A Bayesian Framework Based On Gaussian Mixture Mod

Geosci. Model Dev. Discuss., doi:10.
5194/gmd-2016-311, 2017
Manuscript under review for journal Geosci. Model Dev.
Published: 17 January 2017
c Author(s) 2017. CC-BY 3.0 License.
A Bayesian Framework Based on Gaussian Mixture Model and
Radial Basis Function Fisher Discriminant Analysis for Flood
Spatial Prediction (BayGmmKda V1.1)
Dieu, Tien Bui1 and Nhat-Duc, Hoang2
5 1
Geographic Information System Group, Department of Business Administration and Computer Science, University College
of Southeast Norway (USN), Hallvard Eikas Plass, N-3800, Bø I Telemark, Norway.

2
Faculty of Civil Engineering, Institute of Research and Development, Duy Tan University, P809 - K7/25 Quang Trung,
Danang, Vietnam.
Correspondence to: Nhat-Duc Hoang ([email protected])
10 Abstract. In this study, a probabilistic model, named as BayGmmKda, is proposed for flood assessment with a study area in
Central Vietnam. The new model is essentially a Bayesian framework constructed a combination of Gaussian Mixture
Model, Radial Basis Function Fisher Discriminant Analysis, and a Geographic Information System database. Experiments
used for measuring the model performance point out that the hybrid framework is superior to other benchmark models
including the adaptive neuro fuzzy inference system and the support vector machine. To facility the model implementation, a
15 software program of BayGmmKda has been developed in Matlab environment. The newly proposed model is shown to be a
useful tool for flood susceptibility evaluation.
Key words: Flood Evaluation; Bayesian Classifier; Gaussian Mixture Model; Discriminant Analysis; Latent Variable.
1 Introduction
Flood stands out as one of the most destructive phenomena featured by its immense scale of damages as well as its large
20 spatial extent (Dottori et al., 2016b). Catastrophic flood events destroy crops, infrastructures, and bring out heavy loss of
human lives. Floods also lead to siltation of the reservoirs and thus limit the capacity of dams designed for flood control
(Sanyal and Lu, 2004). This natural hazard is known to negatively affect 170 million people around the globe annually
1
Geosci. Model Dev. Discuss., doi:10.5194/gmd-2016-311, 2017
(Kazakis et al., 2015;Judi et al., 2011). On average, people in more than 90 countries are victimized by catastrophic flooding,
resulting in more than 170,000 deaths (Kazakis et al., 2015).
25 Because of monsoonal rainfalls and cyclonic patterns, regions in Southern Asian are considerably affected by flood (Loo
et al., 2015). Particularly in Vietnam, floods are often triggered by tropical cyclones and this country has undergone
destructive consequence of flooding in many provinces. According to Tien Bui et al. (2016b), more than 70% of the
population and roughly 60% of the area in this nation are negatively affected by flood hazards. Based on a report done by
Kreft et al. (2014), in the last two decades, flood accounts for destructions that cost approximately $2.9 billion in Vietnam.
30 In addition, the occurrences of flood in Vietnam are expected to soar in the near future due to the increases of poorly
planned infrastructure developments and urbanization near watercourses, as well as an increased activity of deforestation and
climate change. Hence, accurate forecasting of flood becomes an increasingly crucial task for land-use planning as well as
disaster mitigation strategies. The reason is that it is possible to forecast flood events and flood-prone area can be identified
by means of intelligent approaches at a region scale (Kia et al., 2012;Tien Bui et al., 2016b).
35 Needless to say, the identification of flood susceptible areas can significantly reduce its damages to the national
economy and human lives by avoiding infrastructure developments and densely populated settlements in the highly flood
susceptible areas (Zhou et al., 2016). The prediction outcomes also help Government agency to issue appropriate flood
management policies and to focus its limited financial resource to construct large-scale flood defense infrastructure in areas
that possess great economic values but are highly susceptible to flood (Bubeck et al., 2012). Therefore, a framework of
40 accurate flood prediction is of great usefulness.
Recently, there is an increasing trend of applying Geographic Information System (GIS) in flood prediction. GIS
technology is demonstrated to be a helpful tool to investigate the multi-dimensional events of flooding (Tien Bui et al.,
2016a;Shahabi and Hashim, 2015;Jia et al., 2016). GIS method is capable of simultaneously analyzing different layers of
information (Candy et al., 2014), such as flood inventory map and various flood conditioning factors (e.g. topological and
45 hydrological features) to yield flood susceptibility evaluation results. In addition, flood forecasting systems based on GIS are
indeed suitable for a participatory approach to flood management; the reason is that this technique facilitates the
communication with the public in a scientific manner.
2
To construct flood susceptibility evaluation models, databases of GIS that contains a set of flood influencing factors and
information of past flood events is established at the first step. At the next step, advanced soft computing models can be
50 utilized to distinguish the flood vulnerable areas from the entire studied region (Tehrany et al., 2015b). In this way, the flood
prediction problem boils down to a supervised classification task. Nevertheless, most models in the current studies can only
yield qualitative outputs of flood prediction outcome (i.e. flood –no flood) (Dottori et al., 2016a); probabilistic evaluations
have rarely been seen in the literature. Given these motivations, this study proposes a novel methodology designed for
enhancing the prediction accuracy as well as deriving probabilistic evaluation of flood susceptibility in a regional scale.
55 The method is relied on Bayesian framework with the Gaussian mixture model (GMM) and the Radial Basis Function
Fisher Discriminant Analysis (RBFDA). GMM is employed for density approximation to calculate the posterior probability
of flood within the Bayesian framework. Furthermore, to boost the classification accuracy of the Bayesian model, RBFDA is
employed to construct a latent variable for from the original GIS database; this latent variable aims at maximizing data
discrimination with respect to the two classes of ‘flood’ and ‘no-flood’.
60 In essence, the proposed integrated framework contains two phases of analysis. RBFDA is first employed for latent
variable construction. The Bayesian approach assisted by GMM is then used to perform probabilistic pattern recognition.
The first level performs pattern discriminant analysis task and the second level carries out the prediction to derive the model
output of flood evaluation. Based on previous studies indicating that hierarchical model structure can yielded improving
prediction accuracy (Chou and Tsai, 2012;Shahangian and Pourghassem, 2016;Hoang and Tien Bui, 2016), the proposed
65 unified framework can potentially bring about desirable flood assessment results. The subsequent parts of this study are
organized in the following order: Related works on flood prediction are summarized in the second section. The next section
introduces the research method of the current paper. The fourth part describes the proposed Bayesian model for flood
susceptibility forecasting. The next part reports the model prediction accuracy and comparison. The last section of the paper
discusses some conclusions on this work.
70 2 A Review of Related Works on Flood Susceptibility Prediction
Because of the criticality of flood prediction, this problem has gained an increasing attention from the academic
community. Following this trend, various flood analyzing tools have been developed, ranging from relatively simple
3
methods to more sophisticated methodologies involving hydrological and hydraulic models (Winsemius et al.,
2013;Papaioannou et al., 2015). In recent years, remote sensing coupled with the advancement of GIS technology has been
75 increasingly shown to be a reliable method for producing synoptic coverage over a large area in a cost effective way (Sanyal
and Lu, 2004;Tien Bui et al., 2016b;Kazakis et al., 2015;Kia et al., 2012).
The new approach based on GIS successfully evades the limitation of the hydrological models and equips decision-
makers with a powerful flood analysis tools. GIS databases integrated with data-driven methods have demonstrated their
effectiveness and accuracy in large scaled flood predictions. An fuzzy logic based algorithm has been used to develop a map
80 of flooded areas from synthetic aperture radar imagery, used for the operational flood management system in Italia, was
established by Pulvirenti et al. (2011). A model based on the frequency ratio approach and GIS for spatial prediction of
flooded regions was first introduced by Lee et al. (2012); the spatial database were constructed by field surveys and maps of
the topography, geology, land cover, and infrastructure.
Prediction models with artificial neural network (ANN) have been employed for flood susceptibility evaluation by
85 various scholars (Kia et al., 2012;Seckin et al., 2013;Rezaeianzadeh et al., 2014;Radmehr and Araghinejad, 2014); previous
works have shown ANN as a capable nonlinear modeling tool. Nevertheless, ANN learning is prone to overfitting and its
performance has been shown to be inferior to that of support vector machine (Hoang and Pham, 2016).
Kazakis et al. (2015) introduced a multi-criteria index to assess flood hazard areas that relies on GIS and Analytical
Hierarchy Process (AHP); in this methodology, the relative importance of each flood conditioning factors for the occurrence
90 and severity of flood were determined via AHP. More recently, Support Vector Machine-based flood susceptibility analysis
approaches have been proposed by Tehrany et al. (2015a) and Tehrany et al. (2015b); the research finding is that SVM is
more accurate than other benchmark models including the decision tree classifier and the conventional frequency ratio
model.
Mukerji et al. (2009) constructed flood forecasting models based on an adaptive neuro-fuzzy interference system
95 (ANFIS), Genetic Algorithm optimized ANFIS, ; experiments demonstrated that ANFIS attained the most desirable
accuracy. Recently, a metaheuristic optimized neural fuzzy inference system, named as MONF, has been introduced by Tien
4
Bui et al. (2016b); the research finding is that MONF is more capable than decision tree, ANN, SVM, and conventional
ANFIS.
As can be seen from the literature review, various data-driven and advanced soft computing approaches have been
100 proposed to construct different flood forecasting models. In most of previous studies, the flood prediction was formulated as
a binary pattern recognition problem in which the model output is either flood or no flood. Probabilistic models have rarely
been examined to cope with the complexity as well as uncertainty of the problem under concern. Therefore, our research
aims at enriching the body of knowledge by proposing a novel Bayesian probabilistic model to estimate the flood
vulnerability with the use of a GIS database.
105 3 Research Method
3.1 Flood inventory map and flood conditioning factors of the study area
3.1.1 The study area
In this research, Tuong Duong district (central Vietnam) is selected as the study area (see Figure 1). This is
recognizably a heavily affected flood region in the country (Reynaud and Nguyen, 2016). The area of the district is
110 approximately 2803 km2 and locates between the longitudes of 18°58'42''N and 19°39'16''N, and between the latitudes of
104°15''58'E and 104°55''57'E. The topographical feature of the Tuong Duong district is inherently complex with
mountainous areas, watersheds, and rivers. Drastic floods often divided the district into several isolated areas which are very
difficult to approach for rescuing or evacuation purposes.
The district is featured by two separated seasons, namely a cold season (from November to March) and a hot season
115 (from April to October). The yearly rainfall of the district is within the range of 1679 mm and 3259 mm. The rainfall amount
is primarily intensified during the rainy period which contributes to roughly 90% of the total annual rainfall. Due to the
district’s location as well as its topographic and climatic features, the study area is highly susceptible to flood events with
immense infliction to human casualty and economic value. An examination carried out by Reynaud and Nguyen (2016)
reported that approximately 40% of families have damaged by floods and roughly 20% of families must be departed from
120 the flooded areas; the average loss of flood goes up to 24% of the family income each year.
5
Figure 1 Tuong Duong district (Central Vietnam)

