Expert Systems With Applications 46 (2016) 60–68

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Hybrid artificial intelligence approach based on metaheuristic and

machine learning for slope stability assessment: A multinational data
analysis
Nhat-Duc Hoang a,∗, Anh-Duc Pham b
a
Institute of Research and Development, Faculty of Civil Engineering, Duy Tan University, P809 - K7/25 Quang Trung, Danang, Vietnam
b
Faculty of Project Management, The University of Danang - University of Science and Technology, 54 Nguyen Luong Bang, Danang, Vietnam

a r t i c l e i n f o a b s t r a c t

Keywords: Slope stability assessment is a critical research area in civil engineering. Disastrous consequences of slope
Slope assessment collapse necessitate better tools for predicting their occurrences. This research proposes a hybrid Artificial
Metaheuristic
Intelligence (AI) for slope stability assessment based on metaheuristic and machine learning. The contribu-
Machine learning
tion of this study to the body of knowledge is multifold. First, advantages of the Firefly Algorithm (FA) and the
Least squares support vector classification
Firefly algorithm Least Squares Support Vector Classification (LS-SVC) are combined to establish an integrated slope prediction
model. Second, an inner cross-validation with the operating characteristic curve computation is embedded
in the training process to reliably construct the machine learning model. Third, the FA, an effective and eas-
ily implemented metaheuristic, is employed to optimize the model construction process by appropriately
selecting the LS-SVM’s hyper-parameters. Finally, a dataset that contains 168 real cases of slope evaluation,
recorded in various countries, is used to establish and confirm the proposed hybrid approach. Experimental
results demonstrate that the new hybrid AI model has achieved roughly 4% improvement in classification
accuracy compared with other benchmark methods.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction can be performed more efficiently (Cheng & Hoang, 2015a; Ghosh,
Bhattacharya, Boccardo, & Samadhiya, 2015).
In many countries, due to the population expansion and eco-
In effect, the slope failure prediction can be formulated as a pat-
nomic development, extensive road networks and residential areas
tern recognition task (Zhao, Yin, & Ru, 2012). To establish a slope as-
have been constructed in the hilly or mountainous regions. This leads
sessment model, historical cases of slope failures in the studied ar-
to the fact that many man-made facilities are susceptible to dam-
eas are first recorded; accordingly, certain features that characterize
ages caused by slope collapses. Slope collapses are complex natural the natural conditions of the areas are extracted for analysis (Wang,
hazards that bring about disastrous consequences (Lu & Rosenbaum,
Xu, & Xu, 2005). Based on the collected database, a machine learn-
2003). Such hazards are responsible for heavy destructions of public/
ing method can be employed to generalize the decision boundary
private property, disruptions of traffic, and losses of human lives ev- that separates the input features of an earth slope into two distintive
ery year (Cheng & Hoang, 2015b; Kang & Li, 2015; Tien Bui, Pradhan,
classes: ‘stable’ and ‘unstable’.
Lofman, Revhaug, & Dick, 2013).
Proposed by Suykens, Gestel, Brabanter, Moor, & Vandewalle,
Hence, to prevent and mitigate the damages, slope stability anal-
2002, the Least Squares Support Vector Classification (LS-SVC) is an
yses are required and better tools for slope assessment are of prac-
advanced machine learning method which possesses many advanced
tical need in the field of civil engineering. The analysis results can
features reflected in its prediction accuracy and fast computation.
be used for identifying collapse-prone areas. Based on such infor-
During the LS-SVC training process, a least squares cost function is
mation, Government agency can acquired better knowledge about
proposed to obtain a linear set of equations in the dual space. Accord-
slope occurrences and the task of allocating financial resources to
ingly, it is only required to solve a set of linear equations to derive the
construct the retaining structures and establishing evacuation plans
model structure. Although successful applications of the LS-SVC have
been reported in a wide span of problem domains (Ghosh, Guha, &
∗
Corresponding author. Tel.: +84 05113827111; fax: +84 05113650443.
Bhar, 2013; Liu & Zhou, 2015; Samui & Kothari, 2011), a few studies
E-mail addresses: hoangnhatduc@dtu.edu.vn (N.-D. Hoang), paduc@dut.udn.vn have investigated and harnessed the capacity of this AI approach in
(A.-D. Pham). slope failure prediction.

http://dx.doi.org/10.1016/j.eswa.2015.10.020
0957-4174/© 2015 Elsevier Ltd. All rights reserved.
 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68 61

