1 Introduction

Worldwide, several geographic information system (GIS)-based applications have been used to map and assess landslide susceptibility heuristically, statistically, or deterministically at local or regional scale (Weirich and Blesius 2007; Hervás and Bobrowsky 2009; Regmi et al. 2010). To accomplish the landslide susceptibility modeling and to highlight advantages and limitations of models, the most important task is to compare the results with a detailed and accurate landslide inventory map (Hammond et al. 1992; Pack et al. 1998; Lee 2005; Hervás and Bobrowsky 2009). In Mexico, despite efforts, there are few detailed landslide inventory maps and landslide susceptibility maps that can be used systematically to compare and contrast the models within any region (Capra and Lugo-Hubp 2006; García-Palomo et al. 2006; Pérez-Gutiérrez 2007; Secretaría de Protección Civil 2010). The lack of systematic comparison of landslide susceptibility models compromises the reliability of the models and can lead to their abuse. To address the above deficiency, three landslide models are evaluated in their use in prediction of landslides: Stability Index MAPping (SINMAP), multiple logistic regression (MLR), and multicriteria evaluation (MCE). Although it is never possible to prove the “validity” of an environmental model (Zaitchik et al. 2003), this paper introduces basic strategies to compare and contrast the models. The validation of landslide susceptibility maps and model performance is conducted by comparing with the inventory map under the system LOGISNET (Legorreta Paulin and Bursik 2008). LOGISNET provides tools to compare the predicted susceptibility map with an inventory map by using a histogram and contingency table. The system generates a histogram to show how the predicted map values are distributed among the inventory maps and a contingency table to show how well a model classifies areas into categories compared to the inventory map. Accuracy of the landslide susceptibility maps was also evaluated through the area under the ROC (receiver operating characteristic) curve (AUC) with SPSS Statistics Software.

The present research is based on studies of the stream system of the Río El Estado watershed on the southwestern flank of Pico de Orizaba volcano, Mexico, to accomplish the comparison between models. The study area has physiographic conditions that are prone to landslides: high rainfall during the wet season, susceptible rock types, a high degree of weathering, and steep slopes (Rodríguez et al. 2011; Legorreta Paulín et al. 2013). In the study area, small landslides and debris flows that exhibit a range of volumes of at least two orders of magnitude between 101 and 102 m3 occur continually along the watershed. These small landslides frequently impact and damage human settlements and disrupt economic activity. The landslide distribution is ascertained through a detailed landslide inventory map of 107 landslides covering 0.088 km2 and a related geodataset created from multitemporal aerial photographs and field investigations (Legorreta Paulín et al. 2014). The technique and its implementation are presented and discussed, as are the implications for the landslide susceptibility associated with the El Estado watershed. The results from the study area suggest that the MLR model has a stronger predictive capability compared to MCE and SINMAP models. The results show that the three models are not perfect in representing existing landslides but tend to improve by using calibrated field data and a systematic sample strategy that focuses on the detection of headscarps.

2 Background

There are a number of inventory, heuristic, statistic, and deterministic methods to assess landslide susceptibility. Each one has its own advantages and limitations (Anbalagan and Bhawani 1996; Van Westen et al. 1997; Maceo-Giovanni et al. 2000; Dai et al. 2002; Lee 2005; Weirich and Blesius 2007; Hervás and Bobrowsky 2009; Regmi et al. 2010; Guzzetti et al. 2012). One main issue in modeling landslide susceptibility has been noted: There is a lack of systematic comparison of methods in both natural and theoretical conditions to fully outline the advantages and limitations of applied methodologies. The lack of systematic comparison of landslide models for different scales, DEM resolutions, sampling strategies, and type of landslides not only compromises the reliability of the models, but also leads to abuse of the models. This is especially true for validation in natural conditions, due to the complexity of concomitant natural and technical problems. Problems such as pixel resolution, interpolation, definition of landslide type, misunderstanding of model requirements, incorrect calculation and/or estimation of topographic, hydrologic, and soil parameters compromise the model efficiency. Nevertheless, efforts to create, test, and validate landslide susceptibility models have been made (Pack et al. 1998; Morrissey et al. 2001; Wawer and Nowocień 2003; Zaitchik et al. 2003; Legorreta Paulin and Bursik 2008). For zones with sparse information, three landslide susceptibility models have been used worldwide: (1) SINMAP, which combines the theory of a hydrologic model (Beven and Kirkby 1979; O’Loughlin 1986) and the infinite slope stability model factor of safety (Hammond et al. 1992) to produce the susceptibility map, (2) MLR, which is based on analyzing the relation between landslide-controlling factors and the distribution of landslides (Ohlmacher and Davis 2003), and (3) MCE, which includes the evaluation and combination of criteria (spatial and nonspatial variables that could potentially trigger landslides according to an expert opinion) to determine landslide susceptible areas. These criteria, expressed as thematic maps, are assessed by combined, standardized, and weighted variables. The maps are then used to follow decision rules and hierarchization of alternatives according to user preferences (Aceves-Quesada et al. 2006; Feizizadeh et al. 2014; Dragičević et al. 2015).

