Landslide susceptibility mapping based on self-organizing-map network and extreme learning machine

Accepted Manuscript Landslide susceptibility mapping based on self-organizing-map network and extreme learning machine

Faming Huang, Kunlong Yin, Jinsong Huang, Lei Gui, Peng Wang PII: DOI: Reference:

S0013-7952(16)30314-3 doi: 10.1016/j.enggeo.2017.04.013 ENGEO 4551

To appear in:

Engineering Geology

Received date: Revised date: Accepted date:

13 September 2016 11 March 2017 16 April 2017

Please cite this article as: Faming Huang, Kunlong Yin, Jinsong Huang, Lei Gui, Peng Wang , Landslide susceptibility mapping based on self-organizing-map network and extreme learning machine. The address for the corresponding author was captured as affiliation for all authors. Please check if appropriate. Engeo(2017), doi: 10.1016/ j.enggeo.2017.04.013

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT Landslide susceptibility mapping based on self-organizing-map network and extreme learning machine Faming Huang1 , Kunlong Yin1 , Jinsong Huang2 , Lei Gui1 , Peng Wang1

2

China University of Geosciences, Wuhan 430074, China ARC Centre of Excellence for Geotechnical Science and Engineering, University of Newcastle, NSW,

PT

1

AC C

EP T

ED

MA

NU

SC

RI

Australia

1

ACCEPTED MANUSCRIPT Abstract Among the machine learning models used for landslide susceptibility indexes calculation, the support vector machine (SVM) is commonly used; however, SVM is time-consuming. In addition, the non- landslide grid cells are selected randomly and/or subjectively, which may result in

PT

unreasonable training and validating data for the machine learning models. This study proposes the self-organizing- map (SOM) network-based extreme learning machine (ELM) model to calculate the

RI

landslide susceptibility indexes. Wanzhou district in Three Gorges Reservoir Area is selected as the

SC

study area. Nine environmental factors are chosen as input variables and 639 investigated landslides

NU

are used as recorded landslides. First, an initial landslide susceptibility map is produced using the SOM network, and the reasonable non- landslide grid cells are subsequently selected from the very

MA

low susceptible area. Next, the final landslide susceptibility map is produced using the ELM model based on the recorded landslides and reasonable non- landslide grid cells. The single ELM model

ED

which selects the non- landslide grid cells randomly, and the SOM network-based SVM model are

EP T

used for comparisons. It is concluded that the SOM-ELM model possesses higher success and prediction rates than the single ELM and SOM-SVM models, and the ELM has a considerably higher

AC C

prediction efficiency than the SVM.

Keywords: landslide susceptibility map; self-organizing- map network; extreme learning machine; support vector machine; Three-Gorges Reservoir.

2

ACCEPTED MANUSCRIPT

1. Introduction Landslides are one of the major geological hazards around the world. Thus, it is important to predict the locations of possible landslides. The landslide susceptibility map provides valuable

PT

information for local government to perform master plans (Fell et al. 2008). Recently, many models

RI

have been developed for landslide susceptibility mapping based on Geographical Information

SC

System (GIS) (Erener et al. 2016). These models can be generally divided into five categories: landslide inventory models, heuristic models, deterministic models, statistical models and machine

NU

learning models (Marjanović et al. 2011). Among these models, the landslide inventory models and

MA

heuristic models depend on the knowledge of the researchers and generally have low accuracy (Ruff and Czurda 2008; Van Westen 2000). The deterministic models estimate the safety factors in a

ED

defined area (Godt et al. 2008; Park et al. 2013). However, the deterministic models are mainly

EP T

regarding the rainfall- induced shallow landslides based on the infinite slope model, coupled with stationary hydrological model.

The statistical models are used more widely than the heuristic and deterministic models. In

AC C

statistical models, the correlations between the environmental factors and the landslides are used to calculate landslide susceptibility indexes in a large area (Erener et al. 2016; Kavzoglu et al. 2015; Lee et al. 2008; Martinović et al. 2016). However, statistical models require the normal distribution of environmental factors, which is not always satisfied, and they are inherently linear (He et al. 2012). To overcome these drawbacks, the machine learning models have been proposed for landslide susceptibility mapping, e.g., artificial neural network (ANN) (Choi et al. 2012), support vector machine (SVM) (Marjanović et al. 2011), and decision tree (Park and Lee 2014). The machine

3

ACCEPTED MANUSCRIPT learning models can use different types of input variables without considering the specific statistical regularity (Lee et al. 2003). Moreover, machine learning models are inherently nonlinear. It is important to select an appropriate machine learning model for landslide susceptibility mapping. The traditional ANN models have limitations of the local optimum and over- fitting (Gong

PT

et al. 2016). In addition, the SVM (Cortes and Vapnik 1995) has drawbacks regarding the low

RI

training speed. Recently, a new ANN called extreme learning machine (ELM) was proposed by

SC

Huang et al. (2006). The ELM model exhibits excellent performance with high training speed and excellent generalization ability. The ELM has been successfully used in many fields, such as

NU

landslide displacement prediction (Huang et al. 2016) and pattern recognition (Shrivastava et al.

MA

2016). However, the ELM has received little attention in landslide susceptibility mapping. For the training and validating of the ELM model, it is necessary to obtain reasonable

ED

non- landslide grid cells. A literature review shows that there are mainly three methods for addressing

EP T

this issue: 1) the seed cell procedure (Nefeslioglu et al. 2008), 2) randomly selecting non- landslide grid cells from the landslide free areas (Felicísimo et al. 2013) and 3) non- landslide grid cells selected based on the argument that landslides are free on the terrains with slope lower than 2°

AC C

(Kavzoglu et al. 2014). However, it is difficult to determine whether the randomly selected non- landslide grid cells really have very low susceptibility. To overcome this drawback, the self-organizing- map (SOM) network (Ritter and Kohonen 1989) is used. SOM network can produce the initial landslide susceptibility map automatically and doesn’t need to select no n- landslide grid cells randomly (Lin 2008). However, the SOM network cannot calculate the landslide susceptibility indexes. Although some other machine learning models such as ELM can also be adopted to select reasonable non- landslide grid cells. But they need to select non- landslide grid cells randomly. The

4

ACCEPTED MANUSCRIPT optimal number of hidden neurons is also required. Therefore, the SOM network is proposed to automatically classify the landslide susceptibility of all the grid cells into five classes: very low, low, moderate, high and very high susceptibility. The reasonable non-landslide grid cells are selected from the very low susceptible area. Then the ELM is used to calculate the landslide susceptibility indexes.

PT

The SOM-ELM model is used to estimate landslide susceptibility indexes in Wanzhou district in

RI

the Three Gorges Reservoir Area (TGRA). The TGRA has been seriously influenced by landslides.