Due to the complex and multi-factorial interactions among factors tween the input pattern and the factor of safety of slopes against fail-
that affect slope stability, the task of slope assessment remains a sig- ure. Zhao et al., 2012 utilized the Relevance Vector Machine (RVM)
nificant challenge for civil engineers. Thus, this research proposes an to explore the nonlinear relationship between slope stability and its
AI framework based on the LS-SVC to establish a novel slope assess- influence factors.
ment model. The LS-SVC is utilized as a pattern recognition technique Forecasting models of slope stability based on the Support Vec-
to classify slope conditions into two classes: ‘stable’ and ‘unstable’. tor Machine (SVM) were developed by Li and Wang, (2010); Li and
Furthermore, to better determine the LS-SVC’s hyper-parameters and Dong, (2012); Cheng, Roy, and Chen, (2012), and Cheng and Hoang,
reliably construct the prediction model, the Firefly Algorithm (FA) (2015b); these studies found that SVM forecasting models has more
(Fister, Fister Jr, Yang, & Brest, 2013; Hassanzadeh & Kanan, 2014) and advantages to slope stability evaluation over ANN model under the
an inner cross-validation with computation of the area under the op- condition of limited data.
erating characteristic curve are employed. The remaining part of this Previous researches have demonstrated that that machine learn-
paper is organized as follows. The second and third sections of this ing can provide a viable tool to establish a structured representation
paper present the literature review and the research methodology. of the slope system, which allows the prediction of the slope stability.
The framework of the proposed model is described in the next sec- This study aims at extending the body of knowledge by investigating
tion. The fifth section reports the experimental results. Conclusions the capability of the LS-SVC in solving the task of interest. In order
of this study are stated in the final section. to construct and confirm the prediction method, a dataset including
real cases of slope evaluation has been collected from the literature.
2. Literature review In addition, the training process of the LS-SVC is enhanced by utiliz-
ing the FA optimization and an embedded cross-validation used for
Due to the criticality of slope stability assessment, various re- model evaluation.
search works have been dedicated in tackling this problem of inter-
est. Currently, expert assessment, analytical methods, and machine
learning are commonly employed for analyzing slope conditions. The 3. Research method and material
first method is based on experts’ experiences and knowledge (Cheng
& Hoang, 2014). Using experts’ judgments, the possible factors that 3.1. Least Squares Support Vector Classification (LS-SVC)
trigger slope collapses can be identified and the condition of a slope
can be evaluated. However, the major disadvantage of the expert This section describes the formulation of the LS-SVC for solving
assessment approach is that it strongly relies on subjective judg- classification problems. In the historical data, given a training dataset
ments and it is infeasible to ensure the consistency of the prediction {xk , yk }Nk=1 with input data xk ∈ Rn where N is the number of train-
outcomes. ing data points, n denotes the data dimension, and the correspond-
The analytical methods are derived from the slope displacement ing class labels is defined as yk ∈ {−1, +1} the LS-SVC formulation
models; based on this approach, one can analyze the slope stability for classification is presented as follows (Suykens et al., 2002):
by identifying of the most dangerous sliding surface and calculating
1 2
N
1 T
the factor of safety (Baker, 2003; Kostić, Vasović, & Sunarić, 2015; Li Minimize Jp (w, e) = w w+γ ek (1)
& Chu, 2015; Salmi & Hosseinzadeh, 2015). The analytical methods 2 2
k=1
require input parameters for every calculation point of the studied
region; this brings about serious problems in data collection as well  
Subjected to yk wT φ(xk ) + b = 1 − ek , k = 1, . . . , N (2)
as in controlling the spatial variability of the input parameters. There-
fore, the analytical methods are only appropriate for evaluating slope where w ∈ Rn is the normal vector to the classification hyperplane
stability in small regions (Song et al., 2012). and b ∈ R is the bias; ek ∈ R are error variables; γ > 0 denotes a regu-
Recently, machine learning has been proved to be a feasible and larization constant.
effective approach for slope assessment. In general, machine learning The Lagrangian is given by:
models are established based on artificial intelligence (AI) techniques
and historical databases (Cheng & Hoang, 2015b; Esmaeili, Osanloo, 
N    
Rashidinejad, Aghajani Bazzazi, & Taji, 2014). Using these models, the L(w, b, e, a) = Jp (w, e) − αk yk wT φ(xk ) + b − 1 + ek (3)
slope stability evaluation can be formulated as a classification prob- k=1
lem in which prediction outputs are either “stable” or “unstable”. By
where α k are Lagrange multipliers; φ (xk ) represents a kernel func-
learning events of slope collapse in the past, machine learning ap-
tion. The conditions of optimality can be stated as follows:
proaches can produce predictive results for unlabeled patterns.
⎧
Thus, applying machine learning approaches for solving the task
⎪ 
N
at hand has been an attractive research theme. Yan and Li, (2011) con- ⎪ ∂∂wL = 0 → w = αk yk φ(xk )
⎪
⎪
⎪
structed a method for predicting the stability of open pit slope based ⎨ k=1
∂L = 0 →  α y = 0
N
on the Bayes Discriminant Analysis (BDA). A hybrid instance-based ∂b k k (4)
⎪
⎪ k=1
classifier relied on the Fuzzy K-Nearest Neighbor was established for ⎪ ∂
⎪ ∂ ek = 0 → αk = γ ek , k = 1, . . . , N
L
slope stability assessment (Cheng & Hoang, 2015a). Wu, Kung, Chen, ⎪
⎩ ∂L
∂αk = 0 → yk (w φ(xk ) + b) − 1 + ek = 0, k = 1, . . . , N
T
and Kuo, (2014) proposed a model for predicting and monitoring
slope disaster by employing the K-means model to derive the weight
The linear system below is obtained after the elimination of e and w:
and classification of disaster factors. Cheng and Hoang, (2014) evalu-
ated slope collapses across mountain roads with the employment of 



the Bayes theorem. 0 yT b 0
= (5)
Lu and Rosenbaum, (2003); Zhou and Chen, (2009); Jiang, (2011); y ω + γ −1 I α 1v
Das, Biswal, Sivakugan, and Das, (2011), and Wang et al., (2005) ap-
plied the Artificial Neural Network (ANN) to predict the slope stabil- which y = y1 , . . . , yN , 1v = [1; . . . ; 1], and α = [α1 ; . . . ; αN ]. And the
ity. The Evolutionary Polynomial Regression (Ahangar-Asr, Faramarzi, kernel function is applied as follows:
& Javadi, 2010) and the Least Squares Support Vector Machine (Samui
& Kothari, 2011) have been used to model the mapping function be- ω = yi y j φ(xk )T φ(x1 ) = yi y j K (xk , x1 ) (6)
62 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68
Begin FA
Define objective function f(x), where x=(x1,...,xd)
Generate an initial population of fireflies
Formulate the light intensity I
Define the absorption coefficient L
While (t < Max_Generation)
For i = 1 to n (all n fireflies)
For j=1 to n (all n fireflies)
If (Ij > Ii), move firefly i towards firefly j
End if
Evaluate new solutions and update light intensity;
End for j
End for i
Rank the fireflies and find the current best
End while;
End FA
Fig. 1. The Firefly Algorithm (FA) pseudo code.
The resulting LS-SVC model is as follows:
 

N
y(x) = sign αk yi K (xk , x1 ) + b
k=1
where α k and b are the solution to the linear system (Eq. 5). The ker-
nel function that is commonly used is Radial Basis Function (RBF) ker-
nel. The RBF kernel is described as follows:
 
x − x 1  2
K (xb , x1 ) = exp − k 2
2σ
where σ is the kernel function parameter.
3.2. Firefly Algorithm (FA)
To commence the training process of the LS-SVC, two tuning pa-
rameters (γ , σ ) are required to be determined appropriately. A proper
setting of these tuning parameters is necessary to achieve a desirable
performance of the prediction model. Thus, in this study, we utilize
the FA as a means for tuning the LS-SVC’s free parameters. The de-
scription of the FA algorithm is provided in the following section of
the article.
In tropical and temperate regions, the flashing lights of fireflies are
an amazing scene in the summer sky. The pattern of flashes is often
unique for a particular species (Yang, 2008). In essence, each firefly
is attracted to brighter ones as it randomly explores while searching
for prey. Based on that phenomenon in nature, the FA is formulated
as a global optimization method. The FA is an advanced swarm in-
telligence that can locate the optimum effectively (Fister et al., 2013;
Yang, 2008).
Superior performance of this algorithm over other meta-heuristic
methods has been proved in solving complex optimization problems
(Amiri, Hossain, Crawford, & Wigand, 2013; Baykasoğlu & Ozsoydan,
2014; Kavousi-Fard, Samet, & Marzbani, 2014; Yang, 2014). Further-
more, various recent works demonstrate that the FA is an effective
metaheuristic for optimizing machine learning models (Alweshah &
Abdullah, 2015; Chou & Pham, 2015; Long, Meesad, & Unger, 2015; Vu
& Hoang, 2015).
The FA utilizes the following rules: (1) all fireflies are unisex, so
each firefly is attracted to other fireflies regardless of their sex, (2)
the attractiveness of a firefly is proportional to its brightness and de-
creases as the distance increases. A firefly moves randomly if no other
firefly is brighter, and (3) the brightness of a firefly is affected or de-
termined by the landscape of the objective function. The FA pseudo
code is illustrated in Fig. 1.
The brightness of an individual firefly can be defined similarly
to the fitness value in the genetic algorithm. The light intensity I(r)
Soil parameters:
u
H
Fig. 2. Characteristics of an earth slope.
varies according to the following equation:
I(r) = Io exp ( − γL × r2 ) (9)
where Io denotes the light intensity of the source. γ L is the light ab-
sorption coefficient. r represents the distance from the source.
As the attractiveness of a firefly is proportional to the light inten-
sity seen by adjacent fireflies, the attractiveness β of a firefly is de-
fined as:
(7) β = βo exp ( − γL × r2 ) (10)
In a D-dimensional space, the distance between any two fireflies i at
xi and j at xj , is the calculated as follows:

