Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

European Journal of Environmental and Civil Engineering

ISSN: 1964-8189 (Print) 2116-7214 (Online) Journal homepage: https://www.tandfonline.com/loi/tece20

Hydrodynamic landslide displacement prediction


using combined extreme learning machine and
random search support vector regression model

Rubin Wang, Kun Zhang, Wei Wang, Yongdong Meng, Lanlan Yang & Haifeng
Huang

To cite this article: Rubin Wang, Kun Zhang, Wei Wang, Yongdong Meng, Lanlan Yang & Haifeng
Huang (2020): Hydrodynamic landslide displacement prediction using combined extreme learning
machine and random search support vector regression model, European Journal of Environmental
and Civil Engineering, DOI: 10.1080/19648189.2020.1754298

To link to this article: https://doi.org/10.1080/19648189.2020.1754298

Published online: 25 Apr 2020.

Submit your article to this journal

Article views: 4

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tece20
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING
https://doi.org/10.1080/19648189.2020.1754298

Hydrodynamic landslide displacement prediction using


combined extreme learning machine and random search
support vector regression model
Rubin Wanga,b,c, Kun Zhanga, Wei Wanga, Yongdong Mengc, Lanlan Yanga and
Haifeng Huangb
a
Key Laboratory of Ministry of Education for Geomechanics and Embankment Engineering, Hohai University,
Nanjing, China; bNational Field Observation and Research Station of Landslides in the Three Gorges Reservoir
Area of Yangtze River, China Three Gorges University, Yichang, China; cKey Laboratory of Geological Hazards
on Three Gorges Reservoir Area, Ministry of Education, China Three Gorges University, Yichang, China

ABSTRACT ARTICLE HISTORY


Many models have been developed for landslide displacement prediction, Received 16 March 2020
but owing to complex landslide-formation mechanisms and landslide- Accepted 6 April 2020
inducing factors, such models have different prediction accuracies. Thus,
KEYWORDS
landslide displacement prediction remains a popular but difficult topic of
hydrodynamic landslide;
research. In this paper, a landslide prediction model is proposed by com- extreme learning machine;
bining extreme learning machine (ELM) and random search support vector support vector regression;
regression (RS-SVR) sub-models. Particularly, the combined model decom- random search;
posed accumulative landslide displacement into two terms, trend and peri- displacement prediction
odic displacements, using a time series model, and simulated and
predicted the two terms using the ELM and RS-SVR sub-models, respect-
ively. The predicted trend and periodic terms are then summed to obtain
the total displacement. The ELM and RS-SVR sub-models are applied to
predict the deformation of Baishuihe landslide in the Three Gorges
Reservoir Area (TGRA) as an example. The results showed that the model
effectively improved the accuracy, stability, and scope of application of
landslide displacement prediction, thus providing a new method for land-
slide displacement prediction.

1. Introduction
Landslide displacement prediction is presently a key problem in international geological hazard research
and is one of the most effective means for realizing systematic forecasting of landslide hazards . Many
theories on landslide displacement prediction models have progressed from event-based prediction
methods, empirical models, and mathematical-statistical models to nonlinear models for smart prediction
and models for systematic comprehensive prediction (Bernardie et al., 2015; Van Tien et al., 2018; Lian
et al., 2014; Huang, 2007; Xu et al., 2008). However, owing to complex landslide-formation mechanisms
and landslide-inducing factors, landslide displacement prediction remains a popular but difficult
research topic.
Existing landslide displacement prediction methods mainly predict the future trend of displacement
by developing observational data (such as landslide displacement, reservoir water levels, and precipita-
tion)-based models using mathematical methods (Zhang et al., 2015). Observational data of landslide

CONTACT Rubin Wang rbwang_hhu@foxmail.com Key Laboratory of Ministry of Education for Geomechanics and
Embankment Engineering, Hohai University, Nanjing, China
ß 2020 Informa UK Limited, trading as Taylor & Francis Group
2 R. WANG ET AL.