SC

Zhu et al. (2014) apply the heuristic models to assess the landslide susceptibility in TGRA. Bai et al. (2010) use the statistical models to calculate the landslide susceptibility indexes in the Zhongxian

NU

segment of Yangtze river. Machine learning models are commonly used in TGRA (He et al. 2012;

MA

Wu et al. 2014a). The produced landslide susceptibility maps in TGRA show that the Yangtze river and its branches are one of the crucial environmental factors for landslide occurrence. However,

ED

there is no study mapping the landslide susceptibility in Wanzhou district, and almost all studies

EP T

select the non- landslide grid cells randomly. There are many landslides in Wanzhou district because of the complex geomorphic and geological conditions, the urbanization development and other environmental factors. The local human’s life and property is threatened seriously. Hence, it is

AC C

significant to use the SOM-ELM model to map the landslide susceptibility in Wanzhou district.

2. Methodologies The SOM-ELM model includes five procedures. First, landslide environmental factors selected by correlation analysis are used as input variables. Second, the SOM network is used to classify the landslide susceptibility of Wanzhou district into five classes. Third, the reasonable non- landslide grid

5

ACCEPTED MANUSCRIPT cells are selected from the very low susceptible area. Fourth, the final landslide susceptibility map is obtained by training and validating the ELM. Finally, the prediction performance of the SOM-ELM is assessed. In addition, a single ELM model that selects the non- landslide grid cells randomly, and the SOM network based SVM model (SOM-SVM) are also used as comparisons. The processes of

PT

the SOM-SVM is the same as the SOM-ELM model. Meanwhile, the same input and output

SC

RI

variables that are used in the SOM-ELM are used again in the SOM-SVM.

2.1. Self-organizing-map network

NU

The SOM network proposed by Ritter and Kohonen (1989) is an important ANN classifier. The

MA

SOM network is designed as two-dimensional arrangement of neurons that maps input variables to two dimensional with two main stages: training stage and classification stage. In this study, the

Suppose that X   x1 , x2 ,

ED

built-in SOM network in MATLAB R2015b is used.

, x p  is the input variables and that ul   ul1 , ul 2 ,

, ulp  is

EP T

the weight vector associated with the node l , p is the number of input variables and ulj is the weight assigned to input variable x j of the node l , with each object of the training data placed into

AC C

the SOM network randomly. The learning law of SOM network is to find the node closest to each training case and to move the “winning” node closer to the training case. The node is moved some proportion of the distance between it and the training case at a special learning rate. The distance d i between the weight vector and the input variables is computed for each object i in the training case. Next, the node with the smallest d i is considered as the winner, and then the weights of the winner node are updated by a learning rule, while the weights of the non-winner nodes are not changed. The Euclidean distance is generally used to calculate the d i .

6

ACCEPTED MANUSCRIPT Suppose that uls is the weight vector for the l th node on the sth step of the SOM network, X i is the input vector for the ith training case, and  s is the learning rate for the sth step. The X i is selected, and then the index q of the winning node is determined on each step by q  arg min uls  X i l

(1)

PT

The update rule of the SOM network for the winner node is given as

RI

uqs 1  uqs 1   s   X i s  uqs   s  X i  uqs  (2)

SC

The uls 1 is set to be uls for all non-winning nodes.

NU

2.2. Extreme learning machine

MA

The ELM (Huang et al. 2006) is a single- hidden layer feed- forward neural network (SLFN) with randomly generated hidden nodes that are independent of training data. Input weights and biases can

(MP) generalized inverse. For

N distinct samples

 X i , Ti  ,

X i   xi1 , xi 2 ,

xin   R n and T

, tim   R m , standard SLFNs with N hidden neurons and activation function g  x  T

EP T

Ti  ti1 , ti 2 ,

ED

be randomly chosen, and output weights can be analytically determined using the Moore-Penrose

AC C

are mathematically modeled as

where wi   wi1 , wi 2 ,

  g w X N

i 1

, win 

T

i

i

j

 bi   o j , j  1, 2,

, N (3)

is the weight vector of the connections from the input neurons to the

i th hidden neuron, i   i1 , i 2 ,

and the output neurons, o j  o j1 , o j 2 ,

, im  is the weight vector connecting the i th hidden neuron T

T

, o jm  is the j th output vector of the SLFNs, and bi is

the threshold of the i th hidden neuron. wi  x j denotes the inner product of wi and x j . The above N equations can be written compactly as: H   O . The H ,  and O are defined as

7

ACCEPTED MANUSCRIPT  g  w1 X 1  b1   H  g w X  b  1 N 1 

g  wN X 1  bN    1T  O1T       ,   ,O    T  OT  g  wN X N  bN    N  N m  N  N m  N N

(4)

where H is called the hidden layer output matrix. The i th column of H is the i th hidde n neuron’s output vector with respect to inputs. To minimize the cost function O  T , the output

PT

weights are based on finding the least-square (LS) solution to the linear system H   T :

RI

ˆ  H †T (5)

SC

where H † is the MP-generalized inverse of matrix H . Determining the optimal number of hidden nodes is one of the major challenges associated with developing appropriate ELM model. The

NU

trial-and-error method (Toth et al. 2000) is proposed to overcome this problem. The optimal number

MA

of hidden neurons is chosen based on the one with the lowest root mean square error of the ELM.

ED

2.3. Support vector machine

SVM (Cortes and Vapnik 1995) was developed based on statistical learning theory and structured

EP T

risk minimization principle. The SVM maps the input variables into a higher dimensional feature space via a nonlinear mapping and then solves a linear regression problem in the higher dimensional

AC C

feature space. The radial basis function is used as the kernel function of the SVM. The SVM has three parameters that should be determined appropriately. The parameter C0 denotes the degree of the penalty, the parameter  denotes a non-sensitive loss function, and the parameter  denotes the parameter of radial basis function. The parameter combinations of C0 ,  and  determined by the cross-validation stage (Kavzoglu et al. 2014).

3. Study Area and Data 3.1. Study area 8

are

ACCEPTED MANUSCRIPT Wanzhou district is located on the upper reaches of TGRA as shown in Figure 1. It lies between longitudes 107°55′22′′ E and 108°53′25′′ E, and latitudes 30°24′00′′ N and 31°14′58′′ N, with an area of approximately 3457 km2 . The elevation of Wanzhou district ranges from 130 m to 1640 m, excluding the Yangtze river. The study area is a hilly area, with the topographic inclination primarily

PT

following an NE-WS direction. The low mountains, hills, mid- low mountains and flat land account

RI

for 1/4 of the whole area.

