Você está na página 1de 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/257993804

Comparison on landslide nonlinear displacement analysis and prediction with


computational intelligence approaches

Article  in  Landslides · October 2013


DOI: 10.1007/s10346-013-0443-z

CITATIONS READS

31 136

5 authors, including:

Zaobao Liu Jianfu Shao


University of Lille Nord de France Université des Sciences et Technologies de Lille 1
39 PUBLICATIONS   307 CITATIONS    527 PUBLICATIONS   5,415 CITATIONS   

SEE PROFILE SEE PROFILE

Hongjie Chen Chong Shi


Huaneng Lancang River Hydropower Inc. Hohai University
8 PUBLICATIONS   94 CITATIONS    26 PUBLICATIONS   152 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Time-dependent behaviors of argilaceous rocks View project

Peridynamics View project

All content following this page was uploaded by Zaobao Liu on 20 June 2014.

The user has requested enhancement of the downloaded file.


Original Paper

Landslides Zaobao Liu I Jianfu Shao I Weiya Xu I Hongjie Chen I Chong Shi
DOI 10.1007/s10346-013-0443-z
Received: 11 February 2013
Accepted: 8 October 2013 Comparison on landslide nonlinear displacement
© Springer-Verlag Berlin Heidelberg 2013
analysis and prediction with computational intelligence
approaches

Abstract Landslide displacement is widely obtained to discover displacement analysis, which depends on the mechanical soil
landslide behaviors for purpose of event forecasting. This article properties, and pore pressure (Romeo 2000). Landslide displacement
aims to present a comparative study on landslide nonlinear was observed by monitoring groundwater composition for the
displacement analysis and prediction using computational purpose of estimating future landslide occurrence (Sakai 2001). Some
intelligence techniques. Three state-of-art techniques, the support regression models were obtained from estimating the coseismal
vector machine (SVM), the relevance vector machine (RVM), and landslide displacement by predicting the Newmark displacement
the Gaussian process (GP), are comparatively presented briefly for (Newmark 1965) in terms of the critical acceleration ratio, critical
modeling landslide displacement series. The three techniques are acceleration ratio and earthquake magnitude, arias intensity and
discussed comparatively for both fitting and predicting the critical acceleration, and arias intensity and critical acceleration ratio
landslide displacement series. Two landslides, the Baishuihe (Jibson 2007). The inverse velocity method derived from
colluvial landslide in China Three Georges and the Super-Sauze displacement curves was successfully applied to predict large range
mudslide in the French Alps, are illustrated. The results prove that slope failures in open pit mines by displacement monitoring and
the computational intelligence approaches are feasible and capable analysis (Rose and Hungr 2007). A typical pattern of landslide
of fitting and predicting landslide nonlinear displacement. The displacement was identified for shallow landslides, debris produced
Gaussian process, on the whole, performs better than the support by the excavation and gabions, metallic walls, anchored bulkheads
vector machine, relevance vector machine, and simple artificial distribution, and slope geometry (Bozzano et al. 2011). These works
neural network (ANN) with optimized parameter values in discussed the LDA with various field-monitored factors such as the
predictive analysis of the landslide displacement. seismic features, material properties, and pore water presence.
Besides the above achievements, slope or landslide displacements
Keywords Landslide . Displacement prediction . Nonlinear . are also analyzed and predicted with some time series techniques,
Computational intelligence . Relevance vector machine . Gaussian including the Verhulst model (Li et al. 1996), exponential smoothing
process model (Liu et al. 2009a), and grey model (Chen and Wang 1988; Liu et
al. 2009b). Comparisons of these models (Yang and Liu 2005; Yi 2007)
Introduction have been discussed and lead to the fact that these models all have
Landslide displacement is the typical signature of its underlying some limitations. Each of the models has a definite form of equation
complex evolution process. Landslide displacement analysis which indicates that the model is mainly valid for landslide
(LDA) is a critical issue of an early warning system and can displacements with similar features. In fact, each of the former types
contribute to prevent suffering property damage and loss of of models involves an exponential function with varying parameters
human lives. It provides a potential physically based approach to be adjusted by observations. These limitations make it necessary to
for landslide prediction. Landslide behavior is generally continue the research work on proper predictive models. Recently,
monitored by various instruments, such as extensometers, some computational intelligence techniques such as the artificial
inclinometers, and clinometers. The most reliable parameter neural network (ANN) (Mayoraza and Vullietb 2002; Chen and Zeng
obtained by these instruments is the landslide displacement. 2012; Lv and Liu 2012; Du et al. 2013), support vector machines (SVMs)
Hence, LDA is widely used in landslide forecasting, especially in (Feng et al. 2004; Dong et al. 2007; Zhu and Hu 2013), and the Gaussian
the short-term forecasting of landslides with creep displacement. process (GP) (Liu et al. 2012) have been successfully applied for
Slope or landslide displacement analysis and prediction are analysis and prediction of landslide displacement and some related
valuable and important in predicting slope failures. Based on subjects (Li et al. 2012; Grelle and Guadagno 2012; Liu et al. 2013; Belle
displacement measurements in the secondary creep range, Saito et al. 2013; Lian et al. 2013). These works showed the potential ability of
(1965) proposed an empirical model to predict slope failure time by the corresponding methods to analyze the engineering problems such
analysis of the constant strain rate from the relative displacement as the landslide displacement prediction.
curve. After that, Saito (1969) extended his theory for impending However, there is almost no work made in view of comparing
forecasting by formulating the relationship between the time left these computational intelligence methods for LDA. Such a
before failure and the displacement. Fukuzono (1985) presented a comparison could be useful to evaluate the performance and
new method for predicting slope failure time using the inverse of validity domain of each method. The present study aims to give
surface displacement velocity based on large-scale laboratory a comparative study on the computational intelligence approaches
experiments. Based on these achievements, slope displacement including the Gaussian process (Rasmussen and Williams 2006),
analysis and prediction have been approached with many other support vector machines (Cortes and Vapnik 1995; Vapnik 1998),
techniques. A predictive model was presented for predicting and relevance vector machines (RVMs) (Tipping 2000, 2001) for
seismically induced landslide displacements. Seismic slope stability predictive analysis of landslide displacement. The three techniques
was measured in terms of critical acceleration derived from are typical computational intelligence approaches widely used in

