landslide susceptibility mapping by neuro-fuzzy approach in a

4164 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 48, NO. 12, DECEMBER 2010

Landslide Susceptibility Mapping by Neuro-FuzzyApproach in a Landslide-Prone Area

(Cameron Highlands, Malaysia)Biswajeet Pradhan, Ebru Akcapinar Sezer, Candan Gokceoglu, and Manfred F. Buchroithner

Abstract—This paper presents the results of the neuro-fuzzymodel using remote-sensing data and geographic information sys-tem for landslide susceptibility analysis in a part of the CameronHighlands areas in Malaysia. Landslide locations in the study areawere identified by interpreting aerial photographs and satelliteimages, supported by extensive field surveys. Landsat TM satelliteimagery was used to map the vegetation index. Maps of thetopography, lineaments, Normalized Difference Vegetation Index(NDVI), and land cover were constructed from the spatial datasets. Eight landslide conditioning factors such as altitude, slopegradient, curvature, distance from the drainage, distance fromthe road, lithology, distance from the faults, and NDVI wereextracted from the spatial database. These factors were analyzedusing a neuro-fuzzy model adaptive neuro-fuzzy inference systemto produce the landslide susceptibility maps. During the modeldevelopment works, a total of five landslide susceptibility modelswere constructed. For verification, the results of the analyseswere then compared with the field-verified landslide locations.Additionally, the receiver operating characteristic curves for alllandslide susceptibility models were drawn, and the area undercurve values were calculated. Landslide locations were used tovalidate the results of the landslide susceptibility map, and theverification results showed a 97% accuracy for model 5, employingall parameters produced in the present study as the landslideconditioning factors. The validation results showed a sufficientagreement between the obtained susceptibility map and the ex-isting data on the landslide areas. Qualitatively, the model yieldsreasonable results, which can be used for preliminary land-useplanning purposes.

Index Terms—Adaptive neuro-fuzzy inference system (ANFIS)model, Cameron Highlands, geographic information system (GIS),landslide susceptibility, Malaysia, neuro-fuzzy, remote sensing.

I. INTRODUCTION

LANDSLIDES are one of the recurrent natural hazard prob-lems throughout most of Malaysia. According to local

newspaper reports (The Star 2008 and 2009), in 2006–2008

Manuscript received October 17, 2009; revised February 3, 2010 andApril 25, 2010. Date of publication July 1, 2010; date of current versionNovember 24, 2010.

B. Pradhan and M. F. Buchroithner are with the Faculty of Forestry, Geo-sciences and Hydrosciences, Institute for Cartography, Dresden Univer-sity of Technology, 01062 Dresden, Germany (e-mail: [email protected]; [email protected]; [email protected]).

E. A. Sezer is with the Department of Computer Engineering, HacettepeUniversity, Ankara 06800, Turkey.

C. Gokceoglu is with the Department of Geological Engineering, HacettepeUniversity, Ankara 06800, Turkey.

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TGRS.2010.2050328

and also in 2009, heavy rainfalls triggered numerous landslidesand mudflows along the east coast highways and CameronHighlands in Peninsular Malaysia, in Sabah (East Malaysia),and in the island state of Penang. The areas that were hitthe hardest were those along the Cameron Highlands (in themountainous state of Pahang in Peninsular Malaysia). Theselandslides caused millions of dollars of property loss and evenled to loss of lives. The extent of the damages could have beenminimized if a long-term early warning system predicting themass movements in the landslide-prone areas was in place.

The landslides that occurred along the New Klang ValleyExpress Highways region in 2003 have alerted the highwayauthorities and other governmental organizations toward theseriousness of landslide management and prevention. TheOctober 2002 landslide in Kuala Lumpur which completelydestroyed a few houses and killed six members of a family isstill in the public’s memory. Landslides in Malaysia are mainlytriggered by tropical rainfalls causing failure of the rock surfacealong fracture, joint, and cleavage planes. The lithological unitsof the country are quite stable, but continuous uncontrolledurbanization leads to deforestation and erosion of the coveringsoil layers, thus causing serious threats to the slopes.

Earthquakes are the major triggers of landslides in moun-tainous terrain. However, Malaysia is not a seismically activeregion. In spite of this situation, Malaysia is surrounded byearthquake-prone areas, and there is always the probability oftransmitting mild shocks, as experienced recently in the westernparts of Peninsular Malaysia in 2006, 2008, and 2009 (The Star2006, 2008, and 2009). Landslides in Malaysia are triggeredmainly by heavy rainfalls. Recently, the Cameron Highlandshas faced numerous landslide and mudflow events, and muchdamage occurred in these areas. However, only a little efforthas been made to assess or predict these events, which re-sulted in serious damages. Through scientific analyses of theselandslides, one can assess and predict the landslide-susceptibleareas. Therefore, understanding the landslides and preventingthem are two of the serious challenges in Malaysia. To achievethis aim, in this paper, landslide susceptibility analyses havebeen performed based on the definition of Varnes [1] andhave been verified in the study area using the adaptive neuro-fuzzy inference system (ANFIS), because the ANFIS model hasnot been previously used for landslide susceptibility mappingpurposes, although it is a suitable and powerful inferencesystem. The study includes three main stages such as landslideinventory, analyses, and verification studies.

