Arabian Journal of Geosciences (2021) 14:838 https://doi.org/10.1007/s12517-021-07156-6

ORIGINAL PAPER

Landslide susceptibility investigation for of using regression analysis and machine learning

Sheelu Jones1 & A. K. Kasthurba1 & Anjana Bhagyanathan1 & B. V. Binoy1

Received: 15 January 2021 /Accepted: 23 April 2021 # Saudi Society for Geosciences 2021

Abstract Kerala is the third most densely populated state in , with 860 persons per square kilometer. The uniqueness and diversity of the state’s topology make it highly vulnerable to natural hazards. Kerala State Emergency Operations Centre Kerala State Disaster Management Authority (2016). This study was initiated in the backdrop of landslides and floods in 2018, which had wreaked havoc in the region. Among the 4728 landslides reported in the state’s ten districts, Idukki was the worst affected with 2219 landslide occurrences. A statistically significant cluster of landslide hotspots was identified within the Idukki district using Getis-Ord Gi* statistics. Landslide susceptibility analysis was carried out using logistic regression (LR) and artificial neural network (ANN). Natural parameters influencing landslides such as slope, elevation, rainfall, geology, distance to drainage, and anthropogenic conditioning factors such as land use, road density, and quarry density were considered in this study. The results indicate that both natural and anthropogenic conditioning factors have a significant influence on landslide occurrences. According to the LR results, about 37.87% and 38.07% of the district’s total area is situated in high and medium landslide susceptibility zones. The results establish that ANN has better predictive performance compared with LR.

Keywords Hotspot analysis . Landslide conditioning parameters . Logistic regression . Artificial neural network

Introduction (lithology and morphology), triggering factors (intense rainfall, earthquake, etc.), and accelerating factors (landscape modifica- Landslides are mass movements of rock, debris, or earth down tions) that result in landslide occurrences (Santini et al. 2009). the slope under the influence of gravity, resulting in a geomor- These factors along with high population density pose an in- phic change of the earth’s surface (Cruden et al. 1996; Guzzetti creased risk of vulnerable situations in developing countries et al. 2005;Phametal.2019; Van Thom et al. 2016). Every like India (Kritikos and Davies 2015). These unpredictable year, hundreds of people worldwide lose their lives in land- and massive natural hazards result in property damages incur- slides (Yalcin et al. 2011). The hydro-geomorphic processes ring financial losses in many parts of the world, impacting local are developed by the combination of predisposing factors and global economy (Suzen and Doyuran 2004; Chang and Chiang 2009; Cogan and Gratchev 2019; Lee et al. 2012; 2 Responsible Editor: Biswajeet Pradhan Nourani et al. 2014). About 15% (~ 0.42 million km )of India’s total landmass falls under the landslide-prone hazardous * Sheelu Jones zone (Kanungo and Sharma 2014; Harilal et al. 2019;National [email protected] Disaster Management Authority 2019; Thennavan and Pattukandan Ganapathy 2020). Rao (1993) identified four A. K. Kasthurba landslide-prone regions in India as Western (Uttar [email protected] Pradesh, Himachal Pradesh, and Jammu and Kashmir), Eastern Anjana Bhagyanathan and North Eastern Himalayas and plateau margins (West [email protected] Bengal, Sikkim, and Arunachal Pradesh), Naga-Arakkan B. V. Binoy Mountain belt (Nagaland, Manipur, Mizoram, Tripura), and [email protected] plateau margins of the and some parts of

1 (Kerala, , , and Department of Architecture and Planning, National Institute of ) (Rao 1993). Technology Calicut, , Kerala, India 838 Page 2 of 17 Arab J Geosci (2021) 14:838

Kerala is vulnerable to natural disasters due to the shifting approach analyzes the relationship between landslide condi- climatic conditions and diverse topographical characteristics. tioning parameters and past landslide events (Zêzere et al. The state’s location with the Arabian Sea in the west and the 2017). The major assumption made in this approach is that steep slope gradient of the Western Ghats on the eastern side, the conditions which accelerated landslides in the past will also and high population density (860 persons per square kilome- trigger future events (Dapples et al. 2002). The three funda- ter), accelerates the vulnerability to disasters (Shaharban and mental data-driven methodologies commonly employed for Rathnakaran 2019; Prakashkumar 2019). According to the the study are the bivariate statistical method, multivariate Kerala State Disaster Management Plan 2016, Kerala is listed method, and artificial intelligence (AI)-based methods as the multi-hazard-prone state in the country as it has been (Nourani et al. 2014;Phametal.2016;Jose2018;Vineesh identified with 39 hazards Kerala State Emergency Operations 2019; Dao et al. 2020). The bivariate statistical method as- Centre Kerala State Disaster Management Authority (2016). sesses the relationship between each independent variable Landslides are one of the frequently occurring natural hazards and landslide occurrences separately. The multivariate ap- in sloping terrains of the state, especially during monsoons. proach considers the collective association of dependent and All districts except (coastal district) are prone to independent variables for the assessment (Mutlu and Goz landslides (Kuriakose et al. 2009b). The discrepancy in cli- 2020). Logistic regression (LR) is one among the multivariate matic conditions along with high population density and land- regression methods which explicate the presence or absence of use changes had accelerated the number of landslides within a phenomenon under consideration (Makealoun et al. 2015). the state in recent decades. Therefore, landslide susceptibility This is one of the most reliable and widely used approaches for studies have gained significant importance in the mountainous landslide susceptibility studies (Guzzetti et al. 2005; Lee and regions of the state, with increasing demand for developmen- Pradhan 2007; Rickli and Graf 2009; Yilmaz 2010;Erenerand tal undertakings and hazard mitigation activities. Düzgün 2012). In recent times, AI is widely used in modeling Landslide susceptibility studies elucidate the possibility of intricate phenomena such as landslide susceptibility mapping landslide occurrences of any chosen region in the future (Nourani et al. 2014; Hong et al. 2017; Chen et al. 2018). The (Hemasinghe et al. 2018). It forecasts “where” landslides are artificial neural network (ANN) model is one of the most com- likely to happen within the study area (Guzzetti et al. 2005). monly used AI techniques for landslide susceptibility study Landslide susceptibility studies are conducted with some as- (Zare et al. 2013;Chenetal.2018; Nguyen et al. 2019;Dao sumptions as: (i) the previous landslide events leave visible et al. 2020). ANN uses a computational mechanism for ana- indications that can be recognized, classified, and mapped; (ii) lyzing data to study the relationship between the dependent physical laws govern occurrence, which can be evaluated ex- and independent variables (Arora et al. 2004). perimentally, statistically, or deterministically; (iii) future This study was conducted in the wake of the massive land- landslides would more likely occur in similar conditions of slides and floods in Kerala during the year 2018. This flood past events; and (iv) the frequency of spatial landslides can garnered the attention of the research community and can be be inferred from exploratory inquiries that are computed from entitled as the worst flood in this century (Mishra et al. 2018). environmental information or physical models (Guzzetti et al. According to the Kerala State Disaster Management Authority 1999b, 2012; Reichenbach et al. 2018). Landslide susceptibil- (KSDMA), 46.93% of landslides reported in Kerala during ity studies can be classified into two methods such as (i) qual- 2018 had occurred in the Idukki district. Therefore special itative and (ii) quantitative (Erener and Düzgün 2012; attention is required to examine and forecast landslides in Kayastha et al. 2013; Nourani et al. 2014;Chaeetal.2017). Idukki due to the sensitive terrain and intensified landscape Qualitative methods are knowledge-driven and purely based modifications (Abraham et al. 2021). Initiating these types of on the opinion or judgment of experts (Van Westen et al. studies helps authorities to make necessary initiatives in con- 1999; Wahono 2010). In this method, the relationship be- nection with landslide hazard zoning and thus controlling de- tween landslide susceptibility and landslide conditioning pa- velopment activities. The initial phase of the study identifies rameters is obtained by assigning weightage to the factors landslide hotspots throughout the district. In the later part of based on expert opinions (Kaur et al. 2018). In recent times, the study, landslide susceptibility modeling was done using the traditional qualitative methods are replaced by advanced two different data-driven methods, LR and ANN (Guzzetti quantitative methods (Chae et al. 2017; Tien Bui et al. 2017). et al. 2005; Lee and Pradhan 2007; Rickli and Graf 2009; Quantitative methods most commonly used for landslide stud- Yilmaz 2010; Erener and Düzgün 2012). LR is applied when ies elucidate the probability of landslide occurrence in an area the dependent variables are binary or dichotomous to identify and can be classified as physically based approaches and data- the strength of the relationship between landslide occurrences driven methods (Kaur et al. 2018). The physically based ap- and their conditioning parameter (Erener and Düzgün 2012). proaches are developed by estimating slope instability using In contrast, ANN deals with interconnected artificial neurons physical models for landslide predictions through on-site or that utilize mathematical computation methods for model de- laboratory test evaluations (Fell et al. 2008). The data-driven velopment (Mutlu et al. 2019). This study uses a perceptron- Arab J Geosci (2021) 14:838 Page 3 of 17 838 type neural network with a back-propagation learning tech- Town and Country Planning 2011; Ground Water nique for landslide susceptibility analysis. Information Booklet of Idukki District, K 2013; Kanungo et al. 2020). The district has a well-connected road network but lacks rail and air connectivity (District Census Handbook Study area 2011; Industrial Potential Survey 2017). There are fourteen mountain peaks exceeding 2000m elevation within the The study was conducted in the Idukki district of Kerala, which Idukki district, including the highest peak of South India, falls in the Western Ghat region located between 9° 15′ and 10° . The district also includes many major tourist desti- 2′ North latitude and 76° 37′ and 77° 25′ East longitudes nations in the state. Almost half of the district is covered with District Census Handbook (2011). This is the second-largest forest (especially the eastern portion), while the remaining part district in the state, having an area of 4358km2 as illustrated in is covered by urban and village settlements merged with plan- Fig. 1. The word Idukku which means narrow tations (District Census Handbook 2011;IndustrialPotential gorge was modified to Idukki. As the name indicates, 97% of Survey 2017). Favorable terrain conditions and land-use mod- the study area consists of rugged mountains and forests. ifications in the Idukki district have accelerated the landslide Physiographically, the district comprises mainly of highlands occurrences over the years. There is a repeated history of land- along with a small strip of midland in the western part (District slide events in the district, resulting in the loss of life and Urbanisation Report Department of properties Kuriakose et al. (2009a, b).

