Unsupervised multispectral image Classification By fuzzy hidden Markov chains model For SPOTHRV Images

Faiza DAKKA, Ahmed HAMMOUCH & Driss ABOUTAJDINE
International Journal Of Image Processing (IJIP), Volume (5) : Issue (4) : 2011 446
Faiza DAKKA dakka_f@yahoo.fr
Faculty of Sciences, Department of Physics,
Mohamed V-Agdal University
Rabat B.P. 1014, Morocco
Ahmed HAMMOUCH
GIT-LGE Laboratory, ENSET,
Mohamed V-Souissi University
Rabat, B.P. 6207, Morocco
Driss ABOUTAJDINE
Faculty of Sciences, Department of Physics,
Mohamed V-Agdal University
Rabat B.P. 1014, Morocco
Abstract
This paper deals with unsupervised classification of multi-spectral images, we propose to use
a new vectorial fuzzy version of Hidden Markov Chains (HMC).
The main characteristic of the proposed model is to allow the coexistence of crisp pixels
(obtained with the uncertainty measure of the model) and fuzzy pixels (obtained with the
fuzzy measure of the model) in the same image. Crisp and fuzzy multi-dimensional densities
can then be estimated in the classification process, according to the assumption considered
to model the statistical links between the layers of the multi-band image. The efficiency of the
proposed method is illustrated with a Synthetic and real SPOTHRV images in the region of
Rabat.
The comparisons of two methods: fuzzy HMC and HMC are also provided. The classification
results show the interest of the fuzzy HMC method.
Keywords: Bayesian Iimage Classification, Markov Chains, Fuzzy Hidden Markov,
Unsupervised Classification.
1 INTRODUCTION
Hidden Markov models are used in many forms and with different spatial structures like
strings, fields or trees [3]. The popularity of these models is illustrated by the multitude and
diversity of applications that have been proposed, in signal processing in speech recognition
and also in image processing. In particular, the model of hidden Markov chains (CMC) has
been successfully used for image classification [3].
In this case, 2D images are first transformed into a 1D vector path through Hilbert-Peano [5]
to fit the one-dimensional structure of a chain Fig. 1 (a).The interest for this model comes
from the fact that when the hidden process X = (X, ..., XN) can be represented by a finite
Markov chain and when the structure of the noise is not too complex, then X can be
reconstructed from the only observed process Y = (Y1, ..., YN), Yn R,, using different criteria
for classification like MAP "Maximum A Posteriori" or MPM Maximal Posterior Mode. It is
sometimes interesting to take into account not only the uncertainty in the observations (often
due to noise), but also their vagueness (fuzzy). This imprecision may come from a motion
sensor at the time of the acquisition of a photo, or the effect of surface / volume part found in
Unsupervised Multispectral Image Classification
By Fuzzy Hidden Markov Chains Model For SPOTHRV
Images

medical imaging and satellite. The latter effect reflects the fact that a pixel of the image is the
result of integration over a surface not necessarily homogeneous, where several components
are mixed in different proportions and unknown. In terms of classification, these pixels have
an ambiguous and sometimes it is better not to get a "harsh response" from classifier (eg,
Class 0 or Class 1) but a "vague response" measured by an interval value ]0,1[. This
response reflects the ambiguity of the pixels can be observed and interesting information.
Indeed, the level of degrees or can be interpreted as the mixing ratio of class 0 versus class 1
in the pixel, or as the degree of belonging to class 0.
This paper is organized as follows: The CMC model is briefly presented in Section 2. In
Section 3 the CMC model for fuzzy image classification is detailed, and also The CMC
Parameter estimation, carried out with an extension of the algorithm Iterative Conditional
Estimation (ICE) [1, 2]. Sections 4 illustrate the comparative results of Classification
Unsupervised. Finally, Section 5 reviews the theoretical and experimental work and offers
some perspectives.
2 UNSUPERVISED CLASSIFICATION USING HMC
This section is intended to give some recalls about the HMC model [6,7] and its use for
unsupervised image classification [8]. The HMC model can be adapted to a 2D analysis
through a Hilbert-Peano scan of the image, see Fig. 1. Hence all estimation and classification
processings are applied on the 1D sequence, and the segmented 2D image is reconstructed
by using a reverse Hilbert-Peano scan from the 1D classified sequences.
FIGURE 1: Hilbert-Peano scan construction for an 8*8 image. This scan is used to transform a
2D image into a 1D signal (y), and conversely.
FIGURE 2: Hilbert-Peano scan for an 8*8 multi-component image (M = number of layers in the
image, N = number of pixels in each layer).
The Hilbert-Peano scan presents the ability to take into account the neighborhood of the pixel
of interest [6]. Fig. 2 shows the scan for a 8 * 8 multi-component image with M layers.
A. The HMC model
An image y = {y1…., yN}, N being the total number of pixels, is considered as a realization of
the 1D observed process Y = {Y1,…,YN}, each Yn is a real-valued random variable. The
segmented image x = {x1,…,xN} is considered as a realization of a hidden process X =
{X1,…,XN}, each Xn = {1,…,K} is a discrete random variable. In classical HMC modeling,
X is assumed to be a Markov chain, i.e. p (xn+1 | xn,…,x1) = p (xn+1 | xn).

