1 Introduction

Landslides cause a lot of loss of life in the world. At the same time, landslides are damaging the environment. In some countries, landslides can cause more lives and property losses than earthquakes, and floods (García-Rodríguez et al. 2008). Therefore, it is very important to identify landslide sensitive areas and to make LS maps (Mohammady et al. 2010). Landslide susceptibility (LS) mapping provide information to the relevant executive administrations to locate the landslide susceptible areas and make decisions about preventive plans (Mahdavifar 1997). Particularly in the last two decades, numerous methods have been developed and applied in LS studies. In recent years, LS methods combined with GIS have been used more frequently. These methods are summarized by Guzzetti et al. (1999). In this study, FR, AHP, ANN and FL methods have been used and compared for LS mapping in Taşkent (Konya, Turkey) Basin.

There have been many studies using the frequency ratios (FR) (Akgun et al. 2008; Dahal et al. 2008; Van Westen et al. 2008; Vijith and Madhu 2008; Oh et al. 2009; Ozdemir 2009; Pirasteh et al. 2009; Youssef et al. 2009; Pradhan and Youssef 2010; Regmi et al.2010; Yilmaz 2010; and Ozdemir and Altural 2013) and AHP method (Barredo et al. 2000; Mwasi 2001; Nie et al. 2001; Yagi 2003; Gorsevski et al. 2006; Yalcin 2008; Intarawichian and Dasananda 2010) in LS studies The fuzzy set based methods (Kanungo et al. 2008; Muthu et al. 2008; Pradhan et al. 2009; Pradhan 2010a, b), and ANNs methods (Melchiorre et al. 2006, 2008; Chen et al.2009; Pradhan et al. 2009; Pradhan and Lee 2010a, b; Poudyal et al. 2010; Pradhan and Pirasteh 2010; Pradhan and Buchroithner 2010; Pradhan et al. 2010a, b; Yilmaz 2010) have been implemented for landslide susceptibility zone (LSZ) studies. Although there are many methods in use on the mapping of landslide susceptibility, there is no consensus on the best method yet.

There are also many scientific studies on the comparison of LS methods (e.g., Suzen and Doyuran 2004; Ayalew et al. 2005; Brenning 2005; Yesilnacar and Topal 2005; Kanungo et al. 2006; Lee 2007; Meisina and Scarabelli 2007; Yalcin 2008; Magliuloü et al. 2009; Rossi et al. 2009; Yilmaz 2009; Miner et al. 2010; Poudyal et al. 2010; Pradhan and Lee 2010b, c; Van Den Eeckhaut et al. 2010; Yilmaz 2010; Akgun 2012; Pradhan 2011; Yalcin et al. 2011; Mohammady et al. 2012; Teimouri and Graee 2012; Zhu and Wang 2009; Ozdemir and Altural 2013;). For example, a comparative study of WOE, AHP, ANN, and Generalized Linear Regression procedures for LS mapping is presented by the Vahidnia et al. (2009). Kanungo et al. (2006) compared ANN, fuzzy combined neural and fuzzy weighting procedures. Lee (2005a, b), Lee and Sambath (2006), Lee and Pradhan (2007) and Oh et al. (2009) compared logistic regression with a FR approach and found out that logistic regression was generally more accurate. Pradhan and Lee (2010a) compared ANNs with FR and regression method. Similarly, Yilmaz (2010) compared ANNs with conditional probability, logistic regression and support vector machine and found logistic regression to be the most accurate. Nefeslioglu et al. (2008) compared logistic regression with an artificial neural network (ANN) and found that while the ANN has over-predicted the susceptibility whereas the logistic regression model has under-predicted. Paulin and Bursik (2009) combined the logistic regression model with an ANN in a GIS framework. Suzen and Doyuran (2003) made a comparison of bivariate and multivariate methods in the same study area.

The main objectives of the study are: (1) to prepare LS maps of Taşkent Basin; (2) to evaluate, compare and verify some LS mapping methods. In this study, AHP, FR, ANN and FL methods were studied and the results were compared; (3) to determine the best methods for evaluating the LS in the research area, and (4) to assess and map the LS areas based on the best method of the considered methods.

2 Study Area

The research area is located approximately between latitudes from 18° 06′ 00″ N to 18° 38′ 24″ N and longitudes from 98° 04′ 12″ E to 98° 38′ 24″ E, covering an area of about 80 km2 in the vicinity of Taşkent, Konya (Fig. 1).

Fig. 1
figure 1

Location map of study area

Full size image

Based on the meteorological data of 22 years in the vicinity of Taşkent region, In the region, the coldest month is February with an average of 6.7 °C and the hottest is August with 23.2 °C. The annual average temperature of the study area is 17.9 °C. According to precipitation measurements between 1975 and 2015, it was determined that precipitation height ranged between 587 and 700 mm. The annual average precipitation of Taşkent town is 838 mm. Precipitation usually falls in winter as snow and rain. In the Taşkent basin, the topographic altitude ranges from 1063 m to 2111 m. Slope angles range from 0° to 63° (average angle 19°). In the northern part of the Taşkent basin, the aspect of the slope is in general to the north and in the southern part of the basin to the north. Dominant vegetation includes grapes and mixed deciduous trees below 1500 m, a mixed pine forest from 1500 to 2000 m alternating with the hills, and an evergreen forest that extends up to 2100 m.

3 Materials and Methodology

The methodology used in this study includes seven steps: (1) Preparation of landslide inventory map of the study area from field surveys; (2) Identification of conditioning factors and the use of GIS to produce the landslide conditioning factor maps; (3) Application of various methods (FR, NN, FL and AHP) for the mapping of LS; (5) Use of GIS to produce the LS maps; (6) Testing reliability of the LS maps produced in a location of Taşkent (Konya-Turkey) vicinity and; (7) Comparison of results obtained from the LS maps.