Palakkad

Karnataka

Tamil Nadu KERALA IDUKKI DISRICT Tamil Nadu Idukki

Kattappana

Tajikistan Afghanistan KottayamChina Pakistan New Delhi k Nepal Bhutan

INDIA Bangladesh Myanmar Legend Thailand Alapurza Towns Idukki Kerala Districts -1 0204010 km Sri Lanka India_States

Fig. 1 Study area 838 Page 4 of 17 Arab J Geosci (2021) 14:838

Landslides of 2018 Survey of India) (Hao et al. 2020). Among the reported land- slides, 4642 (98.18%) landslides occurred in the Western In 2018, perilous rainfall and flooding severely affected all Ghats region. Also, 3903 (84.08%) fall within the aspects of human lives in Kerala. This high-intensity rain trig- Environmentally Sensitive Area (ESA) of Western Ghats. gered many landslides in weak zones of Kerala’s hilly regions These landslides and floods resulted in the death of 483 per- along with massive floods. This has resulted in obtrusion of sons and large-scale property loss. Around 5.4 million people infrastructure, agriculture, transportation, socioeconomic con- got affected by this havoc, and more than 1.4 million people dition, and livelihood of people (Shaharban and Rathnakaran got displaced (Joy et al. 2019;Marthaetal.2019). 2019;Mishraetal.2018). Kerala state, with an average annual precipitation of about 3000 mm, has received 2346.6 mm of rainfall from June to August 2018.This had exceeded the ex- Data and methods pected rainfall, which was 1649.5 mm (Joy et al. 2019;Mishra et al. 2018). A total of 4728 landslides were reported from the The study was conducted in four phases, beginning with land- state in 2018. Idukki was the worst affected district in the state, slide data collection and study area selection; the second phase with 2219 landslides (Fig. 2) (includes landslides reported by dealt with the hotspot identification, selection of landslide Kerala Disaster Management Authority and Geological conditioning parameters, and understanding their correlation.

Fig. 2 Spatial distribution of landslides Thrissur

Ernakulam Munnar

Thodupuzha Idukki

Kattappana Tamil Nadu

Kottayam Kumily

Legend

Idukki Landslide Locations

Tow ns

Idukki Pathanamthitta

Kerala Districts 010205km Alapurza India States Arab J Geosci (2021) 14:838 Page 5 of 17 838

The later phase models landslide occurrences using LR and Logistic regression ANN methods. The last phase of the study consists of a com- parison of the models and selection of the best model for the LR is one of the multivariate regression methods used for landslide susceptibility study. The methodology used for this predicting the probability of occurrence of a characteristic or study is given in Fig. 3. phenomena when the dependent variable is dichotomous or This study used landslide inventory data (dependent vari- binary (Makealoun et al. 2015). The presence or absence of able) prepared by the Department of Geoinformation Science the dependent variable was identified based on a set of inde- and Earth Observation (ITC), University of Twente, the pendent variables associated with the phenomenon (S. Lee Netherlands, in collaboration with the Geological Survey of and Pradhan 2007). LR coefficients were used to estimate India and University of Kerala, Thiruvananthapuram (Hao the influence of every particular independent variable used et al. 2020). The landslide inventories obtained are located in the model. The association between the dependent and in- at the crown of the landslide spots that exemplify landslide dependent variables in LR can be quantitatively communicat- origins. Also, eight landslide conditioning parameters (inde- ed as: pendent variables) were identified from the literature for p ¼ 1=ðÞ1 þ e−z model development. Land-use map was prepared from Sentinel-2 satellite image with 10m resolution downloaded where p is the probability of occurrence of the phenomena . from United States Geological Survey (USGS) website The (here p is the probability of landslide occurrence) Lee and land-use map was verified using secondary data and field Pradhan (2007). The probability forms an S-shaped curve, . surveys ASTER Gdem of 30m resolution was the elevation and it varies from 0 to 1. z is the linear combination and it data used in the study. DEM data was downloaded freely could be expressed as: from the USGS website. The slope of the study area was z ¼ β þ β x þ β x þ … þ β x ¼ itðÞ p generated from this DEM using GIS. Average rainfall data 0 1 1 2 2 n n log  was collected for sample locations from the Indian p Metrological Department (IMD) and interpolated for the ¼ Ln −p study area. Geology and drainage data were acquired from 1 Kerala State Land Use Board, Department of Planning and where β0 is the model intercept; coefficients i (i =1,2,..,n) Economic Affairs, . The quarry data are representative of the contribution of single independent for the study was obtained from the Forest Health Division, variables x , p is the probability of landslide occurrence, and p Kerala Forest Research Institute Peechi Alex and STV is named as odds or likelihood ratio (Lee and Pradhan (2017). Road networks used in the study were downloaded 1−p 2007;Mousavietal.2011; Hemasinghe et al. 2018). freely from Open Street Maps (OSM).

