IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL.

19, 2022 3001505

Landslide Susceptibility Prediction Using Sparse

Feature Extraction and Machine Learning Models
Based on GIS and Remote Sensing
Li Zhu , Member, IEEE, Gongjian Wang, Faming Huang , Yan Li, Wei Chen, and Haoyuan Hong

Abstract— Landslide susceptibility prediction (LSP) is a useful difficult task [1], [2]. Landslide susceptibility prediction (LSP)
technology for landslide prevention. Due to the complex nonlinear plays an important role in accurately locating the potential
correlations among environmental factors, traditional machine landslide. Therefore, it is necessary to carry out in-depth
learning (ML) models have unsatisfactory LSP accuracies. In this
research on the LSP [3].
letter, a sparse feature extraction network (SFE+) is proposed for
LSP. First, the landslides and environmental factors are collected, Recently, based on the remote sensing (RS) and geographic
and frequency ratios of environmental factors are calculated as information system (GIS), data-driven LSP models have been
the model inputs. Second, the input data are passed through developed deeply, which can be divided into heuristic [4],
the input layer with the dropout, and then, the features are mathematical–statistical [5], and machine learning (ML) mod-
passed through the hidden layers, that is, the k% lifetime sparsity els [6]. Among them, heuristic and mathematical–statistical
layers. The hidden layers are employed to further sparse these models have been widely employed in LSP, such as analytical
factors to obtain the independent and redundant prediction
hierarchy process and statistical index. The above models
features as much as possible. Finally, certain classifiers are used
to realize the LSP in the study area. SFE-support vector machine can predict the level of landslide susceptibility to a certain
(SVM), SFE-logistic regression (LR), and SFE-stochastic gradient extent. However, the accuracy of susceptibility results calcu-
descent (SGD) models are built. For comparison, principal com- lated by simple linear statistical methods is low. Meanwhile,
ponent analysis (PCA)-SVM, PCA-LR, PCA-SGD, SVM, LR, and it is difficult to truly reflect the nonlinear coupling effect of
SGD models are also built for LSP in Shicheng County, China. environmental factors on landslide susceptibility [7].
Results show that the SFE-based ML models, especially the With the rapid development of ML, various ML models
SFE-SVM, can effectively extract the sparse nonlinear features
of environmental factors to improve LSP accuracies and have
have been successfully used in LSP, including logical regres-
promising prospects for LSP. sion (LR), support vector machine (SVM), and so on [8].
Considerable studies show that compared with heuristic and
Index Terms— Geographic information system (GIS), landslide mathematical–statistical models, ML models have a higher
susceptibility prediction (LSP), neural network, remote sens-
ing (RS), sparse feature extraction (SFE). LSP accuracy [9].
However, when constructing an LSP model, it is urgent to
solve the problems of feature learning and optimization of
input data by ML models. These problems specifically are:
I. I NTRODUCTION 1) ML models require a lot of prior knowledge in the features
learning processes, while the models cannot automatically
L ANDSLIDES have been widely developed worldwide.
Generally, accurately locating the landslide sites is a extract features from data [10] and 2) the abovementioned ML
methods cannot extract more representative features from huge
Manuscript received October 15, 2020; revised January 4, 2021; accepted input data [11]. To address these problems, a new algorithm,
January 18, 2021. Date of publication February 8, 2021; date of current version namely, sparse feature extraction network (SFE+), is proposed
December 16, 2021. This work was supported in part by the National Natural
Science Foundation of China under Grant 41807285, in part by the Natural to extract the features from the input data. Then, the SFE-based
Science Foundation for Outstanding Young Scholars of Jiangxi Province under LSP model is built. Generally, SFE+ is a simple and efficient
Grant 2018ACB21038, in part by the Natural Science Foundation of Jiangxi unsupervised feature extraction method, which improves the
Province of China under Grant 20192BAB216034, in part by the Postdoctoral
Science Foundation of China under Grant 2019M652287, and in part by the
discriminative power of features by extracting superior sparse
Jiangxi Provincial Postdoctoral Science Foundation under Grant 2019KY08. features and then improves the LSP performance. SFE has also
(Corresponding author: Faming Huang.) been widely used in other fields, such as image classification,
Li Zhu, Gongjian Wang, and Yan Li are with the School of face recognition, and radar signal recognition [12]. However,
Information Engineering, Nanchang University, Nanchang 330031,
China (e-mail: [email protected]; [email protected];
the SFE+ has not been developed for LSP; hence, it is
[email protected]). significant to introduce this algorithm into LSP.
Faming Huang is with the School of Civil Engineering and To sum up, this study constructs the SFE-SVM, SFE-
Architecture, Nanchang University, Nanchang 330031, China (e-mail: LR, and SFE-stochastic gradient descent (SGD) models to
[email protected]).
Wei Chen is with the College of Geology and Environment, Xi’an implement LSP. Meanwhile, the single SVM, LR, and SGD
University of Science and Technology, Xi’an 710126, China (e-mail: models without feature extraction are used for comparisons.
[email protected]). Furthermore, the principal component analysis (PCA) algo-
Haoyuan Hong is with the Department of Geography and Regional
Research, University of Vienna, 211500 Vienna, Austria (e-mail:
rithm is also introduced to build the PCA-SVM, PCA-LR,
[email protected]). and PCA-SGD models for comparisons. The Shicheng County,
Digital Object Identifier 10.1109/LGRS.2021.3054029 China, is used as the study area, and its landslide susceptibility
1558-0571 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Hyderabad IG Memorial Library. Downloaded on September 13,2022 at 04:36:42 UTC from IEEE Xplore. Restrictions apply.
 3001505 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 19, 2022