In Mexico, during the last decade, a general framework and guidance for a state and city atlas of landslide susceptibility, hazards, and risks have been prepared by the National Center for Prevention of Disasters (Centro Nacional de Prevención de Desastres 2004). Local or regional landslide mapping has been carried out by using GIS and remote sensing (Capra and Lugo-Hubp 2006; Pérez-Gutiérrez 2007; Secretaría de Protección Civil 2010). In Veracruz State, an atlas of geological and hydrometeorological hazards was created in 2010 by the Secretary of Civil Protection of Veracruz State in collaboration with other federal and state government agencies. In 2011, the federal government published an atlas of natural hazards at municipal level based on the use of GIS (Secretaría de Protección Civil 2010; SEDESOL 2011).

In the study area, most research has focused on the volcanic history of Pico de Orizaba volcano to establish eruptive styles and the potential hazard of volcanic events and flank collapse (Siebe et al. 1992; Carrasco-Núñez and Rose 1995; Macías 2005; Carrasco-Núñez et al. 2006). Using computer simulations, GIS, and remote sensing, maps have been created to show the risk of catastrophic voluminous lahar movement along stream systems of Pico de Orizaba (Hubbard 2001; Sheridan et al. 2001; Hubbard et al. 2007). On the southwestern flank of Pico de Orizaba volcano, a multitemporal inventory map and a landslide susceptibility map were created at regional scale by means of MLR and SINMAP. The model was implemented along the Barranca del Muerto-Río Chiquito watershed (11 km2) whose main tributary is El Estado river. The results illustrate that SINMAP and MLR tend to over-predict. In spite of this over-prediction, SINMAP is only able to match 45.39 % of the inventory map by using field and published geotechnical values, and MLR succeeds in predicting 72.33 % of landslide areas by using cartographic variables (Legorreta Paulín et al. 2013). In 2014, a landslide inventory and a susceptibility map using MLR were created at local scale for El Estado watershed (5.2 km2). The results illustrate that the model succeeds in predicting 79.81 % of landslide areas (Legorreta Paulín et al. 2014).

3 Study area

The three methods for GIS-based landslide susceptibility mapping have been tested in a small watershed in a volcanic mountainous area, located in the eastern part of the Trans-Mexican Volcanic Belt (TMVB) physiographic province. The Río El Estado watershed is located at 18°55′23″–18°59′36″ N and 97°16′17″–97°14′56″ W, on the southwestern flank of Pico de Orizaba volcano within Puebla and Veracruz states, Mexico (Fig. 1). The river is a sub-basin of the Río Chiquito-Barranca del Muerto watershed, which flows into the Gulf of Mexico with the name of Río Blanco. The watershed covers 5.2 km2 with an elevation range from 2677 to 4248 m a.s.l. and hillslopes between ~17° (inner valleys of relatively flat plains) and >45° (mountainous terrain).