  
D
r i j =  xi − x j  = (xi,k − x j,k )2 (11)
(8) k=1
Because a specific firefly xi is attracted to the brighter one xj , the
movement of the ith firefly can be expressed as:
xi = xi + βo × exp ( − γL × ri2j ) × (xi − x j ) + α × (ω − 0.5) (12)
where γ is the light absorption coefficient, γ L varies from 0.1 to 10;
β 0 represents the attractiveness at rij = 0; α denotes a trade-off con-
stant to determine the random behavior of movement; ω represents
a random number drawn from the Gaussian distribution.
3.3. Historical cases of earth slopes
In this research, a database consisting of 168 data samples, col-
lected from previous research works (Li & Wang, 2010; Lu & Rosen-
baum, 2003; Sah, Sheorey, & Upadhyaya, 1994; Yan & Li, 2011; Zhou
& Chen, 2009), are used to establish and verify the new slope assess-
ment method. Based on previous researches, six influencing factors,
including unit weight (kN/m3 ), soil cohesion (kPa), internal friction
angle (o ), slope angle (o ), slope height (m), and pore pressure ratio,
are employed to characterize an earth slope (see Fig. 2).
Table 1 provides the information of the influencing factors and
their statistical descriptions. The whole 168 data samples are shown
in the Appendix where the output of Y = 1 indicates a stable slope
and the output of Y = 0 represents an unstable slope. It is noted that
in the database, there are 84 stable slopes and 84 unstable slopes.
The slope condition, either ‘stable’ or ‘unstable’, is recorded by field
surveys. Specifically, if there is no significant movement of the soil
in the slope surface that affects the safety of road traffic or work-
ing/residential areas, a slope condition is determined to be ‘stable’;
otherwise, the condition of a slope is ‘unstable’.
Furthermore, to enhance the visualization of the collected dataset,
the principal component analysis technique (Jolliffe, 2010) is uti-
lized to transform the original data from the original space to a two-
dimensional space. The graphical demonstration of the transformed
historical data is provided in Fig. 3. From the figure, it is recogniz-
able that there are overlapping domains between the two classes of
slope conditions (‘stable’ and ‘unstable’). Such result indicates that
the decision boundary that separates the two aforementioned classes
 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68 63

Table 1
Influencing factors.

Factors Notation Definition Max Average Std. Min

X1 Unit weight (kN/m ) 3

31.30 21.76 4.13 12.00
X2 C Soil cohesion (kPa) 300.00 34.12 45.82 0.00
X3
Internal friction angle (o ) 45.00 28.72 10.58 0.00
X4 B Slope angle (o ) 59.00 36.10 10.22 16.00
X5 H Slope height (m) 511.00 104.19 132.68 3.60
X6 Ru Pore pressure ratio 45.00 0.48 3.45 0.00

tors appear to contribute significantly to the determination of the

0.6 Stable Slope classification label. Meanwhile, the other three factors of the current
Unstable Slope
dataset including the slope height (X5 ), the slope angle (X4 ), and the
internal friction angle (X3 ) seem to demonstrate lower contributions
Principle Component 2

0.4 to the slope condition, reflected in their smaller values of mutual

information.

