Remote Sensing and GIS-based landslide

susceptibility mapping using bivariate statistical

approach of Chhukha District
Sangay Gyeltshen (  [email protected] )
Jigme Namgyel Polytechnic https://orcid.org/0000-0003-1566-0354
Phub Dem
Royal University of Bhutan College of Natural Resources
Vasker Sharma
Jigme Namgyel Polytechnic
Rabindra Adhikari
Institute of Forestry

Research Article

Keywords: Bivariate Probabilistic Frequency Ratio, Logistic Regression, landslide susceptibility index,
remote sensing, Bhutan

Posted Date: February 9th, 2023

DOI: https://doi.org/10.21203/rs.3.rs-2380168/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Additional Declarations:
Tables are available in Supplementary Files section.
 1 Remote Sensing and GIS-based landslide susceptibility mapping using
2 bivariate statistical approach of Chhukha District

3 Abstract

4 Chhukha, a southern district of Bhutan remains susceptible to landslides due to excessive temporal
5 rainfall variability and land instability aggravated by anthropogenic factors. This has led to multiple
6 fatalities, substantial financial losses, and damages to infrastructure, farmland, and transportation
7 networks. This study developed the district scale Landslide susceptibility index (LSI) by a bivariate
8 statistical approach called Probabilistic Frequency Ratio (FR) and logistic regression (LR) with the help
9 of a geospatial technology system. A total of 236 historical landslide inventories were identified through
10 field deputation and google earth interpretation with the rationing of 70:30. 70% of the existing
11 landslides were used to train the models, while the remaining 30 % of them were used for model
12 validation. The FR model outperformed the LR model with an accuracy of 88.3% and 83.2%
13 respectively. The AUC model verification shows satisfactory agreement to predict landslide
14 susceptibility at the district scale in the Himalayan region. Both models indicated that the central and
15 northern parts of the district account for the least susceptibility, while the southern portion of the
16 Chhukha district accounts for the highest susceptibility to landslides. These authentic findings of the
17 research enable the local government and other decision-making bodies in developing policies,
18 implement innovative measures, and disseminate awareness and preparedness for the consequences of
19 landslide disasters.

20 Keywords: Bivariate Probabilistic Frequency Ratio, Logistic Regression, landslide susceptibility

21 index, remote sensing, Bhutan

22 1. Introduction

23 Landslides, an aggressive geological hazard characterized by the downward movement of soil, rocks,
24 and debris under the direct influence of gravity, pose a considerable risk to life, property, and the
25 foundation economy worldwide. In geographically fragile areas, this large mass displacement of
26 geological features occurs as a result of geo-morphological and hydrological parameters exceeding the
27 resistive force (Anbazhagan and Ramesh 2014; Walker and Shiels 2013). In addition, preparatory
28 factors such as road construction, urbanization, and deforestation trigger shallow landslides through the
29 removal of vegetation and the reduction of soil water infiltration (Nakileza and Nedala 2020). The
30 increase in groundwater table height and convergence of surface flow subsequently also decreases the
31 soil stability (Hennrich and Crozier 2004).

32 While often natural in origin, many landslides are triggered and aggravated by human-induced
33 factors. Destruction of infrastructure, disruption of transportation and telecommunications networks,
34 washing away of fertile soil, and fatalities are some serious risks associated with landslides in Bhutan
35 (Pasang and Kubíček 2020). To mitigate the landslide consequences, disaster management is explicitly
36 recognized as a part of sustainable development goals. The United Nations (UN) sustainable
37 development goal 15 which focuses on life on land and its 15.3 target aims to achieve zero land
38 degradation by 2030 by combating desertification and restoring degraded land (UN 2015). This shows
39 that landslides, one of the main drivers of land degradation, are obviously linked to the socio-economic
40 dimension of a country through reverberated impacts of production loss, lower biomass return, food
41 insecurity, and poverty in the end. The severity of the landslide's effects extends to impediments to the
42 provision of essential services, opportunities, and social activities (Winter et al. 2016). In this
43 anthropogenic era where economic development overtakes environmental conservation efforts, it is
44 indispensable to assess, map, and address the risk and vulnerability of landslides. Especially in a fragile
45 mountainous area like Bhutan, to mitigate the complex effect of landslides, assessing its likelihood and
46 risk are imperative.

47 Landslide susceptibility index (LSI) and mapping are crucial for risk assessment, land use
48 planning, and for mitigation of risks associated with landslides (Chen et al. 2018). It is the categorization
49 of homogeneous areas based on the magnitude of the risk caused by landslide events (Bera,
50 Mukhopadhyay, and Debasish 2019). To investigate the LSI, the qualitative approach, which develops
51 correlation between the landslide occurrences and its associated causative factors, and the quantitative
52 or statistical approach which develops statistical relationships among the factors are commonly used
53 (Fisiha 2020). The statistical approach is further subdivided into the deterministic and probabilistic
54 approach, where the deterministic approach is based on the law of conservation of mass, energy and
55 momentum while the probabilistic approach overcomes the limitation challenged by the deterministic
56 approach (Mark and Cees J. 1995). Artificial Neural Network (ANN) (Sevgen et al. 2019; Valencia
57 Ortiz and Martınez-Grana 2018), Frequency Ratio (FR) (Hidayat, Pachri, and Alimuddin 2019;
58 Rahaman et al. 2020), Logistic Regression (LR) (Chang et al. 2019), Analytic Hierarchy Process (AHP)
59 (Bera, Mukhopadhyay, and Debasish 2019), Weight of Evidence (WoE) (Cao et al. 2021; Elmoulat and
60 Brahim 2018; Pradhan, Oh, and Buchroithner 2010; Rahaman et al. 2020), and Random Forest (RF)
61 (Sevgen et al. 2019) are some widely used probabilistic approaches. In essence, the performance
62 analysis of landslide in this paper was carried out using frequency ratio (FR) and Logistic regression
63 (LR) as FR and LR models have the ability to attain the rank of the causative factors with respect to
64 landslide occurrence. In addition, it determines whether the given causative factor values are significant
65 enough to cause landslides or not.

