Development of Hybrid Computational Approaches for Landslide Susceptibility Mapping Using Remotely Sensed Data in East Sikkim, India

Indrajit Chowdhuri, Paramita Roy,

Rabin Chakrabortty, Subodh Chandra Pal,

Biswajit Das, and Sadhan Malik

Introduction

A landslide is a natural disaster, which causes a loss of social as well as economic property. In recent decades human civilization has been increasing, but not in a scientific way and also in the process destroying the ecology of mountainous regions and enhancing the likelihood of landslides. A landslide is one type of slope instability process and also a mass movement (such as a rock fall, earth flow and debris fall) (Borgatti and Soldati 2010). It is also the outcome of the relationship between shear stress and the materials’ shear strength on the slope (Yalcin 2007). In hilly and mountainous regions, landslide occurrences are very common and frequent in nature, but the time of their happening is not yet possible to determine (Glade and Crozier 2005). The presentation of landslide susceptibility mapping is an area where various rapid mass movement happens, and such an area is identified according to various geo- environmental factors (like slope, aspect, elevation, profile curvature, plan curvature, topographic wetness index (TWI), stream power index (SPI), geology, lithology, soil, distance to thrust, lineament density, land use and land cover (LULC), the normalized differences vegetation index (NDVI), road density, rainfall, and geomorphology) and present landslide area features. Fell et al. (2008) wrote that a susceptibility map shows the spatial distribution, thus the known and unknown landslide areas based on local terrain units. The preparation of a landslide susceptibility map of any region helps to forecast future landslide prone areas or vulnerable zones (Ahmed 2015; Pal et al. 2019). Traditionally, statistical approaches were used for large areas and recently trend machine learning models have also been widely used with their ensembles for occurrences and also for validation (Bui et al. 2016). The statistical methods include various approaches, such as logistic regression (LR), multivariate regression, a statistical index, an analytical hierarchy process (AHP). an evidential belief function (EBF), and certainty factors (Chakrabortty et al. 2018, 2020; Chowdhuri et al. 2020; Das and Pal 2019a, 2019b, 2020; Das et al. 2019a, 2019b; Pal et al. 2019). And therefore machine learning is an artificial intelligence branch which considers computer-based algorithms for making accurate predictions from the dominant data. Various previous literature reviews informed about used machine learning models for landslide prediction areas. These studies included an artificial neural network (ANN); a neural network model; decision trees (DTs); a support vector machine (SVM); boosted regression trees (BRTs); random forest (RF) biogeography-based optimization (BBO); eco-biogeography-based optimization (EBO); data mining; neuro-fuzzy and hybrid neuro-fuzzy, decision-making models; and big data analysis models (Das et al. 2015, 2018, 2019c, 2020a, 2020b, 2020c, 2020d; Dey et al. 2018, 2019; Rout et al. 2020).

Nowadays, hybrid machine-learning models are exploring the model in a scientific way. The hybrid models used in previous studies are ANN-fuzzy logic, stepwise weight assessment ratio analysis (SWARA), an adaptive neuro-fuzzy inference system (ANFIS), ANFIS with frequency ratio (FR), EBF-fuzzy logic, and our recent research w'ork considers hybrid biogeography-based optimization ensembles with a main operator, and which are migrations and mutations. From a previous literature review we were assured that every approach has its own perspective statistical methods and has great influence on factor classes, while machine learning showed correlation between a landslide area and conditioning factors. Hybrid machine learning has two such types of implementation. For this reason its capability for identifying landslide-prone areas increases day by day. This hybrid model has another important capability: to evaluate landslide-related dependent and independent variables. The accuracy level and prediction power for identifying susceptible zones is higher than for the above two models (statistical and machine learning).