Landslides
Original Paper
engineering applications. They are derived from three different approaches. As given below, different approaches make the
points of views as described below in detail. In order to perform mapping f in varying ways.
this comparative study, two representative landslides, one
mudslide in France and one colluvial landslide in China, are Support vector machines
illustrated. The three computational intelligence methods The SVM is based on Mercer's condition. Given the training
mentioned above will be applied to predict in situ displacements samples D=(X,S)=(xi, si)in=1, n is the sample number. The SVMs
in these two slopes. The results obtained from the three methods make mapping of the samples with a linear regression function
will be analyzed and compared.
sðxÞ ¼ ðw⋅ϕðxÞÞ þ b ð2Þ
Landslide displacement prediction formulation and methods

Mathematical formulation where w is the weight vector, b is bias, and ϕ(x) is the
Landslide displacement obtained by instruments in field nonlinear mapping from the input space to the output space.
monitoring is usually a series of scatter numbers which can be In addition to performing linear mapping, SVMs can
treated as time series data. This kind of data can be easily denoted efficiently perform nonlinear mapping using the kernel trick,
mathematically in the following way. implicitly mapping their inputs into high-dimensional feature
Let si (i=0, 1, 2,…, n) be the ith observed value at time xi, spaces.
and then, the landslide displacement domain can be written Suppose all the observed displacement samples can be mapped
as S=(s 0 , s 1 ,…, s i ,…, s n ) versus the corresponding time well with a linear function with precision ε. The purpose of SVM
domain which can be written as X=(x0, x1,…, xi,…, xn). The training is to minimize the following equation:
purpose of the LDA is to find an appropriate mapping f
!
between S and X, i.e., 1  T X n
min ϕðw; ξi ; ξi  Þ ¼ min ww þ C ξi ð3Þ
2
S ¼ f ðX Þ ð1Þ i¼1

Then, prediction of LDA can be executed by mapping with new where the constant C>0 is the penalty factor denoting the
input xm (m=n+1, n+2,…) in order to produce the unknown punishing level of the samples with errors over ε, ξi, and ξi* which
displacement sm where n is the number of training samples. are the nonnegative slack variables.
The accuracy of the predicted output sm depends completely on Therefore, using the Lagrangian, this optimization problem can
the ability of the mapping f and quality of data. In practical be converted into a dual form which is a quadratic programming
analysis, we think that the observed data are in good quality for problem where the objective function is solely dependent on a set
analyses. In this way, the target is to find the optimal mapping(s), of Lagrange multipliers αi; the problem can be rewritten as (John
which is the potential ability of computational intelligence 1998)

( )
1X X     X X
n n n n
max − ðαi −αi  Þ α j −α j  K xi ; x j −ε ðαi −αi  Þ þ si ðαi −αi  Þ
2 i¼1 j¼1 i¼1 i¼1
Xn
S:t ðαi −αi  Þ ¼ 0 ðαi ; αi  ∈½0; cÞ ð4Þ
i¼1

where n is the number of training sample; K(xi,xj)=ϕ(xi)ϕ(xj) is where ∈i is a zero-mean white noise process with variance σ2, i.e.,
the kernel function. There are many kernel functions, such as the p(∈i| σ2)=Ν(∈i| 0,σ2).
multinomial kernel, sigmoid kernel, and radial basis function Considering noise precision β instead of its variance σ2, i.e.,
(RBF) kernel. posing β=σ2, and assuming the independence of the samples ti, the
In this way, the prediction model of the SVMs can be obtained as likelihood of the complete training data set is