Fig. 3 Methodology 2018 Landslides

Study area selection

Landslide conditioning Landslide susceptibility Hotspot analysis parameters Slope Elevation Rainfall Geology Data-driven method Land-use Road density Distance to drainage Quarry density Logistic Regression Artificial Neural Network

Result interpretation

Best model 838 Page 6 of 17 Arab J Geosci (2021) 14:838

Quantitatively, the relationship of landslide occurrences and Results controlling parameters could be expressed as: Hotspot analysis p =1/(1+e−z) Hotspots identify the statistically significant clustering in the The value of p (landslide probability) varies from 0 to 1 on data. In this study, village-level hotspot areas are identified an S-shaped curve. The probability for landslide occurrence through the aggregation of landslide locations. Landslide increases when the value is close to 1 and decreases when it is hotspots are concentrated in the central region of the study close to 0. area, as shown in Fig. 4. Among the six hotspot villages, five fall under 99% confidence and one in 95% confidence. The villages included in landslide hotspots are , Artificial neural network Koonathady, Upputhoda, Vathikudy, Thankamani, and Idukki. No cold spots were identified in the study area. Artificial neural network (ANN) is a data-generating pro- cess used to develop a model concerning a network of Natural and anthropogenic landslide conditioning input data provided. The output is predicted based on the parameters inputs provided. ANN is used to analyze complex data of varied scales (Sangchini et al. 2015; Mutlu and Goz Landslide susceptibility study depends on the relationship be- 2020). Therefore, it is considered a useful approach to tween areas of landslide occurrences and the landslide condi- replace regression models (Mutlu et al. 2019). ANN tioning parameters. Identification of suitable conditioning pa- works based on information obtained from data with rameters for landslide prediction requires precise information known characteristics by deriving weightage to the pa- about previous landslides (Ma et al. 2013). This study used rameters Lee et al. (2003). ANN “learns” by balancing eight landslide conditioning parameters which could be clas- the weight of the neurons in connection with the errors sified as natural and anthropogenic parameters. Natural con- between actual output values and target output values ditioning parameters included in the study are slope, elevation, (Valencia Ortiz and Martínez-Graña 2018). The back- rainfall, geology, and drainage distance. In contrast, anthropo- propagation learning algorithm is the most frequently genic conditioning parameters used in the study include used multi-layered neural network method. This consists human-modified activities such as land use, road density, of an input layer, hidden layers, and an output layer. The and quarry density. The correlation existing between indepen- analysis of ANN is carried out in three significant steps: dent variables is exemplified in Fig. 5. initial step deals with the generation of a database (train- Table 1 and Fig. 6 illustrate the landslide conditioning pa- ing stage), followed by the creation of the matrices (de- rameters and their respective landslide occurrences. Slope and termining weights for input, target, and sample), and fi- elevation are the most significant natural landslide condition- nally establishing training parameters for simulation ing parameters, and these are directly proportional to the prob- (classification stage) (Lee et al. 2003; Valencia Ortiz ability of landslide occurrence (Wang et al. 2016). In the study and Martínez-Graña 2018). The back-propagation algo- area, about 44% of landslides have occurred in regions with a rithm executes the network until minimal error is slope ranging from 20 to 30 degrees. Around 65% of the achieved between the anticipated and real output values landslide occurrences fall in the elevation range of 500– of the network (Guzzetti et al. 1999a). 1000m. The spatial intersection of rainfall and landslides The error E representing one input training pattern t is a shows that half of the landslide occurrences are in regions of function of the anticipated output vector d and the real output maximum rainfall (875–905 mm). Since all the landslides vector o, for node k is: considered in the study have been reported during the south- st th E ¼ 1=2 ∑ ðÞdk −0k west monsoon period (1 June to 30 September), average k rainfall during this period is considered in the study. Geology was another natural condition parameter used in the The error thus obtained is minimized by varying the weight study. The geological unit has its own properties of action in between layers. Thus, the weight is expressed as: ÀÁ geomorphic processes like landslides (Lee et al. 2001; wijðÞ¼n þ 1 ηδj−0i þ αΔwij Pradhan et al. 2006; Youssef et al. 2015). The spatial analysis identified that about 65% of landslide spots are located in where η represents the learning rate parameter, δj is the migmatite complex rocks. Heavy rainfall could result in cata- index of the rate of change of the error, and α is the momen- strophic failure along joints, bedding, and exfoliation plains. tum parameter. This process of back-propagation of the error According to our study, about 46% of landslides have oc- is repeated till the error of the whole network is minimized. curred in locations having less than 75m drainage distance. Arab J Geosci (2021) 14:838 Page 7 of 17 838

Fig. 4 Hotspot Analysis Palakkad

Thrissur Marayoor

KantalloorKottakamboor

Vattavada Ernakulam Manakulam Idukki District

Anaviratty

KunnjithannuyBisonvally Vellathooval Rajakumari

Kodikulam Rajakadu kumaramangalamNeyyassery Kanjikuzhy Sandhampara Gandhipara Karimannur todupuzha karikode Vathikudy Chadhuramgapara Upputhoda Alakode karinkunnam Parathodu Velliamattom Kalkunthal muttam Idukki Arakulam

Elappilly Kattappana Tamil Nadu Vandanmettu

Vagomom Anakara Anavilasom

Kottayam

Kokkayar Peernade Legend Kumily

Kerala Districts Manjamala Puruvanthanam India States Landslide hotspot Villages Cold Spot - 99% Confidence Milapra Cold Spot - 95% Confidence Cold Spot - 90% Confidence