where η represents an estimated support vector set, and (η)c

represents its complement. In order to estimate the activation
statistics for specific hidden units in all samples, minibatch
statistics are used in this letter. Meanwhile, a threshold is found
to realize a sparse rate of k%. k is defined as
k = μ × batch_size (3)
where μ represents the degree of sparsity and batch_size
represents the size of the minibatch. Only a small number of
nonzero elements exist in each row of the feature matrix.
3) Output Layer: The sparse loss function of SFE+ can be
formulated as
s2
 
Fig. 1. SFE+ structure.
L sparse (W, b) = L(W, b) + β KL μμ̂ j
maps (LSMs) are produced. The prediction rate curve and j =1

statistical indicators are used to evaluate the LSP. 1 

m
= L(W, b; x i , R i )
II. M ETHODOLOGY m i=1
The modeling processes of SFE-based ML models can be n i −1  j +1
λ
si s
 l 2
described as follows. + W ji
1) Recorded landslides and geospatial analysis of environ- 2 l=1 i=1 j =1
mental factors are conducted.  s2  
μ 1−μ
2) The effects of these factors on landslide occurrence are +β μ log + (1 − μ) log
calculated by frequency ratio (FR). j =1
μ̂ j 1 − μ̂ j
3) Spatial training and testing data sets are built. (4)
4) SFE+ is used to extract nonlinear sparse features from
the initial input data for building ML models. where W represents the weight, b represents the bias, m
5) SFE-SVM, SFE-LR, and SFE-SGD models are represents the number of inputs, and λ represents the para-
employed to produce LSM of Shicheng County. meters of the weight attenuation. L(W, b) is the loss  function

6) PCA-based and single ML models are used for compar- of the network without the sparse constraint. KL μμ̂ j is
isons. the Kullback–Leibler (KL) dispersion between the Bernoulli
7) LSP accuracies are evaluated. random variables with mean values of μ and μ̂ j , where μ is
the degree of sparsity and μ̂ j is the average activation of the
A. Proposed SFE Network hidden unit j . β represents the weight of the penalty factor,
An SFE-based ML model is proposed, as shown in Fig. 1. and S2 represents the quantity of neurons in the hidden layer.
This model consists of the input layer, SFE of the hidden In the nonsparse  network, a penalty factor item
layer, and traditional classifiers of ML. The FR values of the β sj2=1 KL μμ̂ j is added to the loss function to achieve
environmental factors are employed as the initial input of this the optimal sparsity of the data. The greater difference  
model; then, the dropout and lifetime sparsity are employed to between μ and μ̂j demonstrates the greater KL μμ̂ j .
extract sparse features. Finally, the LSP models are constructed When μ = μ̂ j , KL μμ̂ j = 0, the extracted feature sparsity
by the extracted optimal sparse features. is optimal. As a result, the parameters of objective function
1) Input Layer: In the input layer, all 14 environmental L sparse (W, b) are continuously optimized so that μ and μ̂ j are
factors x i (i = 1, 2, . . . , 14) are used as the raw data. x i is then infinitely close. Ultimately, the independent and nonredundant
manipulated into  x i by dropout. Dropout deletes a portion of optimal features can be extracted. The total loss function of
neurons in the network randomly, which prevents overfitting SFE+ is mean square error (MSE), which can be formulated
and gradient disappearance effectively. It also reduces the as
complex coadaptation between neurons and facilitates the
1
n
following nonredundant optimal feature extraction. L M S E (y, ŷ) = (yi − ŷi )2 (5)
2) Hidden Layer: The SFE+ includes two hidden layers n i=1
and lifetime sparsity for SFE. Meanwhile, the rectified linear
where ŷ, y, and n represent the output data, the original input
unit (ReLU) nonlinear activation function is applied in the
data, and the number of samples, respectively. Backpropaga-
hidden layer that increases the nonlinearity of the model,
tion is carried out to calculate the loss of the SFE+, and the
as well as reducing the interdependence between parameters.
parameters are optimized in the training process.
The formula of the ReLU function can be expressed as (1).
Then, the k% largest activations of Ri are found, and the rest
of activations are set to zero B. Model Performance Evaluation

