Introduction

Landslides, the most common geohazard occurring in mountainous terrain, can cause serious damage to natural resources and infrastructures, as well as loss of life and property (Pourghasemi et al. 2012a). The landslide hazard of an area is determined by the combined effects of intrinsic as well as extrinsic variables (Dahal and Dahal 2017). The intrinsic variables that influence the susceptibility to landslides include the geology of the bedrock, the geomorphology, the soil type and depth, the slope aspect and angle, the convexity and concavity of the slope, the elevation, the land-use pattern, the drainage pattern, and the engineering properties of the slope material (Dahal et al. 2008). The extrinsic variables that trigger landslides are heavy rainfall, earthquakes, volcanoes (Dahal et al. 2008), and human activities (Trianawati and Rahmafitria 2016). Human activities such as excavation work and tunnelling can change the natural equilibrium of slopes and lead to slope failures (Achour and Pourghasemi 2020). The duration of rainfall triggers landslides in a different manner. Slow-moving deep-seated landslides (e.g., earthflows, slumps) are activated by prolonged rainfall which lasts for several days or weeks, whereas fast-moving shallow landslides (e.g., debris flow) are triggered by individual high-intensity rainfall events (Sidle 2007).

The most landslide-vulnerable areas in India include the Meghalayan plateau, the Western Ghats, and the Himalayan Mountain regions, which together constitute about 15% of the total land area (NDMA 2009). This is a matter of concern because of the rapid population expansion in the mountainous region (van Westen et al. 2012). A large number of landslides occurred in the Western Ghats region in Kerala during the 2018 and 2019 monsoons and caused substantial loss of life and property damage. The damage to roads, buildings, and livestock was particularly high. One of the best strategies to reduce landslide-related damage and loss is the preparation of a landslide susceptible zone map. Susceptible zone maps are important in mitigating landslide risks as they show the zones of potential slope failures. The land use data should be updated every 1–10 years, which depends on the dynamics of land use change in the area (van Westen et al. 2008). Land use data is an important factor used in susceptible zone mapping and, therefore, the map too needs to be updated whenever changes in land use occur.

The methods used to create the susceptibility maps can be either quantitative or qualitative (Guzzetti et al. 1999). Qualitative (subjective) methods depend on expert knowledge (Aleotti and Chowdhury 1999) and quantitative (objective) methods use numerical tools to assess the correlation between landslides and the causative/conditioning factors (Guzzetti et al. 1999). Quantitative methods include weights of evidence (Pradhan et al. 2010a; Prasannakumar and Vijith 2012; Polykretis and Chalkias 2018; Quinn et al. 2010), logistic regression (Hemasinghe et al. 2018; Lombardo and Mai 2018; Oh et al. 2018), artificial neural networks (Lee 2007, Ortiz and Martínez-Graña 2018, Pascale et al. 2013, Shahri et al. 2019), and AHP (Achour et al. 2017; El Jazouli et al. 2019; Nguyen and Liu 2019; Yalcin 2008). Researchers have also used methods such as fuzzy logic (Fatemi Aghda et al. 2018; Sharma et al. 2013), Shannon’s entropy (Sharma et al. 2012, 2015), random forest (Catani et al. 2013; Lai et al. 2018; Taalab et al. 2018; Youssef et al. 2016), decision tree (Nefeslioglu et al. 2010; Zhang et al. 2017), support vector machine (Ballabio and Sterlacchini 2012; Lee et al. 2017; Pourghasemi et al. 2013), and Naive Bayes classifier (Bui et al. 2012; Lee et al. 2019) to produce landslide vulnerability/susceptibility maps.

The AHP (Saaty 1980) is a semi-quantitative analytical method for decision making where decisions are made using weights by relative comparison in pairs. This method has been used by many researchers to delineate landslide susceptible zones. El Jazouli et al. (2019) prepared the landslide susceptibility map of the Oum Er-Rbia high basin in Morocco using the AHP method. They used eight conditioning factors such as slope gradient, elevation, slope aspect, land cover, lithology, distance to road, distance to fault, and distance to drainage network. Semlali et al. (2019) created the landslide susceptible zone map of the Oued Laou watershed in Morocco using the same method using the factors slope, aspect, land cover, drainage density, lithology, distance from river, distance from fault, and distance from road. FR is a bivariate statistical method widely employed in demarcating susceptible zones (Jana et al. 2019; Oh et al. 2017; Sharma et al. 2014; Silalahi et al. 2019; Yilmaz 2009).