0196-2892/$26.00 © 2010 IEEE

PRADHAN et al.: LANDSLIDE SUSCEPTIBILITY MAPPING BY NEURO-FUZZY APPROACH 4165

II. PREVIOUS WORKS

There have been many studies carried out on landslidesusceptibility evaluation using geographic information system(GIS); for example, Guzzetti et al. (1999) summarized manylandslide hazard evaluation studies based on geomorphologicalrelationships between landslide types and pattern and basedon the morphological, lithological and structural settings [2].Recently, there have been studies on landslide susceptibilityevaluation using GIS, and many of these studies have appliedprobabilistic models [3]–[6]. One of the statistical modelsavailable, which is logistic regression analysis, has also beenapplied to landslide susceptibility mapping [6]–[15]. Otherconventional methods such as the geotechnical model, thechange detection, and the safety factor model have been appliedin different areas [16]–[19]. Moreover, landslide prediction hasalso been carried out using satellite remote-sensing data andgeospatial data sets. Chang et al. (2007) applied the multisourcedata fusion approach for landslide classification using gener-alized positive Boolean functions [19]. As a new approach tolandslide susceptibility evaluation using GIS, expert system,fuzzy logic, and artificial neural network models have beenapplied [20]–[26]. Lee et al. (2007) have assessed landslidesusceptibility analysis at Youngin, Korea using frequency ra-tio, logistic regression, and the ANN model [26]. In a recentpaper, Choi et al. (2009) have applied the neural networkmodel in landslide susceptibility mapping and have validatedthe model using the existing landslide data [27]. They haveapplied the neural network model at three study areas in Koreaand have cross-applied their weight for landslide susceptibilitymapping to achieve a reasonable prediction accuracy (81.36%).In the neural network method, Nefeslioglu et al. (2008) showedthat ANNs give a more optimistic evaluation of landslidesusceptibility than logistic regression analysis [13], whereasMelchiorre et al. (2008) improved the predictive capability androbustness of ANNs by introducing a cluster analysis [28].Moreover, Kanungo et al. (2006) showed that the landslidesusceptibility map that is derived from combined neural andfuzzy weighting procedure is the best among the other weight-ing techniques [29]. Ermini et al. (2005) compared two neuralarchitectures (probabilistic neural network and multilayeredperceptor), obtaining a better prediction with the latter [30].Also, the temporal hazard was estimated with ANNs via thetranslation of the state of activity in recurrence time and,hence, probability of occurrence by [31]. Ermini et al. (2005)and Catani et al. (2005) have used unique condition unitsfor the terrain unit’s definition in ANNs analysis [30], [31].Chen et al. (2009) used genetic algorithms and neural net-works for the interpretation of rainfall-induced landslides inTaiwan [32]. Lui et al. (2006) assessed the landslide hazardusing ANNs for a specific landslide typology (debris flow) [33],considering the following triggering factors: frequency offlooding, covariance of monthly precipitation, and days withrainfall higher than the critical threshold. Finally, Muthu andPetrou (2007) applied a rule-based expert system for landslideearly warning and alert maps using geological and meteoro-logical data in GIS [34], [35].

Previously, not much work has been done on landslidesusceptibility and hazard analysis in Malaysia. Pradhan and

Lee (2007) have performed landslide susceptibility and riskanalysis for the Penang Island using a frequency ratio andthe logistic regression model [36]. Recently, Pradhan and Lee(2010) have compared three landslide susceptibility maps gen-erated by frequency ratio, multivariate logistic regression, andneural network model for the Penang Island and Selangorarea in Malaysia [37]–[42]. In the last few years, landslidesusceptibility evaluation using GIS and soft computing tech-niques such as fuzzy logic, and artificial neural network modelshave been applied by researchers in different countries [43],[45]. Recently, Pradhan and Lee (2009) have used the ANNmodel with different training sites for landslide hazard and riskanalysis at the Penang Island, Malaysia [46].

The main difference between the present study and theapproaches described in the aforementioned publications is thata neuro-fuzzy model was developed and applied for the firsttime in landslide susceptibility analysis.

III. STUDY AREA

In this paper, the Cameron Highlands was selected forthe application of landslide susceptibility analysis due to thefrequent occurrence of landslides. The study area (Fig. 1) islocated in the Cameron Highlands, which is undergoing rapiddevelopment, with land clearing for housing estates, hotels, andapartments, resulting in erosion and landslides. The CameronHighlands is a district in Pahang, which is one of the 13 statesof Malaysia. The study area covers an area of 26.7 km2 and islocated near the northern central part of Peninsular Malaysia.

The district of the Cameron Highlands is located in theeastern flank of the main range, which is composed of granite.However, scattered outliers (roof pendants) of metasedimentsare also present (Fig. 1). The granite in the Cameron High-lands is classified as megacrystic biotite granite [47], [48].Cobbing et al. (1992) mentioned that some of the granite andthe associated microgranite may contain muscovite and may bemineralized [49]. The metasediments consist of schist, phyllite,slate, and limestone [50]. Minor intercalations of sandstoneand volcanics exist as well. The regional geology map of thestudy area and its surrounding areas are shown in Fig. 1. Post-Triassic–Mesozoic granite comprises most of the granite rocks,whereas there are a few patches of metamorphic rocks, mostlycomposed of Silurian–Ordovician schist, phyllite, limestone,and sandstone.

The annual rainfall of the Cameron Highlands, like alltropical hilly regions, is very high, averaging between 2500and 3000 mm per annum. There are two pronounced wetseasons from September to December and from February toMay in each year. The rainfall in the Cameron Highlandspeaks between March and May and also from November toDecember. The highest single-day rainfall that was recordedranged from 87 to 100 mm. It is during such times that manystreams and rivers in the Cameron Highlands may overflow,flooding the surrounding areas, and landslides such as debrisflow may occur along the river valleys. The intensity of the rainis another factor that affects the fill slopes, causing severe sheet,rill, and gully erosion. During such times, many of the naturaland man-made slopes are marginally stable. The hillslopes are


Fig. 1. Location and geological map of the study area and its surroundings (Source: Department of Minerals and Geoscience, Malaysia).

drained by numerous small streams, resulting in deeply incisederosion.

Several field observations have been carried out in the monthof April/June/September in 2002, 2007, and 2008 for takingground data and for the verification of the landslide locationsand types using GPS survey.

IV. DATA PREPARATION

A spatial database that is related to landslide susceptibil-ity modeling (Table I) has been acquired for the CameronHighlands. This database includes maps, images, tabular data,vectors, and rasters from different data custodians. These datawere in different formats, scales, accuracies, and geometriesand were almost unusable for direct application to this research.

For this reason, these data were converted into appropriate datainput for analysis and neuro-fuzzy modeling. Eight landslideconditioning factors were identified based on the previous paperpublished by Pradhan and Lee (2010). In the studied area,it is also observed that most of the landslides have takenplace immediately after a heavy rainfall [51]. Hence, rainfallis considered as the triggering factor. Landslide susceptibil-ity maps only include the preparatory or conditioning factorsbecause these maps show the landslide susceptibility zones.To obtain a landslide hazard map, the return periods of thetriggering factors such as heavy rainfall or earthquake shouldbe considered. For this reason, the precipitation map is notconsidered in this analysis. The GIS and remote-sensing datathat were used in the present study were shown in Table I.Remote-sensing methods were used to obtain the historical


TABLE ILIST AND SOURCE OF THE DATA USED IN THIS STUDY

TABLE IILIST OF SATELLITE DATA USED IN THIS STUDY

records of the landslides. Archived 1:10 000–1:50 000 aerialphotographs, SPOT 5 panchromatic satellite image, IKONOS,RADARSAT, and landslide reports over the past 21 years wereused for the visual detection of landslide occurrences in thestudy area. These aerial photographs were taken during theperiod of 1981–2002 and were acquired by the MalaysianCenter for Remote Sensing. The satellite data that were usedin this paper are listed in Table II. In addition, all historicallandslide reports, newspaper records, and archived data havebeen assembled for the period under examination. The sourcematerial varies in quality with respect to the precise locationof the landslide event. Based on the site description, archiveddatabase, and aerial photo interpretation, the locations of theindividual landslides were drawn on 1:25 000 maps, and the lo-cation was plotted as close as possible. Field observations wereused to confirm the fresh landslide locations (scars) and types.In the aerial photographs and high-resolution satellite images,historical landslides could be observed as breaks in the forestcanopy, bare soil, or geomorphological features, like head- andside scarps, flow tracks, and soil- and debris deposits below ascar. These landslides were then classified and sorted out basedon their modes of occurrence. The landslide inventory map wasvery helpful in understanding different triggering factors thatcontrol different slope movement types. Most of the landslidesare shallow rotational, and there are a few translational and flowtypes. However, during the analyses that were performed inthe present study, only the rotational failures are considered,and the other types of failures were eliminated because theoccurrence of the other types of failures is rare and ignorable.Also, a few landslides that occurred in slightly inclined areas