Xn  1 
f ðx Þ ¼ ðαi −αi  ÞK ðxi ; xÞ þ b ð5Þ pðsjX; w; βÞ ¼ ð2πβ−1 Þ−N=2 exp − β ks−Φwk2 ð7Þ
i¼1 2
Relevance vector machines where Φ=[8(x1),8(x2),…,8(xn)]t is a N×(N+1) matrix.
The RVM regression has a similar form as the SVM; it employs Using Bayes rule and the properties of Gaussian functions, the
model (2) with an additive noise term to link the vector input xi posterior distribution of the weight can be also described by the
and scalar target variable si Gaussian as follows:

si ¼ f ðxi ; wÞ þ ∈i ð6Þ pðwjX; s; α; β Þ ¼ N ðwjm; Σ Þ ð8Þ

Landslides
where α=[α0, α1,…, αN]t, a N+1 vector of hyperparameters representing where mi is the ith element of the estimated posterior weight w,
the precision on the parameters; m=β ∑Φ t t , Σ = (A + β Φ t Φ ) −1, with and Σii is the ith diagonal element of the posterior covariance
A=diag (α0,…, αN), a diagonal matrix of precisions. matrix Σ.
Values of α and β can then be obtained iteratively, using the Once the iteration has converged to the “most probable” values
following update rules (Tipping 2001): αMP and βMP, the distribution of the target value t* for a new data
input x* is also Gaussian and can be estimated through (Phillips et
1−αi Σ ii kt−Φmk2 al. 2011)
αnew
i ¼ ; ðβnew Þ−1 ¼ XN ð9Þ
m2i N− i−1 ð1−αi Σ ii Þ

 
pðs* jX; s; αMP ; β MP Þ ¼ ∫pðs* jX; w; β MP ÞpðwjX; s; αMP ; βMP Þdw ¼ N s* jmt ϕðx Þ; σ2* (10)

Gaussian process The result is not a single value, but a probability distribution of
The Gaussian process represents the posterior distribution over predictions. This advantage can be used to obtain prediction
functions based on training data and prior distribution. The intervals describing a degree of confidence of the predictions
Gaussian process f(x) is a collection of random variables, any finite (Liu et al. 2012).
set which has a joint Gaussian distribution (Rasmussen and Williams
2006). Its statistical characteristics are completely specified by its Result and comparison
mean function m(x) and covariance function k(x,x'), i.e., Each of the above techniques has some parameters to be adjusted,
for example, the kernel type and penalty parameter c in the SVM,
the kernel type and width value in the RVM, and the covariance
f ðxÞeGPðmðxÞ; kðx; x0 ÞÞ and length scale in the GP. In practical problems, these methods
mðxÞ ¼ Ε½ f ðxÞ ð11Þ are trained first with the training data to optimize the model
kðx; x0 Þ ¼ Ε½ð f ðxÞ−mðxÞÞ−ð f ðx0 Þ−mðx0 ÞÞ parameters and then to give predictions on the test data set with
the optimized models. The predictions of training data can be
treated as the curve fitting results, while the output of test data
Given the observations and the predictive input x* (also named can be thought as the forecasting. Both the curve fitting and
test input), the goal of the Gaussian process modeling is to obtain forecasting are discussed as follows for the nonlinear
the output y* for the distribution P(y*|X,S,x*). Suppose the prior displacements of two landslides. These above strategies are
distribution of observation target s satisfies y∼N(0,k(x,x')) and the executed in Matlab2012a.
independent noise ε obeys ε~N(0,σn2), thus, the covariance of
noisy observations is obtained
Baishuihe landslide
    The Baishuihe landslide, in the Zigui town on the south of Yangtze
cov sp ; sq ¼ k xp ; xq þ σ2n δpq or covðsÞ ¼ K ðX; X Þ þ σ2n I ð12Þ
River and 56 km to the west of the Three Gorges Dam, occurred in
July 2007. The landslide spreads an area of 4.2×105 km2 with an
where K(X,X) is a positive definite covariance matrix with size estimated volume of 12,600 m3. It has a maximum length of 780 m
n×n, and its elements denote the correlations of different
observation samples. 14

Consequently, the joint distribution of the observed targets and 13


the predictions can be signified as ANN SVM
12
Displacement (×102mm)

RVM GP

    11
 Observed
y K ðX; X Þ þ σ2n K ðX; X  Þ
N 0; ð13Þ 10
f e K ðX  ; X Þ K ðX  ; X  Þ
9

8
For notation simplicity, K=K (X, X), K*= K(X, X*); the regression
equation of GP for noisy observed target is obtained 7