xi , 
 xi > 0 Several measures, including the negative predictive rate
Ri = (1) (NPR), positive predictive rate (PPR), total accuracy (TA), and
0,  xi ≤ 0
prediction rate curve, are employed to evaluate the overall LSP
(η)c
Ri =0 where η = suppk (R i ) (2) accuracy for model comparisons.

Authorized licensed use limited to: University of Hyderabad IG Memorial Library. Downloaded on September 13,2022 at 04:36:42 UTC from IEEE Xplore. Restrictions apply.
 ZHU et al.: LSP USING SFE AND ML MODELS BASED ON GIS AND RS 3001505

Fig. 2. Environmental factors of Shicheng area. (a) DEM. (b) Slope.

(c) Distance to river. (d) NDVI. (e) Total surface radiation. (f) Lithology
(other eight environmental factors are not present).

III. S TUDY A REA AND M ATERIALS

A. Shicheng County and Landslides Inventory
Shicheng County is located in the southeast part of Jiangxi
Province, China. Its area covers about 1581.5 km2 with a
latitude from 25◦ 57’47” to 26◦ 36’13”N and a longitude from
116◦ 05’46” to 116◦ 38’03”E. The terrain features are mainly Fig. 3. LSMs, respectively, produced by (a) single SVM, (b) single SGD,
hills and mountains. Shicheng County resides in a subtropical (c) single LR, (d) SFE-SVM, (e) SFE-SGD, (f) SFE-LR, (g) PCA-SVM,
(h) PCA-SGD, and (i) PCA-LR.
monsoon humid climate zone with an annual average rainfall
of about 1748.6 mm. More and more landslides occurred in 14 environmental factors are acquired according to the land-
Shicheng County due to the seasonal heavy rainfall, complex slide features in Shicheng County and related references [17].
terrain features, and disorder engineering constructions. A total As shown in Fig. 2, these factors selected as the input of
of 369 landslides are investigated by the Land and Resources the LSP model include topography (DEM, slope, aspect,
Department of Jiangxi Province from 1970 to 2012. The relief amplitude, plan curvature, and profile curvature), geol-
quaternary silty clay intercalated with crushed stones form ogy (lithology), hydrology [wetness index (TWI), distance
the accumulation of landslides, which can be regarded as to the river, and modified normalized difference water index
shallow soil landslides. These landslides can be characterized (MNDWI)], and land-cover factors [normalized difference
as a movement type of clay/silt slide with small scale and vegetable index (NDVI), total surface radiation, population
group occurrence, with area varying from 1.0 × 103 m2 to density index, and normalized difference building index
1.6 × 104 m2 and the thickness varying from 2 to 8 m [13]. (NDBI)]. In this case, when the slope is 3.9◦ ∼ 14.9◦,
NDBI is greater than 0.373, NDVI is 0.205 ∼ 0.593, TWI is
B. Landslide Environmental Factor Analysis 6.1 ∼ 9.6, MNDWI is 0.145 ∼ 0.612, metamorphic rock is
In this study, 14 environmental factors were mainly derived 1.023 and 1.192, carbonate rock is 1.192, and the FR value of
from: 1) lithology distribution map with a scale of 1: 100 000; landslide occurrences is greater than 1, which is conducive to
2) Landsat TM8 RS image with a resolution of 30 m for the occurrence of the landslide. Therefore, the FR values of
extracting land-cover factors; and 3) the digital elevation the subclasses of the environmental factors are widely used as
model (DEM) with 30-m resolution used to extract topo- input data of ML models because the FR analysis can get a
graphic and geomorphic factors. Slope and grid units are general knowledge about the effects of environmental factors
widely used prediction units. Grid units have advantages of on landslide occurrence [18].
regular shape, convenient for rapid subdivision, and high
model calculation efficiency, which are widely used for IV. L SP R ESULTS OF S HICHENG C OUNTY
LSP [14]. The application of 30-m resolution grids can not
only effectively characterize the topographic features but also A. Preparation of the Spatial Data Set
avoid the excessive computation of the model due to too many The spatial data set building is a significant LSP modeling
grids [15]. process. The feature extraction results of FR values of the
Environmental factors can be mainly classified as topog- 14 environmental factors are used as the final input data of
raphy, geology, hydrological, and land-cover factors [16]; ML models, and the landslide susceptibility indexes (LSIs) are