No other researchers attempted to demarcate landslide susceptibility zones of this data-scarce region in the Western Ghats, which employed two different models (AHP and FR). There are no validated susceptibility maps available for the Western Ghats region, which are created using multi-criteria decision-making methods. Abrams et al. (2018) pointed out that the models like AHP and FR are suitable for data-scarce regions. Also, updated and validated landslide susceptibility maps prepared employing two different methods, which were proved effective in other regions, are helpful in choosing the best one for planning and policy-making purposes. The impact of road cuttings on slope failures in this region was not assessed by other researchers, and hence the findings of this study will be valuable reference material for researchers, engineers, geologists, and the government in preparing proper guidelines with suitable regulations for the betterment of the communities. This study is intended to develop a framework to demarcate the landslide susceptible zones using GIS techniques for the Western Ghats region of Kollam district, Kerala, India, and to assess the impact of road cuttings on slope failures. For this study, ten intrinsic factors have been selected and their influence on landslide initiation will be discussed. The conditioning factors include slope angle, elevation, slope aspect, stream buffer, lithology, soil, road buffer, normalized difference vegetation index (NDVI), land use and land cover (LULC), and lineament buffer. Both the AHP and FR methods will be used to determine the weights, and the outputs will be compared to determine the prediction capability.

Materials and methods

Study area

This area is located in the southern Western Ghats (WG), a biodiversity hotspot in India (Fig. 1). It lies between longitudes of 76˚50’0"E and 77˚15’0"E and latitudes of 8˚40’0"N and 9˚15’0"N, and is around 775.18 sq. km. This area is bound by the Thiruvananthapuram district in the south, Tamil Nadu state in the east, and the Pathanamthitta district in the north. The Thenmala Dam and Reservoir are located within the study area.

Fig. 1
figure 1

Location of the study area

Full size image

Framework for the landslide susceptibility assessment

The framework for the landslide susceptibility assessment is categorized into three phases: desk studies, analysis, and field studies (Fig. 2). This framework was a modified one from Roslee et al. (2012). The literature review, data collection (satellite images, secondary data like published maps, etc.), selection of methods and factors (suitable factors will be selected based on the geo-environmental conditions of the study area, and methods will be selected based on the quality and quantity of data), and creation of landslide inventory (data on landslide affected sites will be collected from available data portals and published documents) are part of the desk studies phase. The analysis phase comprises the creation of thematic layers using the GIS software; reclassification of thematic layers and input of weights/ratings to the factor classes, input of weights derived using the methods adopted to the factors; creation of a landslide susceptibility map; classification or categorization; and validation of the susceptibility map. The field study phase comprises ground truthing of the LULC classification and collection of locations of landslide affected sites using GPS. The AHP and FR based landslide susceptibility modelling are mentioned in the following sections:

Fig. 2
figure 2

Framework for the landslide susceptibility modelling

Full size image

Data used

The landslide susceptibility modelling involves the following steps:

  1. i.

    The 10 conditioning factors were acquired from a variety of sources (Table 1). ArcGIS 10.4 and ERDAS Imagine 9.2 were employed to prepare the thematic layers of these conditioning factors.

  2. ii.

    After deriving the layers, the landslide incidence data provided by the National Remote Sensing Centre (Hyderabad, India) and verified using Google Earth Pro was used as a validation dataset.

  3. iii.

    The AHP and FR models were used to produce the susceptibility maps (Fig. 3).

    Fig. 3
    figure 3

    Flowchart – landslide susceptibility modelling

    Full size image
  1. iv.

    The ROC curve method was used to validate the created susceptibility maps. The ROC curves were created using the SPSS software.