66 Adopting a wide range of techniques, a plethora of landslide susceptibility assessment studies have
67 already been carried out around the world. However, due to versatility in handling a variety of spatial
68 data, simplicity in the decision-making process, and capability processes, RS and GIS has shown to be
69 an instrumental tool and technique for landslide susceptibility evaluation (Chang et al. 2019). Bhutan,
70 deep in the eastern Himalayas with a fragile ecosystem, increases vulnerability to monsoon
71 precipitation-driven landslides (Sarkar & Dorji 2019). However, in the areas where landslide incidents
72 are ubiquitous, landslide risk reduction via susceptibility assessments are scanty. The worldwide fatal
73 landslide database developed by Froude and Petley (2018) shows 64 deaths in Bhutan, the cause
74 attributed to non-seismic landslide events from 2004 to 2016. Due to the country’s fragile orography
75 and intense monsoonal rainfall, landslides, the devastating natural disasters in Bhutan are increasing.
76 Considering the country’s geo-climatic condition and past landslides incidents, a systematic landslide
77 study covering multiple causative factors was found urgent. In line with the aforementioned limitations,
78 the following objectives were specified; (i) to evaluate and map the landslide causative factors, (ii).
79 Generate landslide susceptibility index (LSI) using FR and LR, and (iii). Comparative analysis of FR
80 and LR models on the predicted LSI.

81 2. Study Area

82 The comprehensive landslide assessment was carried out in Chhukha, a district in southwestern Bhutan
83 with an elevation ranging from 144 m to 4426 masl. The regions consist of 1802 km2 of aerial coverage
84 geographically confined by latitudes 26˚45’0” to 27˚15’0” N and longitudes 89˚15’0” E to 90˚10’0” E
85 (Dikshit et al., 2020). With its 11 sub-districts, Chhukha currently serves as a main trade link for Bhutan
86 connecting India (Gariano et al. 2019). The vegetation is dominated by broad-leaf subtropical evergreen
87 forest and geological settings are composed of weak meta-sedimentary rocks in the lower part and high-
88 grade crystalline in upper parts and folded phyllite rocks turning to clay (Penjore, Sharma, and Chhetri
89 2017). The area receives 4000– 6000 mm annual rainfall and up to 800 mm/day, one of the highest
90 along the entire Himalayan front(Prokop and Walanus 2017). Characterized by weak geological settings
91 and receiving around 80% of the annual rainfall between June and September, rainfall induced
92 landslides are highly visible in the area threatening life of residents and commuters and risking
93 transportation of goods and services to the other part of districts (Figure 1). Furthermore, two (2) major
94 hydro-power plants (Chhukha & Tala) and two industrial estates, Phuentsholing and Pasakha industrial
95 estates are located within the district contributing immensely to people’s well-being and national
96 economic growth (Ministry of Economic Affairs (MoEA) 2005). Considering the socio-economic
97 importance of the place, its vulnerability to the landslide, and to address the likelihood of future
98 landslides based on the causative factors, the area was chosen for the current susceptibility assessment
99 pertaining to landslides (Figure 2).

100 3. Material and Methods

101 In developing the landslide susceptibility map (LSI), landslide inventory datasets preparation, causative
102 factors construction and application of frequency ratio (FR) and logistic regression (LR) were carried
103 out in the GIS and R programming environment. The overall method deployed for the study is shown
104 in Figure 3.

105 3.1 Datasets

106 Topographic, hydrological, and geographical factors are among the dominant factors in the occurrence
107 of landslides. Determination of such causative factors is a fundamental step in the susceptibility
108 modeling (Dou et al. 2015). Inferring the past landslide incidences and underlying geo-morphometric
109 conditions, 12 triggering factors collected from various organizations were considered in this study for
110 developing the LSI map using the FR and LR approach as tabulated below.

111 3.2 Landslide Inventory database

112 Understanding previous and contemporary landslide occurrences is crucial for predicting future
113 landslides (Das et al. 2013; Wu et al. 2016). Therefore, the first step in determining landslide
114 susceptibility is to compile the known landslide inventory. To do so, a total of 236 known landslide
115 inventories obtained through intensive field survey and google earth interpretation, were used to
116 produce the landslide susceptibility map of Chhukha district. Detecting the landslide under dense
117 vegetation and in an inaccessible area especially under mountainous regions were challenging.
118 Therefore, the manual field deputation was deployed to collect the spatial distributions of landslides.
119 Figure 2 represents the spatial distribution of landslides obtained through field survey and from google
120 earth platform, respectively.

121 Landslide training and testing samples are needed to develop landslide susceptibility maps utilizing
122 bivariate and soft computing approaches (Pradhan, Oh, and Buchroithner 2010). Where, the training
123 samples are used to develop the model while testing samples are used for model validation. These
124 samples are chosen based on the completeness and distribution of the landslide inventory, the size of
125 the study area, and the type of approach used. Although there are no set protocols for choosing training
126 and testing samples, it is crucial that the samples be chosen in a way that ensures their independence
127 from one another (Yilmaz 2009).Basheer and Hajmeer 2000 have documented different splitting ratios
128 used for various statistical methods. According to their summary, (Swingler 1996) and (Gopal and
129 Woodcock 1996) used 80: 20 ratios for training and testing of their models. Most recently, (Guzzetti et
130 al. 2012) used 70: 30 ratios for delineating landslides from high-resolution images and applied by (Tien
131 Bui et al. 2013) for predicting landslides using soft computing techniques. These studies have used
132 random sampling techniques, which are commonly employed in predicting landslides. In order to make
133 the models computationally robust, a standard rationing of 70:30 rationing was used. The spatial
134 locations of the landslides are constructed using different sampling strategies. The points vector format
135 was used to portray the spatial location of landslides in the study.
136 3.3 Causative factors database

137 The extent, distribution and density of landslides events are primarily controlled by topography,
138 hydrology, geology and anthropogenic factors (Khan et al., 2019). The selection of the causative factors
139 depends on the scope, time periods and availability of datasets (Mersha and Meten 2020; Rabby and Li
140 2020). To produce quality LSI, a total of 12 causative factors covering geological factors such as
141 lithology, soil and fault distance; topographical factors such as slope, aspect, TWI and curvature;
142 hydrological factors such as distance to stream, and anthropogenic features such as LULC and road
143 distance respectively were used.

