Introduction

The world’s development infrastructure is at risk from landslides and their consequences in many areas across the globe. There is evidence that landslide disaster risk is increasing in developing countries (Anderson and Holcombe 2013). With rapid urbanization, the growth of densely populated communities in mountainous and hazardous locations is continuing and this leads to the instability of slopes and thus increases the potential for landslides. Therefore, developing landslide susceptibility maps, risk analyzing, and adopting suitable land use policies for different environmental settings are urgently needed.

Landslides occur due to complex and relate to various factors such as geology, topography, hydrogeological conditions, vegetation, rainstorm, and human activities (Cruden 1991; Montgomery and Dietrich 1994; Wu and Sidle 1995; Guzzetti et al. 1999; Gorsevski et al. 2006; Ardizzone et al. 2007).

A wide variety of methods and techniques have been proposed and applied for landslide susceptibility mapping (LSM). The most common approaches proposed in the literature are frequency ratio (FR), statistical index (SI), weights of evidence (WoE), logistic regression, multivariate regression, discriminant analysis, index of entropy, spatial multi-criteria evaluation, analytical hierarchy process, evidential belief function, decision tree, artificial neural network, fuzzy logic, neuro-fuzzy, and support vector machines (Gorsevski et al. 2003, 2005; Nefeslioglu et al. 2008; Gorsevski and Jankowski 2010; Pourghasemi et al. 2012, 2013; Ozdemir and Altural 2013; Pradhan 2013; Jebur et al. 2014; Jaafari et al. 2014, 2015a, 2017a; Pham et al. 2016, 2017).

Each method differs in terms of input process, calculations, output process, and predictive reliability. Although many comparative studies that have been carried out on predictive ability of different methods (e.g., Pradhan 2013; Pourghasemi et al. 2013; Jebur et al. 2014; Hong et al. 2015; Pham et al. 2016), decision makers and engineers involved in slope management and land use planning still need to select the best method for different environmental settings. Among the commonly used geographic information system (GIS)-based models for susceptibility modeling, FR, SI, and WoE have been widely investigated in the literature (e.g., Mohammady et al. 2012; Pourghasemi et al. 2013; Ozdemir and Altural 2013; Regmi et al. 2014; Jaafari et al. 2014; Youssef et al. 2016). While, in these models, input process and calculations are very straightforward and can easily be implemented within a GIS environment (Mohammady et al. 2012; Jaafari et al. 2014), their results prove high level of accuracy for prediction of future landslides that give rise to appealing qualitative and quantitative maps for the landslide-prone areas (Mohammady et al. 2012; Youssef et al. 2016). These advantages have motivated us to use these three models to assess landslide susceptibility for a part of Mazandaran Province of northern Iran. Our study aims to compare the predictive ability of FR, WoE, and SI methods for LSM in the study area. Since this is a prototype study performed in one of the characteristic landslide susceptible regions of northern Iran, the findings can be utilized for regions that show similar geoenvironmental and landslide characteristics.

Study area

The study area is located about 80 km southwest of Sari between 36°11′43″N to 36°18′14″N latitude and 51°14′37″E to 52°58′49″ E longitude in the Mazandaran Province of northern Iran. It covers an area of about 270 km2 and encompasses farmlands, mountainous terrains of forest, rangelands, and residential areas. The area represents a variable and rough topography, with slope gradients between 0° (flat) and 62° and altitudes between 1173 and 3536 m. The climate in the study area is Mediterranean with a mean annual precipitation of 800 mm that occurs in the form of snow during the winter. The lithology of the study area consists of several geologic units such as dolomite, siltstone, sandstone, marl, and conglomerate (Table 1). The soil texture consists of several types including sand, loamy sand, sandy clay loam, sandy loam, silty clay, silty loam, and silty sand, of which sandy clay loam occupies about 78% of the total area. In this area, deforestation and inappropriate land use practices contribute to natural disasters such as flooding, soil erosion, and landslides in the past several years. The most important fault in this area is the Alborz Fault that is a reverse fault and follows the west–east orientation (Jaafari et al. 2017a). Some reports suggest that this fault is the main source of most earthquakes and landslides that occur in the Mazandaran Province (Darvishzadeh 2004; Jaafari et al. 2017a).