6
Dec-06
Oct-06

Jan-07

May-07

Jul-07
Nov-06

Nov-06

Apr-07

Apr-07

Jun-07
Sep-06

Feb-07

Mar-07


  

f 
X; y; X  eN  f  ; cov f 
h
i −1
Time

ð14Þ
f  ¼ Ε f 
X; y; X  ¼ K  T K þ σ2n I y
  −1 Fig. 1 Observed displacements and fitted values of different methods for point
cov f  ¼ K ðX  ; X  Þ−K  T K þ σ2n I K  ZG93 in the Baishuihe landslide

Landslides
Original Paper
and the width of 700 m. It extends from an elevation of 75 m up to Prediction comparison
390 m and displays a superficial cracking and distinct ground One significant purpose of the LDA is to predict the landslide
displacement. The landslide is a typical retrogressive landslide behaviors according to the obtained data. Thus, it is more
(Du et al. 2013). The displacement of the point ZG93 measured interesting of the prediction than the fitting of landslide
by GPS is shown in Fig. 1. It can be observed apparently that the displacement with practical considerations. Hence, the landslide
displacement of the Baishuihe landslide shows a typical nonlinear displacement is analyzed in the predictive point of view so as to
characteristic. Especially in the imminent period (3 months before show the predictive performance of the mentioned strategies.
occurrence), the landslide displacement increases intensively, up The cross-validation technique has been applied to utilize
to 200 mm/month. The above mentioned computational substantially the few data points since only 11 months of data are
techniques are applied to give fitting values and predictions of obtained before the landslide occurrence. The leave-one-out
the observed displacement series. method (Rasmussen and Williams 2006) is used in the modeling,
i.e., each time using d(1<d<n) samples with d−1 samples for
Displacement fitting comparison training and the dth sample for testing. In this analysis, d=5 after
Firstly, in order to show the outstanding performance of the pilot calculations.
discussed approaches, the displacement series are analyzed as In the SVM analysis, the RBF kernel is used with optimal
the curve fitting problem as in the previous work (Du et al. 2013) penalty parameter c=500; in the RVM analysis, the Gaussian kernel
where the BP ANN is used. The results of all the approaches are is used with the optimal width value of 3.10; in the GP analysis, the
obtained as shown in Fig. 1. In the SVM modeling, the RBF kernel square exponential covariance is used with the optimal length
is used with the penalty parameter c=500. In the RVM modeling, scale of 3.0. These parameters are adjusted with pilot calculations.
the Gaussian kernel is used with width value of 1.10. In the GP The predicted results are given in Fig. 3. We can find in Figs. 1 and
modeling, the square exponential covariance is used with the 3 that these models perform not so well in this analysis as those in
length scale of 2.0. As it is shown in Fig. 1, the fitting results of the former displacement fitting. Especially in the prediction of the
SVM, RVM, and GP are very close to the corresponding observed catastrophe points, none of the methods can generate excellent
displacement values, even at the catastrophe points of the results as that in fitting problems. However, this is reasonable
displacement series. On the contrary, either the RVM or the GP because we are making predictions here.
produces an excellent fitting at the last observation. However, the Comparing the predictive results of the three methods, we can
SVM performs still much better than the ANN. Thus, the three conclude that the GP performs best, and the RVM performs better
methods perform much better than the ANN in fitting the than the SVM. The predictive percentage errors in Fig. 4 have
landslide displacement. shown this more obviously. During the period of September 2006
The model performance of each method is shown in Fig. 2 and April 2007, the RVM produces the smallest predictive
quantitatively with regression R2 value of the fitted values versus percentage error values, i.e., no more than 2 %. The SVM also
monitored displacements. It is shown apparently that the SVM, produces values of percentage errors less than 2 % before April
RVM, and GP perform much better than the ANN in fitting the 2007 compared with those of GP which are less than 4 %. Thus, the
landslide displacement given in the literature. Also, as it is shown RVM performs much better than the SVM and GP at the steady
in Fig. 2, the GP performs best and the RVM performs better than points of the displacement curve. However, the predictive error
the SVM. Hence, the GP shows slightly superior ability than the bars of the GP are much smaller than those of the SVM and RVM
RVM and SVM in fitting the landslide displacement. at the catastrophe points as shown in Fig. 4. This ability is more
useful for the purpose of landslide imminent forecasting.
In many landslide forecasting models, the displacement rate is
one of the widely used variables. As we can see in Fig. 3, the GP
13
ANN SVM
prediction curve and the RVM prediction curve follow well the
GP RVM trends of the observed displacement curve at the imminent period
12 Linear (ANN) Linear (SVM) of the landslide, which should imply that they have similar
Linear (GP) Linear (RVM)
Predicted Displacement((×102mm)

11 14
GP: R² = 0.9998
ANN: R² = 0.9415 13 Observed
10
Displacement((×102mm)