SC

There are two major types of lithology in the Wanzhou district: the Jurassic system and the Triassic system . The Jurassic system (including J1 z, J2 z, J2 x, J2 s,J2 xs, J3 xj, J3 p, and J3 s) mainly

NU

consists of mudstone, sandstone, siltstone and shale. The Triassic system (including T2 b, T1 j, T3 xj,

MA

and T1 d) is primarily composed of limestone, clay stone, sandstone, siltstone, and coal. The study area is located within the subtropical climate belt, which receives frequent heavy rainfall events

ED

during the summer season. The annual average precipitation is 1181.2 mm between 1960 and 2015.

EP T

Many rivers are spread all over the study area in a tree- like form, most of which belong to the Yangtze river system. The landslide inventory map of Wanzhou district displays the locations, morphological characteristics, and other detailed information regarding the landslide s occurrence. To

AC C

validate the accuracy of landslide inventory, fieldwork was performed at randomly selected landslide sites. A total of 639 distinct landslides that occurred from 1970 to 2013 have been recorded in the landslide inventory map . Figure 1 shows that the landslides are mainly distributed in the middle area of Wanzhou district, especially along the Yangtze river and its tributaries. The 639 landslides in Wanzhou district can be divided into rock landslides and soil landslides. Nearly half of the 639 landslides occurred in the last two decades and the rest landslides were the reactivation of old landslides. The rock landslides are mainly controlled by sedimentary bed due to

9

ACCEPTED MANUSCRIPT weak interlayers. Soil landslides are generally triggered by heavy rainfall or the fluctuation of reservoir water level. Human engineering activities such as cutting slope at toe, are becoming more and more important triggering factor. The area of the smallest landslide is 2428 m2 , the area of the largest one is 969620 m2 and the average area is about 44000 m2 . The average depth of the sliding

RI

PT

mass is approximately 18 m. Hence, the mean total volume of the landslides is about 8×105 m3 .

Data

NU

3.2.

SC

Figure 1.Geographical location of the Wanzhou district

MA

It is important to correlate the environmental factors with the registered landslides in the inventory to analysis the landslide susceptibility. In this study, a landslide inventory map is obtained

ED

based on the field investigation and government reports as shown in Figure 1. The Digital Elevation

EP T

Model (DEM), geologic maps, remote sensing images and field survey excluding the Yangtze river are used as the data sources of the environmental factors. Ten environmental factors are extracted from the data sources (as shown in Figure 2 and Table 1): elevation, slope, aspect, profile curvature,

AC C

plan curvature, relief amplitude, lithology, geological structure, Normalized Difference Vegetation Index (NDVI) and distance to river. These environmental factors are all converted to the raster format with a grid cell resolution of 30 m × 30 m. This grid cell resolution is small enough to capture the spatial characteristics of landslide susceptibility and large enough to reduce computing complexity (He et al. 2012). As a result, the size of the image corresponding to each environmental factor is 3317×2272. It is important to assume that future landslides are more likely to occur under the same

10

ACCEPTED MANUSCRIPT conditions as present landslides . The original continuous input variables (environmental factors) cannot be used directly in the ELM or SVM model. It is necessary to divide each continuous input variable into several subclasses to get a general knowledge about the effects of continuous input variable on landslide occurrence. The frequency ratios of the subclasses of the continuous input

PT

variable are commonly used to reflect the effects of input variables on landslide occurrence as shown

RI

in Table 1 (He et al. 2012; Romer and Ferentinou 2016; Wu et al. 2014a) . It can be seen from Table

SC

1 that the aspect has little effect on the landslide occurrence in Wanzhou district, while the other nine environmental factors are strongly correlated to the landslide occurrence.

NU

Based on the influencing factors analysis of the landslide occurrence, the correlation coefficients

MA

between the nine environmental factors are calculated using the SPSS 21 statistical program as shown in Table 2. The results show that there are weak linear correlations between the environmental

ED

factors because all of the correlation coefficients are less than 0.428. Hence, the nine environmental

EP T

factors are selected as independent input variables to predict the landslide susceptibility indexes.

Figure 2. Environmental factors, (a):DEM, (b):Slope, (c):Aspect, (d):Profile curvature maps, (e):Plan curvature, (f):Relief amplitude, (g):Lithology, (h):Geological structure, (i):NDVI and (j):Distance to water. The geological

AC C

zones include a:jiajiaoshan anticline, b:qumahe syncline, c:tiefengshan anticline, d:wanxian syncline, e:huangbaixi syncline, f:dachiqianjin anticline, g:fengdouzhongxian anticline, h:fangdoushan anticline, i:ganchang syncline, j:matouchang syncline.

Table 1.Description of environmental factors and Frequency ratios of all environmental factors Table 2. Correlation coefficients between the eleven environmental factors

4. Landslide Susceptibility Mapping 4.1.

Reasonable non-landslide grid cells selection using SOM network

11

ACCEPTED MANUSCRIPT The SOM network is used to identify reasonable non- landslide grid cells. It is necessary to perform the standardization process for the nine environmental factors, and then the standardized environmental factors are used as the input variables of the SOM network. The output variables of the SOM network are five different susceptibility classes. Hence, a SOM network consists of an

PT

input layer with nine neurons representing nine standardized environmental factors, and a mapping

RI

layer with five neurons representing five different susceptibility classes. For the training of SOM

SC

network, the incremental training is used. The learning rate is initialized as 0.5 and is linearly reduced to 0.01 during the training process. The maximum number of iterations is set as 300, and the

NU

entire data set is used in each iteration. The convergence criterion is set as 0.001. The training

MA

process stops when maximum number of iterations or convergence criterion is satisfied. The landslide susceptibility map produced using the SOM network are automatically classified

ED

into five classes, as shown in Figure 3: very low (18.0%), low (9.7%), moderate (30.1%), high (13%)

EP T

and very high (29.2%). The frequency ratios are shown in Table 3. Table 3 reveals that the frequency ratios of the very high, high and very low susceptibility classes are 3.500, 1.040 and 0.300, respectively. It is reliable that the 101074 grid cells manually selected from the very low susceptible

AC C

area are the reasonable non-landslide grid cells.

Figure 3.Classification map of landslide susceptibility using SOM network Table 3. Frequency ratios of the five susceptibility classes of SOM network

4.2. Landslide susceptibility mapping using the SOM-ELM and SOM-SVM models To map the landslide susceptibility using the SOM-ELM and SOM-SVM models, the correlated

12

ACCEPTED MANUSCRIPT environmental factors are normalized into [0, 1] to prevent large values from overriding small values. The recorded landslide grid cells are assigned the value of 1, and the same number of reasonable non- landslide grid cells selected using the SOM network are assigned a value of 0 . The 101074 recorded landslide grid cells and 101074 reasonable non-landslide grid cells are randomly divided

PT

into two data sets: 80% of these grid cells are used as training data sets, and the remaining 20% of

SC

RI

these grid cells are used as validating data sets.