144 3.3.1 Geological Factors

145 The geological setup of the area inherently drives the landslides occurrence. Especially in
146 tectonically active areas, lithology is considered the most influential factor in triggering landslides
147 because of its influence on the geo-mechanical characteristics of a terrain (Martínez-Álvarez et al.
148 2013). A total of six lithological units were mapped as shown in table 2 and figure 4(a). The active and
149 dormant landslides can be observed in phyllite and quartzite (Group II) litho-unit followed by Group
150 IV, VI and I, while the rest of the litho-unit (Group III and V) are absent of landslide features.

151 The lineament of tectonic fault results in fractured strata and accelerates the weathering process.
152 The previous studies have also revealed the spatial distribution of the landslides highly extend along
153 the major tectonic faults (Cao et al. 2021). Considering the huge influence of faults on landslide
154 distribution, distance to faults was constructed in the GIS platform as represented in figure 4b. Soil
155 texture, the relative proportion of sand, silt and clay, is an important landslide predisposing factor that
156 controls the water holding capacity of soil and infiltration rates. The stability of the slope depends on
157 the soil strength which in turn depends on the local lithology (Rabby and Li 2020). A 250m grid soil
158 texture map (Figure 4c) was extracted from the ISRIC-World Soil Information website. Six different
159 soil types such as acrisol, alisol, cambisol, leptosol, luvisol and vertisol were mapped for the
160 susceptibility assessment.

161 3.1.1 Anthropogenic Factors

162 While often natural in origin, many landslides are triggered by human activities. Catalyzed by
163 rainfall, the occurrence and reactivation of landslides are influenced by land use and its changes(Karsli
164 et al. 2009). Shallow landslides, such as those observed in Gondo, Switzerland, to large rock avalanches
165 such as Elm, Switzerland, or Frank slide, Canada are all triggered by human activities (Michoud et al.
166 2011). Land use changes driven by infrastructure development and road construction are activating
167 landslide incidences everywhere. In Bhutan, landslides are expected to increase due to road construction
168 and deforestation activities for infrastructural development, which is directly linked to the increasing
169 population(Dikshit et al. 2020). LULC change and distance to road were considered as anthropogenic
170 factors influencing landslide occurrences (figure 4d).

171 Land cover strongly controls the distribution of landslides. Lands with vegetative cover have
172 strong root systems which stabilizes the slope reducing mechanical and hydrological actions(Khan et
173 al. 2019). The slope cutting due to road construction disturbs the original topography and destabilizes
174 the slope (Demir, Aytekin, and Akgun 2015). The LULC map prepared by the Ministry of Agriculture
175 and Forest with 10 land cover classes was used as one anthropogenic causative factor. The distance to
176 roads map was reclassified into seven classes: 0–1500 m, 1500–3000 m, 3000–6000 m, 6000–8000 m,
177 8000-12000 m, 12000-14000 m, and >14000 m respectively (figure 4e).

178 3.1.2 Hydrological Factors

179 Hydrological elements have a strong relationship to landslide incidences. Both closer distance to rivers
180 and high amounts of precipitation triggers landslides through undercutting and eroding, reducing the
181 slope stability(Rong et al. 2020; Wang et al. 2015). Chhukha district receives the highest rainfall,
182 accounting for 80% of the annual rainfall during monsoon season (June -September) along the
183 Himalayan front (Prokop and Walanus 2017). The landslide inventory data indicates the higher
184 landslide incidents in regions with higher amounts of annual precipitation. To determine the extent to
185 which the precipitation affects the slopes stability and landslide incidences, eight classes of distance to
186 river were created using the Euclidean distance tool in the GIS environment as shown in figure 4f.

187 3.1.1 Topographical Factors

188 Topographical factors including the derivatives of digital elevation model, including slope features
189 (gradient and aspect) and curvature dimensions are considered as dominant factors towards landslide
190 incidences(Alkhasawneh et al. 2013). Its variation along the area indicates variation in geological
191 strength and hydrodynamical actions. Slope gradient (Figure 5a and 5b) is recognized as an important
192 input parameter in susceptibility studies due to its direct impact on landslide formation (Wang et al.
193 2017). However, it is not always imperative to consider the occurrences of landslides directly related to
194 slope gradient, since landslides are also being contributed by the combination of other inducing and
195 triggering factors such as hydrological and geological factors.

196 The derivation of Topographic Wetness Index (TWI) is significant in landslide susceptibility
197 mapping which estimates the accumulation of water through topographic differences (Ró&zgrave;ycka,
198 Migoń, and Michniewicz 2016). It is mathematically expressed as a function of slope in radian and
199 upstream catchment area (m2). It is also used as a proxy to estimate the soil moisture contents (Kopecký,
200 Macek, and Wild 2021), mathematically represented as follow:

(𝐴
201 𝑇𝑊𝐼 = 𝑙𝑛 𝑡𝑎𝑛𝛽) (1)
202 Where, A represents the upstream catchment area and tanβ is the local slope in radian. According to
203 (Saleem et al. 2019), the interpolation of digital elevation model (DEM) undermines quantifying and
204 estimating the distribution of TWI for depressed landscapes (Figure 5c). Several researchers across the
205 globe have used TWI, as a conditioning factors for landslide susceptibility analysis, for instances, in
206 China (Yu, Zhu, and Liu 2017), South Ethiopia(Shano, Raghuvanshi, and Meten 2021), Hindu
207 Kush(Rahaman et al. 2020), Bangladesh (Rabby and Li 2020), Cambodia (Lee and Sambath 2006) and
208 Pakistan (Khan et al. 2019). High TWI values reflect high susceptibility to landslides and vice (Grimm,
209 Tahmasebi Nasab, and Chu 2018).