displacement are relatively independent of those of its influencing factors, and the response relationship
between the two series of data is difficult to describe using conventional analytical methods. Thus, many
researchers investigated temporal variations in cumulative landslide displacement using time series ana-
lysis methods (Zhu et al., 2018; Wen et al., 2017; Xu et al., 2011; Lian et al., 2016). Wen et al. (2017)
decomposed cumulative displacement into two terms, trend and periodic displacements, using a time
series analysis method and then predicted landslide displacement by combining a least-square support
vector machine model and a genetic algorithm. Xu et al. (2011) decomposed landslide displacement into
trend and periodic terms using a double moving average method and then predicted the two terms
using a (1,1) grey model and automatic regression model, respectively. Huang et al. (2017) revealed the
chaos nature and inducing factors of the time series of reservoir landslide displacement and then pre-
dicted reservoir landslide displacement using a multivariate chaos extreme learning machine (ELM) model
. However, the nonlinear time series based modeling reviewed above involved apparent subjectivity in
neural network parameterization and landslide displacement decomposition and extraction, which
severely compromised landslide displacement prediction accuracy. Thus, the nonlinear models for land-
slide prediction need to be further improved.
To resolve the above problems, the present study developed a model for predicting the displacement
of hydrodynamic landslides by combining an ELM sub-model and a random-search support vector regres-
sion (RS-SVR) sub-model. The combined model is then tested by applying it to the Baishuihe landslide in
the Three Gorges Reservoir Area (TGRA) as a case study. Particularly, the accumulative displacement of
the landslide observed using global positioning system (GPS) is decomposed into trend and periodic
terms using a time series model. With appropriate influencing factors identified, the trend and periodic
terms of the potential displacement of the landslide are then predicted using the ELM and RS-SVR sub-
models, respectively. The predictions for the two terms are then stacked to obtain the total displacement
of the landslide.

2. Methodology
2.1. Time series theory
According to previous studies on time series additive models (Xu et al., 2011), accumulative displacement
of landslide can be expressed as follows:
X ðtÞ ¼ ;ðtÞ þ gðtÞ þ eðtÞ (1)
where t is the discrete time; X(t) is the displacement time series; and ;ðtÞ, gðtÞ, and eðtÞ are the functions
for the trend term, periodic term, and random term of the displacement, respectively.
The accumulative displacement of landslide is mainly caused by the combined action of its internal
geological factors (such as lithological properties, engineering activities, geological structure, and topo-
graphic features, etc.) and external inducing factors (such as rainfall, reservoir water level, etc.) (Wang &
Han, 2014). Landslide displacement caused by geological factors, or the trend term of the total displace-
ment, can be approximated as a function that monotonically increases with time and reflects the overall
trend of accumulative displacement. The displacement caused by inducing factors, or the periodic term
of total displacement, can be approximated as a periodic function (Zhu et al., 2018; Wen et al., 2017; Lian
et al., 2016). Owing to limitations in the monitoring technique, it is difficult to obtain effective observa-
tional data of landslide displacement caused by random factors (such as wind and vehicle loads), or the
random term of total displacement. Thus, the random term is not considered in landslide displacement
prediction studies (Lian et al., 2014). In conclusion, Equation (1) can be simplified as follows:
X ðtÞ ¼ ;ðtÞ þ gðtÞ (2)

2.2. Extreme learning machine model


ELM is a type of single-hidden-layer feedforward neural network (SLFN) (Huang et al., 2011). Basic ELM
algorithm consists of three layers: the input, hidden, and output layers, as shown in Figure 1.
The weight matrix of the input layer and error of the hidden layer is randomly configured for comput-
ing the output matrix of the hidden layer, which is then used together with the target output to
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 3

Figure 1. Extreme learning machine.

compute the weight matrix of the output layer through least-square linear regression. The learning
model as represented by Equations (3) - (4) consists of a training set (xi, ti,), hidden node output function
G(w,b,x), and the number of hidden nodes L:
fL ðxj Þ ¼ yj , 8j (3)
X
L
 
bi G wi , bi , xj ¼ tj , j ¼ 1, 2, :::, N (4)
i¼1

where xj is the input parameter; yj is the learning objectives; wi is the weight vector connecting the i-th
hidden and input nodes; bi is the bias of the i-th hidden node; bi is the weight vector connecting the i-
th hidden and output nodes.
ELM model is trained through the following three steps. First, the hidden node parameters, ai and bi,
are configured randomly. Second, the output matrix of the hidden layer, H, is computed using Equation
(5). Finally, the output weight b is obtained using the following equation: b ¼ HþT, where Hþ is the
Moore–Penrose generalised inverse matrix of the hidden layer output matrix, H, and can be obtained
through orthogonal projection or singular value decomposition:
2 3
Gðw1 , b1 , x1 Þ    GðwL , bL , x1 Þ
6 .. .. .. 7
Hðw1 , :::, wL ; b1 , :::, bL ; x1 , :::, xN Þ ¼ 4 . . . 5 (5)
Gðw1 , b1 , xN Þ    GðwL , bL , xN Þ NL