Table 1 Lithology of the study area
Full size table

Database construction

Our database for LSM involves an inventory map of landslides occurred within the study area and a set of landslide conditioning factors. The inventory map of existing landslides of the study area was compiled using aerial photography interpretation and extensive field surveys. In total, 105 landslides were mapped as georeferenced points (Fig. 1). The detected landslides were then divided into two subsets: The training dataset that contained 70% of the landslide inventory (74 landslides) was used in the training phase of landslide models, and the validation dataset with 30% of the data (31 landslides) was used for the validation purpose (Pourghasemi et al. 2013; Hong et al. 2015; Jaafari et al. 2015a).

Fig. 1
figure 1

Location and landslide inventory map of the study area

Full size image

Given the multiplicity of the landslide conditioning factors, they are usually selected based on the landslide types, the failure mechanisms, the map scale of analysis, and the characteristics of the study area and data availability (Glade and Crozier 2005). A total number of 13 landslide conditioning factors were considered to be used in this study. They are slope degree, slope aspect, altitude, plan curvature, stream power index (SPI), topographic wetness index (TWI), sediment transport index (STI), topographic roughness index (TRI), lithology, distance from streams, faults, roads, and land use type (Figs. 2, 3, 4, 5; Table 1). The significance of these factors on landslide occurrence has explicitly been acknowledged by other authors (e.g., Dai et al. 2001; van Westen et al. 2008; Yalcin et al. 2011; Guillard and Zezere 2012; Mohammady et al. 2012; Kayastha et al. 2012; Ozdemir and Altural 2013; Pourghasemi et al. 2013; Jebur et al. 2014; Jaafari et al. 2014, 2015a; Pham et al. 2017). Slope degree, slope aspect, altitude, and TRI were selected to represent the effect of topographic factors on slope stability analysis. Topographic factors generally control several characteristics (e.g., topographic heterogeneity, land use/cover, shear forces, amount of rainfall, terrain humidity, and erosion–weathering processes) of the landscape that may modulate occurrence of landslides (Guzzetti et al. 1999). Plan curvature and water-related factors (i.e., SPI, STI, and TWI) were chosen due to their influence on hydrogeological conditions that, in turn, exert effect on the surface runoff and infiltration. We also used lithology and distance from faults factors to represent the influence of geomorphological processes and tectonic factors on the occurrence of landslides (Jaafari et al. 2015a; Pham et al. 2017). Since the under-cutting actions and erosion process of streams can trigger landslides (Xu et al. 2012; Jaafari et al. 2015a), the distance from streams was also included to our analysis. Types of land use/cover, which reflect the human activities on the landscape, can significantly affect the susceptibility of a given landscape (Catani et al. 2013). Road development, especially in mountainous regions, is usually the main cause of erosion and slope failure (Nefeslioglu et al. 2008; Jaafari et al. 2015a, b). Thus, distance from roads is a commonly used conditioning factor in landslide susceptibility modeling (e.g., Pourghasemi et al. 2013; Hong et al. 2015; Pourghasemi and Kerle 2016; Pham et al. 2016).

Fig. 2
figure 2

Topographic parameter maps of the study area; a slope degree, b slope aspect, c altitude, d plan curvature, e stream power index, f topographic wetness index, g sediment transport index, h topographic roughness index

Full size image
Fig. 3
figure 3

Lithology map of the study area

Full size image
Fig. 4
figure 4

Distance factors; a distance from streams, b distance from faults, c distance from roads

Full size image
Fig. 5
figure 5

Land use types of the study area

Full size image

A digital elevation model (DEM) with 20 × 20 m grid size (Projection: UTM 39N; Datum: WGS1984) and the data prepared by Geological Survey of Iran were used as the main input data to generate the aforementioned conditioning factors (Table 2). All the calculations and data processing were hosted in ArcGIS 9.3 and SAGA GIS 2.1.4. The map of all of the conditioning factors was generated and converted into raster format with the size of 20 × 20 m (Kayastha et al. 2012; Ozdemir and Altural 2013; Jaafari et al. 2014, 2015a, b; Pham et al. 2017). To implement the FR, WoE, and SI models, the conditioning factors were classified into different classes. The different classes were selected based on previous landslide studies (e.g., Pourghasemi et al. 2013; Jaafari et al. 2014) that were further informed by our surveys and observations in the study area.

Table 2 Data source for the landslide conditioning factors
Full size table

Multi-collinearity analysis of landslide conditioning factor

In the final step of preparing a spatial database for landslide modeling, we examined whether any of the selected conditioning factors exhibited multi-collinearity. Tolerance and variance inflation factors (VIF) (Hair et al. 2006) are the most common metrics to check for multi-collinearity among factors (e.g., Hong et al. 2015; Jaafari et al. 2017b) that a violation of their critical values (VIF > 5 and tolerance <0.2) indicates a potential problem with multi-collinearity (Hair et al. 2006). In our study, the computation of these metrics showed that no significant multi-collinearity existed among the conditioning factors (Table 3) and therefore all factors can be used in the modeling process.