210 Curvature is linked to the rate of change of the slope (Wang et al. 2017). General Curvature, the
211 curvature of the surface itself, plan curvature, the horizontal plane curvature formed due to intersection
212 between the horizontal plane and the surface, and the profile curvature, the vertical plane parallel to the
213 slope direction, affects the acceleration and deceleration of down-slope flows and influences erosion
214 and deposition (Alkhasawneh et al. 2013; Kannan, Saranathan, and Anabalagan 2013). In this study,
215 the general, plan and profile curvature were derived from the 30 m resolution DEM and were
216 categorized into convex (+values), concave (-ve values) and linear or flat (0) surface (figure 5d, 5e and
217 5f).

218 3.1.2 Probabilistic Frequency Ratio (FR)

219 The probabilistic frequency ratio (FR), a frequent and effective bivariate statistical and quantitative
220 modeling technique (Khan et al. 2019), was used to determine the presence of landslides on individual
221 causative factors by rationing existence of landslide to its respective class (Mahalingam, J. Olsen, and
222 S. O’Banion 2016). This approach helps in determining the strength of spatial correlation between the
223 distribution and location of each landslide and each of causative factors based on the FR values obtained
224 through rationing(Lee and Pradhan 2006; Reis et al. 2012). FR value for each class can be calculated
225 using the following Equation 2 (Rabby and Li 2020):

𝑁𝑖𝑗
𝑁𝑡𝑜𝑡𝑎𝑙
226 𝐹𝑅𝑗 = 𝐴𝑖𝑗 (2)
𝐴𝑡𝑜𝑡𝑎𝑙

227 Where:

228 ● Nij = total landslide pixels within the jth subclass of factor i;
229 ● Ntotal = number of landslide pixels in the area;
230 ● Aij = total pixels of the jth subclass of factor i;
231 ● A Total = the total number of pixels in the study area.

232 The landslide susceptibility index (LSI) was derived by multiplying and summing the FR values with
233 their corresponding conditioning factors, mathematically expressed as in equation 3:
234 𝐿𝑆𝐼 = 𝑐1 . 𝐹1 + 𝑐2 . 𝐹2 + 𝑐3 . 𝐹3 + ⋯ . . +𝑐𝑛 . 𝐹𝑛 (3)

235 Where:

236 ● LSI = the landslide susceptibility index study area;

237 ● -c1, c2,c3, ….cn=represent the coefficient of the factors derived using frequency ratio model;
238 ● F1, F2, F3…..Fn = represents the landslide inducing and triggering factors;
239 ● N = the total number of factors.

240 Generally, FR model value represents the association between causative factors and landslide
241 occurrence such as 1 represents average correlation, >1 indicates strong, and <1 demonstrates weak
242 correlation respectively(Lee and Pradhan 2006; Rabby and Li 2020).

243 3.1.3 Logistic Regression (LR)

244 The multivariate regression relationship between a dependent variable and numerous independent
245 variables can be evaluated using logistic regression. It is employed when the response variable is a
246 categorical variable involving binary 0/1 (Rasyid, Bhandary, and Yatabe 2016). It models non-normal
247 distribution and uses the logit link function given by:

𝑦
248 𝑙𝑜𝑔 𝑙𝑜𝑔 (
1−𝑦
) = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ + 𝛽𝑛 𝑥𝑛 + 𝜀 (4)

249 Where y is the probability of occurrence of landslide event while (y/1-y) is the odds ratio and log(y/1-
250 y) are the log odds ratio. β0 is the intercept and β1, β2…. βn are coefficients that measure the
251 contribution of each independent variable (x1, x2….xn) which are landslide influencing factors and ε is
252 the error.

253 3.6.7 LSI validation using Receiver Operating Characteristics (ROC) Curve

254 The probability metric curve which defines the degree of separability (AUC) was used to determine the
255 performance of the proposed models generally plotted between specificity (True Positive Rate) and
256 sensitivity (True Negative Rate). Sensitivity defines the location of existing landslides being correctly
257 classified into susceptible zones, while specificity determines the proportion of non-landslide points
258 being classified under low susceptible zones(Mahalingam, J. Olsen, and S. O’Banion 2016). The AUC
259 value ranges in between 0-1, depicting the level of preciseness of the model, i.e., 0.9–1, excellent; 0.8–
260 0.9, very good; 0.7–0.8, good; 0.6–0.7, average; and 0.5–0.6, poor respectively (Bourenane, aid
261 Guettouche, and Bouhadad 2016; Mahalingam, J. Olsen, and S. O’Banion 2016). In this study, 30%
262 (testing samples) of total historical landslides data were used to validate the LSM.
263 4. Results and Discussion

264 4.1 Descriptive analysis of independent conditioning variables

265 The degree of correlation between the landslide causative factors and the existing landslides
266 distribution trained by 70% of the training samples, randomly distributed, was used for determining the
267 LSM of the region. The FR values of individual sub-class and causative factors are indicated in Table
268 3. The FR model indicates a strong correlation between landslide inventory, LULC, soil type, geology
269 and river distance accounting 33.6, 31.8, 27.4 and 19.2, respectively. Remaining causative factors such
270 as fault distance, slope, aspect, river distance, road distance, and TWI shows moderate correlation with
271 the existing landslide events ranging from 5.8-9.8. Exceptionally, the general, profile and plan curvature
272 shows average correlation with FR with value1, 1.3, and 1.4 respectively. Using the coefficient
273 generated from FR model, equation 3 is remodified as (equation 5):

274 𝑳𝑺𝑰 = 𝟔. 𝟐 ∗ 𝐹𝑎𝑢𝑙𝑡 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 + 𝟖. 𝟓 ∗ 𝑆𝑙𝑜𝑝𝑒 + 𝟐𝟕. 𝟒 ∗ 𝑮𝒆𝒐𝒍𝒐𝒈𝒚 + 𝟓. 𝟖 ∗ 𝐴𝑠𝑝𝑒𝑐𝑡 + 𝟏