12 SVM
SVM: R² = 0.9976 11 RVM
9 GP
10

9
8 RVM: R² = 0.999
8

7 7

6
Nov-06
Oct-06

Nov-06

Jan-07
Dec-06

Jul-07
Apr-07
Mar-07

Apr-07

Jun-07
May-07
Sep-06

Feb-07

6
6 7 8 9 10 11 12 13
Monitored displacement(×102mm) Monitoring Time

Fig. 2 Fitting performance of SVM, RVM, and GP on Baishuihe landslide displacements Fig. 3 Predictive performance of SVM, RVM, and GP on Baishuihe landslide displacement

Landslides
30
extensometer located close to the infiltration area of the mudslide,
25
is shown in Fig. 5. The displacement series have shown the key
features of the mudslide before its occurrence. As it is shown in
SVM
Fig. 5, the monitored mudslide displacements show a typical
Percentage Error(%)

20 RVM
GP creeping characteristic. The displacements of the mudslide
15 increase progressively with a growing displacement rate as the
monitoring time lasts. We apply the SVM, RVM, and GP methods
10 for predicting the landslide displacement series.

5 Displacement fitting comparison


We first apply the GP, SVM, and RVM to fit the observed
0
displacements of the Super-Sauze mudslide. The results are shown
Jan-07

Jul-07
Apr-07
Feb-07

Mar-07

Jun-07
May-07
in Fig. 5. Besides the mentioned methods, the poly regression is
Monitoring Time also applied to give an overall fitting of the displacement. As can
be seen from Fig. 5, all these methods perform very well in fitting
Fig. 4 Percentage errors of the SVM, RVM, and GP on the prediction of Baishuihe the observed displacements of the Super-Sauze landslides. They
landslide displacement can fit the tendency of the displacement curve satisfactorily.
Especially the GP, it almost obtains perfect fitting results in both
displacement rate values. This ability is particularly useful if we tendency fitting and point fitting.
apply these methods together with those landslide forecasting During the fittings of the poly regression methods, we have
models above. We can use these methods to give predictions of found that the six-order poly regression equation performs much
the displacement before the next observation can be obtained. better than the other order (e.g., two-, three-, four-, or five-order)
More practically, supposing we are in June 2007, we make poly regression equations. This phenomenon, on one hand, shows
prediction of the displacement in July 2007 with the GP or RVM; that the landslide displacements should be fitted by the
thus, we can get the same displacement rate in advance to forecast polynomials with six orders. On the other hand, it indicates that
the landslide occurrence. the fitting of the displacement needs to pilot calculations in order
Hence, in short, the GP performs best. The RVM shows better to obtain good fitting performance with the polynomial functions.
ability than the SVM in the LDA of the Baishuihe landslide. This Moreover, the fitting equations of the polynomial regression are
phenomenon is stated above not only in fitting the displacement varying from one displacement series to another. This flaw
curve but also in predicting and tracking the trends of the restricts the applicability of the polynomial regression techniques.
monitored displacement curve of the Baishuihe landslide.
Prediction comparison
Super-Sauze mudslide We again utilize the LOO cross-validation technique in the analysis
The Super-Sauze mudslide, consisted of a silty sand matrix mixed so as to compromise between the predictive performance and the
with moraine debris in the southern French Alps, is a flow-like predicted step. The length of the subset is given by d=11. In the
landslide characterized by a complex vertical structure with a slip SVM analysis, the RBF kernel is used with optimal penalty
surface and a viscoplastic plug. The landslide displacements were parameter c=10; in the RVM analysis, the Gaussian kernel is used
monitored by different techniques, like the benchmark technique, with optimal width value of 5.96; in the GP analysis, the square
extensometers, inclinometers, and terrestrial laser scanning during exponential covariance is used with length-scale value of 1.0. These
5 days from July 10 to 14, 2007 with one scan acquisition per day parameters are adjusted with pilot calculations.
(Travelletti et al. 2008). The displacement, obtained using an The predictive errors are also shown in Fig. 6 for each method
(the SVM, RVM, and GP). We can observe that the percentage
140
Monitoring date(July-2007)
135 9 10 11 12 13 14 15 16 17 18 19 20 21 22
GP Fitting SVM Fitting 3.5 140
RVM Fitting Monitored displacement
130 SVM Error 135
6-PolyRegression 3
RVM Error
Displacement (cm)

125 130
Predictive pencentage error (%)

Monitored displacement(cm)

y = -9E-06x6 + 0.0014x5 - 0.0747x4 + 1.851x3 - 23.852x2 2.5 GP Error


120 125
+ 154.43x - 305.42 R² = 0.9958 Monitored displacement
115 120
2
115
110
1.5
110
105
1 105
100
100
95 0.5
95
90
9 11 13 15 17 19 21 23 0 90
9.4 10.1 11.3 12.6 13.4 14.5 15.4 16.7 17.7 18.7 19.5 20.3 21.1
Monitoring date(July-2007) Monitoring day (July-2007)