Table 3 Multi-collinearity analysis for the conditioning factors
Full size table

Methodology

Frequency ratio model

As a bivariate statistical technique, FR model is a robust geospatial assessment tool for computing the probabilistic relationship between dependent and independent variables (Oh et al. 2011). For the purpose of LSM, FR considers the impact of each conditioning factor on landsliding and assigns the weights very precisely (Lee and Pradhan 2007). If the weight be more than 1, it means a greater correlation, whereas the weights less than 1 represent a minor correlation (Lee and Min 2001). The calculation process of FR model is very straightforward and can be readily realized as follows (Regmi et al. 2014; Jaafari et al. 2014):

$$FR = \frac{E/F}{M/L}$$
(1)

where E is the number of pixels with landslide for each conditioning factor; F, the number of total landslides in study area; M, the number of pixels in the class area of the conditioning factor; and L, the number of total pixels in the study area.

Weights of evidence model

WoE model is a data-driven method based on the Bayesian probability framework (Bonham-Carter et al. 1989). This model is suitable for LSM because its uncertainty is connected with landslide events and their associations with the complex landscape (Chung and Fabbri 1998; Gorsevski et al. 2003; Regmi et al. 2014; Jaafari et al. 2015a, b). This model is based on the determination of positive (W +) and negative weights (W −). The WoE calculates the weight for each class of conditioning factors (B) based on the presence or absence of the landslides (L) within the area as follows (Bonham-Carter et al. 1989):

$$W_{i}^{ + } = \ln \frac{{P\left\{ {B|L} \right\}}}{{P\left\{ {B|\bar{L}} \right\}}}$$
(2)
$$W_{i}^{ - } = \ln \frac{{P\left\{ {\bar{B}|L} \right\}}}{{P\left\{ {\bar{B}|\bar{L}} \right\}}}$$
(3)
$$C = W^{ + } + W^{ - }$$
(4)
$$S(C) = \sqrt {S_{{}}^{2} (w_{{}}^{ + } ) + S_{{}}^{2} (w_{{}}^{ - } )}$$
(5)
$$S_{{}}^{2} (w_{{}}^{ + } ) = \frac{1}{{N\left\{ {B \cap L} \right\}}} + \frac{1}{{\left\{ {B \cap \bar{L}} \right\}}}$$
(6)
$$S_{{}}^{2} (w_{{}}^{ - } ) = \frac{1}{{N\left\{ {\bar{B} \cap L} \right\}}} + \frac{1}{{\left\{ {\bar{B} \cap \bar{L}} \right\}}}$$
(7)
$$W = \frac{C}{S(C)}$$
(8)

where \(W_{i}^{ + }\) is positive weight; \(W_{i}^{ - }\), negative weight; ln, natural log; P, conditional probability; B, the presence of a potential conditioning factor; \(\bar{B}\), the absence of a potential conditioning factor; L, the presence of a landslide; \(\bar{L}\), the absence of a landslide; C, weight contrast; S 2 (w +) and S 2 (w −), variances of positive and negative weights; W, studentized contrast (final weight); and S (C), the standard deviation of the contrast.

Statistical index model

SI model is a bivariate statistical method proposed by van Westen (1997) for the purpose of LSM. This method is based on the following equation:

$$W_{\text{SI}}^{{}} = \ln \left( {\frac{{F_{ij}^{{}} }}{F}} \right) = \ln \left( {\frac{{L_{ij}^{{}} /L_{\text{T}}^{{}} }}{{P_{ij}^{{}} /P_{\text{L}}^{{}} }}} \right)$$
(9)

where W SI is the weight given to a certain class i of factor j; F ij , landslide density within class i of factor j; F, total landslide density within the entire map; L ij , number of landslides in a certain class i of factor j; P ij , number of pixels in a certain class i of factor j; L T, total number of landslides in the entire map; and P L, total pixels of the entire map.

A positive value of W SI demonstrates the existence of a relationship between the presence of the factor class and landslide distribution, the stronger the relationship the higher the score. The W SI is negative when the presence of the factor class is not relevant in landslide development (Pourghasemi et al. 2013).