275 ∗ 𝐺𝑒𝑛𝑒𝑟𝑎𝑙 𝑐𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒 + + 𝟏. 𝟒 ∗ 𝑃𝑙𝑎𝑛

276 𝑐𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒 + 𝟏. 𝟑 ∗ 𝑝𝑟𝑜𝑓𝑖𝑙𝑒 𝑐𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒 + 𝟔. 𝟑 ∗ 𝑇𝑊𝐼 + 𝟏𝟗. 𝟐 ∗ 𝑅𝑖𝑣𝑒𝑟 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 + 𝟑𝟑. 𝟔

277 ∗ 𝑳𝑼𝑳𝑪 + 𝟗. 𝟖 ∗ 𝑅𝑜𝑎𝑑 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 +

278 𝟑𝟏. 𝟖 ∗ 𝑆𝑜𝑖𝑙 𝑇𝑦𝑝𝑒 (5)

279 As indicated in table 2, the geology group (I), (II) and (IV) demonstrated a strong relationship
280 with the frequency of existing landslides with FR values greater than 1, weak correlation for group (VI)
281 with value less than 1 and no associated relationship with group (III) and (V), respectively. On the
282 contrary, fault distance (m) less than 1500 m and greater than 7500 m from the actual fault locations
283 demonstrated strong correlation with respect to landslide occurrence, however, weak correlation was
284 observed between 1500-7500 m. Leptosols (having shallow depth soil profile), acrisols (clay-rich
285 subsoil) and cambisols (weak weathering parent materials-absence of clay) as per the nomenclature of
286 ISRIC/FAO soil grids datasets have shown strong correlation with the existing landslide’s distribution
287 with FR values greater than 1, while luvisols with moderate correlation and alfisols and vertisols with
288 no correlation with the existing landslide distributions having FR value 0.

289 Considering the landslide’s distribution towards the topographic factors, slope and aspect have
290 strong influences over landslide occurrence. The maximum correlation of FR values greater than 1 are
291 observed under 0–20-degree slope gradients followed by 20–60-degree gradients and weak association
292 for slopes greater than 60 degrees. It is obvious that the landslide distribution over flat surface is null;
293 however, the topography facing towards southwest, east, northeast and south have strong correlation
294 and weak correlation towards west, south, southeast, northwest and southwest, respectively. Although,
295 the curvature has shown moderate correlation between the landslide occurrences, concave and convex
296 curvature under profile curvature, linear and convex under flat under general curvature have
297 demonstrated strong correlation while, rest of the category have shown either weak or no correlation
298 with the existing landslides distribution.

299 The anthropogenic factors, such as land use management, are critical for slope stability(Singh
300 et al. 2020). Higher density of vegetation ensures high bonding capabilities between the soil particles,
301 thus increasing the soil stability(Fattet et al. 2011). For example, the current study, which consists of 6
302 land use types, were categorized as per the guidelines prescribed by Department of Forestry and Park
303 Services (DoFPS). Water bodies, shrub and cultivated agriculture provided highest correlation with the
304 existing landslides distribution, while vegetation coverage like meadows and built-up have shown no
305 correlation and moderate for forests with the existing landslides distribution. Additionally, the widening
306 activities of roads have developed an extra threat, and triggered the landslides occurrences. In this
307 study, the distance buffered with 0-1500 m and 1500-3000 m have strong influences, while 12000-
308 14000 m with average correlation, and the rest of the buffered distance with no or weaker correlation
309 with the existing distribution of landslides, respectively.

310 4.2 Landslide susceptibility mapping

311 The landslide susceptibility Index (LSI) depicts the correlation between the existing distribution of
312 landslides and the conditioning factors. The application of probabilistic FR and LR models aids in
313 determining the weight of each causative factor and its sub-layer, which later summed to derived LSI
314 of the region (figure 6). The landslide susceptibility index was prepared based on statistical significance
315 and correlation coefficient derived using FR and LR model.

316 The produced LSI values were categorized based on a natural break algorithm to classify the
317 landslides susceptibility zone, i.e., Very High, High, Moderate, and Low and Very Low susceptibility
318 index. Both models provide comparable areas under various susceptibility categories. Table 4
319 demonstrates the area statistics of both models along with the level of landslide susceptibility and the
320 frequency of existing landslides.

321 4.3 Validation of susceptibility map

322 For the validation of LSI, the models were trained using 70% of the total landslides distributed in the
323 area, while 30% of the total landslides was used for the validation of LSI using AUC/ROC model. The
324 plot was constructed against true positive rate and false positive rate to arrive at model accuracy for the
325 predicted LSI’s. The prediction rate of LSI using FR model exceeded the LR model with 0.883 and
326 0.832 respectively as demonstrated in figure 7.

327

328
329 4.4 Landslide susceptibility mapping for sub-district (geog) levels

330 In order to disseminate the level of risk possessed by landslides, and to help develop the mitigation
331 measures, the district level susceptibility was further narrowed down to the sub-districts level. Hence,
332 figure 8 depicts the spatial pattern of susceptible category under Chhukha Districts with the amount of
333 land falling within each susceptible zone.

334 Samphelling (J), Dala (D), and Phuentsholing (I) accord the highest coverage of the predicted area under
335 very high susceptibility for the events based on predicted LSI using FR model, with respective coverage
336 rates of 71%, 63%, and 60%. The remaining geogs have exposure rates ranging from 2% to 22% of the
337 total area under the high susceptible zone. Logchina account the highest susceptible in high category
338 followed by Dala and Geling geogs with 36%, 29% and 26% of the total area favored due to the river
339 flowing through within the respective boundary. Under very low to low category, Dungna, Metap, and
340 Getana geogs accounts highest areal coverage which is least susceptible to landslide with ranging from
341 31%-38% of the total area.