Fig. 1
figure 1

Location of the study area

Full size image

Climate is classified as subtropical semicold (Cb’(w)) at 3000–4400 m a.s.l. and subtropical temperate, subhumid (C(w1) and C(w2)) at <3000 m a.s.l. (García 2004) (Fig. 2a). Seasonal rainfall averages 1000–1100 mm/year at >4000 m a.s.l. and 927 mm/year at <1500 m a.s.l. and is most abundant in the wet season between May and November (Palacios et al. 1999). The stream system of Río El Estado watershed erodes andesitic and dacitic Tertiary and Quaternary lavas, pyroclastic flows, and fall deposits. Pyroclastic fall deposits and lahar deposits constitute about 50.2 % of the total the watershed area. Massive dacite lava flows cover 34.4 % of the watershed area, whereas the area covered by Quaternary andesitic block lava and deposits is 15.1 %. Only 0.2 % of the area is covered with basaltic–andesitic blocky brecciated lava flows (Fig. 2b). The study area is prone to landslide due to its large area of collapsible, weathered or disjointed volcanoclastic material at high altitudes and under high seasonal rainfall. Also, the area is prone to landslides due to decades of deforestation and modification of the slopes in favor of agricultural activities. One hundred and seven landslides are mapped from aerial photographs and field verification in the study area. In the study area, episodic evacuation of debris by shallow mass movement followed by slow refilling with alluvium and colluvium takes place along the fluvial system, where small buffers of vegetation around the valleys are unable to stop the disruption of the slope. The steep hills capped with ash and pyroclastic deposits are affected by active and dormant deep-seated landslides, and where the stream erodes lava flows and lahar deposits, rock falls have occurred (Legorreta Paulín et al. 2014).

Fig. 2
figure 2

a Climate map of the study area. b Geology map of the study area

Full size image

4 Methods

In order to compare SINMAP, MLR, and MCE landslide susceptibility models, an inventory map created from two sets of aerial orthophotographs, fieldwork, and thematic layers (Legorreta Paulín et al. 2014) is used (Fig. 3a). Three levels of data management are used to incorporate topographic and thematic information into the modeling. During the first level, analogue topographic paper map at scale 1:50,000, geological paper map at scale 1:35,000, and paper maps of land use, climate, and hydrology at scale 1:250,000 were converted to a 10-m raster format, georeferenced, and incorporated as GIS thematic layers. In the second level, five layers are derived from the topographic elevation data: slope angles, slope curvature, contributing area, flow direction, and saturation. In the third level, seven more thematic maps are derived from the previous two level of data: hypsometric map, energy relief map, vertical erosion map, horizontal erosion map, morphographic map, potential erosion map, and a reclassified slope map. The thematic layers and geotechnical field data are used to feed the models based on their technical requirements.

Fig. 3
figure 3

a Landslide inventory map, b SINMAP susceptibility model, c MLR susceptibility model, d MCE susceptibility model

Full size image

The first model, SINMAP, combines the theory of a steady-state hydrology model (Beven and Kirkby 1979; O’Loughlin 1986) with the infinite slope stability model for factor of safety (Hammond et al. 1992) to produce the stability index (SI). The model is designed to detect only shallow landslides. The index is expressed by using six broad classes with subjective break points that require judgment and interpretation in terms of class definition (Table 1).

Table 1 SINMAP, Definición de clases de estabilidad (Pack et al. 1998)
Full size table

SINMAP combines topographic (slope, contributing area, and flow direction), hydrographic (soil transmissivity/net rainfall ratio), and soil variables (soil density, internal friction angle, and cohesion) to predict potential landslide zones with sparse information. The geotechnical properties are used to characterize one unstable layer at the potential failure location within a zone of initiation to obtain the SI (Pack et al. 1998). However, little work has been done in clarifying the importance of the geotechnical properties of different soil layers in producing landslides. To palliate the problem, SINMAP incorporates uncertain parameters through the use of uniform probability distributions, and lower and upper bounds are set on uncertain parameters (Pack et al. 1998). The evaluation of SINMAP in natural conditions has led to the claim that the SINMAP approach fairly defines areas that intuitively appear to be susceptible to landsliding if the model is fed with calibrated geotechnical parameters and accurate DEMs (Pack et al. 1998; Wawer and Nowocień 2003). In this study, stratigraphic columns are sampled to obtain geotechnical properties. Each stratigraphic column is described by the procedures suggested by Gardiner and Dackombe (1983) and Compton (1985). Samples are collected from 20 landslide areas. In the field, transmissivity is calculated with a field minidisk portable tension infiltrometer. The mean annual precipitation and the potential evapotranspiration per year are obtained from a technical evaluation of the Río Chiquito-Barranca del Muerto watershed (Rodríguez et al. 2011), and a weather station located in Atzitzintla town (9 km SW of El Estado watershed). The above information is used to obtain the soil transmissivity/net rainfall ratio. Cohesion and the internal friction angle are calculated in the laboratory by using a direct shear apparatus. The resulting landslide susceptibility map (Fig. 3b) is created by using the following parameters: (1) DEM of 10 m pixel size, (2) soil thickness of 1 m, (3) soil density of 1850 kg/m3, (4) T/R minimum of 282.39 m, (5) T/R maximum of 489.48 m, (6) cohesion minimum of 0 (unitless), (7) cohesion maximum of 0.38 (unitless), (8) internal friction angle minimum of 32.3°, and (9) internal friction angle maximum of 38.1°.