3.1.2 Flood inventory map
Prediction of flood zones can be based on an assumption that future flood events are governed by the very similar
125 conditions of flooded zones in the past. Therefore, it is a reasonably strategy to analyze past records of flood occurrences
(Tien Bui et al., 2016b;Tehrany et al., 2015b). The first step of this analysis is to establish a flood inventory map for the
region under investigation.
The flood inventory map established in the current research stores documentations of past flood events (see Figure 1).
The map was constructed by gathering information of the study area, field works at flood areas, and analyses from results of
130 the Landsat 8 Operational Land Imagery (from 2010 to 2014) with the resolution of 30m (retrieved from
http://earthexplorer.usgs.gov). Furthermore, the location of flood events was also verified by field works carried out in 2014
with handhold GPS devices. In summary, the total number of flood locations during the last five years was recorded to be 76.
6
3.1.3 Flood influencing factors
In order to construct a flood prediction model, besides the flood inventory map, it is crucial to determine the flood
135 influencing factors (Tehrany et al., 2015a). It is worth to notice that the selection of the flood governing factors varies due to
different characteristics of study areas and the availability of data (Papaioannou et al., 2015).
Table 1 Flood influencing variables and their category descriptions
Factors Notation Description of factor categories

1 (0 to 0.5); 2 (0.5 to 2); 3 (2 to 5); 4 (5to 8); 5 (8 to 13); 6 (13 to 20);
Slope (o) IF1
7 (20 to 30); 8 (>30)
1 (<1); 2 (1 to 2); 3 (2 to 3); 4 (3 to 4); 5 (4 to 5); 6 (5 to 6); 7 (6 to 7);

Elevation (100m) IF2
8 (7 to 10); 9 (10 to 13); 10 (>13)
Curvature IF3 1 (<-2); 2 (-2 to -0.05) ; 3 (-0.05 to 0.05); 4 (0.05 to 2); 5 (>2)
Topographic Wetness Index 1 (<6.5); 2 (6.5 to 7.5); 3 (7.5 to 8.5); 4 (8.5 to 9.5); 5 (9.5 to 10.5);
IF4
(TWI) 6 (10.5 to 11.5); 7 (11.5 to 12.5); 8 (>12.5)
1 (<1); 2 (1 to 3); 3 (3 to 5); 4 (5 to 7); 5 (7 to10); 6 (10 to 15);

Stream Power Index (SPI) IF5
7 (15 to 20); 8 (20 to 30); 9 (30 to 50); 10 (>50)
1 (<40); 2 (40 to 80); 3 (80 to 120); 4 (120 to 200); 5 (200 to400);