342 5. Conclusion

343 The Landslide Susceptibility Map of Chhukha District was produced with 12 causative factors using
344 FR and LR models, i.e., topographical factors (slope, aspect, TWI and curvature), geological factors
345 (lithology and fault distance), hydrological factors (river distance), anthropogenic factors (road distance
346 and LULC) and soil type among others. According to the FR model, the strongest correlation between
347 the causative factors and the existing inventory of landslides was provided by LULC, followed by soil
348 type, geology, and river distance. In contrast, the LR model demonstrated a significant correlation
349 between the existing landslide distribution and geology, land use, slope, river distance, road distance,
350 and profile curvature. Owing to the great performance of the FR model, the regional LSI was further
351 narrowed down to sub-district level to assess the level of susceptibility for individual sub-district (geog).
352 Dala, Phuentsholing and Samphelling geog are highly vulnerable to landslide events followed by
353 Bjachho, Chapcha and Logchina area, while rest of the geogs are with low to moderate risk of
354 landsliding events. The LSI of the region demonstrates the spatial distribution of existing landslides and
355 its probability of occurrence in future. Therefore, the regional LSI map shall aid the engineer, planner,
356 and the local government bodies to come up with innovative measures to mitigate the landsliding events.

357 6. Acknowledgement

358 The authors would like to acknowledge, Department of Geology and Mines, MoEA for providing the
359 regional scale structural geological datasets and RNR Statistic, MoAF for providing regional land use
360 map and National Land Commission Secretariat (NLCS) for the basic datasets such as road, river
361 network etc…
362 7. Funding

363 The authors declare that no funds, grants, or other support were received during the preparation of this
364 manuscript.

365 8. Competing Interests

366 The authors have no relevant financial or non-financial interests to disclose.

367 9. Author’s contribution

368 S.G. conceived the core idea of research. S. G, P.D. V.S. and R. A. contributed their expertise in
369 mapping, reviewing, writing and organizing the literature. Every author has provided their tremendous
370 support for the entire research tenure for the successful completion of the research works.

371 10. References

372 Alkhasawneh, Mutasem Sh et al. 2013. “Determination of Important Topographic Factors for Landslide
373 Mapping Analysis Using MLP Network.” The Scientific World Journal 2013(May 2014).

374 Anbazhagan, Siddan, and Veerappan Ramesh. 2014. “Landslide Hazard Zonation Mapping in Ghat
375 Road Section of Kolli Hills, India.” J. Mt. Sci. 11(5): 1308–25.

376 Basheer, I.A., and M. Hajmeer. 2000. “Artificial Neural Networks: Fundamentals, Computing, Design,
377 and Application.” Journal of Microbiological Methods 43: 3–31.

378 Bera, Amit, B.P. Mukhopadhyay, and Das Debasish. 2019. “Landslide Hazard Zonation Mapping Using
379 Multi‑criteria Analysis with the Help of GIS Techniques: A Case Study from Eastern
380 Himalayas, Namchi, South Sikkim.” Natural Hazard, Springer.

381 Bourenane, Hamid, Mohamed aid Guettouche, and Youcef Bouhadad. 2016. “Landslide Hazard
382 Mapping in the Constantine City, Northeast Algeria Using Frequency Ratio, Weighting Factor,
383 Logistic Regression, Weights of Evidence, and Analytical Hierarchy Process Methods.” Arab
384 J Geosci 9(154).

385 Cao, Yanbo et al. 2021. “Landslide Susceptibility Assessment Using the Weight of Evidence Method:
386 A Case Study in Xunyang Area, China.” PLoS ONE 16(1 January): 1–18.
387 Chang, Kuan Tsung et al. 2019. “Evaluating Scale Effects of Topographic Variables in Landslide
388 Susceptibility Models Using GIS-Based Machine Learning Techniques.” Scientific Reports
389 9(1): 1–22.

390 Chen, Wei et al. 2018. “Landslide Susceptibility Modelling Using GIS-Based Machine Learning
391 Techniques for Chongren County, Jiangxi Province, China.” Elsevier.

392 Das, H. O., H. Sonmez, C. Gokceoglu, and H. A. Nefeslioglu. 2013. “Influence of Seismic Acceleration
393 on Landslide Susceptibility Maps: A Case Study from NE Turkey (the Kelkit Valley).”
394 Landslides 10(4): 433–54.

395 Demir, Gokhan, Mustafa Aytekin, and Aykut Akgun. 2015. “Landslide Susceptibility Mapping by
396 Frequency Ratio and Logistic Regression Methods: An Example from Niksar–Resadiye (Tokat,
397 Turkey).” Arabian Journal of Geosciences 8(3): 1801–12.

398 Dikshit, Abhirup et al. 2020. “Spatial Landslide Risk Assessment At.” Geosciences 10(131): 1–16.

399 Dou, Jie et al. 2015. “Optimization of Causative Factors for Landslide Susceptibility Evaluation Using
400 Remote Sensing and GIS Data in Parts of Niigata, Japan.” PLoS ONE 10(7).
401 http://dx.doi.org/10.1371/journal.pone.0133262.

402 Elmoulat, Meryem, and Lahcen Ait Brahim. 2018. “Landslides Susceptibility Mapping Using GIS and
403 Weights of Evidence Model in Tetouan-Ras-Mazari Area (Northern Morocco).” Geomatics,
404 Natural Hazards and Risk 9(1): 1306–25.

405 Fattet, M. et al. 2011. “Effects of Vegetation Type on Soil Resistance to Erosion: Relationship between
406 Aggregate Stability and Shear Strength.” Fuel and Energy Abstracts 87: 60–69.

407 Fisiha, Betelhem Tesfaye. 2020. “Landslide Susceptibility Mapping Using GIS Based Modified
408 Frequency Ratio and Information Value Method- a Case Study of Debretabor to Alember Road
409 Section and Its Surrounding Arera, North Western Ethiopia.” PhD Thesis. UNIVERSITY OF
410 GONDAR. http://hdl.handle.net/123456789/3158.

411 Froude, Melanie J., and David N. Petley. 2018. “Global Fatal Landslide Occurrence from 2004 to
412 2016.” Natural Hazards and Earth System Sciences 18(8): 2161–81.