Fig. 2. Landslide inventory map compiled in this paper.

were not considered and thus were eliminated in the analysis.Consequently, the susceptibility maps that are produced in thispaper are valid for the shallow rotational failures. To assemblea database to assess the surface area and number of landslidesin the study area, a total of 70 shallow rotational failures weremapped in the study area having a 0.5091-km2 surface area.The landslide inventory map that was compiled in the presentstudy is shown in Fig. 2.

In order to develop a method for the assessment of landslidesusceptibility, determination of the conditioning factors for thelandslides is crucial [45]. In this paper, there were a total ofeight landslide conditioning factors that were considered in theanalyses performed. The basic landslide conditioning factorssuch as altitude, slope gradient, lithology, distance from thedrainage, distance from the faults, and plan curvature wereemployed. As a result of the field observations, it was observedthat the landslides have a close relation with the distancefrom the roads. For this reason, the distance to the road wasconsidered as a landslide conditioning factor for the study area,in addition to the basic landslide conditioning factors. The land-use and/or land-cover parameter has been used generally whenproducing landslide susceptibility maps. In this paper, the Nor-malized Difference Vegetation Index (NDVI) was consideredas a landslide conditioning parameter in order to characterizethe land-cover characteristics of the study area. Particularly,the land-cover characteristics have a crucial importance on theoccurrence of the shallow landslides. All of the factors that wereemployed in this paper were transformed into a vector-typespatial database using the GIS (Fig. 3). For the digital elevationmodel (DEM) creation, 10-m interval contours and survey basepoints showing the elevation values were extracted from the1:25 000-scale topographic maps. This is used to generate theDEM with 10-m pixel size and triangulated irregular networkfrom which the altitude, slope gradient, and plan curvatureare derived [Fig. 3(a)–(c)]. In the present study, a substantial


Fig. 3. Maps of the input data layers. (a) Altitude. (b) Slope gradient. (c) Plan curvature. (d) Distance from the drainage. (e) Lithology. (f) Distance fromthe faults.

attention has been given for slope conditions because thereis a physical relation between landslide occurrence and slopegradient. An increase in slope gradient results in an increase

of the driving forces. For this reason, slope configuration andsteepness play an important role on the susceptibility of aslope to landsliding. This makes slope as an important factor


Fig. 3. (Continued.) Maps of the input data layers. (g) Distance from the road. (h) NDVI.

in preparing the landslide susceptibility map. The slope mapwas reclassified into four classes, following the standard classi-fication scheme set by the Ministry of Science, Technology andEnvironment of Malaysia for hill land: 1) < 15◦; 2) 16◦–25◦;3) 26◦–35◦; and 4) > 35◦ [Fig. 3(b)]. Another reason for classi-fying slope is that this classification categorizes the slope anglesgreater than 35◦ in a single group, which is an advantage in thisparticular study area as the studied area has only 3.04% of thetotal area under the very high slope category. Further breakingup of the slope into smaller classes shows a very minimalarea or breaks in the subcategory. Similarly, the very gentleslope category also aggregates the area under the less slopeangle of 15◦. This forms the largest category, with an area of11.42 km2, i.e., 42.77% of the total area under study. Further-more, dividing the slope into smaller categories deprives thesignificance of the higher elevation on which the very gentleslope is situated.

The term curvature is generally defined as the curvature ofa line formed by the intersection of a random plane with theterrain surface [52]. The influence of plan curvature on the landdegradation processes is the convergence or divergence of waterduring downhill flow. In addition, this parameter constitutesone of the main factors controlling the geometry of the terrainsurface where landslides occur [53]. In the case of the curvature,negative curvatures represent a concave surface, zero curvaturerepresents a flat surface, and positive curvatures represent aconvex surface. The plan curvature map was prepared using theavenue routine in ArcView 3.2 [Fig. 3(c)].

Proximity to the drainage pattern was an important factor inthe evaluation of landslide susceptibility, as streams could ad-versely affect the stability by either eroding the toe or saturatingthe slope or both [14]. However, there is no a consensus onwhich buffer interval should be used in such kind of studies[13]. For this reason, various authors [10], [13], [14], [29],[45] have considered different approaches. The distance fromthe drainage was calculated using the topographic database.

The drainage buffer was calculated at 50-m intervals, as shownin Fig. 3(d). The inclusion of the drainage channels in thesusceptibility map is useful in delineating probable travel pathsdown the slope from susceptible initiation areas [54]. It can befound that, as the distance from the drainage lines increases,landslide frequency generally decreases. This can be attributedto the fact that terrain modification caused by gully erosionmay influence the initiation of landslides. There are manysecond-order streams in the studied area, and they dischargedirectly to the main river, which is in the north eastern partof the Cameron Highlands (not part of the studied area). Thebuffer settings (50 m) are assigned depending on the distancebetween the crest of the landslides related with the drainagenetwork [45].

Lithology and structure of the area play a major role indetermining the sites of failure. Secondary weaknesses thatare observed in these rocks make these more susceptible tosliding because of material weakening, stress accumulation, ortectonic activity in different distances [45]. Lithology is oneof the most important factors controlling landslides [55]. Thelithology map was prepared from a 1:63 300-scale geologicalmap [Fig. 3(e)], and the distance from the fault was calcu-lated based on the Euclidean distance method in ArcGIS 9.0[Fig. 3(f)]. Lineaments and faults are zones of weakness,which are prone to instability. Faults form a line or zone ofweakness characterized by fractured rocks. Proximity (buffers)to these structures increases the likelihood of the occurrenceof landslides. Selective erosion and movement of water alongfault lines/planes promote such phenomena. In the study area,the major fault lines are in the NE–SW, NW–SE, and E–Wdirections, although a slight relationship between the landslidedistribution and fault location is observed [Fig. 3(f)].