Fig. 5 Observed displacement and fitting values of different methods (Super- Fig. 6 Monitored displacement and predictive errors of different approaches
Sauze mudslide) (Super-Sauze mudslide)

Landslides
Original Paper
errors of all the three methods are no more than 3.5 %, which 1 Average percentage error (%) Average absolute error (cm)
indicates that the three methods can perform well in predictive
modeling of the Super-Sauze mudslide displacement. Also, all the 0.8
three techniques have big values of predictive errors in the
fluctuating part of the displacement curve. For example, the 0.6
predictive percentage errors between 18 July and 19 July are much
bigger than those of the other periods in Fig. 6. This characteristic 0.4
shows that these techniques perform better in predicting the
0.2
Super-Sauze mudslide displacement for the steady part than that
for the fluctuating part.
0
The predictive performances of the SVM, RVM, and GP are GP SVM RVM
shown in Fig. 7. The R2 values of the regression results are given in
Fig. 7 for the corresponding predicted values by the SVM, RVM, Fig. 8 Predicted errors of different approaches (Super-Sauze mudslide)
and GP versus the monitored displacement. Again, the R2 values
indicate that, on the whole, all the three techniques perform very
well in predicting the Super-Sauze mudslide displacement. Beside mapping S=f(X) in slightly varying ways. The SVM establishes the
the R2 values, the average percentage errors and the average mapping by transferring the normal nonlinear data into linear
absolute errors are given in Fig. 8 for the predictive results of the data in higher space with the kernel trick and converts the problem
displacements obtained by the three methods. The two types of into a quadratic programming problem based on the Mercer's
error bars in Fig. 8 show that the GP performs best of all among condition. It involves many parameters to be specified. Among
the three methods in generating predictions of the Super-Sauze those parameters, the penalty parameter is user defined and plays
mudslide displacements. an important role for the SVMs' performance. Also, the kernel type
and kernel functions need to be optimized during the
Discussions implementation of the SVMs. These factors result in the
As known, the ANN has a very good nonlinear fitting ability, which uncertainty of the SVMs' performance. The RVM derives the
depends, however, on the condition that the network has an mapping under the Bayesian framework for regression and utilizes
appropriate structure and is well trained. In practical problems, the sparse Bayesian kernels with probabilistic significance. It has
the structure of the neural network is difficult to identify properly, no restrictions defined by the Mercer's condition. Its performance
and the neural network often suffers from the local solution depends largely on the parameters of the kernels.
problems. Moreover, the neural network is data driven and has On the contrary, the GP derives the mapping from probability
no convincing equation for prediction. On the contrary, each of theory with prior and covariance functions. Theoretically, its
the SVM, RVM, and GP is derived based on a strict mathematical performance depends on the prior and covariance functions. In
theory as demonstrated in the section “Landslide displacement practical problems, the influences of the prior and covariance types
prediction formulation and methods.” Each of them thus has a can be usually redeemed and manifested by the parameters of the
prediction equation of the definite form of formulation. Hence, the prior or covariance functions. Therefore, only the length-scale
SVM, RVM, and GP outperform the ANN in fitting the Baishuihe parameter needs to be optimized, which facilitates the implementation
landslide displacement as shown in Fig. 1. of the GP. In our practical modeling of the landslide displacement, it
As stated in the section “Landslide displacement prediction turns out that as a whole, the GP performs slightly better than the SVM
formulation and methods,” the SVM, RVM, and GP obtain the and RVM as shown in Figs. 4 and 8. The reasons may lie in two aspects.
Mathematically, the GP is originally derived for regression problems;
140 the SVM and RVM, however, are initially proposed for classification
GP predicted problems and applied for regression problems with some tricks which
135
result in more parameters. Furthermore, the prediction of the GP is a
SVM predicted
130 probabilistic distribution, for the point-like predictions of the
RVM predicted GP: R² = 0.9978 landslide displacement, so the GP can apply the mean of the
125
Predicteddisplacement(mm)

SVM: R² = 0.9967 distribution as point predictions to avoid robust point predictions like
120 that in the SVMs. In practice, the implementation process also affects
the performance of these methods. As stated above in practice, the GP
115
RVM: R² = 0.9942 involves only the length-scale parameter; in contrast, the SVM and
110 RVM involve more parameters. This factor indeed makes it not easy to
find the best performances of the latter two.
105
It is shown apparently in Figs. 1, 3, 5, and 6 that the mentioned
100 methods can obtain very good results in both fitting and
predicting the nonfluctuating parts of the displacement series of
95
the two illustrated landslides. Meanwhile, these methods can also
90 perform well in fitting the nonfluctuating parts of the
90 95 100 105 110 115 120 125 130 135 140
Monitored displacement(mm)
displacements. However, due to the high nonlinearity and
complexity of the landslide displacement, prediction of the
Fig. 7 Predictive performance of different approaches (Super-Sauze mudslide) fluctuations of landslide displacements is still very difficult since