Authorized licensed use limited to: University of Hyderabad IG Memorial Library. Downloaded on September 13,2022 at 04:36:42 UTC from IEEE Xplore. Restrictions apply.
 3001505 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 19, 2022

Fig. 4. Prediction rate curves of each model. (a) SFE-SVM, single SVM and PCA-SVM. (b) SFE-LR, single LR and PCA-LR. (c) SFE-SGD, single SGD
and PCA-SGD.

the model output. First, the 369 landslides were converted into TABLE I
2709 landslide grid units by the ArcGIS software, and then, L ANDSLIDE P REDICTION P ERFORMANCE
the 2709 landslide grid units in the data set were randomly
divided into a training set (70%) and testing set (30%).
Meanwhile, nonlandslide grid units with the same number are
randomly selected from the landslide-free area, which was also
divided into training and test data sets according to the above
ratio. The landslide and nonlandslide grid units were set to the
labels “1” and “0,” respectively.

B. LSP Using Single ML Model

The LSIs of all grid units in Shicheng County were calcu-
lated by the trained single SVM model, and then, the predicted 0.0001, and the liblinear method was selected for optimization.
LSIs were imported into the GIS software to produce LSM. Then, the trained SFE-LR model was used to calculate the
The generated LSM was divided into five susceptibility levels LSIs and generate LSM [see Fig. 3(f)].
using the natural breakpoint method: very high (12.13%), high 3) SFE-SGD Model: In the same way, L2 regularization
(20.32%), moderate (29.43%), low (23.15%), and very low was used to build the SFE-SGD model, and the penalty
(14.97%) [see Fig. 3(a)]. In addition, the trained single SGD coefficient and the stopping criterion were set to 0.001 and
and single LR models were also applied to predict the LSIs 0.001, respectively. Then, the trained SFE-SGD was applied
with the LSMs shown in Fig. 3(b) and (c), respectively. to calculate the LSIs and produce LSM [see Fig. 3(e)].

C. LSP Using SFE-Based ML Models D. LSP Using PCA-Based ML Models

The proposed LSP model was trained and tested based PCA was used for linear feature extraction of environmental
on the environment of TensorFlow 1.14 [11]. The optimal factors. The extracted 11 principal components were used as
parameters of the SFE+ in the experiments are the learning input data to construct PCA-SVM, PCA-SGD, and PCA-LR
rate of 0.005, the sparsity of 0.32, the hidden layers’ number model with LSMs shown in Fig. 3(g) and (h).
of 2, the hidden units’ number of 512, the batch_size of 400,
and the training iterations of 8000. A number of experiments
have been conducted in order to select the batch size, which is E. Model Prediction Performance Analysis
a significant parameter for the deep learning model. Minibatch 1) Predictive Rate Accuracy and Statistical Index Accuracy:
gradient descent is used, and the results show that 400 is an The prediction rate curve is used to evaluate the fit of the
appropriate value of batch size for the SFE+ networks. All LSP model to the test data [20]. First, the calculated LSIs
the experiments were carried out in a computer equipped with of 17 57 337 grid units in Shicheng County were sorted in
Intel Core i7-7820X CPU and NVIDIA TITAN XP GPU. The descending order, and the sorted data were divided into
nonlinear features extracted by SFE+ were used to construct 20 equal parts by GIS software. Then, the prediction rate curve
SVM, LR, and SGD models. was drawn by calculating the proportion of the landslide grid
1) SFE-SVM Model: For the SFE-SVM modeling, units falling into the equal grade in the test data. The areas
the cross-validation approach [19] was adopted to determine its under the prediction rate curve (AUC) of LSP models are
stop standard parameter and the penalty coefficient that were shown in Fig. 4, indicating that AUC values of the SFE-based
set to 0.001 and 4, respectively. Similarly, L2 regularization models are higher than the PCA-based and single ML models.
was selected for optimization. Then, the LSIs of all grid units The statistical measurement results of all the models are
were calculated by the trained SFE-SVM model, and the shown in Table I, indicating that the SFE-based model achieves
generated LSM is shown in Fig. 3(d). better TA value than other models, and SFE-SVM achieves
2) SFE-LR Model: The SFE-LR model also uses L2 reg- the highest prediction accuracy of 74.52%. The above results
ularization to avoid overfitting in the training process. The show that the SFE-based model has better LSP performance
penalty coefficient was 1000, the stopping criterion was set to than the PCA-based and single ML model.