Distance to river (m) IF6
6 (400 to 700); 7 (700 to 1500); 8 (>1500)
Stream density (km/km2) IF7 1 (<1); 2 (1 to 3); 3 (3 to 5); 4 (5 to 7); 5 (7 to9); 6 (>9)
Normalized Difference 1 (<0.3); 2 (0.3to 0.35); 3 (0.35 to 0.4); 4 (0.4 to 0.45); 5 (0.45 to0.5);
IF8
Vegetation Index (NDVI) 6 (0.5 to 0.55); 7 (0.55 to 0.6); 8 (>0.6)
1 (Q); 2 (Nkb); 3 (Jmh); 4 (T3npb); 5 (T2); 6 (C-bslk); 7 (D-ntdl);
Lithology (rock type) IF9
8 (S2-D1hn); 9 (O3-S1sc3); 10 (O3-S1sc2); 11 (O3-S1sc1); 12 (PR2bk)
1 (<1.82); 2 (1.82 to 1.92); 3 (1.92 to 2.02); 4 (2.02 to 2.12); 5 (2.12 to 2.22); 6
Rainfall (1000mm) IF10
(2.22 to 2.32); 7 (2.32 to 2.42); 8 (>2.42)
In this study, the flood conditioning factors were selected by literature review and field data. Accordingly, ten factors
140 are chosen to analyzing flood vulnerability and a GIS database consisting of the flood inventory map and the chosen factors
was established. The ten flood governing factors can be listed as: slope (IF1), elevation (IF2), curvature (IF3), topographic
wetness index (TWI) (IF4), stream power index (SPI) (IF5), distance to river (IF6), stream density (IF7), normalized
difference vegetation index (NDVI) (IF8), lithology (IF9), and rainfall (IF10). The information of ten conditioning factors of
7
flood occurrence employed in this study is summarized in Table 1. The distributions of the ten factors within the studied
145 region are illustrated in Figure 2.
Figure 2 Flood influencing factors: (a) Slope, (b) Elevation, (c) Curvature, (d) Topographic wetness index
8
150
9
155 Figure 2 (Cont.) Flood influencing factors: (e) Stream power index, (f) Distance to river, (g) Stream density, (h) Normalized
Difference Vegetation Index, (i) Lithology, (j) Rainfall
3.2 Bayesian Framework for Pattern Classification
To cope with the complexity as well as the uncertainty of the problem of interest, Bayesian framework is employed in
this study to evaluate the flood susceptibility of each data sample. The Bayesian framework provides a flexible way for
160 probabilistic modeling. This method features a strong ability for dealing with uncertainty and noisy data (Theodoridis,
2015;Cheng and Hoang, 2016). Nevertheless, previous studies have rarely examined the capability of this approach for
inferring flood susceptibility.
Basically, pattern classification aims at assigning a pattern to one of M distinctive class labels Ck (k = 1,…, M). To
recognize an input pattern based on the information supplied by its feature vector X, we need to attain the pose probability
165 P(Ck|X), which indicates the likelihood that the feature vector X falls into a certain group Ck. Based on such information, the
pattern will be categorized to the group with the highest posterior probability. The posterior probability P(Ck|X) is calculated
as follows (Webb and Copsey, 2011):
p ( X | C k )  P (C k )
P (C k | X )  (1)
p( X )
10
where P(Ck | X ) denotes the posterior probability. p( X | Ck ) represents the likelihood which is also called the class-
170 conditional probability density function (PDF). P(Ck ) denotes the prior probability, which implies the probability of the
class before any feature is measured. The denominator p ( X ) is the evidence factor; this quantity is merely a scale factor for
guaranteeing that the posterior probabilities are valid; it can be calculated as follows:
M
P ( X )   p ( X | C k )  P (C k ) (2)
k 1
Generally, the prior probabilities P(Ck) can be calculated by computing the ratio of training instances in each class.
175 Thus, the bulk of establishing a Bayesian classification model is to calculate the likelihood p(X|Ck). This likelihood
expresses the density of input patterns in the learning space within a certain group of data. In most of situations, p(X|Ck) is
unknown and must be estimated from the available data. In this research, the Gaussian mixture model is utilized for
computing the p(X|Ck) quantity.
3.3 Gaussian Mixture Model for Density Estimation
180 3.3.1 Gaussian Mixture Model (GMM)
Generally, density estimation can be defined as the problem of approximating a PDF given a finite number of data
instances (Scott, 2015). GMM have been shown to be an effective parametric method for modeling of data distribution
especially in high dimensional space (McLachlan and Peel, 2000;Theodoridis and Koutroumbas, 2009). Previous studies
(Paalanen, 2004;Figueiredo and Jain, 2002;Gómez-Losada et al., 2014;Arellano and Dahyot, 2016) point out that any
185 continuous distribution can be approximated arbitrarily well by a finite mixture of Gaussian distributions. Due to their
usefulness as a flexible modeling tool, GMMs have received an increasing attention from the academic community (Zhang et
al., 2016;Khanmohammadi and Chou, 2016;Ju and Liu, 2012).
In a d-dimensional space the Gaussian PDF is defined mathematically in the following form:
1 1
N(x | )  exp{ ( x   )T  1 ( x   )} (3)
(2 ) d /2
||
1/ 2
2
11
190 where  denotes the vector of variable mean and  represents the matrix of covariance; and   { , } denotes a set of
distribution parameter.
A GMM is, in essence, an aggregation of several multivariate Normal distributions; hence, its PDF for each data sample
is computed as a weighted summation of Gaussian distributions (see Figure 3):
k k
p ( x | )    i pi ( x |  i )    i N ( x |  i ) (4)
i 1 i 1
195 where   {1 ,  2 ,..., k , 1 ,  2 ,...,  k } . { 1 ,  2 ,..., k , } is called the mixing coefficients of k Gaussian components
k
and 
i 1
i 1.
p(x(t)|Θ )
Σ
α1
αi αk
... ...
p(x(t)|θ1) p(x(t)|θi) p(x(t)|θk)
x(t)
Figure 3 Structure of Gaussian Mixture Model (GMM)
200 Accordingly, the PDF for all data samples can be expressed as follows (Ju and Liu, 2012):
n
p ( X | )   p ( x t | )  L (  | X ) (5)
t 1
Identifying a GMM’s parameters  can be considered as unsupervised learning task within which a dataset of
independently distributed data points X  {x1 ,..., x N } generated from an integrated distribution dictated via the PDF
p( X | ) . The goal is to find the most appropriate value of  , denoted as e , that maximizes the log-likelihood function:
12
n n k
205  e  arg max log( L( X , ))  log( p ( x t | ))   log(   i pi ( x t |  i )) (6)
 t 1 t 1 i 1
Practically, instead of dealing with the log-likelihood function, an equivalent objective function Q is optimized (Ju and
Liu, 2012):
n k
Max. Q   wit log[ i pi ( x t |  i )] (7)
t 1 i 1
where wit is a posteriori probability for the ith class, i  1,...,k and wit satisfies the following conditions:
 i pi ( x t |  i ) k
and  w
210 wit  k it 1 (8)
 s p s ( xt |  s )
s 1
i 1
The expectation maximization (EM) algorithm is a standard method to compute e . Besides the EM method, this study
employs an unsupervised learning approach for determining GMMs proposed by Figueiredo and Jain (2002). These two
algorithms are briefly reviewed in the next section of the paper.
3.3.2 Learning of finite Mixture Model with the Expectation Maximization (EM) Algorithm
215 The EM method is a statistical approach to fit a GMM based on historical data; this method converges to a maximum
likelihood estimate of model parameters (McLachlan and Krishnan 2008). It can be recapitulated as follows (McLachlan and
Peel, 2000). Commencing from an initial parameter o , an iteration of the EM algorithm consists of the E-step in which the
current conditional probabilities pi ( xt |  i )  N ( xi | i ,  i ) that xt generated from the ith mixture component are
calculated, and the M-step within which the maximum likelihood estimates of  i are updated. The iteration of EM algorithm
220 terminates when the change value of the objective function is lower than a threshold value.
These two steps of the EM procedure are stated as follows:
E-step: estimating the expected classes of all data samples for each class wit based on Eq. (8).
M-step: calculating maximum likelihood given the data’s class membership distribution using the following equations:
13
1 n
 inew   wit
n t 1
(9)
w it xt
225  inew  t 1
n
(10)
w t 1
it
w it ( x t   inew )( x t   inew ) T
 inew  t 1
n
(11)
w
t 1
it
3.3.3 Unsupervised learning of finite mixture model
The EM algorithm increases the log-likelihood iteratively until convergence is detected; and this approach generally can
derive a good set of estimated parameters. Nonetheless, EM suffers from low convergence speed in some data sets, high
230 sensitivity to initialization condition, and sub-optimal estimated solutions (Biernacki et al., 2003). Moreover, additional
efforts are required to determine an appropriate number of Gaussian distributions within the mixture.
As an attempt to alleviate such drawbacks of EM, Figueiredo and Jain (2002) put forward an unsupervised algorithm
for learning a GMM from multivariate data. The algorithm features the capability of identifying a suitable number of
Gaussian components autonomously; and by experiments, the authors show that the algorithm is not sensitive to
235 initialization. In other words, this unsupervised approach incorporates the tasks of model estimation and model selection in a
unified algorithm. Generally, this method can initiate with a large number of components. The initial values for component
means can be assigned to all data points in the training set; in an extreme case, it is possible to distribute the component
number equal to the data point number. This algorithm gradually fine-tunes the number of mixture components by casting
out element of Normal distributions that are irrelevant for the data modeling process (Paalanen, 2004).
240 Furthermore, Figueiredo and Jain (2002) employed the Minimum Message Length (MML) criterion (Wallace and
Dowe, 1999) as an index for model selection; the application of this criterion for the case of GMM learning leads to the
following objective function (Figueiredo and Jain, 2002):
14
N n C n C ( N  1)
( | X )   ln( i )  nz ln( )  nz
2 i:i 0 12 2 12 2
 ln L( X , ) (12)
where n denotes the size of the training set, N represents the number of hyper-parameters needed to construct a Gaussian
245 distribution, and Cnz is the number of Gaussian distribution component featuring nonzero weight (  i  0 ). Accordingly,
the EM method is then utilized to minimized Eq. 12 with a fixed number of Cnz .
In detail, the EM algorithm is employed to estimate  i as follows:
n
N
max{0, ( wit )  }
2
inew  k t 1
n (13)
N
j 1
max{0, ( 
t 1
w jt )  }
2
Accordingly, the parameters  i  inew are updated based on Eq. 10 and 11, respectively. The algorithm stops
new
and
250 when the relative decrease in the objective function  (  | X ) becomes smaller than a preset threshold (e.g. 10-5).
3.4 Radial Basis Function Fisher Discriminant Analysis for Latent Variable Generation
In machine learning, discriminant analysis presents a highly useful technique to construct relevant input patterns from
the original data set. This technique aims at unraveling the underlying structure of the data which is helpful for pattern
recognition. Introduced by Mika et al. (1999), the Radial Basis Function Fisher Discriminant Analysis (RBFDA) is an
255 extension of the Fisher Discriminant Analysis for dealing with data nonlinearity. RBFDA can be conveniently utilized to
project the feature from the original learning space to a projected space that expresses a high degree of class separability
(Theodoridis and Koutroumbas, 2009). Using this kernel technique, the data from an input space I is first mapped into a high
dimensional feature space F. Hence, discriminant analysis tasks can be performed nonlinearly in I.
Herein,  (.) is defined as a transformation from an input space I to a high dimensional feature space F, to compute w
260 (the projecting vector), it is necessary to maximize the Fisher discriminant ratio as follows:
w T S B w
J ( w)  (14)
w T SW w
15
where S B  ( m1  m2 )( m1  m2 ) T (15)
C Nk
SW   ( ( x i )  mk )( ( x i )  mk ) T (16)
k 1 i 1
1 Nk
m k   ( xik )
Nk i 1
(17)
265 To obtain w, the kernel trick is applied. Thus, one only needs to establish a formulation of the algorithm which only
requires dot-product  ( x ). ( y ) of the training data and employ kernel functions which calculate  ( x ). ( y ) . The widely-
employed Radial Basis Kernel Function (RBKF) is expressed in the following formula (with σ denotes the kernel function
bandwidth):
 x y 2 
K ( x, y)  exp   (18)
 2 2 
 