0.2 4. Metaheuristic-optimized least squares support vector

classification (MO-LSVC) for slope assessment

0.0 This section describes the proposed slope assessment model

named as MO-LSVC. The model is established by a hybridization of
the LS-SVC and FA algorithms. The LS-SVC is a supervised learning
-0.2 technique for performing classification tasks. Moreover, the FA opti-
mization algorithm is utilized to simultaneously identify the optimal
values of the LS-SVC’s tuning parameters: the regularization param-
-0.6 -0.4 -0.2 0.0 0.2 0.4
eter (γ ) and the RBF kernel parameter (σ ). The source code of the LS-
Principle Component 1
SVC is provided in the LS-SVMlab toolbox (De Brabanter et al., 2010).
Fig. 3. Illustration of the collected dataset. Meanwhile, the FA algorithm is programmed in Matlab environment
by the authors.
Table 2 The structure of the hybrid framework is shown in Fig. 4. The his-
Influencing factor ranking with mutual information. torical database used in the research contains 168 real cases of slope
Factors Notation Definition Mutual information Rank assessment. Before the experiment, the database is randomly divided
into two sets: dataset 1 (including 90% of the samples) used to train
X1 Unit weight (kN/m3 ) 0.23 2
the model and dataset 2 (including 10% of the samples) reserved for
X2 C Soil cohesion (kPa) 0.16 3
X3
Internal friction angle (o ) 0.04 6 testing the model performance.
X4 B Slope angle (o ) 0.04 5
(1) Tuning parameter initialization: at this step, all of the model free
X5 H Slope height (m) 0.09 4
X6 Ru Pore pressure ratio 0.54 1
parameters including the regularization parameter (γ ) and the
RBF kernel parameter (σ ) of the LS-SVC are randomly gen-
erated within the range of lower and upper boundaries. The
lower and upper boundaries of parameters are 10−5 and 105 ,
is definitely nonlinear and complex. This provides a good opportu- respectively.
nity to apply AI methods for generalizing the complex classification (2) Model training process: in this step, the LS-SVC is deployed to
boundary. learn the decision boundary to separate input data into two
Additionally, to prevent the situation in which inputs with greater classes. The classes of ‘stable’ and ‘unstable’ slopes are denoted
magnitudes dominate those with smaller magnitudes, the influenc- as the positive class (1) and the negative class (0), respectively.
ing factors have been normalized into the range of [0, 1] using the (3) Slope assessment process: after the training process terminates,
following formula: the LS-SVC has constructed the decision boundary that sepa-
Xo − Xmin rates the input space into two distinctive regions (‘stable’ and
Xn = (13) ‘unstable’); based on such decision boundary, the model is able
Xmax − Xmin
to produce the class label of each input data.
where Xn is the normalized data. Xo is the original data. Xmax and Xmin (4) Model performance evaluation: the goal of the optimization pro-
denote the maximum and minimum values of the factor, respectively. cess is to identify the most suitable set of tuning parame-
Additionally, to demonstrate the contribution of each influencing ters. To evaluate the classification performance, the receiver
factor to the classification output of slope conditions, the mutual in- operating characteristic (ROC) curve (Chawla, Bowyer, Hall, &
formation (Kwak & Choi, 2002; Peng, Fulmi, & Ding, 2005) is em- Kegelmeyer, 2002) is employed. Using this approach, the area
ployed. It should be noted that a larger value of the mutual informa- under ROC curve, or the area under the curve (AUC), is calcu-
tion indicates a stronger relevance between the influencing factor of lated. For more details regarding the establishment of the ROC
interest and the class label. The detailed process to estimate the mu- curve and the calculation of the AUC, the readers are guided
tual information are comprehensively provided in the previous works to previous research works of van Erkel and Pattynama, (1998)
of Kwak and Choi, (2002) and Peng et al., (2005). The mutual infor- and Fawcett, (2006). It is noted that a higher AUC value indi-
mation of all the influencing factors used in this study is reported in cates a better predictive performance. Thus, the following ob-
Table 2. jective function is used in the step of fitness evaluation:
From Table 2, it can be observed that the pore pressure ratio (X6 ),
1
the unit weight of soil (X1 ), and the soil cohesion (X2 ) demonstrate Ff itness = 5 (14)
comparatively high values of mutual information. These three fac- k=1 AUCk
64 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68
1. Tuning Parameter
Initialization
2. Model Training
Process
Training Data
3. Slope Assessment
Process
4. Performance Evaluation
Fig. 4. Metaheuristic-Optimized Least squares Support Vector Classification (MO-LSVC) for slope assessment.
where AUCk denotes AUC values obtained from the prediction results
of a five-fold cross validation process, respectively. Herein, the dataset
1 (training data) is divided into five mutually exclusive subsets. In
each run, one subset is used as a validating set; meanwhile, the other
subsets are used for constructing the model.
(1) Firefly algorithm optimization: the FA optimization algorithm
is utilized to explore the various combinations of the tuning
parameters. At each generation, the algorithm carries out its
searching process to guide the population of fireflies to the
fittest solution. By evaluating the fitness of each firefly, the al-
gorithm discards inferior combinations of the regularization
parameter (γ ) and the RBF kernel parameter (σ ), and permits
robust combinations of these parameters to be passed on the
next generations.
(2) Stopping condition: the FA’s optimization process terminates
when the maximum number of generation is achieved. When
the program terminates, the optimal set of tuning parameters
can be identified and the model is ready to predict unlabeled
input patterns.
5. Experimental results
In the experiment, 90% of the historical data is used for construct-
ing the MO-LSVC. Meanwhile, 10% of the historical data is reserved for
the testing process. The slope condition of testing data points are un-
known for the prediction model. Therefore, the testing data samples
have the role of new slope evaluation tasks which need to be assessed
and they can be employed to verify the trained model. However, be-
cause of the randomness in selecting testing samples, the evaluation
of model performance can be biased.
Accordingly, to avoid the bias in data selection, the whole dataset
(containing 168 cases) is randomly divided into ten data folds in
which each fold in turn serves as testing samples; and the perfor-
mance of the prediction model can be appraised via average results
of the ten folds. This process is the ten-fold cross validation (Zhang,
1993) which is commonly used for verifying the model performance.
Since all of the subsamples are mutually exclusive, this approach can
reliably estimate how accurately the proposed MO-LSVC performs in
practice.
Tunining 5. Firefly Algorithm
Parameters Optimization
Yes
6. Stopping Condition
No
Optimized Slope
Assessment Model
In addition, to better verify the capability of the proposed hy-
brid model, its performance is compared to results obtained from
the benchmark methods including: the Support Vector Classifier
(SVC) (Mathworks, 2015; Vapnik, 1998), the Artificial Neural Network
(ANN) (Beale, Hagan, & Demuth, 2012), the Relevance Vector Machine
(RMV) (Tipping, 2001; Tipping, 2002), and the Bayes Discriminant
Analysis (BDA) (Krzanowski, 1988).
It is noted that besides the commonly used Levenberg–Marquardt
algorithm (Hagan & Menhaj, 1994), the scaled conjugate gradi-
ent (Moller, 1993) and the resilient back-propagation (Riedmiller
& Braun, 1993) algorithms are also employed as the methods for
training the ANN structures. The ANN models with the Levenberg–
Marquardt, scaled conjugate gradient, and resilient back-propagation
training algorithms, are denoted as the LM-ANN, the SCG-ANN, and
the RBP-ANN, respectively.
To construct the ANN model, it is needed to specify the network
structure and the learning rate. These tuning parameters of the ANN
are usually selected via a trial-and-error process. Based on experi-
ments, the network configuration is set as follows: the number of
hidden layers is set to be 1; the number of neurons in the hidden layer
is set to be 6; the learning rate is 1; the number of learning epochs is
5000.
Moreover, to set the tuning parameters of the SVC (including the
penalty and the radial basis kernel function parameters) and the RVM
(the radial basis function width), the training dataset is randomly
spitted into two sets: Set I (80%) used for constructing the model and
Set II (20%) used for validating the model. For the SVC, its two param-
eters are determined via the grid search method (Hsu, Chang, & Lin,
2010). Meanwhile, for the RVM, the radial basis function width is al-
lowed to range from 0 to 1 with the step size of 0.1. For both the SVC
and RVM models, the values of the parameters which can maximize
the average AUC of the Set I and Set II are selected.
Table 3 provides the confusion matrices of the MO-LSVC obtained
from the 10-fold cross validation process. As can be analyzed from
the Table 3, on average, the proposed model only committed 2.3 false
classified cases. In more detail, the average false ‘stable’ and false ‘un-
stable’ cases are 0.8 and 1.5, respectively. Moreover, the result of the
10-fold cross validation process for all methods is provided in Table 4.
To benchmark model performances, the Area Under the Curve (AUC)
and the Classification Accuracy Rate (CAR) are employed.
 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68 65

Table 3
Confusion matrices of the MO-LSVC.

Fold 1 Predicted Fold 6 Predicted

Stable Unstable Stable Unstable

Actual Stable 8 1 Actual Stable 8 1
Unstable 1 5 Unstable 2 6
Fold 2 Predicted Fold 7 Predicted

Stable Unstable Stable Unstable

Actual Stable 7 2 Actual Stable 5 2
Unstable 0 8 Unstable 0 10
Fold 3 Predicted Fold 8 Predicted

Stable Unstable Stable Unstable

Actual Stable 6 2 Actual Stable 7 1
Unstable 2 7 Unstable 0 9
Fold 4 Predicted Fold 9 Predicted

Stable Unstable Stable Unstable

Actual Stable 8 2 Actual Stable 8 0
Unstable 0 7 Unstable 2 7
Fold 5 Predicted Fold 10 Predicted

Stable Unstable Stable Unstable

Actual Stable 7 2 Actual Stable 5 2
Unstable 0 8 Unstable 1 9

Table 4
Result comparison.