Not Significant Pathanamthitta Hot Spot - 90% Confidence

Hot Spot - 95% Confidence 010205km Alapurza Hot Spot - 99% Confidence

Land use indicates the activities occurring in a particular landslides have occurred in areas having 0.06–1.5 road densi- parcel of land. Land use is a vital parameter used for landslide ty. More than half of the landslides are located in areas with 0– studies as each land-use activity impacts soil erosion and sta- 0.00025 quarry density. Figure 5 illustrates the spatial distri- bility, which are directly related to landslides (Baeza and bution of landslide conditioning parameters. Corominas 2001). Seven classes of land-use activities were identified in the study area. Land-use activities identified in- LR and ANN clude built-up, cropland (paddy, banana, tapioca, etc.), plan- tation (tea and ), other plantation (rubber, coconut, etc.), The LR results show the probability of landslide occurrence in forest plantation (pine, eucalyptus, etc.), forest, and others the study area. The dependent variable is classified as binary, (grassland, rock, water bodies, etc.). The spatial intersection with “0” indicating “the absence of landside and “1” indicat- of landslide locations and land use proves that about 56% of ing the “presence of landslide.” The model coefficient extrac- landslides are reported in plantation area (including planta- tion in LR was done using the maximum likelihood estimation tions, other plantations, and forest plantations). Roads are con- method. The model improvement over the null model was structed in hilly areas by removing materials from the foot of evaluated using the variation in the −2 log-likelihood value. slopes. The construction of roads along with road widening The lower value of −2 log-likelihood indicates the best step fit could adversely affect soil stability. Thirty-nine percent of the of the model with respect to the data. Hosmer-Lemeshow test 838 Page 8 of 17 Arab J Geosci (2021) 14:838

Fig. 5 Correlation of independent variables

illustrates the goodness of fit of the equation (Zhu and Huang that Nagelkerke’s R2 is 0.624 and Cox and Snell’s R2 is 0.466. 2006)(Table2). Cox/Snell’s Nagelkerke’s and Cox/Snell’s R2 Both results indicate the model’s good fit (Zhu and Huang could be used for measuring the usefulness of the model. 2006; Althuwaynee et al. 2014). Table 3 illustrates the regres- Compared to linear regression, the coefficients of R2 can be sion results along with the coefficient values. Coefficients of relatively small in LR, which does not invalidate the model significant variables were substituted to obtain a linear com- (Chau and Y. F. T. 2004; Mousavi et al. 2011). Table 2 shows bination as:

Ln LUn Z ¼ −33:625 þ :122 Slope þ :036 Rainfall þ ∑ 3:326 Lithology þ ∑ :214 Landuse i¼L1 j¼LU1 þ 0:64 Road Density−:001ðÞ Elevation −:002 Distance to drainage−549:617 Quarry Density

Therefore, the probability of landslide occurrence could be expressed as:

  − : þ: þ: þ∑Ln : þ∑LUn : þ : −: ðÞ−: − : 33 625 122 S 036 R ¼ 3 326 L ¼ 214 LU 0 64 RD 001 E 002 DD 549 617 QD p ¼ 1= 1 þ e i L1 j LU1

where p is the probability of landslide occurrence, S elevation, DD indicates the distance to drainage, and QD indicates slope, R respective rainfall, L represents lithology, is quarry density. Also, i indicates the various lithological

LU indicates land use, RD is road density, E represents groups (L1 is the first lithological group and Ln is the last Arab J Geosci (2021) 14:838 Page 9 of 17 838

Table 1 Landslide conditioning parameters and percentage Sl. No Landslide conditioning parameters Class % of landslide occurrence landslide occurrences Natural conditioning parameters 1Slope 0–10° 4.24 10–15° 8.47 15–20° 17.31 20–30° 44.03 30–73° 25.96 2 Elevation < 100m 0.09 100–500m 10.86 500–1000m 64.98 1000–1500m 11.54 1500–2382m 12.53 3Rainfall 658–725mm 0.68 725–780mm 1.17 780–825mm 2.75 825–875mm 46.19 875–905mm 49.21 4 Lithology Acidic rocks 1.62 Charnockite group of rocks 13.20 Khondalite group of rocks 1.08 Migmatite complex 64.67 Peninsular gneissic complex 19.33 Tank/WB/river 0.09 5 Distance to drainage 0–75 46.01 75–150 31.82 150–300 20.73 300–481 1.44 Anthropogenic conditioning parameters 6 Land use Built-up 0.23 Cropland 0.18 Plantation 27.85 Other plantation 23.21 Forest plantation 4.73 Forest 24.38 Others 19.42 7 Road density 0.06–1.5 39.09 1.5–2.3 13.77 2.3–324.21 3–3.7 11.51 3.7–5.7 11.42 8 Quarry density 0–0.00025 52.25 0.00025–0.0008 29.44 0.00080–0.002 12.93 0.002–0.004 3.55 0.004–0.0121 1.83

lithological group in the study area), j indicates the various area). The prediction performance of LR was 84.7%. land-use activities (LU1 is the first land-use and LUn is the Landslide susceptibility map developed based on the results last land-use activities according to Table 3 within the study of LR is shown in Fig. 7. 838 Page 10 of 17 Arab J Geosci (2021) 14:838

Palakkad Palakkad

Thrissur Thrissur

Ernakulam Ernakulam Munnar Munnar

Thodupuzha Thodupuzha Idukki Idukki

Kattappana Kattappana Tamil Nadu Tamil Nadu

Legend Kottayam Kumily LegendKottayam Kumily Towns Towns India States India States Kerala Districts Kerala Districts Slope DEM (in degrees) (Elevation in meters) 0 - 10 < 100

10 - 15 100 - 500 Pathanamthitta Pathanamthitta 15 - 20 500 - 1,000

20 - 30 1,000 - 1,500 010205km 010205km Alapurza Alapurza 30 - 90 1,500 - 2,685

Palakkad Palakkad

Thrissur Thrissur

Ernakulam Ernakulam Munnar Munnar

Thodupuzha Thodupuzha Idukki Idukki

Kattappana Kattappana Tamil Nadu Tamil Nadu

Legend Legend Kumily Kumily Kottayam TownsKottayam

Towns Kerala Districts

India States India States

Kerala Districts Geology Rainfall Average Acidic rocks Basic Rocks (in mm) Charnockite group of rocks 653 - 725 Khondalite Group of rocks 725 - 780 Migmatite Complex 780 - 825 Pathanamthitta Pathanamthitta Pegmatite/Aplite/Quartz vein

825 - 875 Penisular Gneissic Complex 010205km 010205km Alapurza Alapurza 875 - 907 Waterbody

Fig. 6 Spatial distribution of landslide conditioning parameters

ANN is a powerful estimation tool widely used in geosci- The data set was randomly split into training and verification ence studies, especially for landslide susceptibility studies. sets in ANN models. The training set accounted for 70% of Arab J Geosci (2021) 14:838 Page 11 of 17 838

Palakkad Palakkad

Thrissur Thrissur

Ernakulam Ernakulam Munnar Munnar

Thodupuzha Thodupuzha Idukki Idukki

Kattappana Kattappana Tamil Nadu Tamil Nadu

LegendKottayam Kumily Kottayam Kumily Legend Towns

India States Towns

Kerala Districts India States Drainage Distance Kerala Districts (in meters) Road Density 0 - 75 0.06 - 1.5

75 - 150 1.5 - 2.3 Pathanamthitta Pathanamthitta 150 - 300 2.3 - 3

300 - 600 3 - 3.7 010205km 010205km Alapurza Alapurza ? 600 - 1,402 3.7 - 5.7 Palakkad Palakkad