The second model, MLR, is designed to describe probability—the model estimates the probability of landslides between 0 and 1—and it is preferred when the outcome is dichotomous (Kleinbaum and Klein 2002). MLR has no mechanical meaning, but it explores the relationship between landslide occurrence in the past and the terrain and landscape variables (Ayalew and Yamagishi 2005; Can et al. 2005). MLR is popular among multivariate statistical approaches for landslide susceptibility assessment because (1) the model is designed to describe probability, (2) it uses the logarithmic (logistic) transformation to map the substantive, nonlinear relationship between dependent and predictor variables, into the form of a linear relationship (Ohlmacher and Davis 2003), and (3) the sigmoidal shape of the logistic function can be interpreted by assuming that the risk of landslide is minimal for low logistic function values until some threshold is reached. The probability of landslide increases fast over a certain range of intermediate logistic function values and remains high once values of the logistic function get large enough (Kleinbaum and Klein 2002). The evaluation of MLR under natural conditions has led to the claim that the model is fairly successful at identifying landslide areas if an adequate sampling strategy, sample size, and set of variables that are strongly related to landslides are used (Ohlmacher and Davis 2003; Can et al. 2005). In this study, we use the MLR landslide susceptibility map of El Estado watershed prepared by Legorreta Paulín et al. (2014). The map was created by using a random sample from nonlandslide and landslide areas for eight independent variables (elevation, slope angle, terrain curvature, flow direction, saturation, contributing area, lithology, and land use). The susceptibility map was created on the basis of the logistic function: 1/1 + Exp-(18.469774 + Elevation map (−0.005463) + Slope map (0.143026) + Flow direction map (−0.025160) + Land use map (0.326680) + Geology map (0.326680) + Curvature map (−0.647303)) (Legorreta Paulín et al. 2014).

The resulting landslide susceptibility map is categorized into 10 probability zone categories (Fig. 3c).

The third model, MCE, is a heuristic approach that seeks in a holistic way to combine and transform geographic data into a resultant landslide susceptibility map. It allows the inclusion into the GIS computer-based systems of qualitative and semiquantitative information at different scales and sources as well as the perception, preferences, expert opinion, and human judgment (Castellanos Abella and Van Westen 2008; Feizizadeh et al. 2014; Dragičević et al. 2015). The approach encompasses the establishment of a goal (create the landslide susceptibility map of an area) which involves a set of quantifiable spatial and nonspatial criteria (expressed as a factor or constraints map). The criteria maps are grouped, standardized, and weighted according to preferences and processed in the GIS to generate a landslide susceptibility map (Aceves-Quesada et al. 2006; Castellanos Abella and Van Westen 2008; Feizizadeh et al. 2014; Dragičević et al. 2015). The evaluation of MCE at small and medium scales has been shown successful in areas where inventory information is highly incomplete and only a few spatial factors can be used to model landslide susceptibility (Günther et al. 2014). In this research, landslide susceptibility is assessed by using the MCE module in the ILWIS GIS. The inputs are a set of seven morphometric maps: (1) The hypsometric map is heuristically classified to highlight altimetric levels (Aceves-Quesada et al. 2013), (2) the energy relief map is calculated to represent the maximum elevation change within 1 hectare (Lugo-Hubp 1988), (3) the vertical erosion map is calculated as the maximum vertical distance between the thalweg and the watershed per unit area (1 km2) (Lugo-Hubp 1988), (4) the horizontal erosion map is calculated as the total thalweg length per unit area (1 km2) (Lugo-Hubp 1988), (5) the morphographic map is derived from an heuristic re-classification of the slope map to highlight volcanic landforms (Aceves-Quesada et al. 2013), (6) the slope map is re-classified to highlight the difference in the relief (Aceves-Quesada et al. 2013), and (7) the potential erosion map is derived from the slope map to define zones prone to intense erosion (Palacio-Prieto 1983). Weights are assigned to the maps by using the ILWIS ranking method in which the researcher ranks the criteria in order of preference. The numerical weights are then assigned as function of the rank, and the final susceptibility map is created in the following way:

Landslide susceptibility map = hypsometric map * 0.044 + energy relief map * 0.073 + vertical erosion map * 0.156 + horizontal erosion map * 0.228 + morphographic map * 0.37 + slope map * 0.109 + potential erosion map * 0.02. The resulting landslide susceptibility map is categorized into zones of “very low,” “low,” “medium,” “high,” and “very high” categories (Fig. 3d).