In hilly regions every year many news reports are published about landslides. If we check the global scale, then we can see almost 66 million people live in highly landslide-prone areas, where 17% of them are affected by this natural disaster (Sassa and Canuti 2008). Sikkim is a small hilly state in India facing in its southerly direction Eastern Himalaya, where most of the topography has developed by erosional process through a large number of perennial and non-perennial streams and springs (Pal and Chowdhuri 2019; Pal et al. 2019; Rawat et al. 2016; Tambe et al. 2012). Out of four districts, the east district is more populated and civilized, where 79.55% of the total population lives in an urban area. The trend line of rapid urbanization is increasing day by day and is concentrated around the capital of Gangtok. The years of major landslides in Sikkim are 1968, 1997, 2007, 2011, and 2018 (Kaur et al. 2019). Choubey (1995) has reported that about 36,000 people were killed by the 1968 landslide alone. In Sikkim in 1997, June 7 was a black day. Due to heavy rainfall on that day 5000 houses were washed away, 50 people were injured, and National Highway 31A was damaged. In 2007 landslides happened in the area of Taktse. And in 2011, Sikkim faced landslides due to an earthquake (Kaur et al. 2019). In the current study area East Sikkim is more susceptible to landslides because of high population pressure, rising urbanization, and/or effective environment conditioning factors; and this steep hilly region is very vulnerable to landslides (Chakraborty et al. 2011). Some landslide susceptibility works have been done in East Sikkim as well as Sikkim but the many landslide causative factors and ensemble machine learning methods have not applied to landslide susceptibility mapping. So the objective of this present research is to prepare a landslide susceptible map of East Sikkim. Hybrid BBO and its ensembles were used to show more vulnerable zone of landslides from which the Sikkim Government can take administrative steps to reduce rapid unscientific civilizations and also can manage increasing urban development in a scientific way.

Study Materials and Methodology

Area of Research Study

East Sikkim is one of the four administrative districts in Sikkim India. It is located in the south-eastern portion of the state. And this district is divided into the three subdivisions of Gangtok. Pakyong, and Rongli. The researched area covered 94 sq. km following the longitudinal and latitudinal extensions 88°27' to 88°56' E and 27°9' to 27°25' N respectively (Figure 4.1). Accordingly to the Indian Meteorological Department (IMD) East Sikkim belongs to the subtropical humid and temperate climate zone. As in the district of Sikkim. Teesta is the major river, and other major

Location of the study area

FIGURE 4.1 Location of the study area.

drainage systems are the Rangpo, Chhu, and Dik Chhu. The average rainfall is 3894 mm which mainly occurs during the period of May to September. Temperature varies in summer from 15°C to 20°C and in winter from 4°C to 10°C. Forest is the main land cover in this district and about 72.66 sq. km is its total area. Four types of forest coverage zones are found: mixed coniferous broad leaved forest, alpine coniferous, shrubs, and alpine meadow. Geomorphology consists of hilly, slope, and valley formations. The major soils are mountain meadow, brown red, yellow, and laterite. Predominant geological formations are recent alluvium Reyang, Gorubathan in the Daling group, migmatitic gneisses, augean gneiss, amphibolites in Darjeeling, and the Kanchenjunga group. This hilly area is mainly formed by gneiss, green schist, and amphibolites; they are all metamorphosed. The lineament of geological formation trends toward five directions: N-S, E-W, NE-SW. ENE-WSW. and NW-SE. Phylite, amphibolites, and schist were formed in the Precambrian era and alluvium along the river valley is from the Quaternary era.

MULTI-COLINEARITY ASSESSMENT (MCT)

By testing the colinearity among various affective factors of landslides we have obtained the independent variables for their occurring. The colinearity test shows the relationship between causative variables and strong prominent factors for landslides, and these factors can be separated from each other by their character. So if we don't do the colinearity test then the linear relationship of the variables reduces the accuracy of models. This process is controlled by two types of factors: the variance inflation factor (VIF) and the tolerance (TOL) index. The result of a multi-colinearity test establishes the linearity between causative factors when the coefficient value of TOL is <1 and the VIF factor is >10. Here, 18 factors are independent and differ in their characteristics, which helps to increase the accuracy of the models to establish a landslide susceptible map of East Sikkim.