Roadcuts are usually sites of anthropological instability. Agiven road segment may act as a barrier, a net source, a netsink, or a corridor for water flow, depending on its location inthe area [38]. It usually serves as a source of landslides. The


road map is derived from the topography map. From the fieldobservation, it has been noticed that most of the landslides haveoccurred along the cut-slopes and roads. The distance to theroad buffer is selected based on the occurrence of landslidesto the proximity of the road. Therefore, a 50-m buffer zone isselected in the studied area [Fig. 3(g)]. The shallow landslidesare affected by the land-cover characteristics. However, there isno detailed and reliable vegetation map of the study area. Forthis reason, the NDVI map was generated from the SPOT 5scene of the January 2005 (2.5-m spatial resolution) satelliteimage [Fig. 3(h)] to characterize the vegetation characteristicsof the study area. The NDVI value was calculated using theformula NDV I = (IR − R)/(IR + R), where IR is the en-ergy reflected in the infrared portion of the electromagneticspectrum and R is the energy reflected in the red portion of theelectromagnetic spectrum. The NDVI is useful in delineatingvegetation.

Fig. 4 shows the flowchart for the landslide susceptibilityanalysis and spatial data flow diagram. All of the landslideconditioning factors were converted to a raster grid with 10 m ×10 m cells with 440 rows by 607 columns for the application ofthe neuro-fuzzy model. GIS ArcGIS 9.0 version software pack-age was used as the basic analysis tools for spatial managementand data manipulation.

V. LANDSLIDE SUSCEPTIBILITY MAPPING

USING THE NEURO-FUZZY MODEL

Although some soft computing techniques such as fuzzymodeling and artificial neural networks have been used toproduce landslide susceptibility maps, the neuro-fuzzy mod-eling, which is one of the soft computing techniques, hasnot previously been employed for such purpose. A neuro-fuzzy system is, in fact, a neural network that is functionallyequivalent to the fuzzy inference model. It can be trained todevelop IF-THEN fuzzy rules and to determine membershipfunctions for input and output variables of the system [56]. Oneof the neuro-fuzzy inference systems is ANFIS. The Sugenomodel was proposed for a systematic approach in generatingfuzzy rules from a given input–output data set [56]. An ANFISmodel uses a hybrid learning algorithm that combines theleast squares estimator and the gradient descent method. In theANFIS training algorithm, each epoch is composed of forwardand backward passes. In the forward pass, a training set ofinput patterns is presented to the ANFIS, neuron outputs arecalculated on the layer-by-layer basis, and rule consequentparameters are identified by the least squares estimator. Inthe Sugeno style fuzzy inference, an output y is a linearfunction [56].

In the present study, a total of five ANFIS models wereconstructed to determine the landslide susceptibility degreesof each pixel, and the flowchart of the applied methodology isshown in Fig. 4. As shown in Fig. 4, the input data files (altitude,slope gradient, lithology, distance from the drainage, distancefrom the road, distance from the faults, plan curvature, andNDVI) and the landslide inventory data were extracted fromArcGIS separately. In the second stage, these separate data fileswere combined, as required by the MATLAB Software for each

Fig. 4. Flowchart showing the landslide susceptibility analyses and spatialdata flow.

models. The input parameters of each ANFIS model were givenin Table III. Considering these inputs, for each ANFIS model,the data matrices were prepared automatically by using a com-puter program developed by ANSI C programming language.In this stage, two types of data matrices for each model wereprepared. One of them includes all of the inputs and outputs,while the other contains only all of the inputs. The data matrixthat includes the output was used for the selection of the traindata set, while the data matrix that includes only the inputswas used for model simulation after model train. As shownin Fig. 5, showing the membership function of model 5 aftertraining, the number of membership functions of each inputwas determined. When determining the number of membershipfunctions, two basic assumptions were considered. The numberof membership functions should represent physically each input


TABLE IIIINPUT PARAMETERS AND THE NUMBER OF MEMBERSHIP

FUNCTIONS USED IN THE ANFIS MODELS

and minimum. If the number of membership functions is high,the number of “if-then” rules would be very high, and hence,the overlearning may exist. Moreover, the model would loss itscalculation efficiency. Another important issue in the training ofANFIS is how to preserve the human-plausible features such asbell-shaped membership functions, completeness or sufficientoverlapping between adjacent membership functions, minimaluncertainty, etc. [57]–[59]. As shown in Fig. 5, a minimumof 50% overlapping was employed to minimize uncertainty.Finally, lithology, distance from the drainage, distance fromthe faults, and plan curvature have two bell-shaped membershipfunctions, while altitude, slope gradient, distance from the road,and NDVI contain three bell-shaped membership functions.Depending on the number of membership functions of eachinput, the numbers of the “if-then” rules vary between 54(model 1) and 1296 (model 5). The general structure of ANFISmodel 1 was shown in Fig. 6. The study area is formed by267 074 pixels, while a total of 5091 pixels include landslide.A total of 2500 pixels were selected from the pixels thatinclude landslides, while 2500 pixels were selected from thepixels that are free from landslides. Consequently, a total of5000 pixels were selected randomly as the train data set.Each train data set includes the pixels having landslides andthe pixels without landslides equally. After constructing theSugeno fuzzy inference models, each model was trained byconsidering 350 epochs. After the training stage, the seconddata matrices of each model were employed for simulation

Fig. 5. Membership functions of the inputs.

by using a MATLAB script written in this paper. In thisstage, a whole data set was used to simulate the model. Theoutput data that were obtained from MATLAB were converted


Fig. 6. General structure of ANFIS model 1 as an example.

to ArcGIS data format automatically using another computerprogram developed by ANSI C programming language. Finally,the landslide susceptibility maps were produced by employingArcGIS [Fig. 7(a)–(e)].

VI. VALIDATION OF THE SUSCEPTIBILITY MAP

A. ROC Plot

For the effective comparison of the ANFIS-derived landslidesusceptibility maps, all of the five models were evaluated bycomparing them separately with the landslide testing data.

Therefore, five landslide susceptibility maps (ANFIS-derivedfive models) were validated against the existing landslide dataset, and they are shown in Fig. 8. Spatial effectiveness of thesefive susceptibility maps was checked by receiver operatingcharacteristics (ROC). The ROC curve is a useful methodof representing the quality of deterministic and probabilisticdetection and forecast systems [60]. The area under the ROCcurve (AUC) characterizes the quality of a forecast systemby describing the system’s ability to anticipate correctly theoccurrence or nonoccurrence of predefined “events” [57]. TheROC curve plots the false positive rate on the X-axis and


Fig. 7. Landslide susceptibility maps obtained from the ANFIS modeling. (a) Model 1. (b) Model 2. (c) Model 3. (d) Model 4. (e) Model 5.

1—the false negative rate on the Y -axis. It shows the tradeoffbetween the two rates [61]. To obtain the relative ranks for eachprediction pattern, the calculated index values of all pixels in

the study area were sorted in descending order. If the AUC isclose to one, the result of the test is excellent. On the contrary,the closer the AUC is to 0.5, the fairer is the result of the test.


Fig. 8. ROC plots for the susceptibility maps produced in this paper.