For learning and predicting of the data samples, the curve fitting of ELM can be optimized by adjust-
ing the regularization parameter and the number of hidden layer nodes. Although highly controversial,
ELMs are faster than conventional SLFNs but are comparable in terms of learning accuracy. This is the
biggest advantage of ELMs over conventional SLFNs.

2.3. Support vector regression model


The support vector regression (SVR) algorithm is proposed by Vapnik (2000) and has been widely applied
for resolving nonlinear problems. In an SVR model, the sample dataset is divided into two parts: one for
training, and one for testing. Pre-selected input vectors (training samples) are then mapped into a high-
dimensional feature space. Fitting is then optimized in the space of an optimal decision function model.
Notably, training samples are also used for validating and analyzing the modeling results.
Here, fxj, yjg is the characteristic vector of sample data, with xj being an influencing factor of yj and xj
¼ fxj1, xj2, … , xjpg; p is the number of values in yj. Regression estimation using SVM can be expressed as
follows:
4 R. WANG ET AL.

f ðxÞ ¼ W T wðxÞ þ b (6)


where wðxÞ is the function for the nonlinear mapping of sample data into the feature space, WT is the
variable adaptive coefficient of the estimation function, and b is the bias of the estimation function.
Here, WT and b can be obtained by solving the minimum of the following equation:
1 C Xn  
Dðf Þ ¼ jjW jj2 þ j¼1 e j ð j Þ
R y ,f x (7)
2 n
where D(f) is the generalized function for the optimal classification plane, considering the minimum num-
ber of incorrect sample points and maximum classification margin; jjW jj2 is the complexity of the model;
C is the penalty parameter; and Re is the control error function of e. Thus, the optimization problem can
be expressed as follows:
1 Xn
minQðW, nÞ ¼ jjW jj2 þ C nj þ nj (8)
2 j¼1

W T wðxj Þ þ b  yj  e þ nj (9)
yj  W wðxj Þ  b  e þ
T
nj (10)
nj  0, nj  0ðj ¼ 1, 2,    , nÞ (11)

where nj and nj are relaxation factors.


Designating the partial derivatives of W, b, and nj as equal to 0, based on the Lagrange equation and
duality theory, the dual optimization problem can be expressed as follows:
1X n
LðW, a, b, e, yÞ ¼ min ðar ar ÞT Hr, j (12)
2 r, j¼1
  X n
  X n
 
ar ar þ e ar ar þ yr ar ar (13)
r, j¼1 r¼1

X
n
 
ar ar ¼ 0, ð0  ar , ar  CÞ (14)
r¼1

Hr, j ¼ K ðxr , xj Þ ¼ ;ðxr ÞT ;ðxj Þ, ðr ¼ 1, 2,    , nÞ (15)

where K(xr,xj) is the kernel function, which is defined in the present study as the radial basis function
(RBF). Thus, the SVR model can be expressed as follows:
X
n
 
f ðxÞ ¼ aj aj K ðxj , xÞ þ b (16)
j¼1

The model is based on statistical learning theory and has many advantages. Particularly, it requires
only a very small sample for learning, has a simple statistical structure, and performs better than trad-
itional backpropagation neural networks. Thus, the model is appropriate for predicting the displacement
of landslides with only short-period observational data.

2.4. Random search algorithm


For training and classification using SVM, the practice shows that, the selection of kernel parameters,
penalty factor, and other parameters greatly impacts the performance of the classifier (Chapelle et al.,
2002). Parameter search methods commonly used for SVM include particle swarm optimization (PSO),
genetic algorithm (GA), grid search (GS), and random search (RS). GA and PSO are heuristic algorithms
and involve complex operations. Additionally, when used for searching global optima, these two methods
do not traverse all possible solutions and thus tend to yield only local optima. GS traverses all possible
solutions for a given parametric range and has higher classification accuracy. Compared with heuristic
algorithms and GS, RS algorithm is comparable in terms of prediction performance but has lower compu-
tation cost (Bergstra & Bengio, 2012).
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 5

Figure 2. Flowchart of landslide displacement prediction.