In the following section, information about the landslide inventory map, used in the investigation, and the factors affecting the landslide formation, are given.

3.1 Preparation of Landslide Inventory Map

There was no landslide inventory map of the study area. Landslide scars areas were mapped by field studies. A total of 101 landslide areas (Table 1, Fig. 2) and 37 non-landslide areas were mapped and used in this study. Landslides combined cover an area of 72913 m2, accounting for 0.09% of the total study area. The minimum, mean and maximum landslide areas are 18 m2, 904 m2 and 20419 m2, respectively. The polygonal areas of landslides and non-landslides were converted into points using GIS software. This resulted in 377 points (276 points for landslides and 101 points for non-landslides). 76 random numbers were generated between 1 and 276, and the landslide point numbers corresponding to these generated numbers were used as the test landslides. Similarly, 38 random numbers were generated between 1 and 101, and the non-landslide point numbers corresponding to these generated numbers were used as the test non-landslide points (Table 2). In other word, the landslide and non-landslides points were randomly sampled from the inventory. In the FR, AHP and FL methods, 276 landslide points (200 for analysis and 76 for testing) have been used. In the ANN method, 263 points (200 landslide and 63 non-landslide points) for training and 114 points (76 landslide and 38 non-landslide points) for testing have been used.

Table 1 Coordinates of landslide points in the study area
Full size table
Fig. 2
figure 2

Landslide inventory map

Full size image
Table 2 Polygone and point landslide numbers and its percent in this study used
Full size table

3.2 Selected Landslide Conditioning Factors Used in the Susceptibility Mapping

The causes of landslides are directly or indirectly related to slope angle, slope aspect, stream power index, topographic wetness index, sediment loading factor, geology, distance from drainage system, distance from lineament, and land use (Poudyal et al. 2010). In this study, 14 factors which are considered to be effective in landslide formation were selected and used. These are elevation, slope gradient, slope aspect, plan curvature, profile curvature, sediment loading capacity index, stream power index, wetness index, drainage density, distance to the drainage, geology, fault density, distance to the fault and land use/land cover. Some of the landslide conditioning factors were derived from the geology map. The geology and land use maps are prepared by digitizing the hard copy maps manually.

3.2.1 DEM and Its Derivatives

The DEM was prepared by digitizing topographic maps of the area. Some landslide conditioning factors, such as elevation, slope aspect, slope angle, plan, and profile curvature, sediment loading factor, wetness index and stream power index were derived from the DEM.

3.2.1.1 Elevation

Due to the fact that precipitation, freezing and thawing phenomena are higher, more landslides can occur in high altitude areas (Menendez Duarte and Marqinez 2002; Lin and Tung 2003). However, due to the presence of more durable lithologies in high altitude terrain, sometimes fewer landslides can also occur in these areas. (Dai and Lee 2002; Zhou et al. 2002; Cevik and Topal 2003; Ayalew et al. 2004; Lan et al. 2004). Some researchers have suggested that topographic height has little contribution to the formation of landslides (Guzzetti et al. 1999; Lineback Gritzner et al. 2001; Ayalew and Yamagishi 2005). However, in the investigated area, there are higher strength lithologies in the high land sections. The range of elevation is from 1063 to 1961 m, gradually increasing from North-East to South-West (Fig. 3a). Landslides below 1508 m are dominant (94%) for lithology of the geological units that have clay compositions.

Fig. 3
figure 3

Some factor maps that affect landslide formation: a elevation; b slope; c slope aspect; d plan curvature; e profile curvature; f sediment loading factor; g stream power index, h wetness index; i drainage density and j distance to drainage

Full size image
3.2.1.2 Slope

The slope angle is the main factor in the formation of landslides (Lee and Min 2001; Fernandez et al. 2008; Mousavi 2017). The high slope angle produces a high shear force, which makes it easier to slip (Jones et al. 1961). Therefore, slope angle is considered as a factor controlling the formation of the landslide in the susceptibility studies (Dai et al. 2001a, b; Cevik and Topal 2003; Lee 2005a; Yalcin 2008; Nefeslioglu et al. 2008). The slope angle map was prepared for angles ranging from 0° to 63° (Fig. 3b).

3.2.1.3 Slope Aspect

Some authors reported that the slope aspect is the important factors in preparing LS maps (Guzzetti et al. 1999; Nagarajan et al. 2000; Saha et al. 2002; Cevik and Topal 2003; Lee et al. 2004; Lee 2005a). In the northern hemisphere, the north-facing slopes have less sunshine, and as a result, these slopes can be colder. On these slopes, snow can melt more slowly and as a result of this, the infiltration of water into the soil may be more (Mousavi 2017). For these reasons, the slope aspect is considered as an effective factor in the formation of landslides (Clerici et al. 2006). In this study, the slope aspect map of the investigated area was produced and used for LS analysis (Fig. 3c).

3.2.1.4 Plan and Profile Curvature

Ercanoglu and Gokceoglu (2002) defined that “the term curvature is generally defined as the curvature of a line formed by the intersection of a random plane with the terrain surface. A value of positive, negative and zero curvature indicate that the surface is upwardly convex, concave and flat at that pixel, respectively. In this study, plan and profile curvature raster maps, produced from DEM were classified in two classes as positive and negative curvature areas (Fig. 3d, e). In the Taşkent Basin, it was determined that the landslides occur more in negative curvature areas.