Table 1 Data source
Full size table

Causative factors

Slope angle

One of the most critical landslide-inducing factors is slope angle. Low gradients are often associated with reduced shear stresses, so gentle slopes should have a low landslide frequency (Akgun et al. 2008). The increase in slope angle increases the probability of the occurrence of landslides (Achour et al. 2017), as the shear stress is directly proportional to the slope angle (Lee et al. 2004). Using ArcGIS spatial analyst (surface) tools, the slope angle was derived from the DEM. The slope (angle) of the area is categorized into 5 classes: 0-8.47°, 8.47–15.89°, 15.89–23.05°, 23.05-31.00°, and 31.00-67.57° (Fig. 4).

Fig. 4
figure 4

Slope angle

Full size image

Elevation

The probability of landslide initiation is greater at higher elevation areas (Achour and Pourghasemi 2020). The elevation was computed from the DEM using the surface tool available with ArcGIS. The elevation of the area is divided into 5 classes: 18–236 m, 236–455 m, 455–727 m, 727–1079 m, and 1079–1759 m. The elevation classes are represented in Fig. 5.

Fig. 5
figure 5

Elevation

Full size image

Slope aspect

Slope aspect, which represents the direction of the highest slope of the terrain (Xu et al. 2012), is another important factor used in stability assessment of hillslopes. The solar influences are greater on the southern and western aspects (Setiawan et al. 2004), and these slopes have less vegetation density and are subjected to a higher rate of erosion (Elmoulat and Ait Brahim 2018). Erosion is further accelerated by the higher intensity of rainfall. The westward and south-westward slopes are more vulnerable to soil erosion since the wind blowing from these directions brings more rain than the wind from the east (Beullens et al. 2014), making these slopes more susceptible to landslides. The slope aspect was extracted from the DEM employing ArcGIS surface tools, and is grouped into nine (Fig. 6).

Fig. 6
figure 6

Slope aspect

Full size image

Stream buffer

Landslides can be triggered by terrain alteration induced by gully erosion (Pradhan et al. 2010b). Streams can wreak havoc on the stability of slope materials by undercutting or saturating them, lowering their shear resistance (Dai et al. 2001; Saha et al. 2002). Therefore, the areas closer to stream channels are more vulnerable to landslides. The stream networks of the study area were digitized from the topographic maps and a 100 m buffer distance was derived using spatial analyst tools (Fig. 7).

Fig. 7
figure 7

Stream buffer

Full size image

Lithology

The strength and permeability of rock depend on the rock type (Achour et al. 2017), the degree of weathering, as well as the presence and type of discontinuities (such as joints or fractures) (Ajin et al. 2016). Stronger rocks are less susceptible to landslides because they are more resistant to the driving forces (Kanungo et al. 2006). The lithology was extracted from the geological map using ArcGIS tools. The rocks (Fig. 8) present in this region includes charnockite, garnet-biotite gneiss with migmatite, garnetiferous biotite, pink granite gneiss, and quartzo-feldspathic gneiss.

Fig. 8
figure 8

Lithology

Full size image

Soil

Soils in the Western Ghats with higher clay content typically have high porosity and low permeability, so they can hold more water. The shear strength of soil gets much reduced when the water pressure in the pores increases; this leads to slope failures (Chawla et al. 2018). Landslides are influenced by the presence of clayey soil, which functions as a potential slip zone (Sartohadi et al. 2018). The soil texture of this area was extracted from the soil map using ArcGIS tools (Fig. 9).

Fig. 9
figure 9

Soil

Full size image

Road buffer

Road cuts are sites of anthropogenic instability (Pradhan et al. 2010b). The slope equilibrium can be affected by excavation for road construction and increased loads due to vehicle movements (Pourghasemi et al. 2012b). The shear strength at the slope’s toe can be reduced by road cuts (Semlali et al. 2019) and can result in landslides. The topographic maps were used to digitize the road networks, which were then updated using Google Earth Pro data. A 100 m buffer distance was derived using ArcGIS spatial analyst tools (Fig. 10).

Fig. 10
figure 10

Road buffer

Full size image

Normalized difference vegetation index

NDVI quantifies the vegetation density was derived from the satellite image using the ArcGIS raster calculator tool. NDVI was derived using Eq. 1 (Rouse et al. 1974) and ranges between − 1 and + 1. Higher NDVI (close to + 1) represents dense vegetation, whereas lower NDVI represents sparse vegetation. The NDVI less than zero indicates water and NDVI close to zero indicates bare land, rock or urbanized areas (Saravanan et al. 2019; Viana et al. 2019). The probability of landslide occurrence is higher in areas with lower NDVI (sparse vegetation). The NDVI of this area (Fig. 11) is categorized into five (–0.06–0.10, 0.10–0.26, 0.26–0.32, 0.32–0.38, and 0.38–0.63).