4.2.1.Landslide susceptibility mapping using SOM-ELM model

NU

The SOM-ELM model is used to calculate the landslide susceptibility indexes. The normalized

MA

environmental factors in the landslide and reasonable non-landslide grid cells are used to train and validate the ELM model. The trial-and-error method is used to determine the optimal number of

ED

hidden nodes of the ELM by varying the number of hidden neurons from 10 to 50. The optimal

EP T

number of hidden neurons for the ELM model is determined to be 45. Hence, a three- layer network that consists of one input layer with nine neurons, one hidden layer with 45 neurons and one output layer with one neuron is used as the ELM structure. The nine input neurons represent nine

AC C

normalized environmental factors and the output layer represents the landslide susceptibility index. The calculated landslide susceptibility indexes are classified into five landslide susceptibility classes. There are many classification methods, such as natural breaks, standard deviations and equal intervals (Ayalew and Yamagishi 2005). In this study, to compare the prediction accuracy of different models, the landslide susceptibility indexes are classified into five classes of very high (10%), high (20%), moderate (40%), low (20%), and very low (10%), based on the natural breaks method and the histogram of landslide susceptibility indexes (Pradhan 2013). Similar category ratios

13

ACCEPTED MANUSCRIPT are also used for the classification of landslide susceptibility maps produced by the SOM-SVM and single ELM models described in the following sections. The landslide susceptibility map produced using the SOM-ELM model are shown in Figure 4 and the frequency ratios are shown in Table 4. It can be seen from Table 4 that the frequency ratios of the very high, high and very low susceptibility

PT

indexes are 4.110, 1.460 and 0.220, respectively.

RI

Figure 4.Classification map of landslide susceptibility using SOM-ELM model

SC

Table 4. Frequency ratios of the five susceptibility classes of SOM-ELM model

NU

The distribution characteristics of the landslide susceptibility indexes in Wanzhou district are

MA

explored from Figure 4. Figure 4 shows that the very high and high susceptibility indexes are mainly distributed in the regions with low elevation. One reason for this observation is that the rivers are

ED

primarily distributed in the regions with low elevation, the closer to the rivers, the higher the

EP T

landslide susceptibility indexes. Another reason is that, there is a relatively low vegetation coverage in these low elevation regions as shown in the NDVI map, suggesting that the unreasonable human engineering activities have changed the geomorphologic features and decreased the stability of slope,

AC C

it is proven by field investigation. Furthermore, the lithological units in these areas mainly include J1 , J2 and J3 , which have low shear strength and reduce the stability of slope. In addition, the very low and low susceptibility indexes are found to be mainly distributed in the regions with high elevations and the regions with lithological units of T1 , T2 and T3 . One reason for this observation is that the high elevation regions are far from the rivers and have high forest cover. Another reason is that the lithological units of T1 , T2 and T3 have a relatively higher shear strength than the lithological units o f J1 , J2 and J3 .

14

ACCEPTED MANUSCRIPT

4.2.2.Landslide susceptibility mapping using the SOM-SVM model The SOM network based SVM model is also used to calculate the susceptibility indexes. The optimum values of C0 ,  and  are determined as 5, 0.1 and 0.18, respectively. The fully trained

PT

SVM model is used to calculate the landslide susceptibility indexes as shown in Figure 5, and the

RI

frequency ratio of each susceptibility class of SOM-SVM model is shown in Table 5. It can be seen

SC

from Table 5 that the frequency ratios of the very high, high and very low susceptibility classes are

NU

3.710, 1.635 and 0.264, respectively.

Figure 5. Classification map of landslide susceptibility using SOM-SVM model

MA

Table 5. Frequency ratios of the five susceptibility classes of SOM-SVM model

ED

4.3. Landslide susceptibility mapping using the single ELM model The single ELM is also used to map landslide susceptibility. The recorded landslide grid cells

EP T

and randomly selected non- landslide grid cells and are used to train and validate the ELM model. The optimal number of hidden neurons for the ELM model is found to be 35 by the trial-and-error

AC C

method. The landslide susceptibility indexes are also classified into five classes as shown in Figure 6 and Table 6. It can be seen from Table 6 that the frequency ratio increases gradually from the very low to the very high susceptibility class.

Figure 6. Classification map of landslide susceptibility using the single ELM model Table 6. Frequency ratios of the five susceptibility classes of single ELM model

4.4. Prediction accuracy and efficiency analysis 4.4.1.Success rate curve analysis 15

ACCEPTED MANUSCRIPT The success rate curve is used to evaluate how the landslide susceptibility calculation results fit the training datasets (Chung and Fabbri 2008). The landslide susceptibility indexes of all the grid cells are sorted in descending order because the recorded landslides are more likely to occur in the grid cells with high susceptibility indexes comparing to the grid cells with low susceptibility indexes.

PT

Then all the landslide susceptibility indexes are divided into 20 equally sized intervals with 5%

RI

cumulative intervals in ArcGIS 10.1. Based on the obtained 20 equally sized intervals, the

SC

percentage of recorded landslide grid cells of the training dataset in each equally sized interval is calculated to evaluate the success rates of the three models. The success rate curves of the three

NU

models are shown in Figure 7.

MA

It can be seen from Figure 7, the former 2 equally sized intervals of the 20 equally sized intervals reflect that 10% of the study area with the highest susceptibility indexes can account for

ED

41.82% of the landslide grid cells for the SOM-ELM, 37.67% for the SOM-SVM and 35.71% for the

EP T

single ELM model. In addition, the former 12 equally sized intervals of the 20 equally sized intervals reflect that 60% of the study area can account for 89.25% of the landslide grid cells for the SOM-ELM, 87.37% for the SOM-SVM, and 84.25% for the single ELM model. Hence, it is

AC C

concluded that the SOM-ELM model has the highest success rate, whereas the single ELM model has the lowest success rate.

Figure 7. Success rate curves of landslide susceptibility indexes calculated using the SOM-ELM, SOM-SVM and single ELM models

4.4.2.Prediction rate curve analysis

16

ACCEPTED MANUSCRIPT The calculation process of the prediction rate is the same as the success rate. The prediction rate curves are obtained through comparing the recorded landslide grid cells in the validating data set with the landslide susceptibility maps produced by the three models as shown in Figure 8 (Pradhan et al. 2010). Figure 8 shows that 10% of the study area with the highest susceptibility indexes can

PT

account for 38.17% of the recorded landslide grid cells for SOM-ELM, 33.21% for SOM-SVM and

RI

31.40 % for single ELM model. And 60% of the study area can account for 86.73% of the recorded

SC

landslide grid cells for SOM-ELM, 84.42% for SOM-SVM, and 81.22% for single ELM model. The

MA

single ELM model has the lowest prediction rate.