The results of the ROC curve test are shown in Fig. 8. Thesecurves indicate that model 5 (Fig. 8) has a higher predictionperformance than the other models. The ROC plot assessmentresults show that, in the susceptibility map using model 5, theAUC was 0.9765 and that the prediction accuracy was 98%.In the susceptibility map using model 1, the AUC was 0.8208,and the prediction accuracy was 82%. The susceptibility mapsproduced using models 2 and 3 show 86.68% and 86.71%,respectively. This indicates that the inclusion of the conditionfactor “distance from the fault” in model 3 did not increasethe prediction accuracy of model 2. Similarly, in the case ofmodel 4, the susceptibility map shows a higher predictionaccuracy (94%) than models 3, 2, and 1. This shows that thecase of models 4–5 is higher than the case of models 1–3.Therefore, from the prediction accuracy graphs (Fig. 8), it isquite evident that the susceptibility map with models 4 and5 shows a higher prediction accuracy of >93%, whereas thecase of model 1 shows a least prediction accuracy of 82%. Itis also important to note that the inclusion of the conditioningfactor “NDVI” increased the prediction accuracy of model 5by 4% (97%–93%). Therefore, one can conclude here that theselection of parameters has an impact on the accuracy of thelandslide susceptibility analysis, as the results with altitude,slope gradient, plan curvature, distance from the drainage, dis-tance from the road, lithology, plan curvature, and NDVI werethe most suitable parameters for landslide susceptibility of thestudy area. Additionally, it is possible to say that distance fromthe fault has an ignorable influence on landslide occurrence inthe study area, although it has been used commonly for thispurpose.

B. Comparison of LSI Classification With the Landslide Data

The landslide susceptibility values produced from ANFISmodel are shown in Fig. 9. The landslide susceptibility mapsthat were produced due to the application of neuro-fuzzy modelwere verified and compared with the aid of existing landslidetest data set. The range of landslide susceptibility mapping(LSI) values is different in five susceptibility models consid-ered in this paper, so the index value needs to set the same

Fig. 9. Frequency ratio plots of various landslide zones derived from theANFIS models.

interval for the comparison. The indexes were classified intofive classes (highest 10%, second 10%, third 20%, fourth 20%,and remaining 40%) based on the area for visual and easyinterpretation. Relative frequency ratio analysis was performedon the classification results and landslide location data [38]. Thelandslide test locations were overlaid on the LSI classificationobtained from the five models ran in ANFIS. In the overlaid im-ages, the boundary of each existing landslide location enclosesa number of pixels allocated to ANFIS-derived landslide sus-ceptibility zones. The total areas and their percentages coveredby various landslide susceptibility zones within the boundaryof each existing test landslides were determined. Similarly, thetotal area, with their percentages occupied by various landslidesusceptibility zones, in the whole image was also determined.Relative frequencies of areas affected by different landslidesusceptibility zones were calculated from the ratio. Ideally,the frequency ratio value should increase from a very lowsusceptible zone to a very high susceptible zone, since thehighest landslide susceptible zones are generally more prone tolandslides than other zones. In plotting these frequency ratiovalues (Fig. 9), it can be seen that there is a gradual andsmooth increase in the frequency from the no susceptible zoneto the very high susceptible zone in five susceptibility mapscomputed by ANFIS models, except model 1. As expected, thecase of models 2 and 3 gives a similar pattern, whereas thecase of model 1 gives an unusual pattern of the susceptibilityzones. From the frequency ratio plots, it can be clearly seenthat model 5 gives the better separation of zones than model4. From this observation, we concluded that model 5 performscomparatively better than the other models and that a numberof locations having high output values have been mapped moreaccurately in the very high susceptibility zones. In addition,model 5 distinguished the LSIs more widely than model 4as expected since the ROC plot shows a higher area undercurve than the later by 4%. The higher prediction accuracy invery high susceptibility zones in model 5 is attributed to thecontribution of NDVI, which plays an important role during thetraining of the ANFIS. This means that the landslide probabilityincreases with the presence of all of the eight landside condi-tioning factors in model 5.

Therefore, considering the correct classifications as shownin Fig. 9, the prediction capacity of model 5 is slightly higherthan that of the other models. Taking all of the analyses intoconsideration, it may be concluded that the most appropriate


landslide susceptibility map for the study area is obtained byusing model 5 [Fig. 7(e)], which has been obtained through theuse of eight parameters such as the following: altitude, slopegradient, lithology, distance to the drainage, distance from theroad, distance from the faults, plan curvature, and NDVI.

VII. CONCLUSION AND DISCUSSION

In the present study, a regional landslide susceptibility as-sessment was performed in a landslide-prone area (a part ofCameron Highlands, Malaysia). The study included three mainstages such as landslide inventory, susceptibility analyses, andverification. The study area has a frequent landslide problemdue to its mountainous character and heavy rainfall. In the firststage of the study, an extensive landslide inventory mappingstudy was performed because a reliable landslide inventory mapis necessary for all indirect landslide susceptibility evaluations.For this purpose, a landslide inventory database that is used toassess the landslide susceptibility of the study area, with a totalof 70 landslides, was mapped in the study area. In the secondstage, the landslide conditioning factors were determined, andthese were prepared for the landslide susceptibility assessment.During the analysis stage, a total of five ANFIS models wereconstructed. Model 1 included only the altitude, slope gradient,lithology, and distance from the road as the inputs, whilemodel 5 was formed by employing the altitude, slope gradient,plan curvature, distance from the drainage, distance from theroad, lithology, distance from the fault, and NDVI as the inputs.Since this study is a methodological attempt, the study areais selected as small as possible. In the selected study area,only two lithologies are cropped out without a simplification.Additionally, all faults, drainage lines, and roads in the studyarea are considered completely. It is possible to increase thenumber of membership functions, but this can result in over-learning of the models developed. To cope with this problem,the number of membership functions is considered optimally.In fact, this is not a simplification. Due to the nature of themembership functions, all intervals are considered completely.When considering the AUC values, it is evident that all ANFISmodels exhibited high performances because the AUC valuesof the models varied between 0.8208 (model 1) and 0.9765(model 5). This was mainly sourced from the very high predic-tion capacity of the ANFIS modeling. Also, the effects of theparameters used in the models were observed clearly. However,when assessing the performances of the landslide susceptibilitymaps, considering only the AUC values may not be sufficientbecause highly conservative maps may give some very highAUC values. For this reason, in addition to the AUC values,a relative frequency ratio analysis was applied on the classi-fication results of the landslide susceptibility maps producedin the present study and landslide location data. The resultsof the frequency ratio analyses showed that there is a gradualand smooth increase in the frequency from the no susceptiblezone to the very high susceptible zone in five susceptibilitymaps obtained from the ANFIS models constructed in thestudy. As expected, models 2 and 3 exhibited almost similarpattern, while model 5 yielded the higher separation of thesusceptibility zones. When considering this observation, it can

be concluded that model 5 performed comparatively better thanthe other models and that a number of locations having highoutput values have been mapped more accurately in the veryhigh susceptibility zones. The higher prediction accuracy invery high susceptibility zones in model 5 is attributed to thecontribution of all of the eight factors playing important rolesduring the training of the ANFIS. In other words, the landslideprobability varies depending on the presence of the vegetationindex value (NDVI) and shape of the slopes characterized byplan curvature in model 5.