To evaluate the models, the “predicted model vs. inventory matching” approach was used (Van Den Eeckhaunt et al. 2006; Saito et al. 2009). The percentage of overlay between each model susceptibility map and the inventory map was the gauge of how well the model predicts the reality. A two-classification scheme (landslide and nonlandslide) was used for the inventory map and SINMAP, MLR and MCE model susceptibility maps to facilitate the comparison. For SINMAP, the six SI classes were re-classified to a two-classification scheme: Landslide area includes SI defended, upper threshold, and lower threshold classes, and the nonlandslide area includes SI stable, moderately stable, and quasi-stable SI classes. For MLR, values higher than 0.5 were classified as landslide and those lower than 0.5 as nonlandslide (Dai et al. 2002). For MCE, landslide area includes high and very high classes and the nonlandslide area includes very low, low, and moderate classes. The breaking points to separate landslide and nonlandslides areas were selected to compare the results of using heuristic, worldwide, accepted pre-established values with values heuristically selected for the study area.

The accuracy in the predictive capability of the SINMAP, MLR, and MCE models is evaluated by using a histogram and a contingency table in the system LOGISNET under ArcInfo GIS software (Legorreta Paulin and Bursik 2008). The histogram shows the amount of pixels in areas with landslide and without landslides predicted by the model and how the prediction is spread between areas with and without landslides in the inventory map. The contingency table shows the amount of overlap between inventory and predicted maps through the use of different statistics such as the overall accuracy (calculated by the total number of correctly classified pixels of landslides and nonlandslides divided by the total number of pixels in the study area), producer’s accuracy (calculated as the ratio of the number of correctly classified pixels in each category to the total number of true pixels for that category), user’s accuracy (calculated as the ratio of the number of correctly classified pixels in each category to the total number of pixels that are classified by the model in that category) (Legorreta Paulin and Bursik 2008), model efficiency (calculated as the ratio of the correctly minus incorrectly indicated landslide pixels to the total number of the true landslide pixels mapped in the inventory map) (Van Den Eeckhaunt et al. 2005). Also, the area under the ROC curve (AUC) is used. The AUC is calculated under SPSS by considering all terrain elements and assigning all landslides in the inventory map as positives and the rest as negatives. The AUC is defined by plotting into a two-dimensional graph pairs of sensitivity (true-positive rate) and specificity (false-positive rate). The true-positive rate is the number of correctly classified pixels in the landslide category to the total number of true landslide pixels, and the false-positive rate is the number of incorrectly classified nonlandslide pixels in the nonlandslide category to the total number of true nonlandslide pixels (Lee 2005; Fawcett 2006; Namdi and Shakoor 2009; Günther et al. 2014). The overall accuracy, producer’s accuracy, and user’s accuracy have a maximum value of 1 or 100 % when a model has predicted all correctly (Legorreta Paulin and Bursik 2008). The model efficiency has a maximum value of 1 when a model has predicted all true landslides without any incorrect landslides. It becomes negative when the number of incorrectly indicated landslides is larger than the number of correctly indicated landslides (Van Den Eeckhaunt et al. 2005). The accuracies and model efficiency are sensitive to class skews since they depend on the distribution or number of pixels that represent landslide or nonlandslide areas (Fawcett 2006). The area under curve has a value of 1 or 100 % if the model is perfect, while a value close to 0.5 indicates inaccuracy in the model (Fawcett 2006; Namdi and Shakoor 2009; Günther et al. 2014). The AUC is not sensitive to class skews since it is based on ratios that do not depend on class distribution (Fawcett 2006).

5 Results

In the study area, a preexisting inventory of 107 landslides covering 0.088 km2 is used to assess and describe landslide distribution (Legorreta Paulín et al. 2014). With the landslide inventory as a framework of landslide locations, the landslide susceptibility maps are validated. Qualitatively, the models show high instability along andesitic volcanic hilly areas capped with pyroclastic mantles and along the valley walls where steep slopes exist. The match between landslide inventory and model predictions is not perfect, and in general, there is over-prediction of landslides. The over-prediction is greater in the upper portion of the watershed where dacitic lavas are covered by pyroclastic falls (between 3960 and 4248 m a.s.l) and along the valley walls where steep slopes exist. Nevertheless, the “over-predicted” areas do have the potential to fail at some point in the future. In the field, it was in fact sometimes observed that these areas contained landslides that had not been included in the inventory.