3.2.1.5 Sediment Transport Capacity (STI), Stream Power Index (SPI), and Topographical Wetness Index (TWI)

These indices are hydrological based terrain indices. Detailed information about these indices is given in Wilson and Gallant (2000) and Saaty and Vargas (2001). Details about the use of these indices in susceptibility maps are given by Pradhan (2010a, b, c), Yilmaz (2010) and Gritzner et al. (2001).

The STI index is used to express the erosion power of surface runoff water. This index gives information about the degree of soil erosion. The values of STI are found to vary between 0 and 232 and STI map is classified into five classes (Fig. 3f). SPI gives information about the eroding power of flowing water from the stream (Moore et al. 1993). The values of SPI are found to vary between 0 and 23420 and SPI map is also classified into five classes (Fig. 3g). TWI provides information on where more surface water can be collected (Nefeslioglu et al. 2008).

3.2.2 Proximity Maps (Distance to Drainage and Distance to Faults)

Two proximity-based factors evaluated in this study: distance to drainage lines and faults. As you approach the drainage lines, the water content of the slope material increases, and even becomes saturated with water. Again, as the drainage lines approach, the erosion increases in the slope toe. For these reasons, drainage lines may adversely affect slope stability (Gokceoglu and Gokceoglu 1996; Dai et al. 2001a; Saha et al. 2002). The drainage lines used in the study were extracted with GIS software (Tarboton 2003). In the investigated area, the distance to the drainage lines varies between 0 and 916 m (Fig. 3j). The discontinuity of the slope material increases in the areas close to the faults. The more discontinuities in the slope material, the easier the slope will slide (Van Westen and Bonilla 1990; Pachauri and Pant 1992; Luzi and Pergalani 1999; Uromeihy and Mahdavifar 2000; Saha et al. 2002). On the other hand, fault density and distance to faults affect the seismicity of the region. Especially in area with active faults or in area close to active faults, landslides may occur depending on the triggering of earthquakes (Mousavi et al. 2011; Mousavi et al. 2018). For this reason, seismicity of the region can also be evaluated as a factor in slope susceptibility studies especially in tectonically active regions. However, the study area is located in the region of lower seismicity in Turkey. Therefore, seismicity was not considered as a factor in this study. The faults have been drawn from the geological map. In the map of proximity to the faults, the distance values were found to be between 0 and 2233 m (Fig. 4b).

Fig. 4
figure 4

Some factor maps that affect the landslide formation: a fault density; b distance to fault; c land use and d geological map

Full size image

3.2.3 Density Maps (Fault and Drainage Density Maps)

Density maps, such as fault and drainage density maps, are a major parameter in LS analysis. They are generally used in producing LS maps (Pachauri et al. 1998, and Nagarajan et al. 2000). Drainage density was considered as an influential parameter (Suzen and Doyuran 2004; Melchiorre et al. 2010) in this study. Drainage lines are extracted from the DEM. The drainage density map was produced by calculating the cumulative lengths of stream segments of the drainage network falling within unit areas of 1 km2. They are grouped into five classes (Figs. 3i, 4a).

3.2.4 Land Use Map

Land use is the most commonly used human-induced factor on the slope instability (Tangestani 2004; Begueria 2006). This map is generally used in producing LS maps. This map provides information about the land cover. Again, this map provides important information about the purpose of the land use. Several researchers (Jakob 2000; Tangestani 2004; Begueria 2006; Bathurst et al. 2010) have revealed a clear relationship between vegetation cover and slope stability, especially for shallow landslides. Slopes lacking vegetation are more unstable than forest areas (Anbalgan 1992; Turrini and Visintainer 1998; Cannon 2000; Nagarajan et al. 2000; Dai et al. 2001a; Glade 2003).Land cover and land use are thought to be an effective factor in landslide formation. Therefore, this factor was chosen and evaluated in this study on LS. This map was prepared using Land Use (wealth, richness) map of Konya Province. The land use map is classified as follows: “residential area”, “forest”, “rocky land”, “pasture land”, “irrigated agriculture land”, and “dry agriculture land” (Fig. 4c).

3.2.5 Geology of the Study Area

Lithology and structure are very important in the formation of landslides. Many researchers have used geology and structure as a factor in LS analysis (Anbalgan 1992; Pachauri et al. 1998; Dai et al. 2001a; van Westen et al. 2003; Ayalew and Yamagishi 2005; Ayenew and Barbieri 2005; Ermini et al. 2005; Lee and Talib 2005).

Distribution of the geological units-formations is given in Fig. 4d. In the Tashkent region, the age of the rocks varies between Upper Devonian and Quaternary. The investigated region geologically consists of volcanic, metamorphic and sedimentary rocks. The detailed study of the geology of the Taşkent basin has been given by Turan (1990).