The weights calculated using the three models were then assigned to the classes of each conditioning factor to produce multi-class weighted maps for all factors, which were overlaid and numerically added according to the following equation in order to calculate the landslide susceptibility index (LSI) maps (Jaafari et al. 2015a):

$${\text{LSI}} = \sum\limits_{i = 1}^{13} W$$
(10)

where W corresponds to the multi-class weighted conditioning factors.

Performance validation and factor effect analysis

Validation is the most important step in a modeling effort, and without validation the model results lack scientific significance (Chung and Fabbri 2003; Jaafari et al. 2017b). The receiver operating characteristics (ROC) curves with the area under curve (AUC) approach are widely used criteria for evaluating the performance of the prediction models (Pourghasemi et al. 2013; Jebur et al. 2014; Jaafari et al. 2015a, b, 2017a, b; Hong et al. 2015; Pham et al. 2016; Nami et al. 2017). ROC curve is a binary classification metric created by plotting the false-positive rate and the true-positive rate for every possible binary classification of a dataset (Zweig and Campbell 1993). The ideal ROC curve passes through the point of (0, 1) with AUC = 1, indicating that there is no prediction error. An acceptable AUC value for a ROC curve should be higher than 0.5 (Yesilnacar 2005). Further, to investigate the reliability of the produced maps, we used landslide density and spatially agreed area analyses within different classes of each susceptibility map. Landslide density indicates the ratio of landslide pixels to the ratio of total pixels (Pham et al. 2017). The spatially agreed area, expressed in pixels, km2, and as percentage of the total area, is calculated as the total area having the same landslide susceptibility zonation on two susceptibility maps (Bijukchhen et al. 2013; Kayastha et al. 2013). These analyses can reveal spatial differences between the maps and improve predictions in the agreed high susceptible zones.

In this research, we also performed an analysis of conditioning factors effect to explore the effect of each factor on the prediction results and to assess their uncertainties. To this end, we first took the method that achieved the highest ROC values and excluded each of the 13 factors in turn during the summation stage of Eq. 10. We next calculated the success and prediction rates for all cases and compared these cases to the case in which all factors were included (Jaafari et al. 2017b).

Finally, for visual interpretation of the LSI maps, the data were classified into categorical susceptibility classes by examining different classifications methods, including quantile, natural breaks, standard deviation, equal interval, and geometrical interval (Ayalew and Yamagishi 2005). The comparison results indicated that the quantile method was able to produce better results than the other methods. Therefore, this method was chosen and the landslide susceptibility index maps were classified into four susceptibility classes that represent low, moderate, high, and very high susceptibility to landslide occurrence across the study area.

Results and discussion

Application of frequency ratio model

Table 4 shows the results of spatial relationship between landslide locations and landslide conditioning factors using the FR model. From this table, it is seen that a slope angle slope >35° has a higher FR value of 3.67, followed by 30–35° (1.17), whereas other slope classes have a very low value of FR. Given the lower shear stresses associated with low gradients areas, gentle slopes were frequently reported to have lower weight values of FR (Yalcin et al. 2011; Mohammady et al. 2012; Jaafari et al. 2014; Youssef et al. 2016). For slope aspect, the value of FR is higher for the areas facing the northwest, east, west, and southeast. The study conducted by Jaafari et al. (2014) in other regions of northern Iran supports the significance of these directions of slope angle on landslide occurrence. The relationship between altitude and landslide probability shows that the class of 0–1500 m has highest FR value (1.99), indicating higher landslide susceptibility at this range of elevation. This finding agrees with the field surveys since the landslides were commonly observed on low-elevation ranges of the study area. In the case of plan curvature, convex and concave areas have higher FR, with values of 1.59 and 1.37, respectively, whereas the flat class has the lowest FR value (0.56). Generally, the slope instability in the curvature areas is related to the high level of soil moisture. As the moisture content of the soil increases, soil stability generally decreases (Jaafari et al. 2014). According to the results of the FR model obtained for SPI, the highest FR value is related to the class of <200 (1.01). Similarly, for TWI, the areas with lowest value of TWI have higher FR values (1.44). In the case of the STI, landslide susceptibility is higher in areas where the STI is >15. In the case of TRI, the class of >12 shows the highest landslide susceptibility with a FR values of 5.44. Since the lithological units of TR 2e and Js represent 18.44 and 67.55% of our study area, the highest FR values were found to be related to these units. In case of distance from streams, the landslide occurrence probability is higher in areas where the distance is 100–200 m, whereas the probability decreases at a distance of more than 300 m from a stream. For distance from faults, the distances of 0–200 and 400–600 m show higher correlation with the landslides. The relationship between the road networks of the study area, and landslide probability shows that the class of 1000–1500 m has highest FR value, followed by the class of 500–1000 m. Regarding the land use type, the FR value was higher for the forested areas (2.02). In our study area, forests are mainly scattered on steep terrains and unstable slopes. These results in line with results reported by Jaafari et al. (2014, 2015a, b) provide a counterexample for the widely held notion that vegetation coverage always contributes to decreasing landslides and increasing general slope stability. Huge hardwood trees in northern forests of Iran and our study area exposed to wind may transmit the forces into the slope and cause landslide. Weight of such trees surcharges the slope, increasing normal and downhill force components, and may cause failure during the monsoon (Ghimire 2011; Jaafari et al. 2014). The final result of FR model was a LSI map, in which the LSI values vary from 5.122 to 29.6.