Quantitatively, the models’ performance was evaluated in terms of statistics (Table 2) derived from the contingency table. The overall accuracy and AUC summarize the error and success in landslide and nonlandslide area. The AUC indicates a greater predictive capability for MLR (83.1 %), followed by MCE (75.6 %) and SINMAP (71 %). The high overall accuracy of MCE (84.08 %) and MLR (83.1 %) compared to SINMAP (54.39 %) is explained by the fact that a high accuracy in the nonlandslide area is weighted more than that in the landslide area because the nonlandslide area (5.112 km2) is larger than is the landslide area (0.088 km2). In the nonlandslide area, MCE and MLR have a large number of correctly classified nonlandslide pixels (43,108 and 36,362 pixels, respectively) compared to SINMAP (27,721 pixels). The overall accuracy is a good measure of the model reliability. The interpretation must be complemented with the producer’s and user’s accuracies because they show the over-prediction or under-prediction of a model. In the landslide areas, the producer’s accuracy shows that the MLR and SINMAP models have better coincidence (79.81 and 74.58 %, respectively) with the landslide inventory map compared to MCE (54.87 %). This “good match” of MLR and SINMAP model inventory is due to over-prediction. This over-prediction leads to more nonlandslide areas being classified incorrectly as landslide in the models. All models over-predict, as is shown by a low percentage of user’s accuracy and negative values in the model efficiency.

Table 2 Model accuracies and model efficiency
Full size table

6 Conclusions

In this paper, we briefly introduce the comparison of SINMAP, MLR, and MCE models for landslide susceptibility assessment in unstable volcanic terrains. The model assessment at El Estado watershed is an attempt to identify and select the best available model to flag potential landslide areas in volcanic regions of Mexico with sparse information. The model validation, derived from a contingency table, shows the attractive property (insensitive to changes in class distribution) of using the AUC compared to the accuracies and model efficiency techniques. However, the latter techniques are useful to detect over-prediction or under-prediction of landslide models. The AUC points out that the MLR model predicts better than does the MCE and SINMAP models. The model validation for El Estado watershed, using the accuracies and model efficiency, illustrates that SINMAP, MLR, and MCE have a moderate to high match with the landslide areas, but they tend to over-predict. This does not mean that the models perform “wrong.” The over-prediction of landslide areas can be justified because of the cartographic generalization at 10 m pixel size of different scale topographic and thematic maps. The generalization makes the shape of the landslides to be topographically smoothed and could bias the models toward the over-prediction. A more accurate model could be obtained when higher-resolution topographic and thematic data become available in Mexico. In addition, the inventory may be incomplete and under-estimate the real susceptible areas even though it was carefully created by expert knowledge. Thus, a certain degree of over-prediction is expected in the models. The over-prediction may be useful to define potential future landslide areas.

In this research, MLR can be said to be preferred over the MCE and SINMAP models because MLR gets similar or better results with existing thematic maps at different scales, with fewer significant variables. MCE requires heuristic integration of variables that are specific for a study area. The result is acceptable. However, the reproducibility is low and practical application is limited for other areas. For SINMAP, results are improved (74.58 % producer accuracy) at local scale with specific geotechnical data compared to previous founding at regional scale (45.39 % producer accuracy; Legorreta Paulín et al. 2013), but obtaining geotechnical properties is time-consuming and expensive.

Using geomorphologic expert knowledge, a specific heuristic break point is selected for El Estado watershed as with MCE, whereas the break points for MLR and SINMAP are based on heuristic knowledge from other areas of the world. As a result, overall accuracy is the highest using the MCE model, followed by MLR and SINMAP. MCE is less effective in landslide areas where there is a low match between inventory model prediction (producer’s accuracy).

For future study, the selection of break points to define landslide and nonlandslide areas needs to be evaluated. The break points could be improved through the analysis of receiver operating characteristic curves (AUC) to produce a map that satisfies both expert knowledge and landslide signal (Van Den Eeckhaunt et al. 2006).