The distribution of X is consequently determined by the distribution of X1, denoted by πk = p
(X1= k), and the set of transition matrices (An)1≤n≤N whose entries are an
ij = p (Xn+1 = j |Xn =
i).
We further assume that X is a stationary Markov chain, i.e. entry an
ij = aij does not depend on
n. With the following additional properties:
(i) Yn are independent conditionally to X, i.e. p (y | x) = π
N
n=1 p(yn | x), and
(ii) p (yn | x) = p (yn |xn), the distribution of the pairwise process (X,Y) can be written as
p (x, y) = πx1 fx1 (y1)
with fxn(yn) = p (yn | xn) the data-driven densities, which are assumed to be Gaussian in this
work [8].
B. Classification:
The estimation of X from Y can be done by applying the MPM Bayesian criterion:
)(maxarg=(y)N],…[1,n n kxMPM
n ξ∈∀
(1)
with ξn(k) = p (Xn = k | y) the marginal a posteriori probabilities.
The HMC model allows explicit computation of the MPM solution using the well-known
Baum’s “forward” αn(k) and “backward” βn(k) probabilities [10], modified by Devijver [9] for
computational reasons:
αn(k) ≈ P(Xn = k , Y1 = y1,…,Yn = yn) (2)
(3)
In the following, we use the numerically stable forward–backward recursions resulting from
these approximations:
Forward initialization:
for 1≤ i ≤ K.
∑≤≤ kj1
1jj
1ii
n
)(yƒ
)(yƒ
=(i)
π
π
α
(4)
Forward induction:
for n =2,…,N and 1 ≤ i ≤ K.
(5)
for 1 ≤ i ≤ K
βN(i)=1
Backward induction:
for n=N-1,…,1 et 1 ≤ i ≤ K
(6)
It can be shown that marginal a posteriori probabilities involved in MPM classification can be
written
ξn(i) = αn(i) βn(i),
(7)
and joint a posteriori probabilities Ψn(i,j) = P(Xn = i, Xn+1 = j| Y = y) as:
(8)
C. Estimation:
Before classification, all parameters involved in the CMC model have to be estimated.
θ = {πk, akl ,fk} k,l
(9)