Table 4 Spatial relationship between each landslide conditioning factor and landslides extracted by using FR, WoE, and SI models
Full size table

Application of weights of evidence model

The WoE model was used to explore the spatial association between the conditioning factors and landslide distribution (Table 2). The studentized value of C and the value of W serve as a guide to the significance of spatial association and act as a measure of relative certainty of the posterior probability (Bonham-Carter et al. 1989). Given that higher values of W indicate a higher level of significance for a specific factor class (Kayastha et al. 2012; Jaafari et al. 2015a), from Table 4 it can be seen that this value is highest for slope degree >35, indicating a significant positive correlation with the landslides. Further, the results revealed that slope degrees of 0–30 were negatively correlated with the landslides, which can further be interpreted to mean that this range of slope degree has disfavored the occurrence of landslides events across our study area. In the case of slope aspect, most of the landslides occurred in northwest (W = 3.306) and east (W = 2.358) facings. North, northeast, south, and southwest aspects were negatively correlated with the landslides, which have disfavored the occurrence of landslides. For altitude, most of the landslides occurred in 1500–2000 m class with W value of 3.181. The negative values of W for the altitude >2500 m revealed that this range of altitude disfavored the occurrence of landslides in our study area. In the case of plan curvature, convex class has W value of 3.097. So, most of the landslides occurred in this class. In the case of the SPI, the W is highest (23.414) for <200 class. The relation between TWI and landslide probability showed that <6 class has highest value of W (3.559). For STI, the class >15 has most W value (3.460). In the case of TRI, the W is highest in >12 class (9.157) and lowest in <4 class (−3.266). In the case of lithology, the W is highest in TR 2e unit (7.495) since 42 cases out of 74 landslides used in training dataset were captured by this unit. In the case of distance from streams, higher W values of 0.702 and 0.222 were found for distances between 100–200 and 200–300 m, respectively. Assessment of distance from faults showed that distance of 400–600 m is more suitable for landslide occurrence with W value of 0.156. Investigation of distance from roads showed that distances of 1000–1500 have highest correlation with landslide occurrence. In the case of land use type, higher W value was seen for forest areas (3.963). The final result of WoE model was a LSI map, in which the LSI values vary from 70.2403 to 260.709.

Application of statistical index model

Similarly, LSI was calculated using SI model. The landslide distribution for each class of landslide conditioning factors was used to calculate the SI values (Table 2). As mentioned earlier, the larger the value, the stronger the relationship between landslide occurrence and the given factor’s class. The final result of SI model was a LSI map, in which the LSI values vary from −13.6014 to 7.93953 (Table 5).

Table 5 AUC values for the landslide index maps produced by using FR model
Full size table

Landslide susceptibility maps

Landslide susceptibility levels across the study area that ranged from low to very high are shown in Figs. 6, 7, and 8. In these three maps, high and very high susceptibility classes cover the northern and southwestern parts of the study area, whereas low and moderate susceptibility classes cover central parts. Further, the results showed that the high and very high susceptibility classes cover approximately 50% of the study area (Fig. 9). To investigate the reliability of the produced maps, we used landslide density analysis and spatially agreed area approach. The results showed that the value of landslide density varied among the classes and ranged from 0.04 to 2.88 (Table 6). In each map, the highest value is for very high susceptibility class which is followed by high class, moderate class, and low class, respectively.