Bedrock in the region is represented by rocks of Gevne Group (Upper Devonian to Lower –Middle Triassic sedimentary rocks), Sinatdağı Group (Upper Permian to Upper Cretaceous), Hocalar Group (Triassic) and Taşkent Group (Mesozoic to Paleocene). These rock sequences are separated by thrust faults. Gevne Group is sub-divided into six formations as Asarlık Yaylasi, Yarıcak, Arpalık, Kuşakdağı, Gökçepınar and Göztaşı Formations. The Asarlık yaylası formation (Da) and its member, the Gölboğazı Limestone (Dag), are the oldest units in the study area. The Lately–Devonian aged Asarlık Yaylası formation consists of sedimentary units, such as dolomitized limestone, sandstone with quartzite, grey and black shale and white limestone. The Yarıcak Formation (Cy, Carbiniferous) and its member, Kirazpınarı limestone (Cyk, Carboniferous), conformably overlies to the Asarlık yaylası Formation. The formation consists of grey-black limestone and an alternating sequence of limestone with sandy, sandstone and limestone. This formation is mapped eastern parts of the investigated area. Yarıcak formation is overlain by the Arpalık Formation (Pa). Arpalık Formation (Lower Permian) mainly consists of greenish-yellow medium to thick limestone bedding with fossils, reddish-black yellow sandstone with quartzite. It is conformably overlain by the Kuşakdağı Formation (Upper Permian). Kuşakdağı Formation (Pk) consists of medium to thick, jointed, hard, grey-black limestone beds, sandy shale, sandy limestone and sandstone with quartzite. As can be seen from Fig. 4d, 35% of the total surface of the study area is covered by this formation. Gökçepınar Limestone (Lower Triassic) conformably overlies to the Kuşakdağı Formation. The Gökçepınar Limestone (Trg) is composed of grey and purple limestone beddings. Gökçepınar Limestone is observed in south western part of the study area. This formation is overlain by the Göztaşı Formation (Trgt, Lower-Middle Triassic). Gevne Group is thrusted over the Sinatdağı Group. Sinatdağı Group is sub-divided into five formations as Kahtepe Formation (Upper Permian), Kartallıca Limestone (Triassic), Sinatdağı Formation (Jurassic-Lower Crecacous), Türbetepe Limestone (Upper Cretaceous) and Söğütyayla Formation (Upper Cretaceous). Kahtepe Formation (Pka) consists of dark grey, partly oolitic and dolomitize limestone, shale and sandstone with limestone intercalations. Kartallıca Limestone unconformably overlies the Kahtepe Formation. Kartallıca Limestone (Trk) is formed of grey, purple crystallized limestone beds. The Sinatdağı formation (Tod) of Jurassic Lower Crecacous unconformably overlies the Kartallıca Limestone and consists of sedimentary units such as reddish sandy and muddy conglomerate, sandstone, siltstone, and mudstone. The Sinatdağı Formation (Jks) starts with a basal conglomerate on an erosional surface of pre-Jurassic rocks and continues upward with an alternation of various lithologies such as sandstone, siltstone, and mudstone. The conglomerates are grey- and yellow in colour. Pebbles are rounded to sub-rounded. At the top, it is overlain unconformably by Türbetepe limestone (Kt). Söğütyayla formation (Krus) starts with detrital sediments, laminated mudstone, conglomerate, sandstone, shale and dark grey clayey mudstone and continues upward with an alternation of mudstone and sandstone beds. At the top, it mainly consists of a flysch matrix and limestone olistolites. Sinatdağı group is thrusted over the Hocalar group. Hocalar Group is sub-divided into two formations as Zindancık meta-olistostrome-limestone blocks (Trhzb, Triassic) and Kayraktepe Quartzite (Trhk). The Triassic meta-olistostrome and limestone rocks mainly consist of a flysch matrix and limestone olistolites. The flysch matrix is also composed of meta-sandstone, phyllite and slate with quartzite, chlorite, marble, sandy mudstone, and recrystallized limestone. Kayraktepe quartzite mainly consists of dark grey, reddish quartzite and siltstones. Hocalar Group is thrusted over the Taşkent Group. Taşkent Group are sub-divided into three formations as Dedemli (Md, Mesozoic), Korualan (Mk, Mesozoic) and Taşkent ofiolitic complex (Kpt, Upper Cretaceous-Paleocene). Dedemli formation is formed of the various rocks such as tuff, tuffite, basic volcanic, limestone, and chert. Hocalar Group is tectonically overlies the Dedemli formation. Korualan formation starts with a medium to thick beds of limestone, clayey limestone intercalations, and continues upward with an alternation of various lithologies such as white dolomite and dolomitic limestone. At the top, it mainly consists of nodular cherty limestones. This formation is thrusted over the Hocalar group. The Taşkent Melange rocks mainly consist of a flysch matrix, serpentine, pyrocsenite, chert and limestone olistolites. The flysch matrix is also composed of sandstone and mudstone intercalations and conglomerate. The Taşkent Melange were folded and fractured. The rocks are intensively fractured. Quaternary alluvial discordantly overlie the Taşkent Melange rocks (Fig. 4d).

3.3 Susceptibility Mapping Models Used

In this paper, four different methods, FR, FL, AHP, and ANN, were applied to LS mapping analysis and their performances were compared. Details of these applications have been presented below.

3.3.1 Frequency Ratio (FR) Model

Bonham-Carter (1994) and Lee and Pradhan (2007) have been described Frequency Ratio (FR) method as follows “the ratio of relative frequency of landslide cells in a category (NLi/NL) to the relative frequency of all landslide cells in the area (NCi/NC).

$$FR = {{(N_{Li} /N_{L} )} \mathord{\left/ {\vphantom {{(N_{Li} /N_{L} )} {(N_{Ci} /N_{C} )}}} \right. \kern-0pt} {(N_{Ci} /N_{C} )}}$$
(1)

where NLi is the number of landslide cells in the class i. NL is total number of landslide cells. NCi is the total number of cells in the class i. NC is the total number of cells.

FR values greater than 1 indicate that landslide formation is high. FR values were calculated for each class of each factor (Table 3). Values of the FR were than summed to calculate the LS index (Lee and Pradhan 2007). Produced LS map was given in Fig. 5a.