An RS algorithm randomly searches the parameters of a prediction model, with all possible parametric
values sampled for each configuration. RS has two advantages over comprehensive search: (1) The
budget independent of the number and possible values of parameters. (2) The efficiency not compro-
mised by the addition of parameters having no impact on performance.

3. Combined ELM-RS-SVR prediction model


The combined ELM-RS-SVR model predicts landslide displacement through the following three steps.
First, the accumulative displacement of landslide is decomposed into two terms, trend term and periodic
term displacements, using a time series model. Subsequently, the trend and periodic terms are fitted and
predicted using the ELM and RS-SVR sub-models, respectively. Finally, the total landslide displacement is
predicted by summing the predicted trend and periodic terms. Figure 2 shows the flowchart of
the model.

4. Case study: Baishuihe landslide


4.1. Data source
The Baishuihe landslide is located at the southern bank of the Zigui County section of the TGRA, and
spreads down toward the Yangtze river in a terraced topography, with the lower and higher ends meas-
uring 70 and 390 m in elevation, respectively. The sliding mass measured approximately 780, 700, and
30 m in length (south–north direction), width (east–west direction), and thickness (by average), respect-
ively, with a total volume of approximately 126  105 m3. The accumulative displacement of the landslide
and precipitation at the location had been monitored using GPS and a rain gauge since June 2003.
Figure 3 illustrates the landslide mass and observation points.
Rainfall and reservoir water level variation are the key factors that induce the deformation of a hydro-
dynamic landslide. Figure 4 shows the time series of the accumulative displacement and reservoir water
level, indicating that the landslide displacement and reservoir water level are strongly correlated.
Particularly, during the period of May–June every year, the landslide displacement increased drastically,
6 R. WANG ET AL.

Figure 3. (a) Location of the study area; (b) Loction of Baishuihe Landslide; (c) Site photos of Baishuihe Landslide; (d) Loction of
GPS ZG118 and XD-01 in Baishuihe Landslide (modified from Zhou et al., 2018).

Figure 4. Relationship curve between cumulative displacement–monthly rainfall-reservoir water level of Baishuihe landslide.

where as the water level of the TGRA decreased from 160 to 145 m. In contrast, during the period of
August–April every year, the landslide displacement curve varied smoothly. Thus, the displacement curve
exhibited an overall trend of stepped development.
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 7

Figure 5. Extracted trend term of displacement.

The GPS observations of the accumulative displacement of the landslide obtained at observation
points ZG118 and XD-01 during the period from January 2009 to December 2018 were selected for sub-
sequent computations; the observations obtained during the period from January 2009 to December
2017 were used for training and those during the period from January to December 2008 were used
for testing.

4.2. Trend displacement prediction


The trend displacement is extracted using the moving average method (Xu et al., 2011) and computed
using the following equations:
Xi ¼ fx1 , x2 , x3 , ::::::, xt , g (17)
xt1 þ xt2 þ    þ xtn
X ðtÞ ¼ (18)
n
where X(t) is the value predicted for the next period, n is the number of periods with moving average
computed, and (xt-1, xt-2, … , xt-n) is the observed displacement from the previous period to the period n
periods earlier. The ELM sub-model then predicted the trend term of future displacement by fitting the
extracted trend displacement.
Considering that the Baishuihe landslide is rainfall-induced, with its existing displacement greatly
impacted by rainfall, the number of periods is defined as n ¼ 12. The accumulative displacement at the
observation points ZG118 and XD-01 is decomposed to obtain the trend term using a simple moving
average method, as shown in Figure 5. The extracted trend term is then used to obtain the non-trend
terms (or the periodic term of displacement, as the random term is not considered) (Lian et al., 2016).
The trend term is simulated using a least-square cubic polynomial as follows (Zhang et al., 2015):
eðtÞ ¼ at3 þ bt2 þ ct þ d (19)
Table 1 shows the coefficients of the trend term yielded by the polynomial simulation. The goodness
of fit, R2, is then computed, thereby obtaining the optimal fitting curve of the least-square cubic
8 R. WANG ET AL.

Table 1. Parameters of trend displacement yielded by polynomial simulation.