Model BDA LM-ANN SCG-ANN RMV SVC RBP-ANN MO-LSVC

Training
CAR (%) 75.07 87.03 88.43 87.52 96.57 86.97 91.41
AUC 0.75 0.87 0.88 0.87 0.97 0.87 0.91
Testing
CAR (%) 75.06 78.67 81.02 83.29 83.37 82.12 86.31
AUC 0.75 0.79 0.81 0.83 0.83 0.84 0.86

It is worth reminding that an AUC is a portion of the area of the
unit squares, its value will always be between 0.0 and 1.0 (Fawcett,
2006). It is worth noticing that higher the AUC value better is the
model prediction performance. The CAR is calculated as the ratio of
correctly classified cases over the total number of classified cases.
Based on Table 4, the average AUC values of the BDA, LM-ANN, SCG-
ANN, RMV, SVC, RBP-ANN, and the proposed MO-LSVC are 0.75, 0.79,
0.81, 0.83, 0.83, 0.84, and 0.86, respectively.
Furthermore, the average CARs of these prediction models in
the testing process are 75.06, 78.67, 81.02, 83.29, 83.37, 82.12, and
86.31%. From the experiment, it can be seen that compared with the
Levenberg–Marquardt training algorithm, the scaled conjugate gra-
dient and resilient back-propagation training algorithms can deliver
better prediction performances. Thus, the scaled conjugate gradient
and resilient back-propagation deem more suitable for construct-
ing ANN-based classifiers than that of the Levenberg–Marquardt
algorithm.
Observably, the RVM (AUC = 0.83) and the SVC (AUC = 0.83)
slightly outperform the SCG-ANN (AUC = 0.81) and are significantly
better than the LM-ANN (AUC = 0.79) as well as the conventional BDA
(AUC = 0.75). The RBP-ANN (AUC = 0.84, CAR = 82.12%) is better than
the RVM and the SVC in terms of the AUC. However, the average CAR
of the RBP-ANN is slightly worse than that of the RVM (CAR = 83.29%)
and SVC (CAR = 83.37%) models.
Based on the experimental results, it can be concluded that the
proposed MO-LSVC is best suited for predicting slope stability since
the model has obtained the best prediction performance. The MO-
LSVC has achieved a roughly 4% in terms of classification accuracy
compared to the second best model (the RBP-ANN). Furthermore, a
better balance between the training and testing outcomes demon-
strates that the cross-validation with the calculation of the area un-
der the operating characteristic curve computation embedded in the
model construction process is really helpful to establish the MO-
LSVC. These facts convincingly demonstrate that the proposed hybrid
approach is a promising alternative for civil engineers to cope with
the slope assessment problem.
6. Conclusion
This article has introduced a novel approach for earth slope as-
sessment named as MO-LSVC. To verify the performance of the pro-
posed method, various machine learning algorithms are used as
benchmark approaches. The experiment has shown that the hybrid AI
method is capable of achieving the most desirable prediction result.
In summary, this research extends the body of knowledge in multi-
fold aspects. First, capabilities of the LS-SVC as a nonlinear classifier
and the FA as an optimization technique are harnessed and combined
to formulate a hybrid slope prediction method. Second, to reliably
construct the machine learning model, this study employs a five-fold
cross-validation with the computation of the area under the oper-
ating characteristic curve embedded in the training process. Experi-
mental results and comparison have pointed out that this framework
can help to construct a model with a high predictive capability. Third,
this research demonstrates that the FA metaheuristics approach can
effectively assist the model selection process of the LS-SVC. Finally,
a database consisting of historical cases of slope conditions has been
collected from various regions around the globe to construct and ver-
ify the proposed AI model.
In practice, the MO-LSVC can be utilized as an intelligent system
to predict the stable/unstable condition of an earth slope when the
information of the slope is available. The model only requires the in-
formation of the six influencing factors including unit weight, soil
cohesion, internal friction angle, slope angle, slope height, and pore
pressure ratio which are relatively easy to be measured. Civil engi-
neers can employ this intelligent system to quickly derive the slope
condition to check the construction safety during excavation projects.
66 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68