As a final conclusion, the results obtained from the studyshowed that the ANFIS modeling is a very useful and powerfultool for the regional landslide susceptibility assessments. Toprevent overlearning, the number of membership functions ofthe inputs and the number of training epochs should be selectedoptimally and carefully. Also, when selecting the number ofthe membership functions of the inputs, the physical meaningsof the inputs should be considered by an expert. Therefore,the results that are to be obtained from the ANFIS modelingshould be assessed carefully because the overlearning maycause misleading results. As a final recommendation, the resultsobtained in this paper showed that the methods followed in thepresent study exhibits a high performance. However, it is notforgotten that the performance of such type maps depends notonly on the methodology followed but also on the quality ofthe available data. For this reason, if the quality of the dataincreases, the performance of the maps produced by the neuro-fuzzy approach could increase.

ACKNOWLEDGMENT

The authors would like to thank the Malaysian Center forRemote Sensing and the National Mapping Agency, Malaysia,for providing the various data sets used in this paper, the twoanonymous reviewers for their very helpful reviews, and C. Ruffor the editorial comments. B. Pradhan would like to thankthe Alexander von Humboldt Foundation (AvH), Germany,for awarding the Visiting Scientist position at the DresdenUniversity of Technology, Dresden, Germany.

REFERENCES

[1] D. J. Varnes, “IAEG Commission on Landslides: Landslide hazardzonation—A review of principles and practice,” UNESCO, Paris, France,1984.

[2] F. Guzzetti, A. Carrarra, M. Cardinali, and P. Reichenbach, “Landslidehazard evaluation: A review of current techniques and their applicationin a multi-scale study, Central Italy,” Geomorphology, vol. 31, no. 1–4,pp. 181–216, Dec. 1999.

[3] S. Lee and M. J. Lee, “Detecting landslide location using KOMPSAT 1and its application to landslide-susceptibility mapping at the Gangneungarea, Korea,” Adv. Space Res., vol. 38, no. 10, pp. 2261–2271, 2006.

[4] B. Pradhan, R. P. Singh, and M. F. Buchroithner, “Estimation of stressand its use in evaluation of landslide prone regions using remote sensingdata,” Adv. Space Res., vol. 37, no. 4, pp. 698–709, 2006.

[5] S. Lee and B. Pradhan, “Probabilistic landslide risk mapping at PenangIsland, Malaysia,” J. Earth Syst. Sci., vol. 115, no. 6, pp. 661–672,Dec. 2006.

[6] S. Lee and B. Pradhan, “Landslide hazard mapping at Selangor, Malaysiausing frequency ratio and logistic regression models,” Landslides, vol. 4,no. 1, pp. 33–41, Mar. 2007.

[7] A. Akgün, S. Dag, and F. Bulut, “Landslide susceptibility mapping fora landslide-prone area (Findikli, NE of Turkey) by likelihood-frequencyratio and weighted linear combination models,” Environ. Geol., vol. 54,no. 6, pp. 1127–1143, May 2008.


[8] M. C. Tunusluoglu, C. Gokceoglu, H. A. Nefeslioglu, and H. Sonmez,“Extraction of potential debris source areas by logistic regression tech-nique: A case study from Barla, Besparmak and Kapi mountains (NWTaurids, Turkey),” Environ. Geol., vol. 54, no. 1, pp. 9–22, Mar. 2008.

[9] T. Can, H. A. Nefeslioglu, C. Gokceoglu, H. Sonmez, and T. Y.Duman, “Susceptibility assessments of shallow earthflows triggered byheavy rainfall at three subcatchments by logistic regression analyses,”Geomorphology, vol. 72, no. 1–4, pp. 250–271, 2005.

[10] S. Lee, “Application of logistic regression model and its validation forlandslide susceptibility mapping using GIS and remote sensing data,”Int. J. Remote Sens., vol. 26, no. 7, pp. 1477–1491, Apr. 2005.

[11] B. Pradhan, S. Lee, S. Mansor, M. F. Buchroithner, N. Jallaluddin, andZ. Khujaimah, “Utilization of optical remote sensing data and geographicinformation system tools for regional landslide hazard analysis by usingbinomial logistic regression model,” Appl. Remote Sens., vol. 2, no. 1,pp. 1–11, Oct. 2008.

[12] T. Gorum, B. Gonencgil, C. Gokceoglu, and H. A. Nefeslioglu, “Imple-mentation of reconstructed geomorphologic units in landslide suscepti-bility mapping: The Melen Gorge (NW Turkey),” Nat. Hazards, vol. 46,no. 3, pp. 323–351, Sep. 2008.

[13] H. A. Nefeslioglu, C. Gokceoglu, and H. Sonmez, “An assessment onthe use of logistic regression and artificial neural networks with differentsampling strategies for the preparation of landslide susceptibility maps,”Eng. Geol., vol. 97, no. 3/4, pp. 171–191, Apr. 2008.

[14] C. Gokceoglu and H. Aksoy, “Landslide susceptibility mapping of theslopes in the residual soils of the Mengen region (Turkey) by deterministicstability analyses and image processing techniques,” Eng. Geol., vol. 44,no. 1–4, pp. 147–161, Oct. 1996.

[15] C. Gokceoglu, H. Sonmez, and M. Ercanoglu, “Discontinuity controlledprobabilistic slope failure risk maps of the Altindag (settlement) region inTurkey,” Eng. Geol., vol. 55, no. 4, pp. 277–296, Mar. 2000.

[16] G. Zhou, T. Esaki, Y. Mitani, M. Xie, and J. Mori, “Spatial probabilisticmodeling of slope failure using an integrated GIS Monte Carlo simulationapproach,” Eng. Geol., vol. 68, no. 3/4, pp. 373–386, Mar. 2003.

[17] A. M. Youssef, B. Pradhan, A. F. D. Gaber, and M. F. Buchroithner,“Geomorphological hazard analysis along the Egyptian red sea coastbetween Safaga and Quseir,” Nat. Hazards Earth Syst. Sci., vol. 9, no. 3,pp. 751–766, 2009.

[18] C. Tarantino, P. Blonda, and G. Pasquariello, “Application of changedetection techniques for monitoring man-induced landslide causalfactors,” in Proc. IEEE Int. Geosci. Remote Sens. Symp., 2004, vol. 2,pp. 1103–1106.