Parameter
Observation point a b c d Goodness of fit (R2)
ZG118 0.0017 0.2685 22.7709 1559.6308 0.9909
XD-01 0.0015 0.2381 30.2495 1901.9871 0.9912

Table 2. Parameters of trend displacement yielded by the ELM sub-model.


Parameter
Observation point C L Goodness of fit (R2)
ZG118 1e16 500 0.9991
XD-01 1e19 600 0.9990

Figure 6. Fitting curves yielded by the polynomial model and the ELM sub-model.

polynomial. However, this method had shortcomings. Particularly, similar to other fitting curves yielded
by conventional least square methods, the sum of squares of errors can be minimized, but the fitting
error at a few individual points may be very large, much larger than that at other points. This is evidently
unacceptable. Thus, the trend term of the total displacement is simulated using the ELM sub-model.
Table 2 shows the parameters of the trend term yielded by the ELM. Figure 6 compares the fitting curves
yielded by the polynomial model and ELM sub-models.
As shown in Tables 1 and 2 and Figure 6, the ELM sub-model is markedly better than the polynomial
model in terms of goodness of fit, thereby confirming that it is appropriate to simulate the trend term of
observed displacement using the ELM sub-model.

4.3. Prediction of periodic displacement

4.3.1. Extraction of periodic term displacement


According to the time series additive model, with the trend term of displacement removed, only the peri-
odic term remains in the accumulative displacement. Figure 7 shows the extracted periodic displacement
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 9

Figure 7. Extracted periodic term displacement.

Table 3. Degrees of correlation between periodic displacement at observation points ZG118 and XD-01 and the two major
influencing factors.
Variation of Variation of
reservoir water reservoir water
Rainfall in the Rainfall in the level in the level in the
previous previous Current reservoir previous previous
Monitoring point one month two months water level one month two months
ZG118 0.78 0.78 0.83 0.84 0.83
XD-01 0.79 0.79 0.83 0.84 0.82

at the two selected observation points, revealing that the periodic displacement fluctuated periodically
with the reservoir water level and rainfall.

4.3.2. Selection of factors influencing periodic term displacement


The selection of factors influencing displacement directly impacts the learning capability of the model.
An analysis of the observed displacement of the Baishuihe landslide, reservoir water level, and precipita-
tion revealed that the displacement of the landslide had been closely related to the seasonal fluctuation
in precipitation and reservoir water level and had been gradually increasing with time. Thus, precipitation
and reservoir water level are considered as the two major factors influencing the periodic displacement
of the landslide.
Rainfall seepage into landslides is a slow process. According to the relationship between precipitation
and landslides in the TGRA (Du et al., 2013), the accumulative precipitation in the past one month and
that in the past two months are used to assess the impact of precipitation on the periodic displacement
of the Baishuihe landslide.
The displacement of the reservoir landslides is closely related to the periodic fluctuation in reservoir
water level. According to the relationship between reservoir water level fluctuation and landslides (Jiao
et al., 2014; Yang et al., 2017), the current reservoir water level, reservoir water level variation in the past
one month, and that in the past two months are used to assess the impact of reservoir water level vari-
ation on the periodic displacement of the Baishuihe landslide.
The correlation between the periodic displacement at the observation points ZG118 and XD-01 and
the two influencing factors, ck, is obtained using grey correlation analysis (Table 3). All the yielded ck
10 R. WANG ET AL.

Table 4. Parametric configurations of periodic displacement yielded by the RS-SVR sub-model.


Parameter
Observation point c g Goodness of fit (R2) Root mean square error
ZG118 6024.25 0.2750 0.9998 0.1993
XD-01 7530.76 0.7245 0.9999 0.0260

Table 5. Parametric configurations yielded by the ELM sub-model.


Parameter
Observation point c L Goodness of fit (R2) Root mean square error
ZG118 1e12 500 0.9421 11.5135
XD-01 1e14 500 0.9089 21.9646

Figure 8. Predictions for periodic displacement at observation point XD-01 yielded by the two sub-models.

values are larger than 0.6, thus confirming that the correlations are close and the influencing factors are
reasonably defined (Yang et al., 2004).