Furthermore, government agencies can utilize the MO-LSVC to

recognize collapse-prone areas across roadway and residential areas
to construct retaining structures and to formulate evacuation plans. No. X1 X2 X3 X4 X5 X6 Y
The proposed model can also be useful to help government agencies 25 19.63 11.97 20 22 12.19 0.41 0
in the task of land-use planning, especially for hilly or mountainous 26 21.82 8.62 32 28 12.8 0.49 0
regions. Furthermore, one notable advantage of MO-LSVC is that the 27 20.41 33.52 11 16 45.72 0.2 0
model can operate independently and change its structure adaptively 28 18.84 15.32 30 25 10.67 0.38 1
29 18.84 0 20 20 7.62 0.45 0
according to updating information. Thus, the intelligent method can
30 21.43 0 20 20 61 0.5 0
be easily employed by users without domain knowledge. 31 19.06 11.71 28 35 21 0.11 0
From the experimental result, the MO-LSVC is the most accurate 32 18.84 14.36 25 20 30.5 0.45 0
prediction method. Nevertheless, one limitation of the current model 33 21.51 6.94 30 31 76.81 0.38 0
is that it can only deliver binary classification output. The reason is 34 14 11.97 26 30 88 0.45 0
35 18 24 30.15 45 20 0.12 0
that in some circumstances, the ability of the model to yield proba- 36 23 0 20 20 100 0.3 0
bilistic or fuzzy prediction results can better assist the decision mak- 37 22.4 100 45 45 15 0.25 1
ing process. Another limitation of the MO-LSVC lies in the collected 38 22.4 10 35 45 10 0.4 0
dataset which features only six slope stability’s influencing factors. 39 20 20 36 45 50 0.25 0
40 20 20 36 45 50 0.5 0
Despite the fact that these factors are easily measured, they cannot
41 20 0 36 45 50 0.25 0
fully characterize the condition of a slope. Therefore, more relevant 42 20 0 36 45 50 0.5 0
factors which determine the slope stability (e.g. slope aspect, vegeta- 43 22 0 40 33 8 0.35 1
tion coverage, effects of typhoon and earthquake) should be incorpo- 44 24 0 40 33 8 0.3 1
rated for possibly enhancing the predictive capability of the current 45 20 0 24.5 20 8 0.35 1
46 18 5 30 20 8 0.3 1
model.
Lu and Rosenbaum (2003)
Regarding the field of expert and intelligent systems, the poten-
47 26.49 150 33 45 73 0.15 1
tial improvements of the current research can be stated as follows: 48 26.7 150 33 50 130 0.25 1
(1) assessing the uncertainty of a classification output can be made 49 26.89 150 33 52 120 0.25 1
possible by incorporating probabilistic and fuzzy theories into the 50 26.57 300 38.7 45.3 80 0.15 0
LS-SVC. (2) Investigating other efficient machine learning algorithms 51 26.78 300 38.7 54 155 0.25 0
52 26.81 200 35 58 138 0.25 1
and metaheuristics can help to identify better tools for slope evalu-
53 26.43 50 26.6 40 92.2 0.15 1
ation. (3) The current model employs the Radial Basis Function ker- 54 26.7 50 26.6 50 170 0.25 1
nel within the LS-SVC; future works may study the effect of other 55 26.8 60 28.8 59 108 0.25 1
kernel functions (e.g. the polynomial kernel or hybrid kernel func- Zhou and Chen (2009)
tions) on slope prediction performance. (4) Pre-processing input data 56 22.4 10 35 45 10 0.4 0
by data transformation methods (Duda, Hart, & Stock, 2001) (e.g. 57 20 20 36 45 50 0.5 0
58 20 0 36 45 50 0.25 0
mapping data points from an original lower-dimensional space to a
59 20 0 36 45 50 0.5 0
higher-dimensional space) can help to classify the dataset more con- 60 22 0 40 33 8 0.35 0
veniently; thus, studying data transformation methods and applying 61 20 0 24.5 20 8 0.35 1
these approaches for solving the slope stability assessment can also 62 27 40 35 43 420 0.25 0
be a promising future research direction. 63 27 50 40 42 407 0.25 1
64 27 35 35 42 359 0.25 1
65 27 37.5 35 37.8 320 0.25 1
Appendix. The Historical Cases
66 27 32 33 42.6 301 0.25 0
67 27 32 33 42.4 289 0.25 1
68 27.3 14 31 41 110 0.25 1
69 27.3 31.5 29.7 41 135 0.25 1
No. X1 X2 X3 X4 X5 X6 Y 70 27.3 16.8 28 50 90.5 0.25 1
71 27.3 26 31 50 92 0.25 1
Sah et al. (1994) 72 27.3 10 39 41 511 0.25 1
1 18.68 26.34 15 35 8.23 0 0 73 27.3 10 39 40 470 0.25 1
2 16.5 11.49 0 30 3.66 0 0 74 25 46 35 47 443 0.25 1
3 18.84 14.36 25 20 30.5 0 1 75 25 46 35 44 435 0.25 1
4 18.84 57.46 20 20 30.5 0 1 76 25 46 35 46 432 0.25 1
5 28.44 29.42 35 35 100 0 1 77 26 150 45 30 200 0.25 1
6 28.44 39.23 38 35 100 0 1 78 18.5 25 0 30 6 0.25 0
7 20.6 16.28 26.5 30 40 0 0 79 18.5 12 0 30 6 0.25 0
8 14.8 0 17 20 50 0 0 80 22.4 10 35 30 10 0.25 1
9 14 11.97 26 30 88 0 0 81 21.4 10 30.34 30 20 0.25 1
10 25 120 45 53 120 0 1 82 25 46 35 46 393 0.25 1
11 26 150.05 45 50 200 0 1 83 25 48 40 49 330 0.25 1
12 18.5 25 0 30 6 0 0 84 31.3 68.6 37 47 305 0.25 0
13 18.5 12 0 30 6 0 0 85 25 55 36 45.5 299 0.25 1
14 22.4 10 35 30 10 0 1 86 31.3 68 37 47 213 0.25 0
15 21.4 10 30.34 30 20 0 1 Li and Wang (2010)
16 22 20 36 45 50 0 0 87 18.66 26.41 14.99 34.98 8.2 0 0
17 22 0 36 45 50 0 0 88 28.4 29.41 35.01 34.98 100 0 1
18 12 0 30 35 4 0 1 89 25.96 150.05 45 49.98 200 0 1
19 12 0 30 45 8 0 0 90 18.46 25.06 0 30 6 0 0
20 12 0 30 35 4 0 1 91 21.36 10.05 30.33 30 20 0 1
21 12 0 30 45 8 0 0 92 15.99 70.07 19.98 40.02 115 0 0
22 23.47 0 32 37 214 0 0 93 20.39 24.91 13.01 22 10.6 0.35 1
23 16 70 20 40 115 0 0 94 19.6 12 19.98 22 12.2 0.41 0
24 20.41 24.9 13 22 10.67 0.35 1
(continued on next page)
(Continued)
 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68 67