Affecting Factors

To prepare a susceptibility map of the landslide area of East Sikkim, we selected 18 types of various geo-environment parameters which were sourced from ALOS PALSER DEM with 12.5 m and LANDS AT 8 at 30 m resolutions, meteorological data, a topographical map at 1:50,000, a thematic map, and locally available data from various Sikkim institutions working on landslide activity.

All considered factors were put into five sub-categories (Figure 4.2). Seven topographical factors (slope, elevation, aspect, plan curvature, profile curvature, TWI, and geomorphology) were sourced from a topographical map at 1:50,000 scale and from DEM at 12.5 m.

Three hydrological factors (Figure 4.3) (drainage density, rainfall, and SPI) are related to properties of hydrology; the information was collected from IMD, a topographical map, and a digital elevation model.

Topographical factors

FIGURE 4.2 Topographical factors.

Hydrological factors

FIGURE 4.3 Hydrological factors.

Soil is another major element where its texture, compactation, and water holding capacity plays an important role in landslides. All soil related data are collected from the National Bureau of Soil Survey and Land Use Planning (NBSS&LUP) and the local survey office of Sikkim.

Environmental factors

FIGURE 4.4 Environmental factors.

Environmental factors incorporating land use and land cover. NDVI, and road density were created from a topographical map and LANDSAT 8 images (Malik et al. 2019; Pal et al. 2019) (Figure 4.4). These factors may also be called anthropogenic influences because environments are affected by both nature as well as human activity.

Geological factors

FIGURE 4.5 Geological factors.

Geology shows us the formation of the area and also the reason for the occurrence of slope-instability-related natural hazards. Distance to thrust, lineament density, and lithologies are all compactly developed main factors of geology (Figure 4.5). The Geological Survey of India (GSI) provided us with information about these factors to establish a geology map of East Sikkim.

Landslide Inventory Map (LIM)

Inventory maps as well as locational maps help to show the location of the researched study area. Here we show the landslide zone of East Sikkim by presenting an inventory map. Using predictive models and training data also shows us the spatial distribution of landslides and provides information for future landslide- vulnerable zones. We collected all data from the Sikkim Government and the GSI to prepare the landslide susceptible map and validated the results through a field survey and the Global Positioning System. Overall 163 landslide points are considered where 114 points (70%) are considered as the training dataset and 49 points (30%) are for validation.

Methodology

Hybrid Biogeography-Based Optimization

The hybrid В BO was evolved to form the fundamental structure of the В BO (Bhattacharya and Chattopadhyay 2010). The main way this was done was by devolving it by combining this algorithm with other ensembles. The BBO model is capable of estimating local exploration in a precise manner, but it is unable to estimate the level of exploration on a global scale (Simon 2008). So, the hybridization of BBO can be considered as a reliable prediction tool which is the best model that can balance global and local application (Savsani et al. 2014). For landslide susceptibility modeling, the landslide occurrence and non-landslide dataset were divided by the nominal data (such as 1 and 0). For Table 4.1, the frequency ratio (FR) value of each factor and sub-factor were needed for hybrid BBO modeling.

Hybridization with Differential Evolution

The diversified mutation scheme is very optimal, which makes differential evolution (DE) very realistic for determining large scale prediction when considering the maximum possible optimal solution. In a small region predictive capability is limited; so for this purpose mutation and migration can be used to complement each other. Beside this, there are several unique approaches that can be considered for the hybridization process.

The DE/BBO Algorithm

The integration of BBO with the DE algorithm (termed the DE/BBO algorithm) was initiated by Gong et al. (2010) (Equations 4.1-4.4). The basis of this optimization is the integration of hybrid BBO migration and DE mutation.

TABLE 4.1

Importance of Landslide Conditioning Factors

Count

Area

Class (%)

Landslide Occurrence (%)

FR

Distance to Structure (m)

0-4619.772

263,064

229.159

24.2125

25

1.03252

4619.772-9349.540

280,884

244.682

25.8527

32.8947

1.27239

9349.540-14079.307

271,031

236.099

24.9458

21.0526

0.84393

14079.307-19689.032

173,270

150.938

15.9479

13.1579

0.82506

19689.032-28048.621

98.298

85.6288

9.0474

7.89474

0.8726

Geology

Gorubathan Fm.