Table 3 Data employed in the analyses and results obtained from theFR [total number cells of study area (NC = 797104), and total number cells of training landslides (NCi = 200). FR = (NLi/NL)/(NCi/NC], and AHP method
Full size table
Fig. 5
figure 5

Landslide susceptibility maps a FR model; b AHP model; c FL model; and d ANN model

Full size image

3.3.2 Analytic Hierarchy Processes (AHP)

The Analytic Hierarchy Process (AHP) is developed by Saaty (1980). Daneshvar and Bagherzadeh (2010) has been determined that “AHP is a multi-objective, multi-criteria decision-making approach to arrive at a scale of preference among a set of alternatives.” Yalcin (2008) describes the use of this method as follows: “this method is widely used in site selection, suitability analysis, regional planning, and landslides susceptibility analysis.” Factor weights for each criterion are determined by a pairwise comparison matrix as described by Saaty (1990, 1994) and Saaty and Vargas (2001). The detailed information for the application of the AHP for LS mapping was given by Barredo et al. (2000), Mwasi (2001), Nie H et al. (2001), Yagi (2003), Yalcin (2008), Moradi et al. (2012) and Teimouri and Graee (2012).

According to the Voogd (1982) the steps of the method are as follows: “(1) the complex unstructured problem is broken down into its component factors. The 14 landslide conditioning factors were chosen in this study; (2) each factor is classified into sub classes based on their relative influence on landslide; (3) the relative importance of each class of factors is quantitatively determined by pairwise comparison; (4) the weights of each class on each factor are calculated; (5) the weights of all of the factors are calculated; (6) values of the landslide susceptibility index (LSI) for each considered factor is calculated by summing each factor’s weight multiplied by the class weight of each referred factor (for that pixel).”

The obtained weights of conditioning factors (Wj) and the ratings (wij) of various classes are also given in Table 3. The LSI map is established according to Eq. 1. In this study, the CR is found to be 0.068, a ratio which indicates a reasonable level of consistency in the pair-wise comparison that was good enough to recognize the factor weights (Intarawichian and Dasananda 2010). For all the cases of the gained class weights, the CRs are less than 0.1, a ratio which indicates a reasonable level of consistency in the pair-wise comparison that was good enough to recognize the class weights (Pourghasemi et al. 2012; Intarawichian and Dasananda 2010).

$$LSI = \sum\limits_{i = 1}^{n} {\left( {W_{i} \, \times \,R_{i} } \right)}$$
(2)

where LSI is the landslide susceptibility index of a given pixel, Ri and Wi are the class weights. The factor weights are given in Table 3. The produced LS map is given in Fig. 5b.

3.3.3 Artificial Neural Network (ANN) Method and Its Implementation

The multi-layer ANN model is typically composed of three types of layers: input hidden, and output layers. ANN algorithms calculate the weights for the input values and for the layer nodes of input, hidden and output layers by introducing the input in a feed forward manner, which propagates through the hidden layer to the output layer (Pajanowski et al. 2002). The hidden layer gives an “activation function. It transforms the sum of the input data into feature using one of many possible functions, including exponential, sine, logistic (sigmoid), hyperbolic, and square root. The Multi-Layer Perceptron (MLP) with a feed-forward Back-error Propagation (BP) type of learning algorithm is a commonly used and are widely available neural network structure in the geospatial analysis (Pradhan et al. 2010a). Modelling of complex and nonlinear functions can be done through MLPs (Mousavi et al. 2016). The BP algorithm randomly selects the initial weights, and then compares the calculated output for a given observation with the expected output for that observation (Pajanowski et al. 2002). The differences between the expected and calculated output values across all observations are summed up using a loss function. After all observations are presented to the network, the weights are modified according to a generalized delta rule (Rumelhart et al. 1986), so that the total error is distributed among the various nodes in the network. These processes of feeding forward the features and back-propagating the errors are repeated iteratively (in some cases, many thousands of times) until the error stabilizes at a low level (Pajanowski et al. 2002).

According to the Chauhan et al. (2010) “there are three stages involved in ANN data processing for a classification problem: the training stage, the weight determination stage, and the classification stage. The training process is initiated by assigning arbitrary values to the connection weights which are constantly updated until an acceptable training accuracy is reached.” The details of the method were given by Paola and Schowengerdt (1995), and Atkinson and Tatnall (1997).

The method was used to the determination prediction of LS in four phases: (1) designing of the network and preparing the inputs data; (2) training the network using the training set; (3) testing the network using the test data set; and (4) using the information from the neural network to forecast landsliding. For this study, the neural networks were constructed using the neural network toolbox of MATLAB (Pao 1999).

In this study, similar to the earlier studies (e.g., Nelson and Illingworth 1990; Haykin 1994; Masters 1994; Dowla and Rogers 1995; Looney 1996; Swingler 1996; Chauhan et al. 2010), we used 70% of the samples in the dataset as the training and 30% as the test sets.. The dataset consisted of 377 points denoting the presence and the absence of landslides. 263 points (200 landslide and 63 non-landslide points), 70% of the dataset) were kept for training (Table 2). The MLP_BP learning algorithm is preferred for the analyses. Learning rate and momentum factor were selected as 0.01 and 0.9, respectively (Polykretis and Chalkias 2018).

Detailed information on the application of the method to LS studies can be obtained from Aleotti et al. (1996), Mayoraz et al. (1996), Lee et al. (2001), Fernandez-Steeger et al. (2002), Ermini et al. (2005) and Bartolomei et al. (2006).

In this study, 14 inputs, 28 hidden, and 1 output neuron structure has been used for the network (Fig. 6). The initial weights were randomly selected. The numbers of epochs in the software were set to 2000, and the root mean square error (RMSE) goal for the convergence criterion was set to 0.01. All of the iterations satisfied the 0.01 root mean square error condition with 2000 epochs. After the training, the model become ready to be used for the study area.