270 Since a solution of the vector w lies in the span of all data samples in the projected space, the transformation vector w
is shown in the following formula:
N
w    i ( x i ) (19)
i 1
1 N Nk
From Eq. (17) and Eq. (19), we have: wT mk  
Nk j 1
 k ( x , x
i 1
j j
k
i )  T Mk (20)
Nk
1
where M k 
Nk
k(x
i 1
j , x ik )
275 Taking into account the formulas of J (w ) , S B , as well as Eq. (20), we can restate the numerator of Eq. (14) in the
following manner:
w T S B w   T M (21)
where M = (M1-M2)(M1-M2)T
16
Similarly, based on the Eq. (17) that defines m k , the denominator of Eq. (14) can be demonstrated in the following
280 way:
w T S W w   T N  (22)
2
where N  K
k 1
k ( I  1lk ) K kT ; Kk denotes a N-by-Nk kernel matrix with a typical element is k ( x n , x mk ) . I represents the
identity matrix and 1l k is a matrix within which all positions are 1/ lk .
Considering all Eq. (14), Eq. (21), and Eq. (22), the solution of RBFDA can be found by maximizing:
 T M
285 J ( )  (23)
 T N
The optimization problem with the objective function expressed in Eq. (23) is found by identifying the primal
eigenvector of N-1M. Based on the optimization results, an input patter in I is projected on to a line defined by the vector w in
the following manner:
N
w. ( x )   i k ( xi , x ) (24)
i 1
290 4 The proposed Bayesian Framework for Flood Susceptibility Prediction
4.1 The established GIS database
To formulate a flood assessment model, the first stage is to construct a GIS database (see Figure 4) within which
locations of past flood events, maps of topographic feature, Landsat 8 data, maps of geological feature, and precipitation
statistical records are acquired and integrated. In this study, the data acquisition, processing, and integration were performed
295 with ArcGIS (version 10.2) and IDRISI Selva (version 17.01) software packages.
Furthermore, a C++ application has been developed by the authors to transform the flood susceptibility indices into a
GIS format for ArcGIS implementation. Accordingly, the compiled outcomes are employed to form a database that includes
the aforementioned flood conditioning features with two class outputs: “flood” and “no-flood”. As mentioned earlier, a total
of 76 flood locations has been recorded. To balance the dataset and reliably construct the flood prediction model, 76
17
300 locations of non-flood areas are randomly sampled and included for analysis. Hence, the total database consists of 152 data
samples.
Topographic
Landsat-8 Geological and Climatological
Map
OLI Imaginery Mineral Map Data
DEM
GIS Data Processing
Distance to Stream
Slope Elevation Curvature TWI SPI NVDI Lithology Rainfall
river density
Field Work
Handhold GPS
Historical Cases of Flood

Occurrences
Landsat-8 Flood Inventory GIS
OLI Imaginery Map Database  Conditioning factors
 State of flood occurrence
Local documents
Figure 4 The established GIS database

305
4.2 The Proposed Model Structure
The proposed model for flood vulnerability assessment that incorporates RBFDA, the Bayesian classification
framework, and GMM is presented in this section of the study. As stated earlier, the model employs RBFDA as the first level
of analysis to generate a latent input factor. This RBFDA-based latent factor aims at expressing a data projection that
310 features the highest data separation. The Bayesian classification framework coupled with GMM is utilized in the next level
of analysis within which this framework analyzes all relevant conditioning variables and the newly created latent variable to
derive the evaluation result. The overall flowchart of the proposed Bayesian framework based on GMM and RBFDA for
flood susceptibility prediction, named as BayGmmKda, is demonstrated in Figure 5.
18
Start
Historical Preliminary Feature Mutual Information