NU

comparison results show that the SOM-ELM model has the highest prediction rate, whereas the

Figure 8. Prediction rate curves of landslide susceptibility indexes calculated using the SOM-ELM, SOM-SVM and

ED

single ELM models

EP T

4.4.3.False Positive and False Negative rates False Positive Rate (FPR) and False Negative Rate (FNR) are also used to assess the prediction performance of the models (Viola et al. 2005). The FPR describes the rate of the non- landslide grid

AC C

cells falsely locating in the high or very high susceptible areas. The FNR describes the rate of the recorded landslide grid cells falsely locating in the low or very low susceptib le areas. In this study, the FPRs of the SOM-ELM, SOM-SVM and single ELM are 10.2%, 11.8% and 14.2%, respectively. The FNRs of the SOM-ELM, SOM-SVM and single ELM are 7.0%, 8.3% and 10.3%, respectively. The results show that the SOM-ELM model has the highest prediction performance.

4.4.4.Prediction efficiency of ELM and SVM models

17

ACCEPTED MANUSCRIPT The computational time of calculating the landslide susceptibility indexes using the ELM and SVM models is recorded. All the predictions are conducted in a server with Intel Xeon CPU [email protected] GHz with 256GB RAM. The computational time of the SVM is 14179.2 seconds, whereas the computational time of the ELM is 10.8 seconds. The results show that the ELM model

PT

requires much less computational time than the SVM model. In addition, the ELM model requires

SC

significantly higher prediction efficiency than the SVM model.

RI

only one parameter, whereas the SVM model requires three parameters. Hence, the ELM model has

NU

5. Conclusion

MA

This study explores the potential application of SOM network based ELM model for mapping landslide susceptibility in Wanzhou district. Nine environmental factors selected by correlation

ED

analysis are used as input variables. The description of a given area as landslide/non- landslide grid cells are assumed to be output variable. And 639 registered landslides in the inventory identified

EP T

from field surveys between 1970 and 2013 are used as recorded landslide grid cells. The reasonable non- landslide grid cells are selected using the SOM network. Next, the SOM-ELM model is applied

AC C

to map the landslide susceptibility based on the input variables, landslide grid cells and reasonable non- landslide grid cells. From the landslide susceptibility map produced by the SOM-ELM, the very high and high susceptibility indexes are mainly distributed in the areas of low elevations and clay stone. On the contrary, the very low and low susceptibility indexes are mainly concentrated in the areas with high elevations and in the areas with lithological units of T1 , T2 and T3 , where are mainly limestone or sandstone. In conclusion, the proposed SOM-ELM model predicts landslide susceptibility indexes more

18

ACCEPTED MANUSCRIPT accurately than the single ELM and SOM-SVM models, and the ELM has much higher prediction efficiency than SVM model. The most important contributions of the SOM-ELM model include the introduce of the SOM network to select the non- landslide grid cells reasonably from the very low susceptible area, and the use of the ELM model to calculate the landslide susceptibility indexes to

PT

overcome the drawbacks of the traditional ANN and SVM models. The locations of present

RI

landslides used as validation dataset are predicted well by the proposed model. Therefore, the

NU

SC

landslide susceptibility map produced in this study can be used to locate the future landslides.

Acknowledgements

MA

This research is funded by the Natural Science Foundation of China (No. 41572292). And thanks to the Department of Wanzhou Geo-environment Monitoring and Prevention for their support

References

EP T

ED

of data sources.

Atkinson, P.M. & Massari, R. 2011. Autologistic modelling of susceptibility to landsliding in the Central Apennines, Italy. Geomorphology, 130, 55-64.

AC C

Ayalew, L. & Yamagishi, H. 2005. The application of GIS-based logistic regression for landslide susceptibility mapping in the Kakuda-Yahiko Mountains, Central Japan. Geomorphology, 65, 15-31. Bai, S.-B., Wang, J., Lü, G.-N., Zhou, P.-G., Hou, S.-S. & Xu, S.-N. 2010. GIS-based logistic regression for landslide susceptibility mapping of the Zhongxian segment in the Three Gorges area, China. Geomorphology, 115, 23-31. Bui, D.T., Pradhan, B., Lofman, O., Revhaug, I. & Dick, O.B. 2012. Spatial pred iction of landslide hazards in Hoa B inh province (Vietnam): a co mparative assessment of the efficacy of evidential belief functions and fuzzy logic models. Catena, 96, 28-40. Choi, J., Oh, H.-J., Lee, H.-J., Lee, C. & Lee, S. 2012. Co mbining landslide susceptibility maps obtained fro m freque ncy ratio, logistic regression, and artificial neural network models using ASTER images and GIS. Engineering Geo logy, 124, 12-23. Chung, C.-J. & Fabbri, A.G. 2008. Pred icting landslides for risk analysis —spatial models tested by a cross -validation technique. Geomorphology, 94, 438-452. Cortes, C. & Vapnik, V. 1995. Support-vector networks. Machine learning, 20, 273-297. Erener, A., Mutlu, A. & Dü zgün, H.S. 2016. A co mparat ive study for landslide susceptibility mapping using GIS -based

19

ACCEPTED MANUSCRIPT mu lti-criteria decision analysis (MCDA), logistic regression (LR) and association rule mining (A RM). Engineering Geology, 203, 45-55. Felicísimo , Á.M., Cuartero, A., Remondo, J. & Qu irós, E. 2013. Mapping landslide susceptibility with logistic regression, mu ltip le adaptive regression splines, classificat ion and regression trees, and maximu m entropy methods: a comparative study. Landslides, 10, 175-189. Fell, R., Coro minas, J., Bonnard, C., Cascini, L., Lero i, E. & Savage, W.Z. 2008. Guidelines for landslide susceptibility, hazard and risk zoning for land-use planning. Engineering Geology, 102, 99-111. Godt, J., Bau m, R., Savage, W., Salciarini, D., Schulz, W. & Harp, E. 2008. Transient deterministic shallow landslide modeling: requirements for susceptibility and hazard assessments in a GIS framework. Engineering Geo logy, 102,

PT

214-226.

Gong, Y., Zhang, Y., Lan, S. & Wang, H. 2016. A Co mparat ive Study of Artificial Neural Net works, Support Vector

RI

Machines and Adaptive Neuro Fuzzy Inference System for Forecasting Groundwater Levels near Lake Okeechobee, Florida. Water Resources Management, 30, 375-391.