Fig. 6
figure 6

Landslide susceptibility map produced by frequency ratio model

Full size image
Fig. 7
figure 7

Landslide susceptibility map produced by weights of evidence model

Full size image
Fig. 8
figure 8

Landslide susceptibility map produced by statistical index model

Full size image
Fig. 9
figure 9

Landslide susceptibility classes delimited by the five ANFIS models

Full size image
Table 6 Landslide density within the landslide susceptibility maps
Full size table

The results of the spatially agreed area revealed that the landslide susceptibility map produced using the FR method has 79.15 and 70.80% agreed area with the maps produced by the WoE and SI methods, respectively (Table 7), whereas the WoE and SI susceptibility maps have 76.56% the same content. These results indicate an average 24.5% mismatch in these three maps that require the planner to pay special attention for the construction works using these maps.

Table 7 Agreed area between FR method, and WoE and SI methods
Full size table

Validation and factor importance

The ability of FR, WoE, and SI model in LSM was examined using the ROC-AUC method with success and prediction rate curves. The success rate was produced by comparing the three susceptibility maps with the training dataset (Fig. 10). The result demonstrated that FR had the highest success rate value of 81.51%. It was followed by SI (81.27%) and WoE (79.43%). Although the success rate shows how well a model could fit to the training dataset, the prediction ability of the model cannot be measured by success rate because it is measured by landslides that have already been utilized for building the model (Bui et al. 2012). In this context, the prediction rate can be used to evaluate the prediction ability of the model. The prediction rates were measured by comparing the landslide susceptibility maps with the testing dataset (Fig. 11). They explain how well the models and conditioning factors predict the landslide. The prediction rate curves showed that the prediction ability of the models is highest for FR (80.44%), followed by SI (79.55%) and WoE (77.94%).

Fig. 10
figure 10

Success rate curves for the susceptibility maps produced in this study

Full size image
Fig. 11
figure 11

Prediction rate curves for the susceptibility maps produced in this study

Full size image

From these results, it is seen that the models employed in this study showed reasonably good accuracy in predicting the landslide susceptibility of the study area. The map produced by FR model exhibited the best result for the purpose of LSM in the study area. When the results of this study are compared with those of reported by other authors it can be stated that the prediction ability of the models are within a similar range. For instance, Mohammady et al. (2012) indicated that the FR model has a success rate of 80.13% and a prediction rate of 75.16%, whereas the WoE model has a success rate of 74.6% and a prediction rate of 69.98%. Regmi et al. (2014) achieved success rates of 76.8, 75.6, and 75.5%, and prediction rates of 75.4, 74.9, and 74.6%, for FR, WoE, and SI, respectively. In a recent paper, Youssef et al. (2016) reported that the FR model has a success rate of 81.3% and a prediction rate of 95%, whereas the WoE model has a success rate of 81.5% and a prediction rate of 95.2%.

Given the highest predictive capability of the FR model compared to the others, this model was used to perform the sensitivity analysis (Table 5). The analysis showed that some factors (i.e., distance from faults and streams, and SPI) were possible source of bias as the AUC values increased when they were omitted from the modeling process. Similarly, excluding slope aspect, plan curvature, and lithology slightly improved the prediction rate compared to the full FR model. Conversely, the exclusion of slope degree, TWI, STI, TRI, and land use type decreased the AUC values, presumably because they have the most information that was not present in the other factors.

Although we performed a multi-collinearity analysis before model building, the results of the sensitivity analysis revealed that a single multi-collinearity analysis may fail to fully protect a landslide modeling process from including slightly useful conditioning factors. Therefore, an integrated framework of model building and sensitivity analysis is desirable to identify those landslide conditioning factors that either introduce null usefulness to the model performance or decrease the reliability of the produced susceptibility maps (Jaafari et al. 2017a).

Conclusion

The field of LSM is one of the most popular areas of research. Various methods have been examined for this field of research by numerous researchers. In this study we applied widely accepted models, i.e., FR, WoE, and SI, for the purpose of production of a reliable map of landslide susceptibility. Using these models with the integration of a GIS provides a relatively flexible and easy-to-use framework to spatial prediction of landslides.

The validation results showed that the susceptibility map produced by the FR model has the highest prediction accuracy (80.44%), followed by the SI model (79.55%) and the WoE model (77.94%). Success rate curve also gives similar result, with FR model the highest AUC value (81.51%), followed by the SI model (81.27%) and the WoE model (79.43%). Further, the sensitivity analysis using the FR model revealed that the modeling process was sensitive to input conditioning factors as the exclusion of each of these factors changed the model performance. Overall, this comparative study showed that three models used have performed reasonably well in predicting the landslide susceptibility. These results can indeed greatly help planners and policy makers to adopt appropriate land use planning policies to guide the future developments of infrastructures in the area.