Data Selection Criterion
BayGmmKda
Grid Search Procedure
Model Selection
Training
Data RBFDA Training
BayGmmKda
Training
GMM Training
Testing BayGmmKda
Data
Bayesian Framework
Prediction
Flood Susceptibility
Evaluation Results
315
Figure 5 The proposed BayGmmKda
To construct and evaluate the prediction model, the whole dataset, including 152 data samples, are separated into two
sets: Training Set (90% or 137 samples) employed for model establishing and Testing Set (10% or 15 samples) used for
320 model testing. It is noted that the input variables of the dataset has been normalized using the Min-Max normalization; the
purpose of data normalization is to hedge against the situation of unbalanced variable magnitudes.
Furthermore, it is beneficial to equip the model with an initial feature selection procedure in which the relevancies of
flood influencing factors are examined. In this research, the mutual information (Kwak and Choi, 2002;Hoang et al., 2016), a
widely employed criterion for feature selection in machine learning, is selected to express the pertinence of each influencing
325 factor to the flood evaluation outcomes (either flood or no-flood). Basically, the mutual information can be defined as a
measure of the mutual dependence between the two random variables; this criterion quantifies the amount of information
that can be attained about one random variable through the information of another one (Qian and Shu, 2015). It is noticed
that the larger the mutual information, the stronger the relevancy between the flood conditioning factor and the class output.
In addition, to establish the BayGmmKda model, it is required to provide the hyper-parameters of the quantity of
330 mixture components (k) used in GMM and the kernel function bandwidth (  ) used in RBFDA. It is worth reminding that to
19
train GMM for density estimation, the model employs two methods: the EM algorithm and the unsupervised algorithm. In
case of the EM algorithm, this study employs the Akaike information criterion (AIC) (Akaike, 1974) to identify an
appropriate the number of k. The value of k is allowed to vary from 1 to 20; AIC is then used to select a model that exhibits a
good fit to the data and concurrently requires a few number of mixture components, which indicates less complexity (Olivier
335 et al., 1999). When the unsupervised GMM learning (Figueiredo and Jain, 2002) is used, the model starts with a maximum
component number of 20, the algorithm autonomously carries out the model selection process by removing irrelevant
mixture components.
On the other hand, a simple grid search procedure is performed to select a suitable value of the kernel function
bandwidth (  ) used in RBFDA. Within this grid search, the available values of  can be one of the following set:
340 {0.01, 0.05, 0.1, 0.5, 1, 5, 10, 50, 100} ; in addition, the Training set is subdivided into Training Subset 1 (90%)
and Training Subset 2 (10%). Thus, the Training Subset 1 plays the role as a training set for constructing BayGmmKda with
a value of  ; the appropriateness of a hyper-parameter is expressed by the rate of correct classification with the Training
Subset 2. The parameter  corresponding to the highest classification accuracy rate is selected for prediction phase.
When a suitable set of tuning parameters, including the number of mixture component k and the kernel function
345 bandwidth  ), are properly specified, the training phases of RBFDA and GMM can be carried out. Based on the whole
Training Set, RBFDA construct a discriminant analysis based latent variable. GMM is trained based on the original input
factors and the RBFDA-based latent factor. Consequently, based on the class conditioning likelihood estimated by GMM,
the Bayesian classification framework is formulated to derive a class output (either flood or no-flood) to a novel input
pattern.
350 4.3 The Developed Matlab Interface of BayGmmKda
It is noted that GMM with the EM training algorithm is implemented with the Matlab statistical toolbox (MathWorks,
2012b); meanwhile, the BayGmmKda performs the unsupervised algorithm with the program code provided by Figueiredo
(2002). The RBFDA algorithm and the unified BayGmmKda model have been coded in Matlab by the authors. In addition, a
software program with a graphical user interface (GUI) (see Figure 6) for BayGmmKda model implementation has been
20
355 coded in Matlab environment by the authors. The GUI development aims at providing a user-friendly system for performing
flood susceptibility predictions.
Figure 6 Main Menu of BayGmmKda
As shown in Figure 6, the program consists of three modules: Data Process and Visualization, Model Training, and
360 Model Prediction. The first module provides basic functions for data inspection and visualization including data
normalization, data viewing, and preliminary feature selection with mutual information. In the second module, the users
simply provide model parameters including the kernel function parameter and the GMM training method. The trained model
is employed to carry out prediction tasks in the third module, within which the model prediction performance is reported.
5. Experimental Results
365 5.1 BayGmmKda Implementation
The outcome of the preliminary examination on the pertinence of flood influencing factors is reported in Figure 7. As
mentioned earlier, the relevancies of influencing factors are exhibited by the mutual information criterion. Based on the
outcome, IF5 (SPI) features the highest mutual dependence, followed by IF7 (stream density) and IF8 (NVDI). Influencing
factors of IF4 (TWI) and IF10 (rainfall) exhibit comparatively low values of mutual information. Because all of the mutual
370 information values are not null, all conditioning factors deem to be relevant and should be retained for the subsequent
processes of model training and prediction.
21
Figure 7 Mutual information of flood conditioning factors
375 It is worth reminding that the BayGmmKda’s training phase are executed in two consecutive steps: RBFDA training and
GMM training. RBFDA analyzes the data in Training Set to establishes a latent factor which is a one-dimensional
representation of the original input pattern. Figure 8 illustrates a typical latent factor constructed by RBFDA. In the next
step of the training phase, GMM is constructed by the original input patterns with their corresponding labels which consists
of ten input factors and with the RBFDA -based latent feature.
380
Figure 8 RBFDA-Based Latent Factor Construction
22
Commonly, the classification accuracy rate, denoted as CAR, is employed to exhibit the rate of correctly classified
instances. In addition, a more detailed analysis on the model capability can be presented by calculating true positive rate
TPR, false positive rate FPR, false negative rate FNR, and true negative rate TNR are also commonly utilized to exhibit the
385 predictive capability of a prediction model (Hoang and Tien-Bui, 2016). The four metrics are calculated as follows:
TP FP FN TN
TPR  ; FPR  ; FNR  ; TNR  (25)
TP  FN FP  TN TP  FN TN  FP
where TP, TN, FP, and FN represents the values of true positive, true negative, false positive, and false negative,
respectively.
Besides, the two indices of TPR and FPR can be graphically summarized by means of Receiver Operating Characteristic
390 (ROC) curve (van Erkel and Pattynama, 1998). The ROC curve basically demonstrates the trade-off between the two
aforementioned TPR and FPR when the threshold for accepting the positive class of ‘flood’ varies. In addition, the area
under the curve of ROC, or AUC for short, can be employed to quantify the classification performance; generally, a better
model is characterized by a larger value of AUC.
As aforementioned, the data set are randomly separated into the Training Set and the Testing Set which occupy 90% and
395 10% of the data samples, respectively. The Training Set is employed to train the mode; meanwhile, the Testing Set is used
for validating the model capability after being trained. Since one selection of data for the Training Set and the Testing Set
may not truly demonstrate the model’s predictive capability, this study carries out a repetitive sub-sampling procedure within
which 30 experimental runs is carried out. In each experimental run, 10% of the data set is retrieved in a random manner
from the database to constitute Testing Set; the rest of the database are included in the Training Set.
400 The testing performance of the proposed Bayesian framework for flood susceptibility is reported in Table 2 and Figure
9, which provides the average ROC curves of the proposed model framework, obtained from the random subsampling
process, with two methods of GMM training. Herein, the two Bayesian models that employ the EM algorithm and the
unsupervised learning algorithms for training GMM are denoted as BayGmmKda-EM and BayGmmKda-UL, respectively.
As can be seen from this table, BayGmmKda with the unsupervised learning algorithm demonstrates clearly better predictive
405 performance (CAR = 89.58%, AUC = 0.94, TPR = 0.96, TNR = 0.91) than that of the BayGmmKda with EM algorithm
23
(CAR = 86.67%, AUC = 0.93, TPR = 0.95, TNR = 0.85). The performances of BayGmmKda-EM and BayGmmKda-UL are
comparable in true positive rates. However, BayGmmKda-UL deems more accurate than BayGmmKda-EM when the two
models predicts patterns with negative class label (no-flood).
Table 2 Prediction Results of BayGmmKda
Data Set CAR (%) AUC TPR FPR FNR TNR

Average
BayGmmKda-EM 86.67 0.93 0.95 0.12 0.15 0.85
BayGmmKda-UL 89.58 0.94 0.96 0.12 0.09 0.91
Standard deviation
BayGmmKda-EM 6.51 0.07 0.05 0.10 0.12 0.12
BayGmmKda-UL 7.22 0.05 0.04 0.11 0.10 0.10
410
Figure 9 ROC plots of the proposed BayGmmKda

5.2 Result Comparison
In this experiment of the current research, the result of the BayGmmKda model is benchmarked with those of other
415 machine learning models including the support vector machine (SVM), the adaptive neuro fuzzy inference system (ANFIS),
and the GMM-based Bayesian Classifier. The above soft computing models are chosen for result comparison because SVM
and ANFIS have been recently verified to be effective tools for predicting flood vulnerability (Tien Bui et al.,
24
2016b;Tehrany et al., 2015b). It is noted that the GMM-based Bayesian Classifier, denoted as BayGmm for short, is the
Bayesian framework for classification which employs GMM for density estimation, but is not integrated with the RBFDA
420 algorithm.
Furthermore, comparison between BayGmmKda and BayGmm is helpful to confirm the advantage of the newly
constructed BayGmmKda and to verify the usefulness of RBFDA in enhancing the discriminative capability of the hybrid
framework. Furthermore, since the performance of BayGmmKda-UL is better than that obtained from BayGmmKda-EM,
the proposed BayGmmKda as well as the BayGmm trained by the unsupervised approach (Figueiredo and Jain, 2002) for
425 GMM learning is selected for accuracy comparison in this section.
To construct the SVM model, the model’s hyperparameters of the regularization constant (C) and the parameter of the
radial basis kernel function (  ) need to be specified. Herein, a grid search process, that is identical to the one used to
identify the kernel function bandwidth used in RBFDA, is employed to fine-tuned such hyperparameters of the SVM. It is
noted that SVM method is implemented in Matlab package (MathWorks, 2012a). Meanwhile, the ANFIS model used in this
430 section is trained with the metaheuristic approach described in the previous work of Tien Bui et al. (2016b).
Table 3 Prediction Result Comparison
Models CAR (%) AUC TPR FPR FNR TNR