References
No. X1 X2 X3 X4 X5 X6 Y
Ahangar-Asr, A., Faramarzi, A., & Javadi, A. A. (2010). A new approach for prediction of
95 21.78 8.55 32 27.98 12.8 0.49 0
the stability of soil and rock slopes. Eng. Computation, 27, 878–893.
96 20.39 33.46 10.98 16.01 45.8 0.2 0
Alweshah, M., & Abdullah, S. (2015). Hybridizing firefly algorithms with a probabilistic
97 19.03 11.7 27.99 34.98 21 0.11 0 neural network for solving classification problems. Appl. Soft Comput., 35, 513–524.
98 17.98 4.95 30.02 19.98 8 0.3 1 Amiri, B., Hossain, L., Crawford, J. W., & Wigand, R. T. (2013). Community detection in
99 20.96 19.96 40.01 40.02 12 0 1 complex networks: multi–objective enhanced firefly algorithm. Knowl.-Based Syst.,
100 20.96 34.96 27.99 40.02 12 0.5 1 46, 1–11.
101 19.97 10.05 28.98 34.03 6 0.3 1 Baker, R. (2003). Sufficient conditions for existence of physically significant solutions
102 18.77 30.01 9.99 25.02 50 0.1 1 in limiting equilibrium slope stability analysis. Int. J. Solids. Struct, 40, 3717–3735.
103 18.77 30.01 19.98 30 50 0.1 1 Baykasoğlu, A., & Ozsoydan, F. B. (2014). An improved firefly algorithm for solving dy-
104 18.77 25.06 19.98 30 50 0.2 0 namic multidimensional knapsack problems. Expert. Syst. Appl., 41, 3712–3725.
105 20.56 16.21 26.51 30 40 0 0 Beale, M. H., Hagan, M. T., & Demuth, H. B. (2012). Neural network toolbox user’s guide.
106 16.47 11.55 0 30 3.6 0 0 The MathWorks, Inc.
107 18.8 14.4 25.02 19.98 30.6 0 1 Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: synthetic
minority over-sampling technique. J. Artif. Intell. Res., 16, 321–357.
108 18.8 57.47 19.98 19.98 30.6 0 1
Cheng, M.-Y., & Hoang, N.-D. (2014). Slope Collapse prediction using bayesian frame-
109 28.4 39.16 37.98 34.98 100 0 1
work with k-nearest neighbor density estimation: case study in taiwan. J. Comput.
110 13.97 12 26.01 30 88 0 0 Civ. Eng., ASCE.
111 24.96 120.04 45 53 120 0 1 Cheng, M.-Y., & Hoang, N.-D. (2015). A Swarm-optimized fuzzy instance-based learning
112 18.46 12 0 30 6 0 0 approach for predicting slope collapses in mountain roads. Knowl.-Based Syst., 76,
113 22.38 10.05 35.01 30 10 0 1 256–263.
114 21.98 19.96 36 45 50 0 0 Cheng, M.-Y., & Hoang, N.-D. (2015). Typhoon-induced slope collapse assessment using
115 18.8 15.31 30.02 25.02 10.6 0.38 1 a novel bee colony optimized support vector classifier. Nat. Hazards, 78, 1961–1978.
116 18.8 14.4 25.02 19.98 30.6 0.45 0 Cheng, M.-Y., Roy, A. F. V., & Chen, K.-L. (2012). Evolutionary risk preference inference
117 21.47 6.9 30.02 31.01 76.8 0.38 0 model using fuzzy support vector machine for road slope collapse prediction. Ex-
118 13.97 12 26.01 30 88 0.45 0 pert. Syst. Appl., 39, 1737–1746.
119 17.98 24.01 30.15 45 20 0.12 0 Chou, J.-S., & Pham, A.-D. (2015). Smart Artificial Firefly Colony Algorithm-Based Sup-
120 22.38 99.93 45 45 15 0.25 1 port Vector Regression for Enhanced Forecasting in Civil Engineering. Comput-
aided Civ. Inf., 30, 715–732.
121 22.38 10.05 35.01 45 10 0.4 0
Das, S. K., Biswal, R. i., Sivakugan, N., & Das, B. (2011). Classification of slopes and predic-
122 19.97 19.96 36 45 50 0.25 0
tion of factor of safety using differential evolution neural networks. Environ. Earth
123 19.97 19.96 36 45 50 0.5 0
Sci., 64, 201–210.
124 20.96 45.02 25.02 49.03 12 0.3 1 De Brabanter, K., Karsmakers, P., Ojeda, F., Alzate, C., De Brabanter, J., Pelckmans, K., De
125 20.96 30.01 35.01 40.02 12 0.4 1 Moor, B., Vandewalle, J., & Suykens, J.A.K. (2010). LS-SVMlab Toolbox User’s Guide
126 19.97 40.06 30.02 30 15 0.3 1 version 1.8. Internal Report 10-146, ESAT-SISTA, K.U.Leuven (Leuven, Belgium).
127 17.98 45.02 25.02 25.02 14 0.3 1 Duda, R. O., Hart, P. E., & Stock, D. G. (2001). Pattern classification (2nd Edition). Hobo-
128 18.97 30.01 35.01 34.98 11 0.2 1 ken, New Jersey: John Wiley & Sons.
129 19.97 40.06 40.01 40.02 10 0.2 1 Esmaeili, M., Osanloo, M., Rashidinejad, F., Aghajani Bazzazi, A., & Taji, M. (2014). Mul-
130 18.83 24.76 21.29 29.2 37 0.5 0 tiple regression, ANN and ANFIS models for prediction of backbreak in the open pit
131 18.83 10.35 21.29 34.03 37 0.3 0 blasting. Eng. Comput., 30, 549–558.
132 18.77 25.06 9.99 25.02 50 0.2 0 Fawcett, T. (2006). An introduction to ROC analysis. Pattern Recognition Letters, 27, 861–
133 18.77 19.96 9.99 25.02 50 0.3 0 874.
Fister, I., Fister Jr, I., Yang, X.-S., & Brest, J. (2013). A comprehensive review of firefly
134 19.08 10.05 9.99 25.02 50 0.4 0
algorithms. Swarm Evol. Comput., 13, 34–46.
135 18.77 19.96 19.98 30 50 0.3 0
Ghosh, A., Guha, T., & Bhar, R. B. (2013). Classification of yarn interlacement pattern in
136 19.08 10.05 19.98 30 50 0.4 0
fabrics using least square support vector machines. Fiber. Polym., 14, 1215–1219.
137 21.98 19.96 22.01 19.98 180 0 0 Ghosh, J., Bhattacharya, D., Boccardo, P., & Samadhiya, N. (2015). Automated geo-spatial
138 21.98 19.96 22.01 19.98 180 0.1 0 hazard warning system geowarns: italian case study. J. Comput. Civ. Eng., 29(5),
Yan and Li (2011) 04014065.
139 20.41 33.52 11 16 45.7 0.2 0 Hagan, M. T., & Menhaj, M. B. (1994). Training feedforward networks with the Mar-
140 18.84 0 20 20 7.62 0.45 0 quardt algorithm. IEEE Trans. Neural Netw., 5, 989–993.
Hassanzadeh, T., & Kanan, H. R. (2014). Fuzzy FA: A modified firefly algorithm. Applied
141 19.06 11.7 28 35 21 0.11 0
Artificial Intelligence, 28, 47–65.
142 18.84 14.36 25 20 30.5 0.45 0
Hsu, C.W., Chang, C.C., & Lin, C.J. (2010). A practical guide to support vector classifica-
143 14 11.97 26 30 88 0.45 0
tion. Technical Report. Department of Computer Science, National Taiwan University.
144 18 24 30.15 45 20 0.12 0 Jiang, J.-P. (2011). BP neural networks for Prediction of the factor of safety of slope sta-
145 22.4 10 35 45 10 0.4 0 bility. In Proceedings of International Conference on Computing, Control and Industrial
146 20 20 36 45 50 0.5 0 Engineering (CCIE), 20–21 Aug. 2011.
147 22.4 100 45 45 15 0.25 1 Jolliffe, I. T. (2010). Principal component analysis (2nd Edition). New York: Springer-
148 27 50 40 42 407 0.25 1 Velag.
149 31.3 68 37 46 366 0.25 1 Kang, F., & Li, J. (2015). Artificial bee colony algorithm optimized support vector regres-