413 Gariano, Stefano Luigi et al. 2019. “Automatic Calculation of Rainfall Thresholds for Landslide
414 Occurrence in Chukha Dzongkhag, Bhutan.” Bulletin of Engineering Geology and the
415 Environment 78(6): 4325–32.
416 Gopal, Sucharita, and Curtis Woodcock. 1996. “Remote Sensing of Forest Change Using Artificial
417 Neural Networks.” Geoscience and Remote Sensing, IEEE Transactions on 34: 398–404.

418 Grimm, Kendall, Mohsen Tahmasebi Nasab, and Xuefeng Chu. 2018. “TWI Computations and
419 Topographic Analysis of Depression-Dominated Surfaces.” Water, MDPI 10(663).

420 Guzzetti, Fausto et al. 2012. “Landslide Inventory Maps: New Tools for an Old Problem.” Earth-
421 Science Reviews 112: 42–66.

422 Hennrich, Kirsten, and Michael J. Crozier. 2004. “A Hillslope Hydrology Approach for Catchment-
423 Scale Slope Stability Analysis.” Earth Surface Processes and Landforms 29(5): 599–610.

424 Hidayat, S., H Pachri, and I Alimuddin. 2019. “Analysis of Landslide Susceptibility Zone Using
425 Frequency Ratio and Logistic Regression Method in Hambalang, Citeureup District, Bogor
426 Regency, West Java Province.” In The 4th International Conference of Indonesian Society for
427 Remote Sensing,.

428 Kannan, M., E. Saranathan, and R. Anabalagan. 2013. “Landslide Vulnerability Mapping Using
429 Frequency Ratio Model: A Geospatial Approach in Bodi-Bodimettu Ghat Section, Theni
430 District, Tamil Nadu, India.” Arabian Journal of Geosciences 6(8): 2901–13.

431 Karsli, F. et al. 2009. “Effects of Land-Use Changes on Landslides in a Landslide-Prone Area (Ardesen,
432 Rize, NE Turkey).” Environmental Monitoring and Assessment 156(1–4): 241–55.

433 Khan, Hawas et al. 2019. “Landslide Susceptibility Assessment Using Frequency Ratio, a Case Study
434 of Northern Pakistan.” The Egyptian Journal of Remote Sensing and Space Science 22(1): 11–
435 24.

436 Kopecký, Martin, Martin Macek, and Jan Wild. 2021. “Topographic Wetness Index Calculation
437 Guidelines Based on Measured Soil Moisture and Plant Species Composition.” Science of The
438 Total Environment 757: 143785.

439 Lee, Saro, and Biswajeet Pradhan. 2006. “Probabilistic Landslide Hazards and Risk Mapping on Penang
440 Island, Malaysia.” Journal of Earth System Science 115(6): 661–72.

441 Lee, Saro, and Touch Sambath. 2006. “Landslide Susceptibility Mapping in the Damrei Romel Area,
442 Cambodia Using Frequency Ratio and Logistic Regression Models.” Environmental Geology
443 50(6): 847–55.
444 Mahalingam, Rubini, Michael J. Olsen, and Olsen S. O’Banion. 2016. “Evaluation of Landslide
445 Susceptibility Mapping Techniques Using Lidar-Derived Conditioning Factors (Oregon Case
446 Study).” GEOMATICS, NATURAL HAZARDS AND RISK, Taylor and Francis Group7 7(6).

447 Mark, T. J, Terlien, and Van Westen Cees J. 1995. “Deterministic Modelling in Gis_based Landslide
448 Hazard Assessment.” Advance in Natural and Technological Hazards Research, Springer 5:
449 57–77.

450 Martínez-Álvarez, F., J. Reyes, A. Morales-Esteban, and C. Rubio-Escudero. 2013. “Determining the
451 Best Set of Seismicity Indicators to Predict Earthquakes. Two Case Studies: Chile and the
452 Iberian Peninsula.” Knowledge-Based Systems 50: 198–210.

453 Mersha, Tilahun, and Matebie Meten. 2020. “GIS-Based Landslide Susceptibility Mapping and
454 Assessment Using Bivariate Statistical Methods in Simada Area, Northwestern Ethiopia.”
455 Geoenvironmental Disasters 7(1).

456 Michoud, Clément et al. 2011. “Classification of Landslide-Inducing Anthropogenic Activities.” 5th
457 Canadian Conference on Geotechnique and Natural Hazards (April): 231–33.

458 Ministry of Economic Affairs (MoEA). 2005. Dzongkhag-Wise Inventory of Resources-Bhutan;

459 Chhukha Dzongkhag. Thimphu: Bhutan. https://www.moea.gov.bt/.

460 Nakileza, Bob Roga, and Shafiq Nedala. 2020. “Topographic Influence On Landslides Characteristics
461 And Implication For Risk Management In Upper Manafwa Catchment, Mt Elgon Uganda.” :
462 1–28.

463 Pasang, Sangey, and Petr Kubíček. 2020. “Landslide Susceptibility Mapping Using Statistical Methods
464 along the Asian Highway, Bhutan.” Geosciences 10(11): 430.

465 Penjore, Tshering, S.R. Sharma, and Lalit Kumar Chhetri. 2017. Integrated Geo-Hazard Risk
466 Assessment of Critical Barsa Watershed under Chukha Dzonkhag. Thimphu, Bhutan:
467 Department of Geology and Mines.

468 Pradhan, Biswajeet, Hyun-Joo Oh, and Manfred Buchroithner. 2010. “Weights-of-Evidence Model
469 Applied to Landslide Susceptibility Mapping in a Tropical Hilly Area.” Geomatics, Natural
470 Hazards and Risk 1(3): 199–223.

471 Prokop, Pawe\textbackslashl, and Adam Walanus. 2017. “Impact of the Darjeeling–Bhutan Himalayan
472 Front on Rainfall Hazard Pattern.” Natural Hazards 89(1): 387–404.
473 Rabby, Yasin Wahid, and Yingkui Li. 2020. “Landslide Susceptibility Mapping Using Integrated
474 Methods: A Case Study in the Chittagong Hilly Areas, Bangladesh.” geosciences, MDPI 10:
475 483.