One well-known solution is to use the EM iterative procedure [11] aiming at optimizing the
log-likelihood of data, according to the steps described in Algorithm 1, with following update
equations
)(
N
1
=,k
1
][
k
N
n
q
n
q
k ∑=
ΠΩ∈∀ ξ
(10)
q
k
N
n
q
n
q
kl
lk
∧
−
=
Π
Ψ
Ω∈∀
∑
N
),(
=â,l,k
1
1][
(11)
q
k
n
N
n
q
nq
k
yk
∧
=
∧
Π
Ω∈∀
∑
N
)(
=,k 1
][ ξ
µ
(12)
q
k
q
kn
N
n
q
nq
k
yk
∧
∧
=
∧
Π
−
Ω∈∀
∑
N
))((
=,k
2
][
1
][
2
µξ
σ
(13)
Algorithm 1 : HMC parameters estimation using EM.
q ← 0: Initialize parameters θ[0]
repeat
q ←q + 1
- Compute “Forward”
][q
nα and “Backward”
][q
nβ probabilities using equations (5) and (6).
- Compute a posteriori probabilities )(][
kq
nξ using eq. (6) and ),(][
lkq
nψ using eq. (7) and (8).
- Estimate CMC parameters θ[q] using equations (10) and (11) for Markov parameters and
equations (12) and (13) for Gaussian data-driven parameters.
until |θ[q] - θ[q-1] | <Threshold.
Iterative estimation of parameters is stopped when parameters do not vary much.
3 UNSUPERVISED CLASSIFICATION USING FUSSY HMC
A. Fuzzy Markov chains model
We consider a multi-component image of M layers. According to the Hilbert-Peano scan, we
get N series of M data, denoted by y = { 1y ,… Ny }, where ny ={
]1[
ny ,…,
][M
ny }
t
, 1≤n ≤ N.
In classical HMC approach, the aim is to classify each yn RM into a set of K classes, the
state space ={w1,…,wK}, in order to obtain the segmented chain x ={x1,…, xN} ( Fig. 2). The
segmented image is then reconstructed from x using an inverse Hilbert-Peano scan.
For the sake of simplicity, we confine our study to the K = 2 case, i.e. = {0,1}.
In fuzzy HMC context, the range of xn is now the interval = [0, 1]. In the following, εn will
denote a realization of random variable Xn and we will adopt the notation:
εn = 0 if the pixel is from class 0,
εn ]0, 1[ if the pixel is a fuzzy one,
εn = 1 if the pixel is from class 1.
B. Probabilities in fuzzy Markov chains context
As stated previously, each xn takes its value in two types of sets: a hard one {0, 1}, and a
fuzzy one defined over the range ]0, 1[.
Let δ0 and δ1 be Dirac weights on 0 and 1, and 1 the Lebesgue measure on ]0, 1[.
By taking v = δ0 + δ1 + 1 as a measure on , the distribution of Xn can be defined by a
density h on with respect to v.
If we assume that X is homogeneous and the distribution of each Xn is uniform on the fuzzy
class, P (Xn = εn) =h(εn) = πεn can be written:
h(εn = 0) = π0;

h(εn = 1) = π1;
h(εn) = π]0,1[; εn ]0, 1[;
with π0 + π1 + π]0,1[ = 1.
We can now detail the new expression for the transition probabilities of the Markov
chain nn
t εε ,1−
:
==== −−−
)/( 11,1 nnnn XXPt nn
εεεε
)()/0( 011 nnnn XXP εδε −− ==
] [ )()/( 1,011 nnnnn XXP ειεε −− ==+
)()/1( 111 nnnn XXP εδε −− ==+
Ω∈∀ − nn εε ,1 and { }Nn ,...,2∈∀
With 0,1
≥− nn
t εε and 111, =−−∫Ω
nnn
dt εεε
∏=
−
Π==
N
n
nn
txXP
2
,11
)( εεε
C. Multi-component fuzzy HMC implementation.
εn is a realization of a random variable Xn, and each yn is a realization of a random vector
ny ={
]1[
ny ,…,
][M
ny }
t
, Thus the problem is to estimate the unobserved realization x of a
random process X from the observed realization y of a random process y = { 1y ,… Ny };
Similarly to classical HMC, multi-component fuzzy HMC based image classification methods
consider the two following assumptions:
H1: the random variables Y1,…,YN are independent conditionally on X;
H2: the distribution of each Yn conditionally on X is equal to its distribution conditionally on Xn.
It is important to note that the random variables (
][m
ny )1≤m≤M are not assumed to be
mutually independent conditionally on Xn.
Assuming that distributions of (Xn,Yn,Xn+1,Yn+1) are independent of n, each state εn of the
state space (i.e. hard classes {0,1}, as well as the fuzzy class ]0,1[) is associated to a
distribution characterizing the M-dimensional observations yn:
fεn(yn) = P (Yn = yn | Xn = εn),
(14)
Given an observed sequence y = {y1,…, yN}, the joint state-observation probability is given by:
)()(),(
2
,1 111 n
N
n
yftyfyYxXP nnn εεεεε ∏=
−
Π===
(15)
In unsupervised classification, the distribution P (X = x,Y = y) is unknown and must first be
estimated in order to apply a Bayesian classification technique (MAP or MPM). Therefore the
following sets of parameters need to be estimated:
1) The set Γ characterizing the fuzzy Markov chain parameters, i.e. the initial probability
vector π = {πε} ε and the transition probabilities Ω∈∀− nnnn
t εεεε ,1_,1
2) The set ∆ regrouping the parameters of the M- dimensional distributions presented in (14),
i.e. the distributions associated with the hard classes and the fuzzy one.
D. Fuzzy Markov Chain parameters estimation
For the estimation of the parameters in ¡, we propose to use an adaptation of the general ICE
algorithm [12] which is an alternative to the well-known Estimation-Maximization (EM)
algorithm. In fact, ICE does not refer to the likelihood, but it is based on the conditional
expectation of some estimators from the complete data (x, y). It is an iterative method which
produces a sequence of estimations θq of parameter θ as follows:
1) initialization θ0, obtained with an initial classification algorithm (k-means algorithm).