379.386

330.488

34.9211

37.7193

1.08013

Undiff. Darjeeling - Kanchenjunga gneiss

565.641

492.738

52.0652

46.4912

0.89294

Kanchenjunga gneiss

15.047

13.1077

1.38502

1.31579

0.95001

Lingste granite gneiss

126.473

110.172

11.6414

14.4737

1.2433

Geomorphology

Moderately dissected hills and valleys

544.920

474.687

50.1958

54.386

1.08348

Alluvial plain

2212

1.9269

0.20376

0

0

Waterbody-river/waterbodies-other

26.176

22.8023

2.41123

2.63158

1.09139

Highly dissected hills and valleys

503.254

438.392

46.3577

42.5439

0.91773

Snow cover

9438

8.22157

0.86939

0

0

Glacial Lake and Nunatak

547

0.4765

0.05039

0.4386

8.70449

Lithology

Banded migmatite. Garnet Bt gneiss, mica schist

535.088

466.123

49.2489

46.9298

0.95291

Quartzite

628

0.54706

0.0578

0.4386

7.58812

Tourmaline granite

72

0.06272

0.00663

0

0

Mylonitic granite gneiss

82.636

71.9854

7.60573

10.9649

1.44167

Basic intrusives

608

0.52964

0.05596

0

0

Chlorite sericite schist and quartzite

390.397

340.08

35.9317

36.4035

1.01313

Quartzite, mica schist, gneiss, calcgranulite

36.848

32.0988

3.39145

3.50877

1.03459

Calc silicate rock

5256

4.57858

0.48376

0

0

Unmapped

29.693

25.866

2.73291

0.4386

0.16049

Quartz arenite. black slate, cherty phyllite

569

0.49566

0.05237

0

0

Mela greywacke

1366

1.18994

0.12573

0.4386

3.48853

Dolimitic quartzite, chert, phvllite, slate

866

0.75438

0.07971

0

0

Biotite quartzite

2520

2.19521

0.23194

0.87719

3.78201

LULC

Agricultural land

182.066

158.6

16.5692

16.2281

0.97941

Barren land

125.131

109.003

11.3877

17.1053

1.50208

Built-up areas

20.594

17.9397

1.87419

2.63158

1.40412

Deciduous broadleaf forest

886

0.77181

0.08063

0

0

Evergreen broadleaf forest

673.446

586.648

61.288

55.7018

0.90885

Grassland

67.857

59.1112

6.17543

6.14035

0.99432

Mixed forest

7245

6.31122

0.65934

0.4386

0.6652

Water bodies

9322

8.12052

0.84836

1.75439

2.06797

Soil Texture

Coarse-loamy mollic udarents

28.344

24.6909

2.6088

2.19298

0.84061

Loamy-skeletal lithic udorthents

48.896

42.594

4.50041

5.26316

1.16948

Loamy-skeletal typic udorthents

13.932

12.1364

1.28231

0.4386

0.34204

Fine-loamy fluventic eutrudepts

32.876

28.6387

3.02592

3.94737

1.30452

Coarse-loamy humic pachic Dystrudepts

176.575

153.817

16.2521

17.9825

1.10647

Coarse-loamy humic dystrudepts

528.617

460.486

48.6542

43.8596

0.90146

Fine-loamy typic paleudolls

20.264

17.6523

1.86511

1.31579

0.70548

Fine-loamy typic argiudolls

102.103

88.9434

9.39761

11.8421

1.26012

Fine-skeletal cumulic hapludolls

63.835

55.6076

5.87541

5.26316

0.89579

Fine-skeletal entic hapludolls

28.607

24.92

2.633

2.19298

0.83288

Coarse-loamy typic hapludolls

42.498

37.0206

3.91154

5.70175

1.45768