SC

He, S., Pan, P., Dai, L., Wang, H. & Liu, J. 2012. Applicat ion of kernel-based Fisher discriminant analysis to map landslide susceptibility in the Qinggan River delta, Three Gorges, China. Geomorphology, 171, 30-41. Huang, F., Yin, K., He, T., Zhou, C. & Zhang, J. 2016. Influencing factor analysis and displacement prediction in

NU

reservoir landslides− a case study of Three Gorges Reservoir (China). Tehnički vjesnik, 23, 617-626. Huang, G.-B., Zhu, Q.-Y. & Siew, C.-K. 2006. Ext reme learning machine: theory and applications. Neurocomputing, 70, 489-501.

MA

Kavzoglu, T., Sahin, E.K. & Colkesen, I. 2014. Landslide susceptibility mapping using GIS -based mult i-criteria decision analysis, support vector machines, and logistic regression. Landslides, 11, 425-439. Kavzoglu, T., Sahin, E.K. & Colkesen, I. 2015. Select ing optimal conditioning fac tors in shallow translational landslide susceptibility mapping using genetic algorithm. Engineering Geology, 192, 101-112.

ED

Lee, C.-T., Huang, C.-C., Lee, J.-F., Pan, K.-L., Lin, M.-L. & Dong, J.-J. 2008. Statistical approach to earthquake-induced landslide susceptibility. Engineering Geology, 100, 43-58. Lee, S., Ryu, J.H., Min, K. & Won, J.S. 2003. Landslide susceptibility analysis using GIS and artificial neural network.

EP T

Earth Surface Processes and Landforms, 28, 1361-1376. Lin, W.T. 2008. Earthquake‐induced landslide hazard mon itoring and assessment using SOM and PROM ETHEE techniques: A case study at the Chiufenershan area in Central Taiwan. International Journal of Geographical Informat ion Science, 22, 995-1012.

AC C

Marjanović, M., Kovačević, M., Bajat, B. & Vo žen ílek, V. 2011. Landslide susceptibility assessment using SVM machine learning algorithm. Engineering Geology, 123, 225-234. Martinović, K., Gav in, K. & Reale, C. 2016. Development o f a landslide susceptibility assessment for a rail network. Engineering Geology, 215, 1-9.

Nefeslioglu, H., Gokceoglu, C. & Sonmez, H. 2008. An assessment on the use of logistic reg ression and artificial neural networks with different sampling strategies for the preparation of landslide susceptibility maps. Engineering Geology, 97, 171-191. Park, H.J., Lee, J.H. & Woo, I. 2013. Assessment of rainfall-induced shallow landslide susceptibility using a GIS -based probabilistic approach. Engineering Geology, 161, 1-15. Park, I. & Lee, S. 2014. Spatial prediction of landslide s usceptibility using a decision tree approach: a case study of the Pyeongchang area, Korea. International Journal of Remote Sensing, 35, 6089-6112. Pradhan, B. 2013. A co mparat ive study on the predictive ability of the decision tree, support vector machine and neuro-fuzzy models in landslide susceptibility mapping using GIS. Computers & Geosciences, 51, 350-365. Pradhan, B., Lee, S. & Buchroithner, M.F. 2010. A GIS-based back-propagation neural network model and its

20

ACCEPTED MANUSCRIPT cross-application and validation for lands lide susceptibility analyses. Co mputers, Environ ment and Urban Systems, 34, 216-235. Ritter, H. & Kohonen, T. 1989. Self-organizing semantic maps. Biological cybernetics, 61, 241-254. Ro mer, C. & Ferentinou, M. 2016. Shallow landslide susceptibility assess ment in a semiarid environ ment — A Quaternary catchment of KwaZulu-Natal, South Africa. Engineering Geology, 201, 29-44. Ruff, M . & Czurda, K. 2008. Landslide susceptibility analysis with a heuristic approach in the Eastern Alps (Vorarlberg, Austria). Geomorphology, 94, 314-324. Shrivastava, N.A., Panig rahi, B.K. & Lim, M. -H. 2016. Electricity price classificat ion using extreme learning machines. Neural Computing and Applications, 27, 9-18.

PT

Toth, E., Brath, A. & Montanari, A. 2000. Co mparison of short -term rainfall predict ion models for real-t ime flood forecasting. Journal of Hydrology, 239, 132-147.

RI

Tsangaratos, P. & Benardos, A. 2014. Estimating landslide susceptibility through a artificial neural network classifier. Natural hazards, 74, 1489-1516.

SC

Van Westen, C.J. 2000. The modelling of landslide hazards using GIS. Surveys in Geophysics, 21, 241-255. Van Westen, C.J., Castellanos, E. & Kuriakose, S.L. 2008. Spatial data fo r landslide susceptibility, hazard, and vulnerability assessment: an overview. Engineering Geology, 102, 112-131.

NU

Vio la, P., Jones, M.J. & Snow, D. 2005. Detecting pedestrians using patterns of motion and appearance. International Journal of Computer Vision, 63, 153-161.

Wu, X., Ren, F. & Niu, R. 2014a. Landslide susceptibility assessment using object mapping units, decision tree, and

MA

support vector machine models in the Three Gorges of China. Environmental Earth Sciences, 71, 4725-4738. Wu, Y., Chen, L., Cheng, C., Yin , K. & Törö k, Á. 2014b. GIS -based landslide hazard predict ing system and its real-t ime test during a typhoon, Zhejiang Province, Southeast China. Engineering Geology, 175, 9-21. Zhu, A.-X., Wang, R., Qiao, J., Qin, C.-Z., Chen, Y., Liu, J., Du, F., Lin, Y. & Zhu, T. 2014. An expert knowledge-based

AC C

EP T

ED

approach to landslide susceptibility mapping using GIS and fuzzy logic. Geomorphology, 214, 128-138.

21

ACCEPTED MANUSCRIPT

ED

MA

NU

SC

RI

PT

Figures

Figure 1. Geographical location of the Wanzhou district (True color image derived from two images of

AC C

path/row 126/39)

EP T

Landsat TM5 with 30 m resolution, one on Sep 2nd, 2009, path/row 127/39 and another one on Aug 29th, 2010,

22

AC C

EP T

ED

MA

NU

SC

RI

PT

ACCEPTED MANUSCRIPT

Figure 2. Environmental factors, (a):DEM, (b):Slope, (c):Aspect, (d):Profile curvature, (e):Plan curvature, (f):Relief amplitude, (g):Lithology, (h):Geological structure, (i):NDVI and (j):Distance to water. The tectonic lines include a:jiajiaoshan anticline, b:qumahe syncline, c:tiefengshan anticline, d:wanxian syncline, e:huangbaixi syncline, f:dachiqianjin anticline, g:fengdouzhongxian anticline, h:fangdoushan anticline, i:ganchang syncline, j:matouchang syncline.