Fig. 6
figure 6

Input and output layers of the ANN network

Full size image

Finally, the LS map was prepared using the training set (Fig. 5d). The values were classified and grouped into four equal areas (first highest 10%, second highest 10%, third highest 20% and the remaining 60%) for visual interpretation. Some properties related to the application of artificial neural networks are also given in Table 4.

Table 4 ANN model characteristics used
Full size table

5000 points were randomly generated for the study area. The produced 5000 points were crossed with 14 factor maps and values of 5000 points on the factor maps have been determined. The newly determined (= 5000 × 14) 70000 data points were used in the simulations performed by the trained network. The sensitivities of the landslides corresponding to 5000 points were determined from the artificial neural network. The determined 5000 data points were interpolated to create a LS map (Fig. 5d). LS map was classified into 4 classes of LS regions: very low (no LS), low, medium and high.

3.3.4 Fuzzy Logic (FL) and Its Application to Landslide Susceptibility

In this paper, LS in the basin of Taşkent was also examined using fuzzy logics. For this purpose, reasoning engine named Mamdani has been utilized. Using the factors controlling the formation of landslides given above, LS map is also created by Fuzzy Logic (FL) model. The analyses were performed using MATLAB fuzzy toolbox and GIS software.

According to the Karkazi et al. (2001) “in order to solve a problem using a knowledge-based fuzzy system, it is necessary to describe and process the influencing factors in Fuzzy terms and provide the result of these processes in a usable form.” The application steps of this method are given by Mendel (1995) as follows: “The basic elements of a knowledge-based FL consist of four major parts: (1) fuzzier, (2) rules, (3) inference engine, and (4) defuzzifier. Firstly, a crisp set of input data are gathered and converted to a fuzzy set using fuzzy linguistic variables, fuzzy linguistic terms and membership functions. The assignment of a membership function to every variable of the problem is called fuzzification process.” The concept of the linguistic variable illustrates particularly and clearly how fuzzy sets can form the bridge between linguistic expression and numerical information (Karkazi et al. 2001). This step is known as fuzzification. Membership functions are used in the fuzzification and defuzzification steps of a FLS, in order to map the non-fuzzy input values to fuzzy linguistic terms and vice versa. A membership function is used to quantify a linguistic term. There are different forms of membership functions such as triangular, trapezoidal, piecewise linear, Gaussian, or singleton (Saraswathi and Senthil 2017). However, the most common types of membership functions are triangular, trapezoidal, and Gaussian shapes. The type of membership function can be context dependent and is generally chosen arbitrarily according to the user experience (Mendel 1995). The functions used in the production of Fuzzy membership values are given in Eastman (2006).

In this study, the triangle shape was selected for all the membership functions after several trials and errors to find out the best function shape. Using the FR coefficients and the identified landslides, the fuzzy membership functions were constructed. The second step in this method is to determine how the inputs and outputs are connected to each other. A fuzzy rule is a simple IF–THEN rule with a condition and a conclusion. The evaluations of the fuzzy rules and the combination of the results of the individual rules are performed using fuzzy set operations.” Zimmerman (1996) discussed a variety of combination rules, and Bonham-Carter (1994) discussed five operators, namely the fuzzy and, fuzzy or, fuzzy algebraic product, fuzzy algebraic sum and fuzzy gamma operator. The mostly- used operations for OR and AND operators are max and min, respectively (Mendel 1995). This study uses the fuzzy OR operator for combining the fuzzy membership functions. After the inputs were fuzzified, If–Then rules have been assigned (Riad et al. 2011). These if–then rule statements are used to formulate the conditional statements that comprise fuzzy logic (Pao 1999). 112 rules (if–then) were assigned in this study, Fuzzy OR operator was used in all the rules. After the inference step, the overall result should be defuzzified to obtain a final crisp output. Defuzzification is performed according to the membership function of the output variable (Alreshoodi et al. 2015). There are different algorithms for defuzzification. The frequently used algorithms are Center of Gravity, Center of Gravity for Singletons, Left Most Maximum, and Right Most Maximum (Mendel 1995). The most popular defuzzification method which was also applied in here is the centroid calculation which returns the center of the area under the curve. As a result of the analysis, a map showing the calculated fuzzy values was provided. These fuzzy values have been “defuzzyfied” into different LS levels. Defuzzyfication means to translate the calculated fuzzy values back to the “real world” (Schernthanner 2007). In LS mapping, fuzzy logic defines the instability factors as members of a set ranging from 1, expressing the highest susceptibility, to 0, expressing no susceptibility of landsliding, allowing different degrees of membership (Schernthanner 2007). Finally, the produced map (Fig. 5c) was compared with the existing landslide location map for verification of the prediction accuracy.

4 Validation of the Models

The validity of the produced LS maps was checked by test landslides which were not used in the production of LS maps. Again, the Receiver Operating Characteristics (ROC) curves method was used in the comparison process. As a result of this comparison, the performances of the LS maps produced by different methods were determined.

ROC values are obtained by plotting the cumulative True Positive Rates (TPR = sensitivity) versus False Positive Rate (FPR = 1–specificity) for each model (Marjanovic and Caha 2011). Figure 7 shows the ROC curve of the FR, FL, AHP and ANN model. The AUC values of the ROC curve obtained from FR, ANN, AHP, and FL models were found to be 0.926, 0.916, 0.899, and 0.842, respectively, with estimated standard errors between 0.023 and 0.039 (Fig. 7, Table 5). LS map produced with FR model has higher accuracy. The models evaluated in this study had a quite reasonable performance considering the results of the similar studies available in the literature (Kanungo et al. 2006; Oh and Lee 2010; Pradhan and Lee 2010a, 2010b; Yilmaz 2010). Obtaining a relatively better results using the FR method (AUC of 92.6%) compared to the other methods is also in agreement with previous studies.