Aspect

-1

20

0.01742

0.00184

0

0

0-22.5

48.897

42.5949

4.50022

4.38596

0.97461

22.5-67.5

76,454

66.6001

7.03642

6.57895

0.93499

67.5-12.5

97.792

85.188

9.00025

5.70175

0.63351

112.5-157.5

135,726

118.233

12.4915

14.4737

1.15868

157.5-202.5

151.908

132.329

13.9808

13.5965

0.97251

202.5-247.5

157.063

136.82

14.4552

19.2982

1.33503

247.5-292.5

168.057

146.397

15.4671

11.8421

0.76563

292.5-337.5

183.439

159.796

16.8827

17.5439

1.03916

337.5-360

67.191

58.531

6.1839

6.57895

1.06388

TABLE 4.1 (Continued)

Importance of Landslide Conditioning Factors

Count

Area

Class (%)

Landslide Occurrence (%)

FR

Distance to Road (m)

0-629.755

512.298

446.27

47.1522

64.9123

1.37666

629.755-1529.405

272.128

237.055

25.0468

17.1053

0.68293

1529.405-2668.962

188.440

164.153

17.3441

12.2807

0.70806

2668.962^1438.274

91.607

79.8001

8.43156

5.26316

0.62422

4438.274-7647.027

22.074

19.229

2.0317

0.4386

0.21588

Drainage Density (sq. km)

0-0.313

325.531

283.575

29.962

26.3158

0.8783

0.313-0.611

271.389

236.411

24.9788

24.5614

0.98329

0.611-0.958

241.052

209.984

22.1866

21.9298

0.98843

0.958-1.371

142.226

123.895

13.0906

11.4035

0.87112

1.371-2.106

106.349

92.6421

9.78842

15.7895

1.61308

Elevation (m)

272-1165

211,720

184.432

19.4856

21.4912

1.10293

1165-1801

293.733

255.875

27.0336

29.386

1.08702

1801-2561

223.326

194.542

20.5537

16.6667

0.81088

2561-3453

136.828

119.193

12.5929

14.4737

1.14935

3453-4669

220.940

192.464

20.3341

17.9825

0.88435

Lineament Density (sq. km)

0-0.0170

573,776

499.824

52.8106

51.7544

0.98

0.0170-0.049

173.369

151.024

15.957

19.7368

1.23688

0.049-0.086

137.984

120.2

12.7001

11.4035

0.89791

0.086-0.122

132,661

115.563

12.2102

11.4035

0.93393

0.122-0.173

68,757

59.8952

6.32843

5.70175

0.90097

NDVI

-0.357 to -0.071

35,069

3.5069

0.37053

0.4386

1.18369

-0.071 to 0.143

385.102

38.5102

4.06894

8.33333

2.04804

0.143-0.400

2.021.402

202.14

21.3579

27.6316

1.29374

0.400-0.605

2,668.625

266.863

28.1963

29.8246

1.05775

0.605-0.857

4.354.238

435.424

46.0063

33.7719

0.73407

Plan Curvature

-16.103 to-1.645

56.047

48.8233

5.15827

5.70175

1.10536

-1.645 to-0.488

243.982

212.536

22.4548

16.6667

0.74223

-0.488 to 0.436

457275

398.339

42.0852

44.2982

1.05259

0.436-1.593

263.892

229.88

24.2872

27.193

1.11964

1.593-13.390

65.351

56.9282

6.01456

6.14035

1.02091

Profile Curvature

-16.123 to -2.106

54.621

47.5811

5.02703

7.01754

1.39596

-2.106 to -0.624

231.429

201.601

21.2995

18.4211

0.86486

-0.624 to 0.453

443.031

385.93

40.7742

41.2281

1.01113

0.453-1.936

291.228

253.693

26.8031

27.193

1.01455

1.936-18.244

66.238

57.7008

6.09619

6.14035

1.00724

Rainfall (mm)