Fig. 11
figure 11

Normalized difference vegetation index

Full size image
$$\text{NDVI}=\frac{(\text{NIR}-\text{R})}{(\text{NIR}+\text{R})}$$
(1)

where NIR and R stand for spectral reflectance measurements acquired in the near-infrared and visible regions, respectively.

Land use and land cover

Areas with thicker vegetation are less prone to landslides. This is due to the fact that vegetation with their root systems can stabilize the slopes (García-Rodríguez et al. 2008) by binding the soil particles, which increases soil shear strength (Turrini and Visintainer 1998). The sparsely vegetated or barren areas are more prone to sliding as these areas are subjected to weathering (Anbalagan 1992). Changes in the land cover on hill slopes can induce slope failures (Kanungo et al. 2009; Karsli et al. 2009). The land use and land cover types were derived from the Landsat image. The maximum likelihood (ML) classification (Mohajane et al. 2018) tool available in the ERDAS Imagine was applied to categorize the satellite image. The land use and land cover types (Fig. 12) of this area include agricultural land, evergreen forest, deciduous forest, barren land, water bodies, built-up areas, scrubland, and mixed vegetation.

Fig. 12
figure 12

Land use/land cover types

Full size image

Lineament buffer

Along faults, landslides are more common (Semlali et al. 2019). These tectonic breaks can decrease the strength of rocks (Devkota et al. 2013). Lineaments initiate the weathering process and thus disintegrate and decompose the rocks (Mandal and Maiti 2014). Under saturated conditions, these weathered materials may lose their consistency and initiate sliding. The lineaments were downloaded from the Bhukosh portal and the 100 m buffer distance was calculated using spatial analyst tools (Fig. 13).

Fig. 13
figure 13

Lineament buffer

Full size image

AHP modeling

The AHP model was developed by Saaty (1980) and is the most widely used technique for decision making in complicated contexts where numerous variables or criteria are taken into account while prioritizing and selecting alternatives. In the AHP modelling, a 1–9 scale is used to construct judgement matrices (Saaty 1980). The crucial processes involved in the modelling are the creation of a pairwise comparison matrix (Table 2), the estimation of Eigen value, Eigen vector, and weighting coefficient (Table 3), and the estimation of the consistency ratio (Table 4) to check the consistency (Nikhil et al. 2021).

Table 2 Pairwise comparison matrix
Full size table

The Eigen vector (Vp), and weighting coefficient (Cp) were determined applying Eqs. 2 and 3 as in Nikhil et al. (2021)

$$\text{V}_\text{p}=\sqrt[k]{W1\times \dots Wk}$$
(2)

Where k = number of factors; W = ratings of the factors.

$$\text{C}_\text{p}=\frac{Vp}{Vp1+ \dots \dots . Vpk}$$
(3)
Table 3 Normalized matrix
Full size table
Table 4 Consistency ratio
Full size table

The matrix normalization, calculation of priority vector [C], overall priority [D], rational priority [E] were done as in Nikhil et al. (2021).

The Eigen value (λmax), consistency index (CI), and consistency ratio (CR) were determined using Eqs. 4 and 5 as in Nikhil et al. (2021)

$$\lambda\text{max}=\frac{\left[E\right]}{k}$$
(4)
$$\text{CI}=({\uplambda } \text{max}-k)/(k-1)$$
(5)
$$\text{CR}=\frac{CI}{RI}$$
(6)

where RI = random index (Saaty 1980).

Saaty (1980) suggests that the CR should be less than 0.1 (i.e., 10%), otherwise the judgements are unreliable and we need to revise the subjective judgements. Hence, the judgements are acceptable in this study because the CR is 0.040 (which is less than 0.1).