Thrissur Thrissur

Ernakulam Ernakulam Munnar Munnar

Thodupuzha Thodupuzha Idukki Idukki

Kattappana Kattappana Tamil Nadu Tamil Nadu

Kottayam Kumily Legend Kottayam Kumily Legend Towns Towns India States India States Kerala Districts Kerala Districts Idukki Landuse Quarry Density Builtup 0 - 0.00025 Cropland Forest Plantation 0.00025 - 0.0008 Forest 0.00080 - 0.002 Pathanamthitta Pathanamthitta Other Landuse

0.002 - 0.004 Other Plantation 010205km 010205km Alapurza Alapurza 0.004 - 0.0121 Plantation Fig. 6 (continued) landslide data, while the rest 30% was used for validation. The 93.1%, indicating that the model had excellent performance. area under the curve (AUC) value obtained for ANN was Figure 8 demonstrates the normalized importance of the 838 Page 12 of 17 Arab J Geosci (2021) 14:838

Table 2 Summary statistics of LR model significant land-use activity in the hotspots. In the present Hosmer and Lemeshow test Chi-square 203.13 study, both LR and ANN methods were applied to identify df 8 landslide susceptible areas. The validation results indicated Sig. 0.00 that ANN had better prediction compared with the LR model. The landslide conditioning parameters (natural and anthropo- −2 log-likelihood 3749.94a genic) were analyzed using LR and ANN along with landslide R Cox and Snell square 0.466 occurrences. R Nagelkerke square 0.624 According to LR results, the highest positive coefficient β a indicates as estimation terminated at iteration number 20 because max- was obtained for lithology with acidic rock characteristics ( imum iterations has been reached = 3.326). Acidic rock type is highly prone to landslides. Also, slope and rainfall were having a significant positive impact on variables. The variable importance obtained from ANN results landslide occurrence. Among anthropogenic conditioning pa- indicates that slope is the foremost important variable, follow- rameters, plantations had highly significant positive coeffi- ed by rainfall, quarry density, elevation, and drainage density. cients. All types of plantations considered in the study trigger The prediction performance of the ANN model is 86.2% landslide occurrences. According to the results obtained, an (Table 4). increase in road density and the built-up area increase the probability of landslide occurrences. Landslide susceptibility map developed based on the results of LR is shown in Fig 7. 4.92% of the total area was classified to be highly susceptible Discussion to landslides. High and medium susceptibility zones representes 37.87% and 38.07% of the total area,while low The spatial distribution of landslides was analyzed using and very low susceptibility zones are of 18.41% and 0.74% hotspot analysis, and the results indicated that only one cluster of the total area. Thus the result indicates that a significant of landslide hotspots was present in the Idukki district. The portion of the district falls under high and medium suscepti- cluster was located in the central part of the district and it bility zones. Only an area below 1% could be classified as a consisted of six villages of the district. All these hotspots very low susceptibility zone. The variable importance obtain- recline on the Western Ghats region. The plantation was the ed from ANN results indicates that among the human-

Table 3 Variables in LR model and their coefficients Conditioning factors B S.E. Wald Df Sig. Exp(B)

Elevation −.001 .000 121.347 1 .000 .999 Slope .122 .005 610.232 1 .000 1.129 Distance to drainage −.002 .000 15.889 1 .000 .998 Rainfall .036 .002 445.464 1 .000 1.036 Lithology Acidic rocks 3.326 .943 12.443 1 .000 27.836 Basic rocks −17.091 9342.869 .000 1 .999 .000 Charnockite group of rocks 1.256 .895 1.968 1 .161 3.510 Khondalite group of rocks 2.727 .950 8.242 1 .004 15.287 Migmatite complex 2.843 .892 10.157 1 .001 17.159 Peninsular −17.378 40192.970 .000 1 1.000 .000 Peninsular gneissic complex 3.185 .897 12.591 1 .000 24.156 Land use Other plantation .214 .138 2.400 1 .121 1.238 Built-up .397 .709 .313 1 .576 1.487 Cropland −0.771 .631 1.490 1 .222 .463 Forest −.365 .127 8.320 1 .004 .694 Forest plantation .568 .196 8.402 1 .004 1.765 Plantation .400 .131 9.320 1 .002 10492.000 Quarry density −549.617 82.245 44.658 1 .000 .000 Road density .064 .047 1.829 1 .176 1.066 Constant −33.625 1.722 375.407 1 .000 .000 Arab J Geosci (2021) 14:838 Page 13 of 17 838

Fig. 7 Landslide susceptibility Palakkad map (LR) Thrissur

Ernakulam Munnar

Thodupuzha Idukki

Kattappana Tamil Nadu

Kottayam Kumily Legend

Towns

India States

Kerala Districts Landslide Susceptibility Very Low

Low Pathanamthitta Medium

High 010205km Alapurza Very High modified activities, quarry density is the most critical param- entire Idukki district. Landslide hotspots were identified in eter resulting in landslide occurrences, followed by road den- six villages of the district, and all the hotspots were located sity, built-up, other plantation, cropland, plantation, and forest in the Western Ghats region of the state. The spatial interac- plantation. Among the natural conditioning parameters, the tions between the regressor and predictors are evaluated using slope was observed as a highly influencing parameter for land- data-driven LR and ANN models. The results indicated that slide occurrences and was followed by rainfall, elevation, dis- both natural and anthropogenic conditioning parameters have tance to drainage, and geology. Results of both methods indi- a vital role in defining landslide susceptibility. According to cated that slope and rainfall have the maximum effect on land- LR, lithology was identified as the most significant natural slide occurrences. conditioning parameter, followed by the magnitude of slope and rainfall. Among anthropogenic conditioning parameters, plantations have the highest positive significant coefficient. Conclusions ANN clearly demonstrates the normalized importance of the variables, and the highest importance was observed for slope This study identified landslide hotspots in Idukki district followed by rainfall, quarry density, elevation, and drainage Kerala, based on 2018 landslides. Both LR and ANN methods density. Among the anthropogenic conditioning parameters, were applied for landslide susceptibility mapping for the the highest percentage importance was allotted for quarry 838 Page 14 of 17 Arab J Geosci (2021) 14:838

Fig. 8 Independent variable importance chart (ANN)

density, followed by road density and built-up areas. performance, the landslide events are considered a geological Landslide susceptibility map generated using the LR indicated engineering problem. It is essential to understand the mecha- a close relationship with the results of hotspot analysis. nism of landslides and analyze the relationship between condi- Figure 7 shows that very high and high landslide susceptibility tioning parameters and landslide occurrences, especially in area is concentrated on the center portion of the district, which landslide-prone areas like Idukki (H. Hong et al. 2017; Zhou exactly lies in the hotspot cluster identified Fig. 4. et al. 2018; M. Hong et al. 2020). ANN models fail to identify The prediction performance of both models was studied the relation between the conditioning parameters and landslide using the validation data set. ANN model was able to predict occurrences, whereas LR could overcome these issues. These 86.20% of landslide occurrences, while the LR model predicted findings provide enough evidence that both LR and ANN can 84.7%, respectively. The results showed that the ANN model be used to predict landslide susceptibility. Landslide suscepti- has the best predictive accuracy in landslide susceptibility bility studies have an important role in suggesting suitable de- study. Even though the ANN model showed better predictive velopmental activities, especially in a tourism-based district