2208.502-2312.940

77.742

67.7221

7.15541

7.01754

0.98073

2312.940-2384.397

216.875

188.923

19.9613

24.1228

1.20848

2384.397-2439.365

414.894

361.42

38.1871

33.7719

0.88438

2439.365-2501.661

238.173

207.476

21.9216

25.8772

1.18044

2501.661-2675.724

138.863

120.965

12.781

9.21053

0.72064

Slope (in Degrees)

0-14.169

150.004

130.671

13.8056

10.9649

0.79424

14.169-23.616

288.534

251.346

26.5551

19.2982

0.72672

23.616-32.176

302.336

263.369

27.8254

27.193

0.97727

32.176-41.623

235.651

205.279

21.6881

29.386

1.35494

41.623-75.276

110.022

95.8417

10.1258

13.1579

1.29944

SPI

5.124-10.372

5978

337.992

35.9969

35.5263

0.98693

10.372-11.603

5878

332.338

35.3947

33.7719

0.95415

11.603-13.158

3319

187.654

19.9855

17.5439

0.87783

13.158-15.878

1213

68.5821

7.30415

8.33333

1.1409

15.878-21.644

354

20.0149

2.13163

4.82456

2.26332

TWI

10.813-12.483

7582

428.68

45.6554

41.2281

0.90303

12.483-13.960

5981

338.161

36.0149

34.6491

0.96208

13.960-16.336

2117

119.694

12.7476

12.2807

0.96337

16.336-19.933

816

46.136

4.91359

6.14035

1.24967

19.933-27.192

246

13.9087

1.4813

5.70175

3.84915

where N is the population, D is the problem dimension, c is the crossover likelihood, and F is the scaling factor of mutation.

DE/BBO can replace the original mutation by considering the hybrid migration. For the purpose of calculating the migrant rate operator, the DE/BBO differs from the original BBO. The DE/BBO can be very optimistic in comparison to DE and BBO.

Local-DE/BBO

The improvement of BBO can be made with the help of local topologies. The DE/ BBO with local topologies can be considered as a reliable predictor of hybrid BBO, which creates three categories of DE/BBO: RingDB, SquareDB, and RandDB. RingDB encompasses ring topology, whereas SquareDB and RandDB are related to square topology and random topology (Zheng et al„ 2019). This triple local algorithm is more optimistic and is an improvement on DE/BBO. It is much more realistic than DE on the dimensional function:

Self-Adaptive DE/BBO

Self-adaptive DE/BBO is an upgraded form of DE, and the main function of this model is in dealing with multiple mutation and choosing the most suitable one. The suitable mutation is very optimistic in functional analysis. The appropriate mutation combinedly works with the relevant parameters. A general approach for obtaining the information from initial values is:

where t is the current, max is the maximum, and min is the minimum generation number of the algorithm.

The obtained probability of the mutation and migration are estimated by:

The general function was replaced by the immigration rate and their associated crosscover operations are:

For increasing the consistency of the parameters, the Gaussian distribution is used: In HSDB, we set Fp to 0.5 and F„ to 0.3, and limit the value F in the range [0.05,0.95]:

Validation of Models

For the validation of the training dataset the AUC values with the receiver operating characteristic (ROC) curve was considered to obtain an authentic result. The AUC values state the model which is perfect for this area. Here, the ROC curve is a technique whose value is able to predict the best model. The scientific calculation is:

where SAUC is considered an ROC, XK is 1 specificity, and SK is the sensitivity of ROC, which considers primary information about recent landslides. The performances of the classifiers identified the optimal model which is generated by different cutting threshed values.

Shortly Structured Methodology

This study consists of four main stages which are described and summarized here: data selection, data preparation, geo-conditioning factors, and a colinearity test using VIF and TOL for independent parameters selection. A landslide susceptibility map uses a hybrid BBO and its ensembles. The result is validated through AUC values of ROC. The prescribed methodology framework is shown in Figure 4.6.

Methodology flow chart

FIGURE 4.6 Methodology flow chart.

 
Source
< Prev   CONTENTS   Source   Next >