150 27 35 35 42 359 0.25 1 sion for system reliability analysis of slopes. J. Comput. Civ. Eng, 04015040.
151 27 37.5 35 38 320 0.25 1 Kavousi-Fard, A., Samet, H., & Marzbani, F. (2014). A new hybrid Modified Firefly Al-
152 27 32 33 42 289 0.25 1 gorithm and Support Vector Regression model for accurate Short Term Load Fore-
153 27 14 31 41 110 0.25 1 casting. Expert. Syst. Appl., 41, 6047–6056.
Kostić, S., Vasović, N., & Sunarić, D. (2015). A new approach to grid search method in
154 27 31.5 29.7 41 135 0.25 1
slope stability analysis using box–Behnken statistical design. Appl. Math Comput.,
155 27 16.8 28 50 90.5 0.25 1
256, 425–437.
156 27 26 31 50 92 0.25 1
Krzanowski, W. J. (1988). Principles of multivariate analysis: a user’s perspective. New
157 27 10 39 41 511 0.25 1 York.: Oxford University Press.
158 27 10 39 40 470 0.25 1 Kwak, N., & Choi, C.-H. (2002). Input feature selection by mutual information based on
159 25 46 35 47 443 0.25 1 Parzen window. IEEE Trans. Pattern Anal. Mach. Intell., 24, 1667–1671.
160 20 20 36 45 50 0.25 0 Li, J., & Dong, M. (2012). Method to Predict Slope Safety Factor Using SVM. In Proceedings
161 19.63 11.97 20 22 21.19 0.4 0 of the Earth and Space 2012: Engineering, Science, Construction, and Operations in
162 25 55 36 44 299 0.25 1 Challenging Environments, Pasadena (pp. 888–899). ASCE.
163 27.3 10 39 40 480 0.25 1 Li, J., & Wang, F. (2010). Study on the forecasting models of slope stability under data
164 25 46 35 46 393 0.25 1 mining. In Proceedings of the Earth and Space 2012: : Engineering, Science, Construc-
165 16.5 11.49 0 30 3.66 0 0 tion, and Operations in Challenging Environments (pp. 765–776). ASCE.
166 25 120 45 53 120 0 1 Li, L., & Chu, X. (2015). Risk assessment of slope failure by representative slip surfaces
167 19.06 11.75 28 35 21 0.11 0 and response surface function. KSCE J. Civ. Eng., 1–10.
Liu, F., & Zhou, Z. (2015). A new data classification method based on chaotic particle
168 18.84 14.36 25 20.3 50 45 0
swarm optimization and least square-support vector machine. Chemometr. Intell.
Lab., 147, 147–156.
68 N.-D. Hoang, A.-D. Pham / Expert Systems With Applications 46 (2016) 60–68
Long, N. C., Meesad, P., & Unger, H. (2015). A highly accurate firefly based algorithm for
heart disease prediction. Expert. Syst. Appl., 42, 8221–8231.
Lu, P., & Rosenbaum, M. S. (2003). Artificial neural networks and grey systems for the
prediction of slope stability. Nat. Hazards, 30, 383–398.
Mathworks. (2015). Statistics and machine learning toolbox. Natick, Massachusetts,
United States: The MathWorks, Inc.
Moller, M. F. (1993). A scaled conjugate gradient algorithm for fast supervised learning.
Neural Networks, 6(4), 525–533.
Peng, H., Fulmi, L., & Ding, C. (2005). Feature selection based on mutual information cri-
teria of max-dependency, max-relevance, and min-redundancy. IEEE Trans. Pattern
Anal. Mach. Intell., 27(8), 1226–1238.
Riedmiller, M., & Braun, H. (1993). A direct adaptive method for faster back-propagation
learning: the RPROP algorithm. In Proceedings of the IEEE International Conference
on Neural Networks: Vol. 1 (pp. 586–591).
Sah, N. K., Sheorey, P. R., & Upadhyaya, L. N. (1994). Maximum likelihood estimation
of slope stability. International Journal of Rock Mechanics and Mining Sciences & Ge-
omechanics Abstracts, 31, 47–53.
Salmi, E., & Hosseinzadeh, S. (2015). Slope stability assessment using both empirical
and numerical methods: A case study. B. Eng. Geol. Environ., 74, 13–25.
Samui, P., & Kothari, D. P. (2011). Utilization of a least square support vector machine
(LSSVM) for slope stability analysis. Sci. Iran., 18, 53–58.
Song, Y., Gong, J., Gao, S., Wang, D., Cui, T., Li, Y., & Wei, B. (2012). Susceptibility assess-
ment of earthquake-induced landslides using Bayesian network: A case study in
Beichuan, China. Comput. Geosci., 42, 189–199.
Suykens, J., Gestel, J. V., Brabanter, J. D., Moor, B. D., & Vandewalle, J. (2002). Least square
support vector machines. Singapore: World Scientific Publishing Co. Pte. Ltd.
Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., & Dick, Ø. (2013). Regional prediction
of landslide hazard using probability analysis of intense rainfall in the Hoa Binh
province, Vietnam. Nat. Hazards, 66, 707–730.
Tipping, M. E. (2001). Sparse bayesian learning and the relevance vector machine. J.
Mach. Learn. Res., 1, 211–244.
Tipping, M.E. (2002). SparseBayes. http://www.miketipping.com/sparsebayes.htm
van Erkel, A. R., & Pattynama, P. M. T. (1998). Receiver operating characteristic (ROC)
analysis: Basic principles and applications in radiology. European Journal of Radiol-
ogy, 27, 88–94.
Vapnik, V. N. (1998). Statistical learning theory. United States of America: John Wiley &
Sons, Inc.
Vu, D. T., & Hoang, N.-D. (2015). Punching shear capacity estimation of frp-reinforced
concrete slabs using a hybrid machine learning approach. Struct. Infrastruct. E. In
Press.
Wang, H. B., Xu, W. Y., & Xu, R. C. (2005). Slope stability evaluation using back propaga-
tion neural networks. Eng. Geol., 80, 302–315.
Wu, C.-I., Kung, H.-Y., Chen, C.-H., & Kuo, L.-C. (2014). An intelligent slope disaster pre-
diction and monitoring system based on WSN and ANP. Expert. Syst. Appl., 41, 4554–
4562.
Yan, X., & Li, X. (2011). Bayes discriminant analysis method for predicting the stability of
open pit slope. In Proceedings of the International Conference on Electric Technology
and Civil Engineering (ICETCE), 22–24 April 2011 (pp. 147–150).
Yang, X.-S. (2014). Nature-inspired optimization algorithms. Oxford: Elsevier.
Yang, X. S. (2008). Firefly algorithm. Bristol, UK: Luniver Press.
Zhang, P. (1993). Model selection via multifold cross validation. The Annals of Statistics,
21, 299–313.
Zhao, H., Yin, S., & Ru, Z. (2012). Relevance vector machine applied to slope stability
analysis. Int. J. Numer. Anal. Meth. Geomech., 36, 643–652.
Zhou, K.-p., & Chen, Z.-Q. (2009). Stability prediction of tailing dam slope based on
neural network pattern recognition. In Proceedings of the Second International Con-
ference on Environmental and Computer Science ICECS ’09, 28–30 Dec. 2009 (pp. 380–
383).