23

SC

RI

PT

ACCEPTED MANUSCRIPT

AC C

EP T

ED

MA

NU

Figure 3. Classification map of landslide susceptibility using SOM network

24

NU

SC

RI

PT

ACCEPTED MANUSCRIPT

AC C

EP T

ED

MA

Figure 4. Classification map of landslide susceptibility using SOM-ELM model

25

NU

SC

RI

PT

ACCEPTED MANUSCRIPT

AC C

EP T

ED

MA

Figure 5. Classification map of landslide susceptibility using SOM-SVM model

26

NU

SC

RI

PT

ACCEPTED MANUSCRIPT

AC C

EP T

ED

MA

Figure 6. Classification map of landslide susceptibility using the single ELM model

27

NU

SC

RI

PT

ACCEPTED MANUSCRIPT

AC C

EP T

ED

MA

Figure 7. Success rate curves of landslide susceptibility indexes calculated using the SOM-ELM, SOM-SVM and single ELM models

28

NU

SC

RI

PT

ACCEPTED MANUSCRIPT

AC C

EP T

ED

MA

Figure 8. Prediction rate curves of landslide susceptibility indexes calculated using the SOM-ELM, SOM-SVM and single ELM models

29

ACCEPTED MANUSCRIPT Tables Table 1. Description of environmental factors and Frequency ratios of all environmental factors Environmental

Description

Values

factors

Frequency ratios

340~440m

1.498

The elevation map shown in Figure 2(a) comes from DEM . The DEM is

440~520m

0.933

also data sources of slope, aspect, relief amplitude, plan curvature and

520~600m

0.641

profile curvature. Some studies show that elevation affects the landslide

600~680m

0.396

stability (M arjanović et al. 2011).

680~770m

0.497

770~860m

0.591

860~

0.189

0~5

0.381

5~10

1.177

10~15

1.280

datum (Wu et al. 2014b). The slope in the Wanzhou district ranges from 0

15~20

1.299

to 72.5° as shown in Figure 2(b). The landslides mainly occur on medium

20~25

1.184

slopes.

25~30

0.930

PT

3.422

SC

RI

Elevation

130~340m

MA

Slope

NU

The slope is the angle between the surface of the earth and a horizontal

ED

Plan curvature

0

0.274 1.010

22.5~67.5

1.082

frequency ratio of 0° is 0.274, while the other eight subclasses are of

67.5~112.5

1.177

almost the same frequency ratios of about 1. Hence, the aspect is not used

112.5~157.5

0.978

in this study.

157.5~202.5

0.894

202.5~247.5

0.995

247.5~292.5

1.186

292.5~337.5

1.039

0~3

1.523

3~6

1.256

The profile curvature describes the complexity of the terrain and it is

6~9

0.959

calculated as the slope of the slope (He et al. 2012) as shown in Figure

9~12

0.772

2(d). A large profile curvature means that the surface is upwardly

12~15

0.631

concave, and a low profile curvature means that the surface is upwardly

15~18

0.530

convex.

18~21

0.440

21~

0.317

0~3

1.523

The plan curvature is the slope of the aspect and describes the horizontal

0~10

1.406

form of the topography (Atkinson and M assari 2011). A relatively large

10~20

1.301

EP T

curvature

0.448

0~22.5,

AC C Profile

0.671

337.5~360

The aspect shown in Figure 2(c) is divided into nine subclasses. The Aspect

30~35 35~65.3

30

ACCEPTED MANUSCRIPT 1.080

30~40

0.893

shown in Figure 2(e).

40~50

0.724

50~60

0.572

60~70

0.548

70~

0.594

0~30

0.393

30~60

1.021

60~90

1.137

90~120

1.149

120~150

0.974

150~180

0.858

180~210

0.719

210~

0.425

J1

0.624

J2

1.265

J3

0.984

T1

0.067

T2

0.540

T3

0.314

1

1.731

2

0.359

3

0.452

4

1.213

5

0.522

6

1.043

7

1.111

8

0.655

9

2.202

10

0.358

11

0.281

12

1.843

~0.36

1.461

0.36~0.41

1.267

0.41~0.46

1.139

0.46~0.49

1.006

0.49~0.52

0.906

0.52~0.55

0.773

0.55~0.58

0.636

The relief amplitude describes the terrain features by calculating the

elevation difference in a certain area (Bui et al. 2012). A large relief amplitude indicates a steep slope as shown in Figure 2(f).

SC

RI

amplitude

20~30

relatively small value d enotes that the surface is upwardly concave as

PT

Relief

plan curvature denotes that the surface is upwardly convex, and a

Lithology has significant effect on landslide occurrence because different

NU

rocks have different mechanical and hydrological properties (Van Westen et al. 2008). There are six lithological units in Wanzhou district: J 1 (quartz sandstone, shale, limestone), J 2 (sandy mudstone, quartz sandstone, shale, feldspatite, siltstone), J 3 (quartz sandstone, aubergine mudstone, lithic

MA

Lithology

sandstone, shale), T 1 (dolomitic limestone, flaglike limestone), T 2 (argillaceous

limestone,

dysaerobic

fauna,

aubergine

mudstone,

EP T

as Figure 2 (g).

ED

calcareous shale), and T 3 (lithic Sandstone, arenaceous shale, coal seam)

There are ten tectonic lines through the Wanzhou district. The whole Wanzhou district is classified into 12 different geological zones using the tectonic lines as Figure 2 (h). The classification criterion is that the

structure

between

different

geological

zones

should

be

AC C

Geological

boundary

perpendicular to the anticline and syncline lines. For example, the boundary between 1st zone and the 8th zone should be perpendicular to the wanxian syncline and dachiqianjin anticline lines. The different geological zones have different influence on the landslide occurrence (Tsangaratos and Benardos 2014).

Land cover changes hydrological characteristics of slope soil (M arjanović et al. 2011). The NDVI in Wanzhou district is derived from the two images of Landsat TM 8 with 30 m resolution, one on Oct 14th, 2015, NDVI

path/row 126/39 and another one on Oct 21th, 2015, path/row 127/39. A low NDVI means that the grid cell is covered by few vegetation as Figure 2(i).

31

ACCEPTED MANUSCRIPT 0.58~

0.497

Rivers play important roles on landslide occurrence because of slope

1

1.590

Distance to

erosion and slope materials saturating. A grid cell near the river has

2

0.979

river

higher water content than a grid cell far from river (He et al. 2012). The

3

0.559

4

0.384

AC C

EP T

ED

MA

NU

SC

RI

PT

distance to river is obtained by buffer analysis of ArcGIS as Figure 2(j).

