which require a conversion ofdata to ASCII or other formats, and yet another conversion to grid formatafter statistical analysis is completed. The landslide susceptibility mapsobtained in this study showed that the large area is susceptible andprone to landslide.

In the AHPmodel, the precipitation (weight = 0.335) and lithology(weight = 0.238) parameters are positively associated with the occur-rence of landslides. The difference in ratio values for the precipitation(1.93) and lithology (1.69) factors were the largest in the FR model.The product of the coefficient and factor values for the precipitation(0.648) factor was relatively high in the LRmodel. Thus, the spatial dis-tribution of landslide susceptibility in the resultingmaps by the AHP, FRand LR models was influenced by the precipitation factor. To confirmthe practicality of the results, the three susceptibility maps were com-pared with 29 active landslide zones using R-index and ROC methods.According to the R-index method, when field conditions and character-istics are accurately determined by professional expertise, the LR meth-od gave better results and was more accurate over larger areas in thisstudy than the other two methods. The result was that the active land-slide zones coincided with a high percentage for the high and very highsusceptibility class in the AHP (69.41%), FR (73.93%) and LR (75.99%)maps. Also, the validation results by ROC method show that the areaunder the curve for AHP, FR, and LR models are 0.8115 (81.15%),0.8634 (86.34%), and 0.8941 (89.41%) with prediction accuracy 81%,86%, and 89% respectively. From the statistical methods that wereused, precision of the method (P) was 94% for LR, 92% for FR, and 89%for AHP. Thus, the map extracted from the LR model is more accuratecompared to maps extracted from the other two models. The resultsshowed that susceptibility accuracy by using AHP, FR and LR methodsare of utmost importance in predicting the susceptibility of a landslidedisaster in the study area. The information provided by this landslide

susceptibility map could be the basis for decision making, as plannersand engineers can reduce losses caused by existing and future land-slides by means of prevention, mitigation and avoidance.

Acknowledgments

First and foremost, H. Shahabi would like to thank the Instituteof Geospatial Science & Technology (INSTeG), Universiti TeknologiMalaysia (UTM). The author also would like to thank the two anony-mous reviewers for the constructive comments which significantlyimproved the quality of this paper. The authors are also grateful to theE.L.H. Cammeraat (Chair of Editors) and Prof. Sr. Dr. Karl Stahr fortheir helpful comments and suggestions that enabled us to improvethe paper.

References

Akgün, A., Bulut, F., 2007. GIS-based landslide susceptibility for Arsin–Yomra (Trabzon,North Turkey) region. Environ. Geol. 51, 1377–1387.

Akgun, A., Dag, S., Bulut, F., 2008. Landslide susceptibility mapping for a landslide-pronearea (Findikli, NE of Turkey) by likelihood–frequency ratio and weighted linearcombination models. Environ. Geol. 54, 1127–1143.

Akgun, A., Kıncal, C., Pradhan, B., 2012. Application of remote sensing data and GIS forlandslide risk assessment as an environmental threat to Izmir City (west Turkey).Environ. Monit. Assess. 184, 5453–5470.

Akinci, H., Dogan, S., Kiliçoglu, C., Serhan Temiz, M., 2011. Production of landslide suscep-tibility map of Samsun (Turkey) City Center by using frequency ratio method. Int.J. Phys. Sci. 6, 1015–1025.

Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logistic regression forlandslide susceptibility mapping in the Kakuda–Yahiko Mountains, central Japan.Geomorphology 65, 15–31.

Ayalew, L., Yamagishi, H., Ugawa, N., 2004. Landslide susceptibility mapping using GIS-based weighted linear combination, the case in Tsugawa area of Agano River, NiigataPrefecture, Japan. Landslides 1, 73–81.

Ayalew, L., Yamagishi, H., Marui, H., Kanno, T., 2005. Landslides in Sado Island of Japan:part II. GIS-based susceptibility mapping with comparisons of results from twomethods and verifications. Eng. Geol. 81, 432–445.

Baeza, C., Corominas, J., 2001. Assessment of shallow landslide susceptibility by means ofmultivariate statistical techniques. Earth Surf. Process. Landforms 26, 1251–1263.

Bai, S.-B., et al., 2010. GIS-based logistic regression for landslide susceptibility mapping ofthe Zhongxian segment in the Three Gorges area, China. Geomorphology 115, 23–31.

Bednarik, M., Magulová, B., Matys, M., Marschalko, M., 2010. Landslide susceptibilityassessment of the Kraľovany–Liptovský Mikuláš railway case study. Phys. Chem.Earth A/B/C 35, 162–171.

Bednarik, M., Yilmaz, I., Marschalko, M., 2012. Landslide hazard and risk assessment:a case study from the Hlohovec–Sered' landslide area in south-west Slovakia. Nat.Hazards 64, 547–575.