4.3.3. Comparison of predictions yielded by different models


After training using the sample data, they are simulated using the RS-SVR and ELM sub-models. The
simulation results are then compared. Note that the parametric configuration of the RS-SVR sub-model
required a pretreatment of the sample data of the influencing factors and periodic displacement, i.e., con-
version into the [-1,1] format. The RS parameters are configured as follows: penalty factor c ¼
[6000,8000]; kernel function parameter g ¼ [0.1,1].
Periodic displacement is simulated and predicted separately using the RS-SVR and ELM sub-models.
Tables 4 and 5 show the final parameter configurations yielded by the two sub-models, respectively.
Figures 8 and 9 show the displacement predictions yielded by the two sub-models, respectively.
As shown in Tables 4 and 5 and Figures 8 and 9, the periodic displacement yielded by the RS-SVR
sub-model better fit the observations than that yielded by the ELM sub-model in terms of goodness of
fit (R2), root mean square error (RMSE), and image. From its principle and realization conditions, ELM sac-
rifices too much useful data for speed. Additionally, it relies entirely on an increased number of hidden
nodes to improve learning accuracy. Thus, ELM is highly risky, particularly in applications of data with
non-simple structures.
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 11

Figure 9. Predictions for periodic displacement at observation point ZG118 yielded by the two sub-models.

Figure 10. Comparison between the measured and predicted cumulative displacements of landslide.

4.4. Prediction of accumulative displacement


According to the time series additive model, the total displacement is predicted by summing the trend
and periodic terms of the displacement predicted using the ELM and RS-SVR sub-models, respectively, as
shown in Figure 10. The results showed that the displacement simulated by the combined ELM-RS-SVR
model is consistent with the observed displacement in terms of temporal variation and overall trend. The
combined model also effectively predicted the displacement of the landslide in the future one-year
period, thus confirming the good performance of the model in displacement prediction.

5. Conclusion
Landslide displacement prediction is an important means for realizing systematic landslide hazard fore-
casting and early warning system. In this paper, a landslide displacement prediction model by combing
12 R. WANG ET AL.

ELM and RS-SVR sub-models is proposed, which is showed better prediction accuracy than the models
using either the ELM or SVR technique. Taking the Baishuihe landslide in TGRA, for a case study, the
Landslide displacement sequence is decomposed into two terms, trend and periodic displacements by
using the time series additive model, and predicted the two terms using the ELM and RS-SVR sub-mod-
els, respectively. The results show that the trend displacement is simulated separately using the polyno-
mial model and ELM sub-model, confirming higher simulation accuracy of the ELM, and the periodic
displacement is simulated separately using the ELM and RS-SVR sub-models, confirming the greater
accuracy of the second sub-model in simulating periodic displacement. The accumulative displacement
simulated and predicted by the combined model demonstrated that the combination of the two sub-
models effectively increased the accuracy, stability, and scope of application of landslide displacement
prediction, providing a new method for landslide displacement prediction.
However, to better investigate landslide displacement in relation to precipitation and reservoir water
level, it is necessary to establish large databases of more accurate real-time observational data. Presently,
landslide hazards are not completely under control, particularly for landslides under the combined action
of strong precipitation and reservoir water level fluctuation. Thus, it is particularly important to monitor
large landslides in hydropower reservoir areas using state-of-the-art monitoring devices.

Acknowledgements
We thank the National Field Observation and Research Station of Landslides in the TGRA of Yangtze
River for their help in providing monitoring data for this study.

Funding
This work is supported by the National Key Research and Development Program of China (No.
2017YFC1501100), the Open Foundation of National Field Observation and Research Station of Landslides
in the TGRA of Yangtze River, China Three Gorges University (2018KTL03), the Fundamental Research
Funds for the Central Universities (No. 2019B13814), the Open Foundation of Key Laboratory of Ministry
of Education for Geomechanics and Embankment Engineering, Hohai University (No. GHXN201905).