Authorized licensed use limited to: University of Hyderabad IG Memorial Library. Downloaded on September 13,2022 at 04:36:42 UTC from IEEE Xplore. Restrictions apply.
 ZHU et al.: LSP USING SFE AND ML MODELS BASED ON GIS AND RS 3001505

[2] W. Chen, X. Chen, J. Peng, M. Panahi, and S. Lee, “Landslide

susceptibility modeling based on ANFIS with teaching-learning-based
optimization and satin bowerbird optimizer,” Geosci. Frontiers, vol. 12,
no. 1, pp. 93–107, Jan. 2021.
[3] A. M. Youssef and H. R. Pourghasemi, “Landslide susceptibility map-
ping using machine learning algorithms and comparison of their per-
formance at Abha basin, Asir region, Saudi Arabia,” Geosci. Frontiers,
vol. 12, no. 2, pp. 639–655, Mar. 2021.
[4] Z. Liang, Y. Tian, L. Wu, and G. Jia, “Comparison and inte-
gration of heuristic and statistical models of landslide suscepti-
bility mapping,” in Proc. 18th Int. Conf. Geoinform., Jun. 2010,
pp. 1–5.
[5] H. Zhang, G. Zhang, and Q. Jia, “Integration of analytical hierarchy
process and landslide susceptibility index based landslide susceptibility
assessment of the pearl river delta area, China,” IEEE J. Sel. Topics
Appl. Earth Observ. Remote Sens., vol. 12, no. 11, pp. 4239–4251,
Fig. 5. Loss and accuracy curves with iterations. Nov. 2019.
[6] N. Kussul, M. Lavreniuk, S. Skakun, and A. Shelestov, “Deep learn-
ing classification of land cover and crop types using remote sensing
2) Analysis of the Feature Extraction of Environmental data,” IEEE Geosci. Remote Sens. Lett., vol. 14, no. 5, pp. 778–782,
Factor: In the training process, model learning and automatic May 2017.
update of parameters are completed by iterating continuously [7] D. Tien Bui, Q.-T. Bui, Q.-P. Nguyen, B. Pradhan, H. Nampak, and
to the optimal values. The loss curve (see Fig. 5) of the P. T. Trinh, “A hybrid artificial intelligence approach using GIS-based
neural-fuzzy inference system and particle swarm optimization for forest
training process can intuitively reflect the learning situation fire susceptibility modeling at a tropical area,” Agricult. Forest Meteorol.,
of the model. It is obvious that the training loss is declining, vol. 233, pp. 32–44, Feb. 2017.
while the testing accuracy is rising. [8] B. Wu, W. Qiu, J. Jia, and N. Liu, “Landslide susceptibility
modeling using bagging-based positive-unlabeled learning,” IEEE
An SFE-based ML model has significant advantages com- Geosci. Remote Sens. Lett., early access, May 11, 2020, doi:
paring to other traditional ML models. First, the PCA algo- 10.1109/LGRS.2020.2989497.
rithm is a simple method to extract linear features and is [9] P. Reichenbach, M. Rossi, B. D. Malamud, M. Mihir, and F. Guzzetti, “A
more dependent on the input data. On the contrary, the SFE+ review of statistically-based landslide susceptibility models,” Earth-Sci.
Rev., vol. 180, pp. 60–91, May 2018.
is more universal and improves LSP performance by mining [10] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, P. A. Manzagol, and
the potential nonlinear features in the input data. Second, L. Bottou, “Stacked denoising autoencoders: Learning useful represen-
the lifetime sparsity is introduced into the SFE+ for extracting tations in a deep network with a local denoising criterion,” J. Mach.
Learn. Res., vol. 11, no. 12, pp. 3371–3408, 2010.
sparse features, which improves the discrimination of features. [11] Y. Wang, Z. Fang, and H. Hong, “Comparison of convolutional neural
Finally, the ReLU activation function and dropout are also networks for landslide susceptibility mapping in Yanshan county, China,”
used during the training process. This function can not only Sci. Total Environ., vol. 666, pp. 975–993, May 2019.
reduce overfitting and the cooperative adaptation between [12] R. Abiantun, U. Prabhu, and M. Savvides, “Sparse feature extraction for