2) computation of ]/),([1
yYyxq
==
∧
+
θθ where ),( yx
∧
θ is an estimator of θ.
3) Stop the algorithm when θQ-1 θQ.
This section is not intended to give a complete description of the ICE algorithm in the HMC
context, interested readers may consult [5]. Similarly to the classical case, parameters in
Γcan be calculated analytically by using the Baum-Welch algorithm [13]:
for the hard classes, the classical normalized Baum-Welch probabilities [14] can be used
directly.
for the fuzzy class, the forward and backward probabilities are defined by:
ςςαξα ξςξ dtyf nnn ,
]1,0[
11 )()()( ∫++ ∝
(16)
ςςβξβ ξξς dyft nnn )()()( 1,
]1,0[
1 ++∫∝
(17)
The integrals above can not be calculated analytically and numerical integration must be
performed. Hence, the continuous interval ]0,1[ is partitioned into a given number F of
subintervals. We thus reduce the domain of fuzzy membership degree ]0,1[ to F attributed
values, corresponding to the medium value of the sub-interval of interest (see Fig. 3), thus
resulting in a quantization of the fuzzy measure ]0; 1[.
We obtained F fuzzy classes "discrete, whose value corresponds to the fuzzy median
subinterval considered.
For example, F = 2 implies ε {0.25, 0.75 }, F = 3 implies x {0. 165,0:5, 0.825}. More F is
large, over the parameter estimation is accurate.
FIGURE 3: Partition of the continuous interval ]0,1[ into F = 4 sub-intervals. The attributed
values ε correspond to the medium value of the sub-interval of interest.
E. Multi-dimensional density estimation
At each ICE iteration, we need to estimate the multidimensional densities fεn(yn). Several
strategies from multivariate data analysis are available, depending on the assumptions made
on the statistical links between the layers, and on the choice of the shape of the one-
dimensional densities. If independence between the layers is assumed, fεn(yn) is the product
of M densities
M
nn
gg εε ,...,1
defined on R:
)()(
1
m
n
M
m
m
n ygyf nn ∏=
= εε
For example, if we consider that fεn(yn) are multidimensional
Gaussian densities, parameters estimation can be easily achieved from the first and second
moments of a M-dimensional sample. Denoting by N(m, σ
2
) the normal distribution with mean
m and variance σ2
, the pdf of the hard classes are then expressed according to:
εn = 0 : N(m0; σ20),
εn = 1 : N(m1; σ21);
The parameters ∆={m0,m1,σ0, σ1} can be estimated by computing the empirical mean of

several estimates according to ),(
1
1
1
yx
L
l
L
l
q
∧
=
+
∑= θθ , where xl is an a posteriori realization
of X conditionally on Y . It can be shown that X | Y is a non homogeneous Markov chain
whose parameters can be computed from the forward and backward probabilities: (16) and
(17).
The definition of the fuzzy measure Ā: “the pixel belongs to class 1” corresponding to the
fuzzy class ]0,1[, and its fuzzy membership function µA, allows us to estimate the fuzzy
parameters of the set ∆ in this new context. The proposed fuzzy membership function µA is
defined by:
[ ]