Fig. 7
figure 7

ROC curves of the FR, FL, AHP and ANN model

Full size image
Table 5 Area under the curve
Full size table

The accuracy of the LS maps has been also checked by overlaying them with the landslide inventory map. Locations of landsides have been intersected with the susceptibility maps and the number of coincident landslides was determined for each susceptibility class. A landslide was classified as the ‘‘correct’’ prediction when at least a part of it was located within a high or very high probability zone (Dai and Lee 2002). Otherwise, the prediction was considered as the ‘‘wrong’’ one. The 58 test landslides (76 pixels), 43 training landslides (200 pixels), and 101 total landslides (276 pixels), were overlaid onto the LS maps produced, and the existing landslide pixels within the different landslide occurrence potential classes were determined. The results of these inspections are given in Fig. 8. The results obtained by overlaying the predicted LS maps produced by AHP, FR, FL, and ANN methods onto all existing landslides, indicated that 87%, 83%, 89%, and 76% of the observed landslides were concentrated in the high and very high LS classes, respectively (Fig. 8). For the test landslides, these percentages were 74%, 60%, 87%, and 33% for AHP, FR, FL, and ANN methods, respectively. During the validation process, it was shown that 93%, 91%, 89%, and 93% of the training or analysis landslides were predicted correctly, as shown in Fig. 8. Poudyal et al. (2010) found that the accuracies of the LS maps produced by the frequency ratio and neural networks methods had been 82.21 and 78.25%, respectively. In addition, the validation results of the study by Mohammady et al. (2012) showed that “the prediction accuracy of the Frequency Ratio, Dempster–Shafer, and Weights of Evidence models had been 80.13%, 78.32%, and 74.60% respectively.” In the determination of landslide sensitivity, The FR method was found to be reliable.

Fig. 8
figure 8

The distribution of the produced landslide susceptibility classes of test, analysis and all landslides

Full size image

The LS maps produced by different models have also been compared with each other (Table 6). Pearson’s correlation coefficients for AHP versus FR, AHP versus FL, and AHP versus ANN were found to be 0.82, 0.63, and 0.54, respectively. The correlation coefficients were found to be 0.70 and 0.48 for FR versus FL and FR versus ANN maps, respectively. The correlation coefficient was found to be 0.45 for FL versus ANN map. Interpretations of correlation coefficients are given by ILWIS, (2001). According to this definition, if the correlation coefficient is between 0–0.2, 0.2–0.4, 0.4–0.7, 0.7–0.9, and 0.9–1, the correlation is evaluated as very weak, weak, medium, high and very high, respectively. As can be seen from the results, there are high correlations between the maps produced by FR-AHP, and FR-FL methods. The susceptibility map produced by the ANN model was compared with the susceptibility maps produced by other models. As a result of this comparison, the relationship was found to be weak. It is evaluated that the low number of landslides (the small of the training set) can be effective in the poor performance of the ANN method. Detailed information on increasing the performance of the ANN method can be obtained from Mousavi et al. (2016) and Pradhan and Lee (2010a). However, when the LS map produced by the FR model was compared with the sensitivity maps produced with other models, it was found that the correlation coefficient was high.

Table 6 Matrix of correlation coefficients obtained from the comparison of landslide susceptibility maps produced by different methods
Full size table

5 Results and Discussions

In this study, four models: (1) the FR, (2) the AHP, (3) the ANN, and (4) the FL have been used to derive the relationships between the landslide distribution and the landslide conditioning factors. The existing and produced landslide inventory maps were compared to each other for the verification of the results. Details of this LS map comparison have been described and given below (Fig. 9a–n).

Fig. 9
figure 9

Distribution of landslides into the classes of conditional factors

Full size image

Many landslides occurred at the elevations between 1063 and 1508 m, concentrated at elevations between 1063 and 1317 m (Fig. 9a). There are no or few landslides at elevations between 1508 and 1961 m. The results were dependent on the geology of the investigated area. In this area, the elevation is in a range of 1063–1508 m. The rocks in this area are generally volcanic rocks with high strength.

Many landslides occurred in the study area at slopes between 14° and 33° (Fig. 9b). There are no or few landslides at gentle (slope angle < 6°) and steep slopes (slope angle > 43°). Due to the small shear stresses, landslides are less common on the gentle slopes (Lee and Talib 2005). The shear stresses increase with increasing slope angle. On very steep slopes, there is generally no soil; there is a high-strength rock.

Therefore the landslide susceptibility of these slopes are low (Jones et al. 1961; Pradhan 2010a, b). The 14°–33° slope categories were the most unstable areas in the study area. It can be explained by the fact that most of the geological units with low strength were outcropped at slope angles between 14° and 33°.

Based upon the landslide distribution, the slopes facing South (113–203) and North (293–338) were considered being prone to landslides more than the others. This condition may be caused by the humidity (Fig. 9c).