32

ACCEPTED MANUSCRIPT

Table 2 Correlation coefficients between the eleven environmental factors Elevation

Slope

Profile curvature

Plan curvature

Relief amplitude

Lithology

Geological zone

NDVI

1

Slope

0.024

1

Profile curvature

0.070

-0.018

1

Plan curvature

0.042

-0.039

0.298

1

Relief amplitude

0.030

0.362

-0.083

-0.111

1

Lithology

0.320

0.027

0.058

-0.009

0.049

Geological zone

0.357

0.040

0.052

0.015

0.037

-0.011

1

NDVI

0.278

0.047

0.115

-0.017

0.033

0.224

0.185

1

Distance to river

0.428

-0.055

-0.019

0.023

0.258

0.114

0.096

RI

PT

Elevation

1

SC

NU

AC C

EP T

ED

MA

-0.002

33

Distance to river

1

ACCEPTED MANUSCRIPT

Table 3. Frequency ratios of the five susceptibility classes of SOM network Percentage of pixels in

Landslide

Percentage of landslide

Frequency

domain

domain (%)

occurrence pixels

occurrence pixels (%)

ratio

Very high

491086

13.0

46058

45.9

3.500

High

1100729

29.2

30659

M oderate

1136991

30.1

14456

Low

365883

9.7

4447

Very low

677313

18.0

5454

34

30.3

1.040

14.3

0.475

4.4

0.454

5.4

0.300

RI

SC NU MA ED EP T AC C

SOM

PT

Pixels in

Classification

ACCEPTED MANUSCRIPT

Table 4. Frequency ratios of the five susceptibility classes of SOM-ELM model Percentage of pixels

Landslide

Percentage of landslide

Frequency

domain

in domain (%)

occurrence pixels

occurrence pixels (%)

ratio

Very high

377200

10

41541

41.1

4.11

High

754400

20

29513

29.2

1.46

M oderate

1508801

40

22944

22.7

0.568

Low

754400

20

4852

Very low

377200

10

2224

RI SC NU MA ED EP T AC C

SOM -ELM

PT

Pixels in

Classification

35

4.8

0.24

2.2

0.22

ACCEPTED MANUSCRIPT

Table 5. Frequency ratios of the five susceptibility classes of SOM-SVM model Percentage of pixels

Landslide

Percentage of landslide

Frequency

domain

in domain (%)

occurrence pixels

occurrence pixels (%)

ratio

Very high

377200

10

37498

37.1

3.710

High

754400

20

33052

32.7

1.635

M oderate

1508801

40

22135

21.9

0.548

Low

754400

20

5721

Very low

377200

10

2668

RI SC NU MA ED EP T AC C

SOM -SVM

PT

Pixels in

Classification

36

5.7

0.283

2.6

0.264

ACCEPTED MANUSCRIPT

Table 6. Frequency ratios of the five susceptibility classes of single ELM model Percentage of pixels

Landslide

Percentage of landslide

Frequency

domain

in domain (%)

occurrence pixels

occurrence pixels (%)

ratio

Very high

377200

10

35275

34.9

3.49

High

754400

20

29918

M oderate

1508801

40

25471

Low

754400

20

7277

Very low

377200

10

3133

RI

SC NU MA ED EP T

ELM

AC C

Single

PT

Pixels in

Classification

37

29.6

1.48

25.2

0.63

7.2

0.36

3.1

0.31

ACCEPTED MANUSCRIPT

NU

SC

RI

PT

Appendix: Fifteen registered landslides which distribute evenly over Wanzhou district.

AC C

EP T

ED

MA

Figure 9. Fifteen registered landslides in the inventory in Wanzhou district

38

ACCEPTED MANUSCRIPT Table 10. The attributions of the fifteen registered landslides in the inventory occurrence ID

Landslide name

Volume

Location

Trigger factors

Lithology

Elevation

Threatened

Threatened

(m )

(m)

people

buildings

Slope (°)

time

3

Sanping village, 1

Qianjiabang

30/6/1995

Rainstorm

J2

15

1170000

622

55

16

25/6/1995

Reservoir water level change

J2

20

1020000

235

210

345

3/3/2003

continually heavy rain

J2

18

104000

710

61

85

20/6/1995

Rainstorm

J2

30

3000

1074

46

40

24/7/1990

continually heavy rain

14/8/2002

Rainstorm

1/9/2008

Reservoir water level change

1/5/2003

7/6/2002

Huangbai town Tangjiao village, 2

Tangjiao Chenjiaba town Xiaowan Changeling town

4

Daoluoba

village, Baitu

RI

Lianmeng

town Hongqiangyuan

J2

15

86700

773

14

20

J2

15

82000

183

15

15

J2

18

527000

158

12

15

Reservoir water level change

J2

15

5800000

235

400

500

continually heavy rain

J2

25

45000

360

140

136

J2

15

1200000

564

134

268

Longju town Yeya village, 6

Zhujiagou

NU

Zouma town Tanshao village, 7

Xiayaozui Xintian town Youfangzui

MA

Wuxi village, 8

Xintian town Laolin village, 9

Tianba

10

Huanghuaping

village,Fenshui

28/6/2003

Kuaile village, 11

Doudibang Sunjia town Huanglong Wujiaping

village,Tanzi

9/6/2005

continually heavy rain

J3

17

280000

840

81

100

16/7/1998

Rainstorm

J1

35

532500

800

16

24

22/8/1998

Rainstorm

J2

15

185625

723

53

51

1/10/1988

Rainstorm

J2

35

535000

285

60

160

28/7/2002

Rainstorm

J2

20

213000

340

641

915

AC C

12

continually heavy rain

EP T

twon

ED

Longsha town Daxing

SC

Chayuan village, 5

PT

Luci village, 3

town

Tianyuan village, 13

Chapanshi

Houshan town

Jizhong village, 14

Shibaozui

Zhoujiaba town Wutu village, 15

Liujiaping Dazhou town

39

ACCEPTED MANUSCRIPT Highlights  Reasonable non-landslide grid cells are selected from the very low susceptible area produced by 

self-organizing-map (SOM) network. SOM-extreme learning machine (ELM) model is successfully used to map landslide susceptibility in Wanzhou district, Three Gorges Reservoir. SOM-ELM model possesses higher success and prediction rates than the single ELM and SOM-support vector machine (SVM) models.



ELM model has much higher prediction efficiency than the SVM for calculating landslide susceptibility indexes.

AC C

EP T

ED

MA

NU

SC

RI

PT



40