Table 4 Classification table and statistical performance of LR and Model Observed Predicted ANN Landslide Percentage correct

01

LR Landslide 0 2419 395 86.0 1 376 1843 83.4 Overall percentage 84.7 ANN Landslide (training) 0 1711 273 86.2% 1 207 1349 86.7% Overall percentage 54.2% 45.8% 86.4% Landslide (testing) 0 717 113 86.4% 1 93 570 86.0% Overall percentage 54.3% 45.7% 86.2% Arab J Geosci (2021) 14:838 Page 15 of 17 838 like Idukki. The results and findings of the current research Chang K-T, Chiang S-H (2009) An integrated model for predicting – – could aid planners, engineers, and developers, especially for rainfall-induced landslides. Geomorphology 105(3 4):366 373. https://doi.org/10.1016/j.geomorph.2008.10.012 land-use planning and slope management. The model consid- Chau KT, Y. F. T. (2004) GIS based rockfall hazard map for Hong Kong. ered in this research could be utilized for future planning and Int J Rock Mech Min Sci 41(3):1–6 developmental activities in the study area. Chen W, Peng J, Hong H, Shahabi H, Pradhan B, Liu J, Zhu A-X, Pei X, Duan Z (2018) Landslide susceptibility modelling using GIS-based machine learning techniques for Chongren County, Jiangxi Province, China. Sci Total Environ 626:1121–1135. https://doi. Author contribution Sheelu Jones: data collection, analyses and interpre- org/10.1016/j.scitotenv.2018.01.124 tation, the conceptualization of manuscript, writing, reviewing, and Cogan J, Gratchev I (2019) A study on the effect of rainfall and slope editing of the manuscript. Dr. Kasthurba A.K: conceptualization, meth- characteristics on landslide initiation by means of flume tests. odology, and reviewing. Dr. Anjana Bhagyanathan: conceptualization, Landslides 16(12):2369–2379. https://doi.org/10.1007/s10346- methodology, and reviewing. Binoy B V: data interpretation, analyses, 019-01261-0 the conceptualization of manuscript, visualization, and reviewing. Cruden DM, Varnes DJ, Cruden DM, Varnes DJ (1996) Landslide types and processes, special report. Transportation Research Board, Availability of data and material Landslide inventory data is prepared by National Academy of Sciences 247:36–75 the Department of Geoinformation Science and Earth Observation (ITC), Dao D, Van Jaafari A, Bayat M, Mafi-Gholami D, Qi C, Moayedi H, Van University of Twente, Netherlands, in collaboration with the Geological Phong T, Ly H-B, Le T-T, Trinh PT, Luu C, Quoc NK, Thanh BN, Survey of India and the University of Kerala. Land-use map was prepared Pham BT (2020) A spatially explicit deep learning neural network from Sentinel-2 satellite image, and elevation data is obtained from model for the prediction of landslide susceptibility. CATENA USGS. The slope of the study area is generated from the Aster GDEM. 188(December 2019):104451. https://doi.org/10.1016/j.catena. Rainfall data is collected from the Indian Metrological Department (IMD) 2019.104451 and interpolated for the whole area. Geology and drainage data were Dapples F, Lotter AF, Van Leeuwen JFN, Van Der Knaap WO, acquired from Kerala State Land Use Board, Department of Planning Dimitriadis S, Oswald D (2002) Paleolimnological evidence for and Economic Affairs, Government of Kerala. The quarry data for the increased landslide activity due to forest clearing and land-use since study was obtained from the Forest Health Division, Kerala Forest 3600 cal bp in the western swiss alps. J Paleolimnol 27(2):239–248. Research Institute Peechi. Road networks are downloaded from Open https://doi.org/10.1023/A:1014215501407 Street Maps (OSM). District Census Handbook (2011) Idukki Operations, D. of C. District Census Handbook Idukki Code availability No codes are used. District Urbanisation Report Thiruvananthapuram Department of Town and Country Planning (2011) District Urbanisation Report Declarations Thiruvananthapuram (Issue January) Erener A, Düzgün HSB (2012) Landslide susceptibility assessment: what are the effects of mapping unit and mapping method? Environ Earth Conflict of interest The authors declare that they have no competing Sci 66(3):859–877. https://doi.org/10.1007/s12665-011-1297-0 interests. Fell R, Corominas J, Bonnard C, Cascini L, Leroi E, Savage WZ (2008) Guidelines for landslide susceptibility, hazard and risk zoning for land use planning. Eng Geol 102(3–4):85–98. https://doi.org/10. References 1016/j.enggeo.2008.03.022 Ground Water Information Booklet of Idukki District, K (2013) Ground water information booklet of Idukki District, Kerala. Ministry of Abraham MT, Satyam N, Pradhan B (2021) Forecasting landslides using Water Resources Central Ground Water Board, December mobility functions: a case study from Idukki district, India. Indian Guzzetti F, Carrara A, Cardinali M, Reichenbach P (1999a) Landslide Geotechnical Journal 12(1):540–559. https://doi.org/10.1007/ hazard evaluation: a review of current techniques and their applica- s40098-020-00490-8 tion in a multi-scale study , Central Italy. Geomorphology 31:181– Alex CJ, STV (2017) Mapping of Granite Quarries in Kerala , India: a 216 critical mapping initiative. Erudite Lecture Series of Prof. Madhav Guzzetti F, Carrara A, Cardinali M, Reichenbach P (1999b) Landslide Gaadgil, March hazard evaluation: a review of current techniques and their applica- Althuwaynee OF, Pradhan B, Park H-J, Lee JH (2014) A novel ensemble tion in a multi-scale study, Central Italy. In Geomorphology 31 decision tree-based CHi-squared Automatic Interaction Detection Guzzetti F, Reichenbach P, Cardinali M, Galli M, Ardizzone F (2005) (CHAID) and multivariate logistic regression models in landslide Probabilistic landslide hazard assessment at the basin scale. susceptibility mapping. Landslides 11(6):1063–1078. https://doi. Geomorphology 72(1–4):272–299. https://doi.org/10.1016/j. org/10.1007/s10346-014-0466-0 geomorph.2005.06.002 Arora MK, Das Gupta AS, Gupta RP (2004) An artificial neural network Guzzetti F, Mondini AC, Cardinali M, Fiorucci F, Santangelo M, Chang approach for landslide hazard zonation in the Bhagirathi (Ganga) K-T (2012) Landslide inventory maps: new tools for an old problem. Valley, Himalayas. Int J Remote Sens 25(3):559–572. https://doi. Earth Sci Rev 112(1–2):42–66. https://doi.org/10.1016/j.earscirev. org/10.1080/0143116031000156819 2012.02.001 Baeza C, Corominas J (2001) Assessment of shallow landslide suscepti- Hao L, van Westen AR, Martha CKSS, TR Jaiswal P, McAdoo B (2020) bility by means of multivariate statistical techniques. Earth Surf Constructing a complete landslide inventory dataset for the 2018 Process Landf 26(12):1251–1263. https://doi.org/10.1002/esp.263 Monsoon disaster in Kerala, India, for land use change analysis. Chae BG, Park HJ, Catani F, Simoni A, Berti M (2017) Landslide pre- Earth System Science Data Discussions 2:1–32. https://doi.org/10. diction, monitoring and early warning: a concise review of state-of- 5194/essd-2020-83 the-art. Geosci J 21(6):1033–1070. https://doi.org/10.1007/s12303- Harilal GT, Madhu D, Ramesh MV, Pullarkatt D (2019) Towards estab- 017-0034-4 lishing rainfall thresholds for a real-time landslide early warning 838 Page 16 of 17 Arab J Geosci (2021) 14:838