pose-tolerant face recognition,” IEEE Trans. Pattern Anal. Mach. Intell.,
hidden units but also enhance the generalization ability of the vol. 36, no. 10, pp. 2061–2073, Oct. 2014.
LSP model. Compared with the linear features extracted by [13] O. Hungr, S. Leroueil, and L. Picarelli, “The Varnes classification of
PCA, the nonlinear features extracted by the SFE+ are more landslide types, an update,” Landslides, vol. 11, no. 2, pp. 167–194,
discriminative and can preferably represent the landslide data. Apr. 2014.
[14] F. Huang, J. Zhang, C. Zhou, Y. Wang, J. Huang, and L. Zhu, “A deep
learning algorithm using a fully connected sparse autoencoder neural
V. C ONCLUSION network for landslide susceptibility prediction,” Landslides, vol. 17,
no. 1, pp. 217–229, Jan. 2020.
In this letter, a type of SFE-based ML model is pro- [15] B. T. Pham, D. Tien Bui, I. Prakash, and M. B. Dholakia,
posed for LSP in Shicheng County. Results show that the “Hybrid integration of multilayer perceptron neural networks and
proposed model has higher LSP accuracies than those of machine learning ensembles for landslide susceptibility assessment
other models. The proposed SFE-based model has several at himalayan area (India) using GIS,” Catena, vol. 149, pp. 52–63,
Feb. 2017.
advantages: 1) this model is universal and practical; 2) the [16] A. Aditian, T. Kubota, and Y. Shinohara, “Comparison of GIS-based
lifetime sparsity is adopted in the SFE+ to improve the landslide susceptibility models using frequency ratio, logistic regression,
discrimination of sparse features; and 3) the SFE+ can prevent and artificial neural network in a tertiary region of ambon, Indonesia,”
overfitting and enhance the generalization ability of the model. Geomorphology, vol. 318, pp. 101–111, Oct. 2018.
[17] L. Zhu et al., “Landslide susceptibility prediction modeling based on
In summary, the SFE-based ML model can effectively extract remote sensing and a novel deep learning algorithm of a cascade-
the high-quality nonlinear features from the input data to parallel recurrent neural network,” Sensors, vol. 20, no. 6, p. 1576,
improve LSP accuracy. These landslide susceptibility-related Mar. 2020.
[18] X. Chen and W. Chen, “GIS-based landslide susceptibility assessment
conclusions can supply the local government with theoretical using optimized hybrid machine learning methods,” Catena, vol. 196,
guidance for land-use planning and landslide risk prevention. Jan. 2021, Art. no. 104833.
In further research, it is significant to use rainfall amounts, [19] T. Kavzoglu, E. K. Sahin, and I. Colkesen, “Landslide susceptibil-
effective antecedent rainfall, and precipitation threshold to ity mapping using GIS-based multi-criteria decision analysis, support
vector machines, and logistic regression,” Landslides, vol. 11, no. 3,
assess landslide hazards, predict the occurrence of landslide pp. 425–439, Jun. 2014.
occurrence, and achieve real-time landslide warning [21]. [20] F. Huang, Y. Wang, Z. Dong, L. Wu, Z. Guo, and T. Zhang, “Regional
landslide susceptibility mapping based on grey relational degree model,”
R EFERENCES Earth Sci, vol. 44, pp. 664–676, Mar. 2019.
[21] A. M. S. Pradhan, H.-S. Kang, J.-S. Lee, and Y.-T. Kim, “An ensemble
[1] F. Huang, K. Yin, J. Huang, L. Gui, and P. Wang, “Landslide susceptibil- landslide hazard model incorporating rainfall threshold for Mt. Umyeon,
ity mapping based on self-organizing-map network and extreme learning South Korea,” Bull. Eng. Geol. Environ., vol. 78, no. 1, pp. 131–146,
machine,” Eng. Geol., vol. 223, pp. 11–22, Jun. 2017. Feb. 2019.

Authorized licensed use limited to: University of Hyderabad IG Memorial Library. Downloaded on September 13,2022 at 04:36:42 UTC from IEEE Xplore. Restrictions apply.