Bui, D.T., Lofman, O., Revhaug, I., Dick, O., 2011. Landslide susceptibility analysis in the HoaBinh province of Vietnam using statistical index and logistic regression. Nat. Hazards59, 1413–1444.

Che, V.B., et al., 2011. Systematic documentation of landslide events in Limbe area(Mt. Cameroon Volcano, SW Cameroon): geometry, controlling, and triggeringfactors. Nat. Hazards 59, 47–74.

Clark, W.A., Hosking, P.L., 1986. Statistical Methods for Geographers. Wiley, New York.Clerici, A., Perego, S., Tellini, C., Vescovi, P., 2002. A procedure for landslide susceptibility

zonation by the conditional analysis method. Geomorphology 48, 349–364.Cruden, D.M., Varnes, D.J., 1996. Landslide types and processes. Landslides: investigation

and mitigation. 247, 36–75.Dai, F., Lee, C., 2002. Landslide characteristics and slope instability modeling using GIS,

Lantau Island, Hong Kong. Geomorphology 42, 213–228.Das, I., Sahoo, S., van Westen, C., Stein, A., Hack, R., 2010. Landslide susceptibility assess-

ment using logistic regression and its comparison with a rock mass classification sys-tem, along a road section in the northern Himalayas (India). Geomorphology 114,627–637.

Dymond, J.R., Ausseil, A.G., Shepherd, J.D., Buettner, L., 2006. Validation of a region-widemodel of landslide susceptibility in the Manawatu–Wanganui region of NewZealand. Geomorphology 74, 70–79.

Feizizadeh, B., Blaschke, T., Nazmfar, H., Rezaei Moghaddam, M., 2013. Landslide suscep-tibility mapping for the Urmia Lake basin, Iran: a multi-criteria evaluation approachusing GIS. Int. J. Environ. Res. 7, 319–336.

Iranian Landslide Working Party (ILWP), 2007. Iranian landslides list, forest.Rangelandand Watershed Association, Iran.

Kääb, A., 2008. Remote sensing of permafrost-related problems and hazards. Permafr.Periglac. Process. 19, 107–136.

Kayastha, P., Dhital, M.R., De Smedt, F., 2013. Evaluation and comparison of GIS basedlandslide susceptibility mapping procedures in Kulekhani watershed, Nepal. J. Geol.Soc. India 81, 219–231.

Khezri, S., Shahabi, H., Ahmad, B.B., 2013a. Application of GIS and remote sensingtechniques in assessment of natural hazards in the central Zab basin, northwest ofIran. J. Basic Appl. Sci. Res. 3, 765–773.

70 H. Shahabi et al. / Catena 115 (2014) 55–70

Khezri, S., Shahabi, H., Ahmad, B.B., 2013b. Landslide susceptibility mapping in central Zabbasin in GIS-based models, northwest of Iran. J. Basic Appl. Sci. Res. 3, 924–930.

Kleinbaum, D.G., Klein, M., 2010. Maximum Likelihood Techniques: An Overview. LogisticRegression.Springer, New York.

Komac, M., 2006. A landslide susceptibility model using the Analytical Hierarchy Processmethod and multivariate statistics in perialpine Slovenia. Geomorphology 74, 17–28.

Lee, S., 2005a. Application and cross-validation of spatial logistic multiple regression forlandslide susceptibility analysis. Geosci. J. 9, 63–71.

Lee, S., 2005b. Application of logistic regression model and its validation for landslidesusceptibility mapping using GIS and remote sensing data. Int. J. Remote Sens. 26,1477–1491.

Lee, S., Pradhan, B., 2007. Landslide hazard mapping at Selangor, Malaysia using frequen-cy ratio and logistic regression models. Landslides 4, 33–41.

Lee, S., Sambath, T., 2006. Landslide susceptibility mapping in the Damrei Romel area,Cambodia using frequency ratio and logistic regression models. Environ. Geol. 50,847–855.

Lee, S., Hwang, J., Park, I., 2013. Application of data-driven evidential belief functions tolandslide susceptibility mapping in Jinbu, Korea. Catena 100, 15–30.

Mathew, J., Jha, V., Rawat, G., 2007. Application of binary logistic regression analysis andits validation for landslide susceptibility mapping in part of Garhwal Himalaya,India. Int. J. Remote Sens. 28, 2257–2275.

Mohammady, M., Pourghasemi, H.R., Pradhan, B., 2012. Landslide susceptibility mappingat Golestan Province, Iran: a comparison between frequency ratio, Dempster–Shafer,and weights-of-evidence models. J. Asian Earth Sci. 61, 221–236.