=
∈∀
−
−
− 10
01
11
,1
elsewhere0
)(
mmm
mm
mm
A mµ
Accordingly, the parameters of the Gaussian pdf for the fuzzy class can then be estimated
by:
mnn mm n ,0)1( εεε +−=
2
1
2
0
2
)1( σεσεσε nnn
+−=
4 MULTI-SPECTRALES IMAGES CLASSIFICATION
This section is intended to evaluate the two local classification approaches, and the benefit of
relaxing the HMCF model.
A. Validation
To validate the proposed approach we consider the synthetic image in Figure 4, we have
considered a hard class(the opaque background and the umbrella), and the
gradient between this class as the fuzzy zone (representing the border).Fig.4 present the
results of two methods of classification based on HMC model and fuzzy HMC model .
We can be seeing that the image of Figure 4 appears to be classified much more precise with
fuzzy CMC. The Taking into account a higher number of classes in the classical model of
CMC did not have much interest because the additional classes tend to specialize on
individual pixels.
FIGURE 4: The synthetic image and their results of classification with CMC model ( 4 class
and 5 class) and fussy CMC model (F=2,and F=3).
B. SPOTHRV Images Classification
1. The studied images
The image chosen for the study is that of the forest region of Rabat Morocco. This forest is an
ecological unit consisting of a typical natural vegetation and reforestation.
Figure 5 shows the satellite images studied, and its multi-spectral images (green, red and
near infrared) of the forest Rabat.

FIGURE 5: The two images studied (a and b) and their multi-spectral images (c, d and e) for
the image (a) and (f, g and h) for the image (b).
HRV image - SPOT XS of this region, was acquired in December 1999 with a spatial
resolution of 20m. in this image we extracted two test areas of size 140x90 pixels occupied
mainly by a natural forest of orkoaks and eucalyptus planting, pine and acacia.
The reality of the ground was obtained from the official plans of the forest, followed by field
verification. These areas are characterized by heterogeneity in the distribution offorest
Stands. Discrimination between different textures is not obvious to human observer.
2. Experimental protocol
In all experiments, parameters initialization was done with a k-means classifier. The ICE
algorithm was stopped after 10 iterations, assuming it has converged.
FIGURE 6: (a) and (c) Maps of the reality of the ground. (b) and (d) Results of classification
with CMC (6 class) .
Figure 6 shows images of the realities ground (a and c) and the results of classification with
CMC on both multi-spectral images of the forest in Rabat.(b and d). Computation times are
339.246seconds.
In this application, the “Sea” or the “Acacia” were considered as crisp class 0 and
“Eucalyptus” or the “Oak-cork” crisp class 1. The fuzzy measure A thus corresponds to: “the
pixel belongs to the free sea class”. The classification results depend on the partition of ]0,1[,
the choice of the F sub-intervals implies different values of fuzzy measure ε, e.g. F = 2
implies ε {0.25, 0.75}, F = 3 implies ε {0.165, 0.5,0.825}, F = 4 implies ε {0.125, 0.375,
0.625, 0.875}
Figure 7 shows images of the realities ground (a and c) and a classification results obtained
with fussy vectorial CMC models for F=2 sub-intervals(b and e) and F=3 sub-intervals (c and
f). Computation times are 545.636 seconds.