Average
BayGmmKda 89.58 0.94 0.96 0.12 0.09 0.91
ANFIS 85.63 0.83 0.84 0.13 0.16 0.87
BayGmm 85.02 0.92 0.82 0.13 0.17 0.88
SVM 83.75 0.82 0.78 0.10 0.22 0.90
Standard deviation
BayGmmKda 7.22 0.05 0.04 0.11 0.10 0.10
ANFIS 6.17 0.05 0.14 0.10 0.14 0.10
BayGmm 7.24 0.08 0.11 0.10 0.11 0.10
SVM 10.33 0.06 0.16 0.11 0.16 0.11
It is noted that a random subsampling with 30 runs is employed for all models in this experiment. The result comparison
between the proposed BayGmmKda and other benchmark methods is reported in Table 3. The experimental outcome shows
435 that out that the proposed BayGmmKda, which is a hybridization of GMM, RBFDA and Bayesian framework, yields the
25
best value of CAR (89.58%, AUC = 0.94). ANFIS (CAR = 85.63%, AUC = 0.83) is the second best model, followed by
BayGmm (85.02%, AUC = 0.92) and SVM (83.75%, AUC = 0.82).
Furthermore, to better confirm the superiority of the proposed method, the Wilcoxon signed-rank test, a non-parametric
test, is employed to analyze the model prediction outcomes. This approach is commonly employed to evaluate whether
440 classification outcomes of prediction models are significantly dissimilar. It is noted that the assumption of normality in the
data is not required for the implementation of the Wilcoxon signed-rank test. It is noted that the significance level of the
statistical test is usually set to be 0.05.
Using this test, the p-values st obtained from experimental results of the four models have can be computed; based on
the threshold value of 0.05, if the p-value falls below the threshold, we can conclude that the prediction outcomes attained
445 from the two models are significantly dissimilar. Outcomes of the Wilcoxon signed-rank test is reported in Table 4. It is
noted that the signs “++”, “+”, “--”, and “-” represents a significant win, a win, a significant loss, and a loss, respectively.
Thus, the test reliably confirms that proposed BayGmmKda achieves significant wins over all other models (ANFIS,
BayGmm, and SVM).
Table 4 Model Comparison Based on the Wilcoxon signed-rank test
BayGmmKda ANFIS BayGmm SVM