The weights obtained from the AHP modelling is shown in Eq. 7.

$$\text{LSZ}=\left(0.291\times \text{Slp.}\right)+\left(0.216\times \text{Ele.}\right)+\left(0.155\times \text{Asp.}\right)+\left(0.110\times \text{SB}\right)+\left(0.077\times \text{Litho.}\right)+\left(0.054\times \text{Soil}\right)+\left(0.038\times \text{RB}\right)+ \left(0.027\times \text{NDVI}\right)+\left(0.019 \times \text{LULC}\right)+(0.014 \times \text{LB})$$
(7)

FR modelling

The FR is the proportion of landslides in a desired class as a percentage of all landslides to the class’s area as a percentage of the total study area (Silalahi et al. 2019). A lower FR value (< 1) implies a lower level of landslide susceptibility, whereas a larger FR value (> 1) suggests a higher level of landslide susceptibility (Silalahi et al. 2019). The FR of each factor and its classes is included in Table 5. The susceptible zones were created by applying Eq. 8.

$$\text{LSZ}=Fr\left(\text{Slp.}\right)+Fr\left(\text{Ele.}\right)+Fr\left(\text{Asp.}\right)+Fr\left(\text{SB}\right)+Fr\left(\text{Litho.}\right)+Fr\left(\text{Soil}\right)+Fr\left(\text{RB}\right)+Fr\left(\text{NDVI}\right)+Fr\left(\text{LULC}\right)+Fr\left(\text{LB}\right)$$
(8)
Table 5 Frequency ratio of factors
Full size table

Validation of the results – ROC curve method

The performance of a classifier is evaluated using ROC curves, with sensitivity plotted on the Y-axis and 1-specificity plotted on the X-axis (Grimnes and Martinsen 2015). The AUC, or area under the receiver operating characteristic (ROC) curve, is a single scalar value that quantifies performance of a binary classifier (Hanley and McNeil 1982). AUC values of 0.5 imply no discrimination, 0.7 to 0.8 are acceptable, 0.8 to 0.9 are excellent, and greater than 0.9 are outstanding (Hosmer and Lemeshow 2000). There was a total of 26 landslides recorded in the study area, which were used for validation. The SPSS 25 software was utilized for plotting the ROC graph and estimating the AUC values.

Results and discussion

In this study, the AHP and FR models and GIS were applied to determine the susceptibility of the Western Ghats region in Kollam district. Ten causative factors, viz. slope angle, elevation, slope aspect, stream buffer, lithology, soil, road buffer, NDVI, LULC, and lineament buffer, were used for susceptibility modeling. Based on the level of susceptibility, the study area is divided into five zones, as shown in Table 6. The landslide-susceptible zone maps are shown in Figs. 14 and 15. During the 2018 southwest monsoon, a total of 26 landslides occurred in this region. The ROC curve analysis confirmed the outstanding AUC value (above 0.90) for the susceptibility maps produced using the AHP (0.906) and FR (0.926) models, respectively (Fig. 16). The higher AUC value for the FR model confirmed that FR has a better prediction accuracy than the AHP model in demarcating landslide susceptible zones. The high and very high susceptibility zones together constitute 33.3% and 23.59% of the study area in the maps produced using the AHP and FR models, respectively. The study by Thomas et al. (2021) also found that the FR model is more efficient than the AHP model in susceptibility zonation of landslides.