References
Bergstra, J., & Bengio, Y. (2012). Random search for hyper-parameter optimization. Journal of Machine
Learning Research, 13(1), 281–305.
Bernardie, S., Desramaut, N., Malet, J. P., Gourlay, M., & Grandjean, G. (2015). Prediction of changes in
landslide rates induced by rainfall. Landslides, 12(3), 481–494. doi:10.1007/s10346-014-0495-8
Chapelle, O., Vapnik, V., Bousquet, O., & Mukherjee, S. (2002). Choosing multiple parameters for support
vector machines. Machine Learning, 46(1/3), 131–159. doi:10.1023/A:1012450327387
Du, J., Yin, K., & Lacasse, S. (2013). Displacement prediction in colluvial landslides, three gorges reservoir,
china. Landslides, 10(2), 203–218.
Huang, R. (2007). Large-scale landslides and their sliding mechanisms in china since the 20th century.
Chinese Journal of Rock Mechanics and Engineering, 26(3), 433–454.
Huang, F., Huang, J., Jiang, S., & Zhou, C. (2017). Landslide displacement prediction based on multivariate
chaotic model and extreme learning machine. Engineering Geology, 218, 173–186. doi:10.1016/j.enggeo.
2017.01.016
Huang, G. B., Wang, D. H., & Lan, Y. (2011). Extreme learning machines: a survey. International Journal of
Machine Learning and Cybernetics, 2(2), 107–122. doi:10.1007/s13042-011-0019-y
Jiao, Y. Y., Zhang, H. Q., Tang, H. M., Zhang, X. L., Adoko, A. C., & Tian, H. N. (2014). Simulating the process
of reservoir-impoundment-induced landslide using the extended dda method. Engineering Geology,
182, 37–48. doi:10.1016/j.enggeo.2014.08.016
Lian, C., Chen, C. L. P., Zeng, Z., Yao, W., & Tang, H. (2016). Prediction intervals for landslide displacement
based on switched neural networks. IEEE Transactions on Reliability, 65(3), 1483–1413. doi:10.1109/TR.
2016.2570540
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 13

Lian, C., Zeng, Z., Yao, W., & Tang, H. (2014). Extreme learning machine for the displacement prediction of
landslide under rainfall and reservoir level. Stochastic Environmental Research and Risk Assessment,
28(8), 1957–1972. doi:10.1007/s00477-014-0875-6
Van Tien, P., Sassa, K., Takara, K., Fukuoka, H., Dang, K., Shibasaki, T., Ha, N. D., Setiawan, H., & Loi, D. H.
(2018). Formation process of two massive dams following rainfall-induced deep-seated rapid landslide
failures in the Kii peninsula of Japan. Landslides, 15(9), 1761–1778. doi:10.1007/s10346-018-0988-y
Vapnik, V. (2000). The nature of statistical learning theory (pp. 138–167). Springer-Verlag.
Wang, X., & Han, M. (2014). Online sequential extreme learning machine with kernels for nonstationary
time series prediction. Neurocomputing, 145, 90–97. doi:10.1016/j.neucom.2014.05.068
Wen, T., Tang, H., Wang, Y., Lin, C., & Xiong, C. (2017). Landslide displacement prediction using the GA-
LSSVM model and time series analysis: a case study of three Gorges reservoir, China. Natural Hazards &
Earth System Sciences, 17(12), 1–20.
Xu, Q., Tang, M., Xu, K., & Huang, X. (2008). Research on space-time evolution laws and early warning-pre-
diction of landslides. Chinese Journal of Rock Mechanics and Engineering, 27(6), 1104–1112.
Xu, F., Wang, Y., Du, J., & Ye, J. (2011). Study of displacement prediction model of landslide based on
time series analysis. Chinese Journal of Rock Mechanics and Engineering, 30(4), 746–751.
Yang, W., Kun-Long, Y., & Guang-Feng, A. N. (2004). Grey correlation analysis of sensitive factors of land-
slide. Rock and Soil Mechanics, 25(1), 91–93.
Yang, B., Yin, K., Xiao, T., Chen, L., & Du, J. (2017). Annual variation of landslide stability under the effect
of water level fluctuation and rainfall in the three Gorges reservoir, China. Environmental Earth Sciences,
76(16), 564. doi:10.1007/s12665-017-6898-9
Zhang, J., Yin, K., Wang, J., & Huang, F. (2015). Displacement prediction of baishuihe landslide based on
time series and PSO-SVR model. Chinese Journal of Rock Mechanics and Engineering, 34(2), 382–391.
Zhou, C., Yin, K., Cao, Y., Intrieri, E., Ahmed, B., & Catani, F. (2018). Displacement prediction of step-like
landslide by applying a novel kernel extreme learning machine method. Landslides, 15(11), 2211–2225.
doi:10.1007/s10346-018-1022-0
Zhu, X., Ma, S. Q., Xu, Q., & Liu, W. D. (2018). A WD-GA-LSSVM model for rainfall-triggered landslide dis-
placement prediction. Journal of Mountain Science, 15(1), 156–166. doi:10.1007/s11629-016-4245-3

You might also like