BayGmmKda ++ ++ ++
ANFIS -- + +
BayGmm -- - +
SVM -- - -
450
5.3 Construction of Flood Susceptibility Map with the Proposed BayGmmKda Model
Experimental outcomes have indicated that BayGmmKda can deliver the best flood susceptibility prediction for the
district of Tuong Duong (in Central Vietnam). Accordingly, in this section, the proposed model is utilized to compute the
flood vulnerability index for all pixels within this studied district. The computed outcomes have been converted to GIS
455 format and used to construct a flood susceptibility map by means of the ArcGIS 10.2 software package.
The flood susceptibility map (see Figure 10) has been combined with the map of flood inventory to compute the
percentage of the flood locations and the percentage of the susceptibility map; accordingly, we can obtain a graphic curve
26
which visualizes five flood hazard levels: very high (10%), high (10%), moderate (10%), low (20%), and very low (50%). A
further examination on the constructed flood vulnerability map reveals that that 10% of the Tuong Duong district was
460 categorized as ‘very high’ and this category covers 73.68% of the total number of flood locations. Meanwhile, the levels of
‘high’ and ‘moderate’ both cover 10% of the region and account for 15.79% and 7.9% of the flooded locations, respectively.
The levels of ‘low’ and ‘very low’ aggregately cover 70% of the map but they only contain 2.63% of the total number of
flood locations. Particularly, 50% of the district, which is categorized to the hazard level of ‘very low’, contains no flood
location. Furthermore, it can be seen that approximately 89% of the recorded floods occurred in the areas of ‘very high’ and
465 ‘high’ levels of flood hazard as indicated by BayGmmKda. This fact strongly verifies the reliability of the proposed
Bayesian framework.
Figure 10 The constructed flood susceptibility map for the study area
27
6. Conclusion
470 This research has developed a new tool for flood susceptibility, named as BayGmmKda, evaluation with data collected
in the Tuong Duong district (in Central Vietnam). The newly constructed model is a Bayesian framework for classification
with an integration of GMM for density approximation and RBFDA for discriminant analysis. A GIS database has been
formulated to train and test the BayGmmKda method. The training phase of BayGmmKda can be break down into two steps:
discriminant analysis with RBFDA in which a latent factor is generated and density estimation using GMM. After the
475 training phase, the Bayesian framework is employed to compute the posterior probability of the two class labels (flood and
no-flood). Furthermore, a Matlab program with GUI has been developed to ease the implementation of the BayGmmKda
model in flood vulnerability assessment.
It is noted that in this study, the GMM training is performed with two methods: the EM algorithm and the unsupervised
learning approach. Furthermore, a repeated subsampling process with 30 experimental runs is carried to evaluate the model
480 prediction outcome. The subsampling process verified by statistical test confirms that the GMM method trained by the
unsupervised learning approach has attained a better prediction accuracy compared with the EM algorithm. Therefore, this
method of GMM learning is strongly recommended for other studies in the same field.
Furthermore, the experiments demonstrate that the latent factor created by RBFDA is really helpful in boosting the
classification accuracy of the BayGmmKda model. This melioration in accuracy of the BayGmmKda stems from its
485 integrated learning structure. As described earlier, the classification task is performed by a hybridization of discrimination
analysis and Bayesian framework. The Bayesian model carried out the classification task by consideration the patterns in the
original dataset and an additional factor produced from the discrimination analysis.
Result comparison pointed out that the BayGmmKda is superior to other benchmark approaches including ANFIS and
SVM. Therefore, the proposed model, featured by high accuracy and the capability of delivering probabilistic outputs, is a
490 promising alternative for flood susceptibility prediction. Future extensions of this research may include the model
application in flood prediction for other study areas, investigations of other flood conditioning factors which may be relevant
for flood analysis, and improving the current model with other novel soft computing methods (e.g. feature selection, pattern
classification, dimension reduction, metaheuristic optimization, etc.).
28
7. Code availability
495 The Matlab code of the BayGmmKda model is given in the Supplement.
8. Data availability
The dataset used in this research is given in the Supplement.
Acknowledgements
Data for this research are from the Project No. B2014-02-21 and were provided by Dr. Quoc-Phi Nguyen (Hanoi University
500 of Mining and Geology, Vietnam)
References
Akaike, H.: A new look at the statistical identification model, IEEE Trans. Automat. Control., 19, 716–723,
10.1109/TAC.1974.1100705, 1974.
Arellano, C., and Dahyot, R.: Robust ellipse detection with Gaussian mixture models, Pattern Recognit., 58, 12-26,
505 http://dx.doi.org/10.1016/j.patcog.2016.01.017, 2016.
Biernacki, C., Celeux, G., and Govaert, G.: Choosing starting values for the EM algorithm for getting the highest likelihood
in multivariate Gaussian mixture models, Comput. Stat. Data. Anal., 41, 561-575, http://dx.doi.org/10.1016/S0167-
9473(02)00163-9, 2003.
Bubeck, P., Botzen, W., and Aerts, J.: A review of risk perceptions and other factors that influence flood mitigation behavior,
510 Risk. Anal. , 32, 1481–1495, 10.1111/j.1539-6924.2011.01783.x, 2012.
Candy, A. S., Avdis, A., Hill, J., Gorman, G. J., and Piggott, M. D.: Integration of Geographic Information System
frameworks into domain discretisation and meshing processes for geophysical models, Geosci. Model Dev. Discuss., 2014,
5993-6060, 10.5194/gmdd-7-5993-2014, 2014.
Cheng, M.-Y., and Hoang, N.-D.: Slope Collapse Prediction Using Bayesian Framework with K-Nearest Neighbor Density
515 Estimation: Case Study in Taiwan, J. Comput. Civ. Eng., 30, 04014116, doi:10.1061/(ASCE)CP.1943-5487.0000456, 2016.
Chou, J.-S., and Tsai, C.-F.: Concrete compressive strength analysis using a combined classification and regression
technique, Autom. Constr., 24, 52-60, http://dx.doi.org/10.1016/j.autcon.2012.02.001, 2012.
Dottori, F., Martina, M. L. V., and Figueiredo, R.: A methodology for flood susceptibility and vulnerability analysis in
complex flood scenarios, J. Flood Risk. Manage., n/a-n/a, 10.1111/jfr3.12234, 2016a.
520 Dottori, F., Salamon, P., Bianchi, A., Alfieri, L., Hirpa, F. A., and Feyen, L.: Development and evaluation of a framework
for global flood hazard mapping, Adv. Water Resour., 94, 87-102, http://dx.doi.org/10.1016/j.advwatres.2016.05.002, 2016b.
Figueiredo, M. A. T.: http://www.lx.it.pt/~mtf/, Access Date: 01/04/2016, 2002.
Figueiredo, M. A. T., and Jain, A. K.: Unsupervised learning of finite mixture models, IEEE Trans. Pattern Anal. Mach.
Intell., 24, 381-396, 10.1109/34.990138, 2002.
525 Gómez-Losada, Á., Lozano-García, A., Pino-Mejías, R., and Contreras-González, J.: Finite mixture models to characterize
and refine air quality monitoring networks, Sci. Total Environ., 485–486, 292-299,
http://dx.doi.org/10.1016/j.scitotenv.2014.03.091, 2014.
Hoang, N.-D., and Pham, A.-D.: Hybrid artificial intelligence approach based on metaheuristic and machine learning for
slope stability assessment: A multinational data analysis, Expert. Syst. Appl., 46, 60–68,
530 http://dx.doi.org/10.1016/j.eswa.2015.10.020, 2016.
Hoang, N.-D., and Tien-Bui, D.: A Novel Relevance Vector Machine Classifier with Cuckoo Search Optimization for
Spatial Prediction of Landslides, J. Comput. Civ. Eng., 30, 04016001, 10.1061/(ASCE)CP.1943-5487.0000557, 2016.
Hoang, N.-D., and Tien Bui, D.: Predicting earthquake-induced soil liquefaction based on a hybridization of kernel Fisher
discriminant analysis and a least squares support vector machine: a multi-dataset study, B. Eng. Geol. Environ., 1-14,
535 10.1007/s10064-016-0924-0, 2016.
29
Hoang, N.-D., Tien Bui, D., and Liao, K.-W.: Groutability estimation of grouting processes with cement grouts using
Differential Flower Pollination Optimized Support Vector Machine, Appl. Soft Comput., 45, 173-186,
http://dx.doi.org/10.1016/j.asoc.2016.04.031, 2016.
Jia, X., Morel, G., Martell-Flore, H., Hissel, F., and Batoz, J.-L.: Fuzzy logic based decision support for mass evacuations of
540 cities prone to coastal or river floods, Environ. Modell. Softw., 85, 1-10, http://dx.doi.org/10.1016/j.envsoft.2016.07.018,
2016.
Ju, Z., and Liu, H.: Fuzzy Gaussian Mixture Models, Pattern Recognit., 45, 1146-1158,
http://dx.doi.org/10.1016/j.patcog.2011.08.028, 2012.
Judi, D. R., Burian, S. J., and McPherson, T. N.: Two-Dimensional Fast-Response Flood Modeling: Desktop Parallel
545 Computing and Domain Tracking, J. Comput. Civ. Eng., 25, 184-191, doi:10.1061/(ASCE)CP.1943-5487.0000064, 2011.
Kazakis, N., Kougias, I., and Patsialis, T.: Assessment of flood hazard areas at a regional scale using an index-based
approach and Analytical Hierarchy Process: Application in Rhodope–Evros region, Greece, Sci Total Environ., 538, 555-
563, http://dx.doi.org/10.1016/j.scitotenv.2015.08.055, 2015.
Khanmohammadi, S., and Chou, C.-A.: A Gaussian mixture model based discretization algorithm for associative
550 classification of medical data, Expert. Syst. Appl., 58, 119-129, http://dx.doi.org/10.1016/j.eswa.2016.03.046, 2016.
Kia, M. B., Pirasteh, S., Pradhan, B., Mahmud, A. R., Sulaiman, W. N. A., and Moradi, A.: An artificial neural network
model for flood simulation using GIS: Johor River Basin, Malaysia, Environ. Earth Sci., 67, 251-264, 10.1007/s12665-011-
1504-z, 2012.
Kreft, S., Eckstein, D., Junghans, L., Kerestan, C., and Hagen, U.: Global climate risk index 2015: Who suffers most from
555 extreme weather events, Report from Germanwatch, 1-31, 2014.
Kwak, N., and Choi, C.-H.: Input feature selection by mutual information based on Parzen window, IEEE Trans. Pattern
Anal. Mach. Intell., 24, 1667-1671, 10.1109/TPAMI.2002.1114861, 2002.
Lee, M. J., Kang, J. e., and Jeon, S.: Application of frequency ratio model and validation for predictive flooded area
susceptibility mapping using GIS, In Proc. of the 2012 IEEE International Geoscience and Remote Sensing Symposium,
560 2012, 895-898,
Loo, Y. Y., Billa, L., and Singh, A.: Effect of climate change on seasonal monsoon in Asia and its impact on the variability
of monsoon rainfall in Southeast Asia, Geosci. Front., 6, 817-823, http://dx.doi.org/10.1016/j.gsf.2014.02.009, 2015.
MathWorks: Bioinformatics Toolbox, The MathWorks, Inc., 2012a.
MathWorks: Statistics Toolbox, The MathWorks, Inc., 2012b.
565 McLachlan, G., and Peel, D.: Finite Mixture Models, Wiley-Interscience; 1 edition, Printed United States 2000.
McLachlan, G., and Krishnan , T.: The EM Algorithm and Extensions, 2nd Edition, Wiley Series in Probability and
Statistics, John Wiley & Sons, Hoboken, New Jersey, USA, 2008.
Mika, S., Rätsch, G., Weston, J., Schölkopf, B., and Müller, K.: Fisher discriminant analysis with kernels, In Proc. of the
1999 IEEE Neural Networks for Signal Processing, Madison, WI , 23 Aug 1999-25 Aug 1999, 41–48,
570 10.1109/NNSP.1999.788121, 1999.
Mukerji, A., Chatterjee, C., and Raghuwanshi, N. S.: Flood Forecasting Using ANN, Neuro-Fuzzy, and Neuro-GA Models,
J. Hydrol. Eng., 14, 647-652, doi:10.1061/(ASCE)HE.1943-5584.0000040, 2009.
Olivier, C., Jouzel, F., and Matouat, A. E.: Choice of the Number of Component Clusters in Mixture Models by Information
Criteria, In Proc. of the Vision Interface ’99, May 18-21 1999, Trois-Rivieres, Quebec, Canada, 74 – 81, 1999.
575 Paalanen, P.: Bayesian classification using Gaussian mixture model and EM estimation: implementations and comparisons,
Technical Report, Department of Information Technology, Lappeenranta University of Technology, 2004.
Papaioannou, G., Vasiliades, L., and Loukas, A.: Multi-criteria analysis framework for potential flood prone areas mapping,
Water. Resour. Manage., 29, 399–418., 2015.
Pulvirenti, L., Pierdicca, N., Chini, M., and Guerriero, L.: An algorithm for operational flood mapping from Synthetic
580 Aperture Radar (SAR) data using fuzzy logic, Nat. Hazards Earth Syst. Sci., 11, 529-540, 10.5194/nhess-11-529-2011, 2011.
Qian, W., and Shu, W.: Mutual information criterion for feature selection from incomplete data, Neurocomputing, 168, 210-
220, http://dx.doi.org/10.1016/j.neucom.2015.05.105, 2015.
Radmehr, A., and Araghinejad, S.: Developing Strategies for Urban Flood Management of Tehran City Using SMCDM and
ANN, J. Comput. Civ. Eng., 28, 05014006, doi:10.1061/(ASCE)CP.1943-5487.0000360, 2014.
30
585 Reynaud, A., and Nguyen, M.-H.: Valuing Flood Risk Reductions, Environ. Model. Assess., 21, 603-617, 10.1007/s10666-
016-9500-z, 2016.
Rezaeianzadeh, M., Tabari, H., Arabi Yazdi, A., Isik, S., and Kalin, L.: Flood flow forecasting using ANN, ANFIS and
regression models, Neural. Comput. & Applic., 25, 25-37, 10.1007/s00521-013-1443-6, 2014.
Sanyal, J., and Lu, X.: Application of remote sensing in flood management with special reference to monsoon Asia: a
590 review, Nat. Hazards, 33, 283–301, 2004.
Scott, D. W.: Multivariate Density Estimation Theory, Practice, and Visualization, 2nd Edition, John Wiley & Sons,
Hoboken, New Jersey, United States, 2015.
Seckin, N., Cobaner, M., Yurtal, R., and Haktanir, T.: Comparison of Artificial Neural Network Methods with L-moments
for Estimating Flood Flow at Ungauged Sites: the Case of East Mediterranean River Basin, Turkey, Water. Resour. Manage.,
595 27, 2103-2124, 10.1007/s11269-013-0278-3, 2013.
Shahabi, H., and Hashim, M.: Landslide susceptibility mapping using GIS-based statistical models and Remote sensing data
in tropical environment, Scientific Reports, 5, 9899, 10.1038/srep09899, 2015.
Shahangian, B., and Pourghassem, H.: Automatic brain hemorrhage segmentation and classification algorithm based on
weighted grayscale histogram feature in a hierarchical classification structure, Biocybern. Biomed. Eng., 36, 217-232,
600 http://dx.doi.org/10.1016/j.bbe.2015.12.001, 2016.
Tehrany, M. S., Pradhan, B., and Jebur, M. N.: Flood susceptibility analysis and its verification using a novel ensemble
support vector machine and frequency ratio method, Stoch. Environ. Res. Risk. Assess., 29, 1149-1165, 10.1007/s00477-
015-1021-9, 2015a.
Tehrany, M. S., Pradhan, B., Mansor, S., and Ahmad, N.: Flood susceptibility assessment using GIS-based support vector
605 machine model with different kernel types, CATENA, 125, 91-101, http://dx.doi.org/10.1016/j.catena.2014.10.017, 2015b.
Theodoridis, S., and Koutroumbas, K.: Pattern Recognition, Academic Press, Elsevier Inc., Printed in the United States of
America, 2009.
Theodoridis, S.: Machine Learning: A Bayesian and Optimization Perspective, Academic Press, Elsevier, Printed in The
United States, 2015.
610 Tien Bui, D., Nguyen, Q. P., Hoang, N.-D., and Klempe, H.: A novel fuzzy K-nearest neighbor inference model with
differential evolution for spatial prediction of rainfall-induced shallow landslides in a tropical hilly area using GIS,
Landslides, 1-17, 10.1007/s10346-016-0708-4, 2016a.
Tien Bui, D., Pradhan, B., Nampak, H., Bui, Q.-T., Tran, Q.-A., and Nguyen, Q.-P.: Hybrid artificial intelligence approach
based on neural fuzzy inference model and metaheuristic optimization for flood susceptibilitgy modeling in a high-frequency
615 tropical cyclone area using GIS, J. Hydrol., 540, 317-330, http://dx.doi.org/10.1016/j.jhydrol.2016.06.027, 2016b.
van Erkel, A. R., and Pattynama, P. M. T.: Receiver operating characteristic (ROC) analysis: Basic principles and
applications in radiology, Eur. J. Radiol., 27, 88-94, http://dx.doi.org/10.1016/S0720-048X(97)00157-5, 1998.
Wallace, C. S., and Dowe, D. L.: Minimum Message Length and Kolmogorov Complexity, Comput. J., 42, 270-283,
10.1093/comjnl/42.4.270, 1999.
620 Webb , A. R., and Copsey, K. D.: Statistical Pattern Recognition, John Wiley & Sons, United Kingdom, 2011.
Winsemius, H. C., Van Beek, L. P. H., Jongman, B., Ward, P. J., and Bouwman, A.: A framework for global river flood risk
assessment, Hydrol. Earth. .System. Sci, 17, 1871–1892, 10.5194/hess-17-1871-2013, 2013.
Zhang, G., Mahfouf, M., Abdulkareem, M., Gaffour, S.-A., Yang, Y.-Y., Obajemu, O., Yates, J., Soberanis, S. A., and
Pinna, C.: Hybrid-modelling of compact tension energy in high strength pipeline steel using a Gaussian Mixture Model
625 based error compensation, Appl. Soft Comput., 48, 1-12, http://dx.doi.org/10.1016/j.asoc.2016.06.007, 2016.
Zhou, Z., Liu, S., Zhong, G., and Cai, Y.: Flood Disaster and Flood Control Measurements in Shanghai, Nat. Hazards Rev.,
Just Released, 10.1061/(ASCE)NH.1527-6996.0000213, 2016.
31