[19] Y. L. Chang, L. S. Liang, C. C. Han, J. P. Phang, W. I. Liang, andK. S. Chen, “Multisource data fusion for landslide classification usinggeneralized positive Boolean functions,” IEEE Trans. Geosci. RemoteSens., vol. 45, no. 6, pp. 1697–1708, Jun. 2007.

[20] M. Ercanoglu and C. Gokceoglu, “Assessment of landslide susceptibil-ity for a landslide-prone area (North of Yenice, NW Turkey) by fuzzyapproach,” Environ. Geol., vol. 41, pp. 720–730, 2002.

[21] S. Lee, J. H. Ryu, K. Min, and J. S. Won, “Landslide susceptibilityanalysis using GIS and artificial neural network,” Earth Surface Process.Landforms, vol. 27, pp. 1361–1376, 2003.

[22] S. Lee, J. H. Ryu, J. S. Won, and H. J. Park, “Determination and applica-tion of the weights for landslide susceptibility mapping using an artificialneural network,” Eng. Geol., vol. 71, no. 3/4, pp. 289–302, Feb. 2004.

[23] S. Lee, “Application and verification of fuzzy algebraic operators to land-slide susceptibility mapping,” Environ. Geol., vol. 52, no. 4, pp. 615–623,Apr. 2007.

[24] B. Pradhan, S. Lee, and M. F. Buchroithner, “Use of geospatial data for thedevelopment of fuzzy algebraic operators to landslide hazard mapping: Acase study in Malaysia,” Appl. Geomat., vol. 1, pp. 3–15, 2009.

[25] Y. Hong, R. F. Adler, and G. Huffman, “An experimental global predictionsystem for rainfall-triggered landslides using satellite remote sensing andgeospatial datasets,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 6,pp. 1671–1680, Jun. 2007.

[26] S. Lee, J. H. Ryu, and I. S. Kim, “Landslide susceptibility analysis and itsverification using likelihood ratio, logistic regression, and artificial neuralnetwork models: Case study of Youngin, Korea,” Landslides, vol. 4, no. 4,pp. 327–338, Dec. 2007.

[27] J. Choi, H. J. Oh, J. S. Won, and S. Lee, “Validation of an artificial neuralnetwork model for landslide susceptibility mapping,” Environ. Earth Sci.(Online First), vol. 60, no. 3, pp. 473–483, Apr. 2010.

[28] C. Melchiorre, M. Matteucci, A. Azzoni, and A. Zanchi, “Artificialneural networks and cluster analysis in landslide susceptibility zonation,”Geomorphology, vol. 94, no. 3/4, pp. 379–400, Feb. 2008.

[29] D. P. Kanungo, M. K. Arora, S. Sarkar, and R. P. Gupta, “A com-parative study of conventional, ANN black box, fuzzy and combined

neural and fuzzy weighting procedures for landslide susceptibility zona-tion in Darjeeling Himalayas,” Eng. Geol., vol. 85, no. 3/4, pp. 347–366,Jun. 2006.

[30] L. Ermini, F. Catani, and N. Casagli, “Artificial neural networks appliedto landslide susceptibility assessment,” Geomorphology, vol. 66, no. 1–4,pp. 327–343, Mar. 2005.

[31] F. Catani, N. Casagli, L. Ermini, G. Righini, and G. Menduni, “Landslidehazard and risk mapping at catchment scale in the Arno River basin,”Landslides, vol. 2, no. 4, pp. 329–342, Dec. 2005.

[32] Y. R. Chen, J. W. Chen, S. C. Hsieh, and P. N. Ni, “The application ofremote sensing technology to the interpretation of land use for rainfall-induced landslides based on genetic algorithms and artificial neural net-works,” IEEE J. Sel. Topics Appl. Earth Obs. Remote Sens., vol. 2, no. 2,pp. 87–95, Jun. 2009.

[33] Y. Lui, H. C. Guo, R. Zou, and L. J. Wang, “Neural network modelling forregional hazard assessment of debris flow in Lake Qionghai Watershed,China,” Environ. Geol., vol. 49, no. 7, pp. 968–976, Apr. 2006.

[34] K. Muthu and M. Petrou, “Landslide-hazard mapping using an expertsystem and a GIS,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 2,pp. 522–531, Feb. 2007.

[35] K. Muthu, M. Petrou, C. Tarantino, and P. Blonda, “Landslide possibilitymapping using fuzzy approaches,” IEEE Trans. Geosci. Remote Sens.,vol. 46, no. 4, pp. 1253–1265, Apr. 2008.

[36] B. Pradhan and S. Lee, “Utilization of optical remote sensing data andGIS tools for regional landslide hazard analysis by using an artificialneural network model,” Earth Sci. Frontier, vol. 14, no. 6, pp. 143–152,Nov. 2007.

[37] B. Pradhan and S. Lee, “Delineation of landslide hazard areas on PenangIsland, Malaysia, by using frequency ratio, logistic regression, and arti-ficial neural network models,” Environ. Earth Sci., vol. 60, no. 5,pp. 1037–1054, May 2010.

[38] B. Pradhan and S. Lee, “Landslide susceptibility assessment and factoreffect analysis: Backpropagation artificial neural networks and their com-parison with frequency ratio and bivariate logistic regression modelling,”Environ. Model. Softw., vol. 25, no. 6, pp. 747–759, Jun. 2010.

[39] B. Pradhan, “Remote sensing and GIS-based landslide hazard analysisand cross-validation using multivariate logistic regression model on threetest areas in Malaysia,” Adv. Space Res., vol. 45, no. 10, pp. 1244–1256,May 2010.

[40] B. Pradhan, S. Lee, and M. F. Buchroithner, “A GIS-based back-propagation neural network model and its cross-application and valida-tion for landslide susceptibility analyses,” Comput. Environ. Urban Syst.,vol. 34, no. 3, pp. 216–235, May 2010.

[41] B. Pradhan, S. Lee, and M. F. Buchroithner, “Remote sensing and GIS-based landslide susceptibility analysis and its cross-validation in three testareas using a frequency ratio model,” Photogrammetrie, Fernerkundung,GeoInformation, vol. 1, pp. 17–32, 2010.

[42] B. Pradhan and A. M. Youssef, “Manifestation of remote sensing dataand GIS for landslide hazard analysis using spatial-based statisticalmodels,” Arab. J. Geosci., 2009, DOI: 10.1007/s12517-009-0089-2, to bepublished.

[43] B. Pradhan, “Manifestation of an advanced fuzzy logic model coupledwith geo-information techniques to landslide susceptibility mapping andtheir comparison with logistic regression modelling,” Environ. Ecol. Stat.,2010, DOI: 10.1007/s10651-010-0147-7, to be published.

[44] B. Pradhan, “Landslide susceptibility mapping of a catchment area us-ing frequency ratio, fuzzy logic and multivariate logistic regression ap-proaches,” J. Indian Soc. Remote Sens., vol. 39, pp. 47–66, 2010.