FIGURE 7: (a) and (d) Maps of the reality of the ground. Results of classification with fussy
CMC for 4 crisp class and 2 fuzzy class (K=4,F=2 ) (b) and (e) and for 3 crisp class and 3
fuzzy class (k=3,F=3 ) ) (c) and (f).
The computational complexity involved in the fuzzy vectorial HMC model is quite higher than
the one involved in the classical one. One can observe that a big number of fuzzy sub-
intervals provide a fine characterization of the observed scene. Furthermore, the
classification results produced by the fuzzy vectorial HMC model are more homogeneous.
The global shape and frontiers of the class seem to be better defined. Nevertheless, this
model clearly characterizes the frontiers between class and produces a reliable classification
of class, and then illustrates the interest of the vectorial fuzzy HMC model.
5 CONCLUSION
In this work, we propose to extend the fuzzy HMC model to unsupervised multi-spectral
image classification. In multi-component case, it is interesting not only to take into account
the uncertainty measure of the noisy observation (characteristic of probabilistic approach in
classical HMC), but also the imprecision measure of this observation (characteristic of fuzzy
approaches [6]) By adding a fuzzy measure in a statistical model, we obtain an original
model, different from classical and models. Indeed, it preserves the robustness of the
statistical classification (based on measures of uncertainty), and enriches it with the fuzzy
characteristic (measure of imprecision).The Experimental results confirm the validity of the
proposed approach. The vectorial fuzzy HMC model seems to be p promising in the field of
multi-component image classification, due to the imprecision measure ability to take into
account the multivariate information. We believe extending the fuzzy model to another model
strictly more general than the CMC, called Markov chain couple [15] could also benefit from
the fuzzy extension proposed in this paper.
REFERENCES
[1] W. Pieczynski. Statistical image segmentation. Mach. Graph. and Vis., 1 : 261-268, 1992.
[2] H. Caillol, W. Pieczynski, A. Hillion. Estimation of fuzzy Gaussian mixture and
unsupervised statistical image segmentation. IEEE Trans. on Im. Proc., 6(3) : 425-
440, 1997.
[3] W. Pieczynski. Modµeles de Markov en traitement d'images. Traitement du Signal, 20(3)
:255{278, 2003.
[4] B. Benmiloud and W. Pieczynski. Estimation des paramètres dans les CMC et
segmentation d'images. Traitement du Signal, 12(5) :433{454, 1995.

[5] W. Skarbek. Generalized Hilbert scan in image printing. In R. Klette and W. G. Kropatsch,
editors, Theoretical Foundations of Computer Vision. Akademik Verlag, Berlin, 1992.
[6] S. Derrode, G. Mercier, and W. Pieczynski, “Unsupervised change detection in SAR
images using a multicomponent HMC model,” in MultiTemp’03, (Ispra, Italy) (2003).
[7] L. R. Rabiner, “A tutorial on HMMs and selected applications in speech recognition,”
Proc. IEEE 77(2), 257–286 (1989). [doi:10.1109/5.18626].
[8] N. Giordana and W. Pieczynski, “Estimation of generalized multisensor hidden Markov
chains and unsupervised image segmentation,” IEEE Trans. Pattern Anal. Mach.
Intell.19(5), 465–475 (1997). [doi:10.1109/34.589206].
[9] P.-A. Devijver, “Baum’s Forward-Backward algorithm revisited,” Pattern Recogn. Lett.
3,369–373 (1985).
[10]L. Baum, T. Petrie, G. Soules, and N. Weiss, “A maximization technique occurring in the
statistical analysis of probabilistic functions of Markov chains,” Ann. Math. Stat. 41(1),
164–171 (1970).
[11]A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from incomplete data
via the EM algorithm,” J. Roy. Stat. Soc. 39(1), 1–38 (1977).
[12]W. Pieczynski, “Statistical image segmentation,” Mach. Graph. and Vis.,vol. 1, pp. 261–
268, 1992.
[13]L. Rabiner, “A tutorial on hidden Markov models and selected applications in speech
recognition,” in Proc. IEEE, vol. 77, no. 2, Feb. 1989, pp. 257–286.
[14]P. A. Devijver, “Baum’s forward-backward algorithm revisited,” Pattern Recognition
Letters, vol. 39, no. 3, pp. 369–373, december 1985.
[15]S. Derrode, C. Carincotte, and J.M. Boucher. Unsupervised image segmentation based
on high-order hidden Markov chains. In IEEE ICASSP, Montreal (Ca), 17-21 mai
2004.

Unsupervised multispectral image Classification By fuzzy hidden Markov chains model For SPOTHRV Images

More Related Content

What's hot (20)

Similar to Unsupervised multispectral image Classification By fuzzy hidden Markov chains model For SPOTHRV Images (20)

Recently uploaded (20)

Unsupervised multispectral image Classification By fuzzy hidden Markov chains model For SPOTHRV Images