Landslides
the start points of the fluctuations are uncertain before their Science Foundation (no. 11272114) is gratefully acknowledged.
occurrence unless the displacement curve manifests a typical The authors are also grateful to the anonymous reviewers for their
periodic feature. Unfortunately, it is not conclusive at present to comments and suggestions.
identify whether there is a period characteristic in the landslide
displacement or how to know the value of the period. Therefore,
further work can be tried to analyze the fluctuating displacement
References
with different components such as the progressive or creep
component, periodic component, mutation component, etc. Belle P, Aunay B, Bernardie S, Grandjean G, Ladouche B, Mazué R, Join J-L (2013) The
These three methods obtain the prediction equations based on application of an innovative inverse model for understanding and predicting landslide
the history information of the landslide displacements as movements (Salazie cirque landslides, Reunion Island). Landslides. doi:10.1007/
s10346-013-0393-5, 1–13
demonstrated in the section “Landslide displacement prediction
Bozzano F, Cipriani I, Mazzanti P, Prestininzi A (2011) Displacement patterns of a
formulation and methods.” Thus, besides the methods' nature, the landslide affected by human activities: insights from ground-based InSAR monitoring.
performances of these methods depend largely on the landslide Nat Hazards 59(3):1377–1396
history displacements. The two landslide displacements shown in Chen M, Wang L (1988) A prediction method by grey system for slope deformation and
Figs. 1 and 5 show, apparently, the progressive characteristics before failure. In: Proceedings of the 5th international symposium on landslides, Lausanne,
pp 577–582
sliding occurrence. This feature makes the findings obtained based
Chen H, Zeng Z (2012) Deformation prediction of landslide based on improved back-
on the two landslides cautious to other landslides of different propagation neural network. Cogn Comput. doi:10.1007/s12559-012-9148-1, 1–7
features. For example, the research findings above may be unsuitable Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273–297
to account for landslide displacements triggered by earthquakes or Dong H, Fu H, Feng W, Deng Z (2007) Landslide displacement prediction based on
other kinds of seismic activities. In these cases, the slope Takens theory and SVM. China J Highw Transp 20(5):13–18
Du J, Yin K, Lacasse S (2013) Displacement prediction in colluvial landslides, Three
displacements often behave to be catastrophic features which are
Gorges Reservoir, China. Landslides 10(2):203–218. doi:10.1007/s10346-012-0326-8
mainly caused by the external accidental factors other than the Feng X, Zhao H, Li S (2004) Modeling non-linear displacement time series of geo-
internal landslide evolutions, and thus, they are out of the range of materials using evolutionary support vector machines. Int J Rock Mech Min
the methods discussed above. That is to say, the research findings are 41(7):1087–1107
appropriate for landslide displacements with progressive features. Fukuzono T (1985) A new method for predicting the failure time of a slope. Tokyo
University Press, Tokyo, pp 145–150, In: Proceedings of the 4th international
conference and field workshop on landslides
Conclusions Grelle G, Guadagno FM (2012) Regression analysis for seismic slope instability based on a
The LDA provides a potential way for landslide time forecasting. double phase viscoplastic sliding model of the rigid block. Landslides. doi:10.1007/
This study demonstrates the state-of-the-art computational s10346-012-0350-8, 1–15
intelligence techniques, i.e., the SVM, RVM, and GP for analysis Jibson RW (2007) Regression models for estimating coseismic landslide displacement.
Eng Geol 91(2–4):209–218
and prediction of the landslide nonlinear displacement. Based on
John C (1998) Sequential minimal optimization: a fast algorithm for training support
the results obtained, conclusions can be made as follows: vector machines. Technical report MSR-TR-98-14. Redmond
Li T, Chen M, Wang L, Zhou Y (1996) Time prediction of landslides using Verhulst
1. The computational intelligence techniques are feasible and inverse-function model. In: Proceedings of the 7th international symposium on
capable for landslide nonlinear displacement prediction. With landslides, Trondheim, pp 1289–1293
optimized parameters and well-mannered data, these Li X, Kong J, Wang Z (2012) Landslide displacement prediction based on combining
method with optimal weight. Nat Hazards 61(2):635–646. doi:10.1007/s11069-011-
techniques can perform well both in fitting and predicting 0051-y
the displacement data. Lian C, Zeng Z, Yao W, Tang H (2013) Displacement prediction model of landslide based
2. The SVM, RVM, and GP perform superior to ANN in modeling on a modified ensemble empirical mode decomposition and extreme learning
the Baishuihe landslide displacement. The GP, on the whole, machine. Nat Hazards 66(2):759–771. doi:10.1007/s11069-012-0517-6
performs a little better than the RVM and SVM in predictive Liu Z, Xu W, Zhang K, LIU K (2009a) Prediction of rock slope deformation based on
optimized exponential smoothing method. J Hohai Univ (Nat Sci) 37(3):313–316
analysis of landslide nonlinear displacement. Liu ZB, Xu WY, Meng YD, Chen H (2009) Modification of GM (1,1) and its application in
3. The computational intelligence methods perform better in analysis of rock-slope deformation. In: Proceedings of 2009 IEEE international
predictive analysis of nonfluctuating displacement series than conference on GSIS, Nanjing, pp 415–419
that of the fluctuating ones. Liu Z, Xu W, Shao J (2012) Gaussian process based approach for application on landslide
displacement analysis and prediction. Comp M Eng Sci 84(2):99–122. doi:10.3970/
cmes.2012.084.099
Liu Z, Shao J, Xu W, Meng Y (2013) Prediction of rock burst classification using the
Nevertheless, the parameters of the SVM, RVM, and GP play an technique of cloud models with attribution weight. Nat Hazards 68(2):549–568.
important role in presenting good predictive performance for doi:10.1007/s11069-013-0635-9
landslide displacement analysis. Hence, the exact effects of these Lv Y, Liu H (2012) Prediction of landslide displacement using grey and artificial neural
parameters are to be discovered in the future work. Meanwhile, the network theories. Adv Sci Lett 11(1):511–514
Mayoraza F, Vullietb L (2002) Neural networks for slope movement prediction. Int J
research findings in this study are appropriate for landslide
Geomech 2(2):153–173
displacements with progressive features and unsuitable to account Newmark NM (1965) Effects of earthquakes on dams and embankments. Geotechnique
for landslide displacements triggered by earthquakes or other 15:139–159
kinds of seismic activities. Phillips CL, Bruno M-A, Maquet P, Boly M, Noirhomme Q, Schnakers C, Vanhaudenhuyse
A, Bonjean M, Hustinx R, Moonen G, Luxen A, Laureys S (2011) “Relevance vector
machine” consciousness classifier applied to cerebral metabolism of vegetative and
Acknowledgments locked-in patients. Neuro Image 56(2):797–808
Financial support from the China 973 Program for Key Basic Rasmussen CE, Williams CKI (2006) Gaussian processes for machine learning. MIT,
Research Project (no. 2011CB013504) and the China Natural Cambridge