[45] M. Ercanoglu and C. Gokceoglu, “Use of fuzzy relations to producelandslide susceptibility map of a landslide prone area (West Black SeaRegion, Turkey),” Eng. Geol., vol. 75, no. 3/4, pp. 229–250, Nov. 2004.

[46] B. Pradhan and S. Lee, “Landslide risk analysis using artificial neuralnetwork model focusing on different training sites,” Int. J. Phys. Sci.,vol. 3, no. 11, pp. 1–15, 2009.

[47] F. Bignel and G. Snelling, “The geochronology of the main rangeBatholith: Cameron Highlands road and Gunong Bujang Melaka,”Overseas Geol. Miner. Resour., vol. 47, pp. 3–35, 1977.

[48] R. Krahenbuhl, “Magmatism, tin mineralization and tectonic developmentof Main Range, Malaysian Peninsula: Consequences for plate tectonicmodel of Southeast Asia based on Rb-Sr, K-Ar and fission track analysis,”Geol. Soc. Malaysia Bull., vol. 29, pp. 12–28, 1991.

[49] E. J. Cobbing, P. Pitfield, D. P. E. Darbyshire, and D. I. J. Mallick, “Thegranites of Southeast Asian tin belt,” Brit. Geol. Survey Oversea Memoir,vol. 10, pp. 1–53, 1992.

[50] W. S. Chow, M. Zakaria, A. Ferdaus, and A. Nurzaidi, “Geologicalterrain mapping. JMG unpublished report,” JMG.SWP.GS, vol. 16,pp. 1–42, 2003.


[51] B. Pradhan and S. Lee, “Regional landslide susceptibility analysisusing backpropagation neural network model at Cameron Highlands,Malaysia,” Landslides, vol. 7, no. 1, pp. 13–30, Mar. 2010.

[52] J. P. Wilson and J. C. Gallant, “Digital terrain analysis,” Soil Sci., vol. 121,pp. 234–241, 2000, pp. 1–27 in eastern Kentucky.

[53] I. S. Evans, “What do terrain statistics really mean?” in LandformMonitoring, Modelling and Analysis, S. Lane, K. Richards, andJ. Chandler, Eds. London, U.K.: Wiley, 1998, pp. 119–138.

[54] J. A. Coe, J. W. Godt, R. L. Baum, R. C. Bucknam, and J. A. Michael,“Landslide susceptibility from topography in Guatemala,” in Landslides:Evaluation and Stabilization, W. A. Lacerda, M. Ehrlich, S. A. B. Fontura,and A. S. F. Sayão, Eds. New York: Taylor & Francis, 2004, pp. 69–78.

[55] N. Nikolaou, “Rain and landslide manifestation correlation in Korinthoscountry, Greece,” in Proc. Int. Symp. Eng. Geol. Environ., Athens, Greece,1997, vol. 1, pp. 919–924.

[56] H. X. Lan, C. H. Zhou, L. J. Wang, H. Y. Zhang, and R. H. Lic, “Land-slide hazard spatial analysis and prediction using GIS in the Xiaojiangwatershed, Yunnan, China,” Eng. Geol., vol. 76, no. 1/2, pp. 109–128,Dec. 2004.

[57] M. Negnevitsky, Artificial Intelligence: A Guide to Intelligent Systems.Harlow, U.K.: Addison-Wesley, 2002.

[58] J. S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference sys-tems,” IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. 665–685,May/Jun. 1993.

[59] C. C. Lee, “Fuzzy logic in control systems: Fuzzy logic controller—Part I,” IEEE Trans. Syst., Man, Cybern., vol. 20, no. 2, pp. 404–418,Mar./Apr. 1990.

[60] J. A. Swets, “Measuring the accuracy of diagnostic systems,” Science,vol. 240, no. 4857, pp. 1285–1293, Jun. 1988.

[61] B. Biggerstaff, “Comparing diagnostic tests: A simple graphic usinglikelihood ratios,” Stat. Med., vol. 19, no. 5, pp. 649–663, Mar. 2000.

Biswajeet Pradhan received the B.Sc. degree ingeology (with honors) from Berhampur University,Berhampur, India, in 1995, the M.Sc. degree inapplied geology from the Indian Institute of Tech-nology (IIT), Bombay, India, in 1998, the M.Tech.degree in civil engineering from the IIT, Kanpur,India, in 2000, and the Ph.D. degree in geographicinformation system (GIS) and geomatics engineer-ing from the University Putra Malaysia, Serdang,Malaysia, in 2005.

He has been an Alexander von Humboldt Re-search Fellow with the Dresden University of Technology, Dresden, Germany,since August 2008. He specializes in remote sensing, GIS application, and softcomputing techniques in natural hazard and environmental problems. He hasmore than ten years of teaching, research, and industrial experience. He haspublished over 50 research articles in referred technical journals and books.

Dr. Pradhan was the recipient of the German Deutscher AkademischerAustausch Dienst (DAAD) and was a Saxony Scholarship Holder from 1999to 2002.

Ebru Akcapinar Sezer was born in Ankara,Turkey, on December 25, 1974. She received theB.Sc., M.Sc., and Ph.D. degrees from the Depart-ment of Computer Engineering, Hacettepe Univer-sity, Ankara, Turkey, in 1996, 1999, and 2006,respectively.

She is currently an Associate Lecturer with theDepartment of Computer Engineering, HacettepeUniversity. Her areas of interest are semantic Webtechnologies and fuzzy system applications.

Candan Gokceoglu was born in Ardahan, Turkey,on July 25, 1966. He received the B.Sc. degreefrom the Department of Hydrogeological Engineer-ing and the M.Sc. and Ph.D. degrees from theDepartment of Geological Engineering, HacettepeUniversity, Ankara, Turkey, in 1989, 1993, and 1997,respectively.

He is currently a Professor with the Applied Ge-ology Division, Department of Geological Engineer-ing, Hacettepe University. He has published over60 research articles in referred scientific journals. His

areas of interest are landslides, rock mechanics, and fuzzy system applications.

Manfred F. Buchroithner received the degree ingeology and paleontology from the University ofGraz, Graz, Austria, the degree in cartography andremote sensing from the International Institute forGeo-Information Science and Earth Observation,Enschede, The Netherlands, and the Ph.D. degreefrom the University of Graz in 1977.

He is currently a Full Professor of cartography andthe Director of the Institute for Cartography, DresdenUniversity of Technology, Dresden, Germany. Out ofhis more than 290 articles, more than 70 have been

published in reviewed journals. He has written three books and has editedthree volumes on remote sensing. His major research interests cover geohazardproblems, true-3-D geodata visualization, and high-mountain cartography.

landslide susceptibility mapping by neuro-fuzzy approach in a

Documents