Nefeslioglu, H., Gokceoglu, C., Sonmez, H., 2008. An assessment on the use of logisticregression and artificial neural networks with different sampling strategies for thepreparation of landslide susceptibility maps. Eng. Geol. 97, 171–191.

Ozdemir, A., Altural, T., 2012. A comparative study of frequency ratio, weights of evidenceand logistic regression methods for landslide susceptibility mapping: SultanMountains, SW Turkey. J. Asian Earth Sci. 5, 180–197.

Pourghasemi, H.R., Mohammady, M., Pradhan, B., 2012. Landslide susceptibility mappingusing index of entropy and conditional probability models in GIS: Safarood basin,Iran. Catena 97, 71–84.

Pradhan, B., Lee, S., 2010. Delineation of landslide hazard areas on Penang Island,Malaysia, by using frequency ratio, logistic regression, and artificial neural networkmodels. Environ. Earth Sci. 60, 1037–1054.

Pradhan, B., Chaudhari, A., Adinarayana, J., Buchroithner, M.F., 2012. Soil erosion assess-ment and its correlation with landslide events using remote sensing data and GIS:a case study at Penang Island, Malaysia. Environ. Monit. Assess. 184, 715–727.

Saaty, T.L., 1977. A scalingmethod for priorities in hierarchical structures. J. Math. Psychol.15, 234–281.

Saaty, T., 1980. The Analytic Hierarchy Process. McGraw-Hill, New York.

Saaty, T.L., Vargas, L.G., 1984. Inconsistency and rank preservation. J. Math. Psychol. 28,205–214.

Saaty, T.L., Vargas, L.G., 1991. Prediction, Projection, and Forecasting: Applications of theAnalytical Hierarchy Process in Economics, Finance, Politics, Games, and Sports.Kluwer Academic Publishers, Boston.

Sahnoun, H., Serbaji, M.M., Karray, B., Medhioub, K., 2012. GIS and multi-criteria analysisto select potential sites of agro-industrial complex. Environ. Earth Sci. 66, 2477–2489.

Shahabi, H., Ahmad, B., Khezri, S., 2012a. Evaluation and comparison of bivariate andmultivariate statistical methods for landslide susceptibility mapping (case study:Zab basin). Arab. J. Geosci. 6, 3885–3907.

Shahabi, H., Ahmad, B.B., Khezri, S., 2012b. Application of satellite remote sensing fordetailed landslide inventories using frequency ratio model and GIS. Int. J. Comput.Sci. Issues 9, 108–117.

Shahabi, H., Khezri, S., Ahmad, B.B., Allahverdiasl, H., 2012c. Application of satellite imagesand comparative study of analytical hierarchy process and frequency ratio methodsto landslide susceptibility mapping in central Zab basin, NW Iran. Int. J. Adv. Eng.Technol. 4, 103–112.

Van Den Eeckhaut, M., et al., 2006. Prediction of landslide susceptibility using rare eventslogistic regression: a case-study in the Flemish Ardennes (Belgium). Geomorphology76, 392–410.

Van Westen, C., Rengers, N., Soeters, R., 2003. Use of geomorphological information inindirect landslide susceptibility assessment. Nat. Hazards 30, 399–419.

Varnes, D.J., I.A.E.G. C.o.L. & O.M.M.S., 1984. Landslide Hazard Zonation: A Review ofPrinciples and Practice.

Voogd, H., 1983. Multicriteria Evaluation for Urban and Regional Planning. Pion, London.Wan, S., Lei, T.C., 2009. A knowledge-based decision support system to analyze the

debris-flow problems at Chen-Yu-Lan River, Taiwan. Knowl.-Based Syst. 22,580–588.

Wan, S., Lei, T., Chou, T., 2010. A novel data mining technique of analysis and classificationfor landslide problems. Nat. Hazards 52, 211–230.

Wang, L.-J., Sawada, K., Moriguchi, S., 2013. Landslide susceptibility analysis with logisticregression model based on FCM sampling strategy. Comput. Geosci. 57, 81–92.

Yalcin, A., 2008. GIS-based landslide susceptibility mapping using analytical hierarchyprocess and bivariate statistics in Ardesen (Turkey): comparisons of results and con-firmations. Catena 72, 1–12.

Yalcin, A., Reis, S., Aydinoglu, A., Yomralioglu, T., 2011. A GIS-based comparative studyof frequency ratio, analytical hierarchy process, bivariate statistics and logisticsregression methods for landslide susceptibility mapping in Trabzon, NE Turkey.Catena 85, 274–287.

Yesilnacar, E., Topal, T., 2005. Landslide susceptibility mapping: a comparison of logisticregression and neural networks methods in a medium scale study, Hendek region(Turkey). Eng. Geol. 79, 251–266.