A Bayesian Framework Based On Gaussian Mixture Mod

Uploaded by

Copyright:

Available Formats

A Bayesian Framework Based On Gaussian Mixture Mod

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Bayesian Framework Based On Gaussian Mixture Mod

Uploaded by

Copyright:

Available Formats

Geosci. Model Dev. Discuss., doi:10.

A Bayesian Framework Based on Gaussian Mixture Model and

Radial Basis Function Fisher Discriminant Analysis for Flood

Spatial Prediction (BayGmmKda V1.1)

Dieu, Tien Bui1 and Nhat-Duc, Hoang2

of Southeast Norway (USN), Hallvard Eikas Plass, N-3800, Bø I Telemark, Norway.

Correspondence to: Nhat-Duc Hoang ([email protected])

useful tool for flood susceptibility evaluation.

resulting in more than 170,000 deaths (Kazakis et al., 2015).

40 accurate flood prediction is of great usefulness.

communication with the public in a scientiﬁc manner.

discrimination with respect to the two classes of ‘flood’ and ‘no-flood’.

discusses some conclusions on this work.

70 2 A Review of Related Works on Flood Susceptibility Prediction

the topography, geology, land cover, and infrastructure.

vulnerability with the use of a GIS database.

105 3 Research Method

3.1.1 The study area

difficult to approach for rescuing or evacuation purposes.

Figure 1 Tuong Duong district (Central Vietnam)

region under investigation.

3.1.3 Flood influencing factors

Table 1 Flood influencing variables and their category descriptions

Factors Notation Description of factor categories

1 (<1); 2 (1 to 2); 3 (2 to 3); 4 (3 to 4); 5 (4 to 5); 6 (5 to 6); 7 (6 to 7);

1 (<1); 2 (1 to 3); 3 (3 to 5); 4 (5 to 7); 5 (7 to10); 6 (10 to 15);

1 (<40); 2 (40 to 80); 3 (80 to 120); 4 (120 to 200); 5 (200 to400);

145 region are illustrated in Figure 2.

Difference Vegetation Index, (i) Lithology, (j) Rainfall

3.2 Bayesian Framework for Pattern Classification

inferring flood susceptibility.

as follows (Webb and Copsey, 2011):

computing the p(X|Ck) quantity.

3.3 Gaussian Mixture Model for Density Estimation

180 3.3.1 Gaussian Mixture Model (GMM)

al., 2016;Khanmohammadi and Chou, 2016;Ju and Liu, 2012).

is computed as a weighted summation of Gaussian distributions (see Figure 3):

Figure 3 Structure of Gaussian Mixture Model (GMM)

algorithms are briefly reviewed in the next section of the paper.

These two steps of the EM procedure are stated as follows:

3.3.3 Unsupervised learning of ﬁnite mixture model

following objective function (Figueiredo and Jain, 2002):

In detail, the EM algorithm is employed to estimate  i as follows:

where S B  ( m1  m2 )( m1  m2 ) T (15)

is shown in the following formula:

identity matrix and 1l k is a matrix within which all positions are 1/ lk .

the following manner:

290 4 The proposed Bayesian Framework for Flood Susceptibility Prediction

4.1 The established GIS database

GIS Data Processing

Historical Cases of Flood

Figure 4 The established GIS database

4.2 The Proposed Model Structure

flood susceptibility prediction, named as BayGmmKda, is demonstrated in Figure 5.

Historical Preliminary Feature Mutual Information

350 4.3 The Developed Matlab Interface of BayGmmKda

flood susceptibility predictions.

Figure 6 Main Menu of BayGmmKda

365 5.1 BayGmmKda Implementation