system in Sikkim, India. Landslides 16(12):2395–2408. https://doi. Ma F, Wang J, Yuan R, Zhao H, Guo J (2013) Application of analytical org/10.1007/s10346-019-01244-1 hierarchy process and least-squares method for landslide suscepti- Hemasinghe H, Rangali RSS, Deshapriya NL, Samarakoon L (2018) bility assessment along the Zhong-Wu natural gas pipeline, China. Landslide susceptibility mapping using logistic regression model Landslides 10(4):481–492. https://doi.org/10.1007/s10346-013- (a case study in Badulla District, Sri Lanka). Procedia Engineering 0402-8 212:1046–1053. https://doi.org/10.1016/j.proeng.2018.01.135 Makealoun S, Eka Putra DP, Wilopo W (2015) Landslide susceptibility Hong H, Pradhan B, Sameen MI, Chen W, Xu C (2017) Spatial predic- assessment of Kokap area using multiple logistic regression. Journal tion of rotational landslide using geographically weighted regres- of Applied Geology 6(2):53–61. https://doi.org/10.22146/jag.7217 sion, logistic regression, and support vector machine models in Martha TR, Roy P, Khanna K, Mrinalni K, Vinod Kumar K (2019) Xing Guo area (China). Geomatics, Natural Hazards and Risk Landslides mapped using satellite data in the Western Ghats of 8(2):1997–2022. https://doi.org/10.1080/19475705.2017.1403974 India after excess rainfall during August 2018. Curr Sci 117(5): Hong M, Jeong S, Kim J (2020) A combined method for modeling the 804. https://doi.org/10.18520/cs/v117/i5/804-812 triggering and propagation of debris flows. Landslides 17(4):805– Mishra V, Aaadhar S, Shah H, Kumar R, Pattanaik DR, Tiwari AD 824. https://doi.org/10.1007/s10346-019-01294-5 (2018) The Kerala flood of 2018: combined impact of extreme rain- Industrial Potential Survey (2017) Idukki District, D. I. A. C Industrial fall and reservoir storage. Hydrology and Earth System Sciences potential survey Idukki District Discussions, September, pp 1–13. https://doi.org/10.5194/hess- Jose T (2018) Kerala post disaster needs assessment floods and landslides 2018-480 August (Issue October) Mousavi SZ, Kavian A, Soleimani K, Mousavi SR, Shirzadi A (2011) Joy J, Kanga S, Singh SK (2019) Kerala flood 2018: flood mapping by GIS-based spatial prediction of landslide susceptibility using logistic – participatory GIS approach, Meloor Panchayat. International regression model. Geomatics, Natural Hazards and Risk 2(1):33 50. Journal on Emerging Technologies 10(1):197–205 https://doi.org/10.1080/19475705.2010.532975 Kanungo DP, Sharma S (2014) Rainfall thresholds for prediction of shal- Mutlu A, Goz F (2020) SkySlide: a hybrid method for landslide suscep- low landslides around Chamoli-Joshimath region, Garhwal tibility assessment based on landslide-occurring data only. The Himalayas, India. Landslides 11(4):629–638. https://doi.org/10. Computer Journal. I 1586104 1007/s10346-013-0438-9 Mutlu A, Goz F, Koksal K, Erener A (2019) Landslide susceptibility assessment using skyline operator and majority voting. Sakarya Kanungo DP, Singh R, Dash RK (2020) Field observations and lessons University Journal of Science, July, 782–787. https://doi.org/10. learnt from the 2018 landslide disasters in Idukki District, Kerala. 16984/saufenbilder.479801 Curr Sci 119(September):1797–1806 National Disaster Management Authority (2019) Annual Report Kaur H, Gupta S, Parkash S, Thapa R (2018) Knowledge-driven method: Nguyen V, Pham B, Vu B, Prakash I, Jha S, Shahabi H, Shirzadi A, Ba D, a tool for landslide susceptibility zonation (LSZ). Geology, Ecology, Kumar R, Chatterjee J, Tien Bui D (2019) Hybrid machine learning and Landscapes 00(00):1–15. https://doi.org/10.1080/24749508. approaches for landslide susceptibility modeling. Forests 10(2):157. 2018.1558024 https://doi.org/10.3390/f10020157 Kayastha P, Dhital MR, De Smedt F (2013) Application of the analytical Nourani V, Pradhan B, Ghaffari H, Sharifi SS (2014) Landslide suscep- hierarchy process (AHP) for landslide susceptibility mapping: a case tibility mapping at Zonouz Plain, Iran using genetic programming study from the Tinau watershed, west Nepal. Comput Geosci 52: and comparison with frequency ratio, logistic regression, and artifi- 398–408. https://doi.org/10.1016/j.cageo.2012.11.003 cial neural network models. Nat Hazards 71(1):523–547. https://doi. Kerala State Emergency Operations Centre Kerala State Disaster org/10.1007/s11069-013-0932-3 Management Authority (2016) Kerala, D. of R. and D. M. G. of Pham BT, Pradhan B, Tien Bui D, Prakash I, Dholakia MB (2016) A ’ ’ Kerala state disaster management plan towards a safer state comparative study of different machine learning methods for land- Kritikos T, Davies T (2015) Assessment of rainfall-generated shallow slide susceptibility assessment: a case study of Uttarakhand area landslide/debris-flow susceptibility and runout using a GIS-based (India). Environ Model Softw 84:240–250. https://doi.org/10. approach: application to western Southern Alps of New Zealand. 1016/j.envsoft.2016.07.005 – Landslides 12(6):1051 1075. https://doi.org/10.1007/s10346-014- Pham BT, Shirzadi A, Shahabi H, Omidvar E, Quoc NK, Lee S (2019) 0533-6 Landslide susceptibility assessment by novel hybrid machine learn- Kuriakose SL, Sankar G, Muraleedharan C (2009a) History of landslide ing algorithms. 1–25 susceptibility and a chorology of landslide-prone areas in the Pradhan B, Singh RP, Buchroithner MF (2006) Estimation of stress and – Western Ghats of Kerala, India. Environ Geol 57(7):1553 1568. its use in evaluation of landslide prone regions using remote sensing https://doi.org/10.1007/s00254-008-1431-9 data. Adv Space Res 37(4):698–709. https://doi.org/10.1016/j.asr. Kuriakose SL, Sankar G, Muraleedharan C (2009b) History of landslide 2005.03.137 susceptibility and a chorology of landslide-prone areas in the Prakashkumar DSMPGR (2019) Impact of landslides on the forest eco- Western Ghats of Kerala, India. Environ Geol 57(7):1553–1568. system in , Kerala with special reference to Floristic https://doi.org/10.1007/s00254-008-1431-9 Wealth Lee S, Pradhan B (2007) Landslide hazard mapping at Selangor, Rao PJ (1993) Landslide management and control in Himalayas. Malaysia using frequency ratio and logistic regression models. International Conference on Case Histories in Geotechnical Landslides 4(1):33–41. https://doi.org/10.1007/s10346-006-0047-y Engineering Lee CF, Li J, Xu ZW, Dai FC (2001) Assessment of landslide suscepti- Reichenbach P, Rossi M, Malamud BD, Mihir M, Guzzetti F (2018) A bility on the natural terrain of Lantau Island, Hong Kong. Environ review of statistically-based landslide susceptibility models. Earth Geol 40(3):381–391. https://doi.org/10.1007/s002540000163 Sci Rev 180(March):60–91. https://doi.org/10.1016/j.earscirev. Lee S, Ryu J-H, Lee M-J, Won J-S (2003) Use of an artificial neural 2018.03.001 network for analysis of the susceptibility to landslides at Boun, Rickli C, Graf F (2009) Effects of forests on shallow landslides - case Korea. Environ Geol 44(7):820–833. https://doi.org/10.1007/ studies in Switzerland. Forest Snow and Landscape Research 82(1): s00254-003-0825-y 33–44 Lee M-J, Choi J-W, Oh H-J, Won J-S, Park I, Lee S (2012) Ensemble- Sangchini EK, Nowjavan MR, Arami A (2015) Landslide susceptibility based landslide susceptibility maps in Jinbu area, Korea. Environ mapping using logistic statistical regression in Babaheydar Earth Sci 67(1):23–37. https://doi.org/10.1007/s12665-011-1477-y Watershed, Chaharmahal Va Bakhtiari Province. Iran İran’ ın Arab J Geosci (2021) 14:838 Page 17 of 17 838