476 Rahaman, Ghani et al. 2020. “Assessment of Landslide Susceptibility Using Weight of Evidence and
477 Frequency Ratio Model in Shahpur Valley, Eastern Hindu Kush.” Natural Hazards and Earth
478 System Sciences.

479 Rasyid, Abdul Rachman, Netra P. Bhandary, and Ryuichi Yatabe. 2016. “Performance of Frequency
480 Ratio and Logistic Regression Model in Creating GIS Based Landslides Susceptibility Map at
481 Lompobattang Mountain, Indonesia.” Geoenvironmental Disasters 3(1): 19.

482 Reis, S. et al. 2012. “Remote Sensing and GIS-Based Landslide Susceptibility Mapping Using
483 Frequency Ratio and Analytical Hierarchy Methods in Rize Province (NE Turkey).”
484 Environmental Earth Sciences 66(7): 2063–73.

485 Rong, G et al. 2020. “Hazard Mapping of the Rainfall – Landslides Disaster Chain Based on
486 GeoDetector and Bayesian Network.” Water 12(9): 2573.

487 Ró&zgrave;ycka, Milena, Piotr Migoń, and Aleksandra Michniewicz. 2016. “Topographic Wetness
488 Index and Terrain Ruggedness Index in Geomorphic Characterisation of Landslide Terrains,
489 on Examples from the Sudetes, SW Poland.” Zeitschrift für Geomorphologie, Supplementary
490 Issues 61.

491 Saleem, Nayyer et al. 2019. “Parameters Derived from and/or Used with Digital Elevation Models
492 (DEMs) for Landslide Susceptibility Mapping and Landslide Risk Assessment: A Review.”
493 ISPRS International Journal of Geo-Information 8(12): 545.

494 Sarkar & Dorji, Kelzang. 2019. “Determination of the Probabilities of Landslide; A Case Study of
495 Bhutan.” Hydrology 6(52).

496 Sevgen, Eray, Sultan Kocaman, Hakan A. Nefeslioglu, and Candan Gokceoglu. 2019. “A Novel
497 Performance Assessment Approach Using Photogrammetric Techniques for Landslide
498 Susceptibility Mapping with Logistic Regression, ANN and Random Fore.” Sensors, MDPI 19:
499 3940.

500 Shano, Leulalem, Tarun Kumar Raghuvanshi, and Matebie Meten. 2021. “Landslide Susceptibility
501 Mapping Using Frequency Ratio Model: The Case of Gamo Highland, South Ethiopia.”
502 Arabian Journal of Geosciences 14(7): 623.
503 Singh, Gaurav, Raj Kumar, Dinesh Jinger, and Dinesh Dhakshanamoorthy. 2020. “Ecological
504 Engineering Measures for Ravine Slope Stabilization and Its Sustainable Productive
505 Utilization, Slope Engineering, Ali Ismet Kanlı.” IntechOpen.

506 Swingler, K. 1996. Applying Neural Networks: A Practical Guide. San Francisco: Morgan Kaufman
507 Publishers, Inc. www.mkp.com.

508 Tien Bui, Dieu et al. 2013. “Regional Prediction of Landslide Hazard Using Probability Analysis of
509 Intense Rainfall in the Hoa Binh Province, Vietnam.” Natural Hazards 66(2): 707–30.

510 UN. 2015. GLOBAL SUSTAINABLE DEVELOPMENT REPORT. United Nations. ADVANCE
511 UNEDITED VERSION.

512 Valencia Ortiz, Joaquın Andres, and Antonio Miguel Martınez-Grana. 2018. “A Neural Network Model
513 Applied to Landslide Susceptibility Analysis (Capitanejo, Colombia).” GEOMATICS,
514 NATURAL HAZARDS AND RISK, Taylor and Francis Group 9(1): 1106–28.

515 Walker, Lawrence R, and Aaron B Shiels. 2013. “Chapter 4 Biological Consequences for Landslide
516 Ecology.” Landslide Ecology: 46–82.

517 Wang, Wenping Li, Wei Chen, and Hanying Bai. 2015. “GIS-Based Assessment of Landslide
518 Susceptibility Using Certainty Factor and Index of Entropy Models for the Qianyang County
519 of Baoji City, China.” Journal of Earth System Science 124(7): 1399–1415.

520 Wang, Qian, Yi Wang, Ruiqing Niu, and Ling Peng. 2017. “Integration of Information Theory, K-
521 Means Cluster Analysis and the Logistic Regression Model for Landslide Susceptibility
522 Mapping in the Three Gorges Area, China.” Remote Sensing 9(9).

523 Winter, Mike G. et al. 2016. “The Economic Impact of Landslides and Floods on the Road Network.”
524 Procedia Engineering 143(Ictg): 1425–34.

525 Wu, Yanli et al. 2016. “Landslide Susceptibility Assessment Using Frequency Ratio, Statistical Index
526 and Certainty Factor Models for the Gangu County, China.” Arabian Journal of Geosciences
527 9(2): 1–16.

528 Yilmaz, Işık. 2009. “Landslide Susceptibility Mapping Using Frequency Ratio, Logistic Regression,
529 Artificial Neural Networks and Their Comparison: A Case Study from Kat Landslides (Tokat—
530 Turkey).” Computers & Geosciences 35(6): 1125–38.
531 Yu, Bin, Yunbo Zhu, and Yang Liu. 2017. “Topographical Factor-Based Shallow Landslide Hazard
532 Assessment: A Case of Dayi Area of Guizhou Province in China.” Geoenvironmental Disasters
533 4(1): 24.

534
Figures

Figure 1

See image above for gure legend

Figure 2

See image above for gure legend

Figure 3

See image above for gure legend

Figure 4

See image above for gure legend

Figure 5

See image above for gure legend

Figure 6

See image above for gure legend

Figure 7

See image above for gure legend

Figure 8

See image above for gure legend

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.
Listoftables.pdf