Profile curvature affects the acceleration and deceleration of flow across the surface (Fig. 9e). Plan curvature affects the convergence and divergence of flow across the surface (Fig. 9d). The landslides in the study area typically had more tendencies to occur in curvature areas with a negative plan (convergence) and with negative profile (convex). The convex areas in the study possess higher LS values than the other areas possess due to curvature. Convex slopes can be exposed to repeated dilation and contraction of loose debris on an inclined surface that can induce creeping or a mudslide due to heavy rainfall (Pradhan 2011). On the other hand, the results of this study are in agreement with those of Havenith et al. (2006), where the following conclusions were made:” (1) landslide scars match convex areas well (Conforti et al. 2012). (2) Convex slopes are less stable under similar hydrogeological conditions (lower factor of safety) because a larger body (larger driving force) is acting on the same sliding surface (equal resistant forces). (3) Convex morphologies might indicate the presence of accumulation material (colluvium) that is characterized by lower shear resistance (Havenith et al. 2006).

LS is higher in areas where the STI ranges from 10 to 60 and lower in areas where the index values are between 0–10 and 60–232 (Fig. 9f). On the other hand, it is determined that landslides are abundant more at the classes of low STI index values than in other classes. In the case of using a stream power index (Fig. 9g), the landslide occurrence was higher for lower stream power index values. High TWI values indicate areas that are more likely (higher probability) to drain by the saturated excess flow. Threshold values for classifying areas where saturation excess overland flow will occur, though based on empirical soil properties, are 0.6 (normalized value) and above (Leh et al. 2008). The landslides had more tendency to occur TWI values between 4 and 8 (0.26–0.53 normalized). For lower or higher values of wetness index, the LS areas are small, whereas for middle values, the areas are large (Fig. 9h). As for the relationship between landslide occurrence and distance from the faults, the landslide frequency generally decreases with increasing distance from fault (Fig. 9m). This study indicates that most landslides occur within a fault distance of 0–250 m, whereas previous studies suggested that most landslides occur within a fault distance of 250–1000 m (Gupta and Joshi 1990; Gokceoglu and Gokceoglu 1996; Pachauri et al. 1998). Khanh (2009) concluded that landslides are more in the zone where the distance to faults is between 250 and 1000 m. However, there are many researchers who state that this determination may not be accurate for every place and every terrain (Gemitzi et al. 2010; Uromeihy and Mahdavifar 2000). The distance to the drainage map shows that landslides are closely located within the 100 m buffer zone (Fig. 9j). As for the relationship between landslide occurrence and the distance from drainage, the landslide frequency generally decreases with the increasing distance from a drainage line. A distance of less than 100 m indicates a high probability of landslide occurrence, whereas distances more than 300 m indicates a low probability. Dai et al. (2001a) has also indicated that “streams can adversely affect the stability by eroding slopes or saturating the lower part of the material until the water level increases. Intarawichian and Dasananda (2011) have concluded that “landslide occurrence probability is high in areas where the distance to drainage is between 1500 and 2000 m and the landslides are likely to occur most often at specific distances from the drainage.” Fourniadis et al. (2007) most likely drew the same conclusion after observing the terrain modification caused by gully erosion and the undercutting of slopes in the study area that might have influenced the initiation of landslides. The fault density values are found to vary between zero and 400.5 m/km2. It is determined that many of the landslides are located in the 0–300 m/km2 fault density zone (Fig. 9l). In the study area, it was determined that there is a linear relationship between drainage density and landslide formation (Fig. 9i). This result is suitable with the findings of Sarkar and Kanungo’s (2004). In mountainous regions, drainage density provides an indirect measure of groundwater conditions which have an important role to play in landslide activity (Sarkar and Kanungo 2004). Landslides are mostly formed in agriculture areas (Fig. 9n). It was determined that the formation of the landslides decreased with the increase in altitude, sediment loading factor, plan curvature, profile curvature, stream power index, distance to drainage, distance to fault, and fault density values. The formation of the landslides is found to increase up to a certain slope, sediment loading factor, and wetness index value, but then decline. Many landslides were formed in Aslakyayla (Da) and Taşkent (KPt) formations (Fig. 9k).

The generated LS maps are divided into 4 subclasses, low, medium, high and very high susceptibility classes. Landslides were also divided into two subclasses as test and analysis landslides. The distribution of the landslides, all the landslides, the test landslides that are not used in the analysis, and the analysis landslides that are used for the analysis are given in Fig. 8 according to the landslide subclasses. It has been determined that 87%, 83%, 89% and 76% of the analysis landslides are located in the high-to-high LS subclasses indicated in the LS maps produced by the AHP, FR, FL and ANN methods, respectively. It was also determined that 74%, 60%, 88% and 33% of the test landslides are located in the high-to-high LS subclasses indicated in the LS maps produced by the AHP, FR, FL and ANN methods, respectively. Similarly, It was also determined that 93%, 91%, 89% and 93% of the analysis landslides are located in the high-to-high LS subclasses indicated in the LS maps produced by the AHP, FR, FL and ANN methods, respectively (Fig. 8).

6 Conclusions

In this study, LS maps were created with four different methods, FR, FL, AHP and ANN, using the data obtained from the same field. The accuracy of the produced susceptibility maps was tested. The highest ROC value was obtained from the sensitivity map produced by the FR method, whereas the lowest ROC value was obtained from the map generated by the FL method. Later on, LS maps produced by different methods were compared with each other. It has been determined that the highest and lowest correlation values were obtained for FR and AHP correlation as 0.82, and ANN and FL correlation as 0.45, respectively. Following these evaluations, it was found out that it was more appropriate to apply the FR as the primary and the AHP as the secondary method in the production of the LS maps. Meanwhile, the FR method is simple, can be easily implemented, and provides very high accuracy. The accuracy comparison of the FR method with LR and WOE methods in the production of the LS maps was performed by Ozdemir and Altural (2013). They found out that the FR method produces more consistent results than the others. As a result, it is recommended to use the FR method in LS studies.