Çaharmahal ve Bahtiyari Bölgesi ’ nde yer alan Baba Haydar Sub District. Wonosobo Regency, Central Java Province, Indonesia Havzası ’ nda lojistik regresyon kullanılarak he 65(1):30–40. https://webapps.itc.utwente.nl/librarywww/papers_2010/msc/aes/ https://doi.org/10.17099/jffiu.52751 wahono.pdf Santini M, Grimaldi S, Nardi F, Petroselli A, Rulli MC (2009) Pre- Wang Q, Li W, Wu Y, Pei Y, Xing M, Yang D (2016) A comparative processing algorithms and landslide modelling on remotely sensed study on the landslide susceptibility mapping using evidential belief DEMs. Geomorphology 113(1–2):110–125. https://doi.org/10. function and weights of evidence models. Journal of Earth System 1016/j.geomorph.2009.03.023 Science 125(3):645–662. https://doi.org/10.1007/s12040-016- Shaharban V, Amritha Rathnakaran (2019) Disaster prevention and manage- 0686-x ment in the era of climate change with special reference to Kerala Flood Yalcin A, Reis S, Aydinoglu AC, Yomralioglu T (2011) A GIS-based 2018. November. https://www.researchgate.net/publication/ comparative study of frequency ratio, analytical hierarchy process, 337444972_DISASTER_PREVENTION_AND_MANAGEMENT_ bivariate statistics and logistics regression methods for landslide IN_THE_ERA_OF_CLIMATE_CHANGE_WITH_SPECIAL_ susceptibility mapping in Trabzon, NE Turkey. CATENA 85(3): REFERNCE_TO_KERALA_FLOOD_2018 274–287. https://doi.org/10.1016/j.catena.2011.01.014 Suzen ML, Doyuran V (2004) A comparison of the GIS based landslide Yilmaz I (2010) Comparison of landslide susceptibility mapping meth- susceptibility assessment methods: multivariate versus bivariate. odologies for Koyulhisar, Turkey: conditional probability, logistic – Environ Geol 45(5):665 679. https://doi.org/10.1007/s00254-003- regression, artificial neural networks, and support vector machine. 0917-8 Environ Earth Sci 61(4):821–836. https://doi.org/10.1007/s12665- Thennavan E, Pattukandan Ganapathy G (2020) Evaluation of landslide 009-0394-9 hazard and its impacts on hilly environment of the Nilgiris District - Youssef AM, Pradhan B, Jebur MN, El-Harbi HM (2015) Landslide a geospatial approach. Geoenvironmental Disasters 7(1). https://doi. susceptibility mapping using ensemble bivariate and multivariate org/10.1186/s40677-019-0139-3 statistical models in Fayfa area, Saudi Arabia. Environ Earth Sci Tien Bui D, Tuan TA, Hoang N-D, Thanh NQ, Nguyen DB, Van Liem N, 73(7):3745–3761. https://doi.org/10.1007/s12665-014-3661-3 Pradhan B (2017) Spatial prediction of rainfall-induced landslides Zare M, Pourghasemi HR, Vafakhah M, Pradhan B (2013) Landslide for the Lao Cai area (Vietnam) using a hybrid intelligent approach of susceptibility mapping at Vaz Watershed (Iran) using an artificial least squares support vector machines inference model and artificial neural network model: a comparison between multilayer perceptron bee colony optimization. Landslides 14(2):447–458. https://doi.org/ (MLP) and radial basic function (RBF) algorithms. Arab J Geosci 10.1007/s10346-016-0711-9 6(8):2873–2888. https://doi.org/10.1007/s12517-012-0610-x Valencia Ortiz JA, Martínez-Graña AM (2018) A neural network model applied to landslide susceptibility analysis (Capitanejo, Colombia). Zêzere JL, Pereira S, Melo R, Oliveira SC, Garcia RAC (2017) Mapping – landslide susceptibility using data-driven methods. Sci Total Geomatics, Natural Hazards and Risk 9(1):1106 1128. https://doi. – org/10.1080/19475705.2018.1513083 Environ 589:250 267. https://doi.org/10.1016/j.scitotenv.2017.02. Van Thom B, Son PQ, Van Hung P, V. A. N. (2016) Research assessment 188 landslide and sedimentation of Hoa Binh hydropower reservoir. Zhou C, Yin K, Cao Y, Ahmed B, Li Y, Catani F, Pourghasemi HR Earth Sci 38(1):131–142 (2018) Landslide susceptibility modeling applying machine learning Van Westen CJ, Seijmonsbergen AC, Mantovani F (1999) Comparing methods: a case study from Longju in the Three Gorges Reservoir – landslide hazard maps. Nat Hazards 20(2–3):137–158. https://doi. area, China. Comput Geosci 112(September 2017):23 37. https:// org/10.1023/a:1008036810401 doi.org/10.1016/j.cageo.2017.11.019 Vineesh (2019) Impact assessment of Kerala flood 2018 & 2019. A Zhu L, Huang J (2006) GIS-based logistic regression method for land- Journal of Composition Theory, XII(Xi), 168–174 slide susceptibility mapping in regional scale. Journal of Zhejiang Wahono BFD (2010) Applications of statistical and heuristic methods for University-SCIENCE A 7(12):2007–2017. https://doi.org/10.1631/ landslide susceptibility assessments: a case study in Wadas Lintang jzus.2006.A2007