Landslides
Original Paper
Romeo R (2000) Seismically induced landslide displacements: a predictive model. Eng Yang Y, Liu M (2005) The present advances and trends of landslide predictions. Soil Eng
Geol 58(3=4):337–351 Found 19(2):61–65
Rose ND, Hungr O (2007) Forecasting potential rock slope failure in open pit mines using Yi S (2007) Status and prospect of the temporal prediction of landslide activity. Chin J
the inverse-velocity method. Int J Rock Mech Min 44(2):308–320 Eng Geophys 4(2):157–163
Saito M (1965) Forecasting the time of occurrence of slope failure. University of Toronto Zhu C, Hu G (2013) Time series prediction of landslide displacement using SVM model:
Press, Montréal, pp 537–542, In: Proceedings of the 6th international conference on application to Baishuihe landslide in Three Gorges reservoir area, China. Appl Mech
soil mechanics and foundation engineering Mater 239–240:1413–1420
Saito M (1969) Forecasting time of slope failure by tertiary creep. In: Balkema AA (ed)
International conference on soil mechanics and foundation engineering, Mexico City,
25–29 Aug, Rotterdam, pp 677–683 Z. Liu : J. Shao : W. Xu : H. Chen : C. Shi
Sakai H (2001) Observation of ground displacement in landslide by monitoring Geotechnical Research Institute,
groundwater composition, vol 42. Railway Technical Research Institute, Tokyo, Hohai University,
Quarterly reports Nanjing, 210098, China
Tipping ME (2000) The relevance vector machine. In: Solla SA, Leen TK, Müller KR (eds)
Advances in neural information processing systems. MIT, Cambridge, pp 652–658 Z. Liu ()) : J. Shao
Tipping ME (2001) Sparse Bayesian learning and the relevance vector machine. J Mach Laboratory of Mechanics of Lille,
Learn Res 1(1):211–244 University of Lille 1- Science and Technology,
Travelletti J, Oppikofer T, Delacourt C, Malet J, Jaboyedoff M (2008) Monitoring landslide 59655, Villeneuve-d’Ascq, France
displacement during a controlled rain experiment using a long-range terrestrial laser e-mail: forward@hhu.edu.cn
scanning (TLS). The International Archives of the Photogrammetry, Remote Sens Spat
Inf Sci XXXVII(B5):485–490 Z. Liu
Vapnik V (1998) Statistical learning theory. Wiley-Interscience, New York e-mail: zaobao.liu@polytech-lille.fr

Landslides

View publication stats

Você também pode gostar