Fig. 14
figure 14

Landslide susceptible zones – AHP method

Full size image
Fig. 15
figure 15

Landslide susceptible zones – FR method

Full size image
Fig. 16
figure 16

The ROC curves

Full size image

Most of the critical areas were observed in areas with a higher slope angle. The areas with higher slope angles are generally unstable. The FR values of 1.39 and 3.94 for slopes ranging from 23.05 to 31.00° and 31.00 to 67.57°, respectively, confirmed the role of steeper slopes in slope failures. The study by Yang et al. (2015) also confirmed higher landslide incidences at slopes ranging between 30 and 40°. The majority of the slope failures were observed in high elevation regions and also in areas bordering moderate and higher elevation regions. The FR value (3.29) confirmed that the elevation ranging between 727 and 1079 m has a higher correlation with landslide occurrence. In their study, Shao et al. (2020) found steeper slopes with higher elevations to be ideal sites for landslide occurrence. The number of landslide incidences was found to be high (18 landslides) in close proximity to stream channels. This is due to increased saturation conditions, stream bank erosion, or toe cutting due to river flow. The rock types present in the study area belong to the crystalline group. And the number of landslide incidences was found to be high in garnet-biotite gneiss rocks (with a FR value of 1.12). A study by Thomas et al. (2021) also found higher landslide incidences in gneissic terrain. Areas with clayey soils are potential slip zones. This was confirmed by assessing the number of landslide incidences in the areas with clayey soils. A total of 17 landslide incidences (with a FR value of 1.71) occurred in areas with clayey soil. Thomas et al. (2021) also reported that the majority of landslide incidences (91.35%) have been recorded in the areas with clayey soil. A considerable number of landslides occurred along road cuttings with a FR value of 1.61. Road cuttings can destabilize the slopes by reducing the shear strength of the toe. The occurrence of failures along the road cuttings is evidence of their anthropogenic origin. A study conducted by Sujatha and Sridhar (2021) in Connoor found that about 74% of the landslides induced as a result of road cuttings.

Table 6 Area and percentage of susceptible zones – AHP and FR method
Full size table

Applying the Fuzzy-AHP model, Sur et al. (2020) created landslide susceptibility zones along a road corridor in the Himalayan region (Kalsi-Chakrata Road corridor), which were frequently affected by landslides. They found 55% of the area to be in very high and high landslide susceptibility zones. Panchal and Shrivastava (2021) created a landslide hazard zone map of National Highway 5 in India employing the AHP model, where they found 21.56% of the area as a very high hazard zone. The study by Getachew and Meten (2021) found road cuttings as one of the crucial reasons for the landslide occurrence in the Kabi-Gebro locality of Ethiopia. The study conducted by Tien et al. (2021) along the Halong-Vandon Expressway in Quang Ninh Province (Vietnam) found that out of 44 landslides recorded during June - August (in the years 2017 and 2018), 35 were shallow debris flows. They found that moderate to high intensity short-period rainfall is the triggering (extrinsic) factor, and that slope cutting for road developments is the major conditioning/causative factor for landslide occurrence.

The shallow slope failures occurring in this area can be minimized by applying nature-based solutions (NbS). When compared to grey engineering solutions, NbS is effective in preventing shallow slope failures (Arce-Mojica et al. 2019) and cost-effective (Chausson et al. 2020; Reguero et al. 2018; Singh et al. 2020). Rahardjo et al. (2014) assessed the performance of both Orange Jasmin and Vetiver grass in reducing slope stability. They found both shrubs effective in minimizing rainfall infiltration on a slope in Singapore. Mohammed et al. (2018) used grasses like Cynodon dactylon and Vetiver for stabilizing a slope along the Wad Medani-Sennar road in Sudan, and found it an excellent solution. Gobinath et al. (2015) found that using lemon grass roots to stabilise landslide-affected soils in the Nilgiri district of India increased shear strength while decreasing permeability, making them suitable for reducing landslide incidences. The results of this study could be improved by utilizing factors such as the Normalized Difference Road Landside Index (Zhao et al. 2018) and topographic ruggedness index. This will be considered in future research.

Conclusions

Due to the geomorphological setting of the area, the Western Ghats region in Kollam district is prone to landslides. This study assessed the spatial relationship between the conditioning factors and landslide incidences using GIS. The AHP and FR models were used to identify the susceptibility. The area of the landslide susceptible zone maps was divided into different zones using the natural breaks method. The critical zones together constitute 33.3% (AHP method) and 23.59% (FR method) of the study area. It was found that slope angle, elevation, streams, soil, and road cuttings are the major inducing factors. From the study, it is found that landslides prevailing in this region are due to natural as well as human-induced reasons. Finally, the prepared map was validated using the incidence data provided by NRSC. The ROC curve analysis proved that the FR method has better prediction accuracy than the AHP method for landslide-susceptible zone mapping. The rainfall (triggering factor) is not considered in this study due to the non-availability of an adequate network of rain gauges or automatic weather stations, and also because the available satellite-derived data and India Meteorological Department gridded data have very low spatial resolution, which will not yield good data for a smaller region like this. The developed framework and created maps are essential to the land use planners, researchers, and the local government for taking proper landslide preventive and mitigation measures in the study area. This will help to reduce the impacts of landslide disasters in the future.