International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
approaches known as "sequential compound decision theory" and 
which attempts to decide the label for one pixel based on the 
observation at all other pixels in the image (khazenie ef al., 1990). 
More recently, another type of approaches called “global 
approaches" have been evolved (Braathen et al, 1993, 
Pieczynski, 2000). These new approaches are two types ; MAP 
(maximum a posterior probability) and MPM (maximum of 
posterior marginal probability) and both present a classic case of 
an ill-posed problem (Maroquin et al. 1987) and their solution 
must be given through stochastic or deterministic optimization 
algorithms. In this paper, we shall solve a MAP problem 
according to a Markov Random Field (MRF) model which 
provides a methodological framework to avoid combinatorial 
problem and effectively incorporate contextual information 
through ICM (Iterated Conditional Modes) which is a 
deterministic optimization algorithm. The supervised contextual 
classifier is applied on data set of SPOT multispectral image 
acquired on February 1986 covering initially, an agricultural region 
sited in northern of Algeria and contains the famous city of Blida. 
Our objective is to obtain an exploitable land cover map to use it 
later for change detection process. 
2. NOTATION 
We assume that a classified image X and observed data Y are 
realisations of stochastic processes X and Y, respectively. 
yz [r! dy y*] are multispectral data observed through K 
spectral bands and are supposed to be acquired on a finite 
rectangular lattice W = ls =(ij):1<s<S}, sis the site of the 
ijth pixel and S is lattice's area. The set y^ wl e where 
s 55V 
k = I, 2, ...K, denotes the data taken at the k/h wavelength, 
where Bez NG} and NG is the number of observable 
grey levels. It is also possible to describe the multispectral data 
with y = Ty <= si where y = bi! Mos is a feature 
vector observed on the site s called also a spectral signature on 
site s. 
3. PUNCTUAL CLASSIFICATION APPROACH 
Image classification can be done visually, by visual interpretation 
of the data, or digitally where numerical procedures, usually 
statistically based decision rules, automate the classification 
process. While a visual classification is superior in the 
interpretation of spatial information (textural and contextual 
information), computers can handle the spectral information more 
efficiently. Conventional digital classifiers, called also punctual 
classifiers, are entirely based on spectral pattern recognition. 
Indeed, n punctual classification, the spectral signature y, which 
represents the observed intensity vector is the only aspect used to 
classify a pixel on site s. The parameters of the distribution are 
learnt from training samples in a supervised classification 
approach, and from test image pixels by suitable clustering 
method in an unsupervised approach. The pixels of the image are 
then classified by calculating, from their observed response, the 
likelihood that they have come from different classes. By this 
procedure, it can be seen that the decision taken for a pixel is 
based solely on the response to that pixel. For this. reason, 
techniques based on this approach have been called “punctual or 
blind approaches” (Braathen er al., 1993). These approaches 
have been widely used for classification and have given fairly 
good results for a wide variety of images (Desachy, 1991). The 
most used supervised punctual method is a maximum likelihood 
method where the analyst supervises the classification by 
identifying representative areas, so called training zones. These 
zones are then described numerically and presented to the 
computer algorithm which classifies the pixels of the entire scene 
into the respective spectral class that appears to be most alike. In 
a maximum likelihood classification, the distribution of the 
response pattern of each class is assumed to be normal 
(gaussian). It means that the feature vector observed y, is drawn 
from a “gaussian distribution”. So, the likelihood probability to 
assign a pixel y, to the class x, is given as fellows: 
E 
Where {ix and Y, are statistic parameters of class xs 
s 
P Gy)» 0-1 Xs Y X (»-i w)- hin E 
  
estimated during training step process. The decision to assign one 
pixel from the analysed scene to a particular class is then given as 
follows: 
pex if P(yfx)) P(vxi) for each jzi (2) 
The accuracy of such methods is very much affected by a “salt 
and pepper" appearance characterizing misclassification of some 
pixels. It means that intensity vector is insufficient and then leads 
to incorrect classification of pixels. In particular of remotely 
sensed data, adjacent pixels are related or correlated, both 
because imaging sensors acquire significant portions of energy 
from adjacent and because ground cover types generally occur 
over a region that is large compared with the size of a pixel. 
Using coherent contextual information for classification efficiency 
and accuracy in remote sensing has long been desired. Contextual 
information is important for the interpretation of a scene. When a 
pixel is considered in isolation, it may provide incomplete 
information about the desired characteristics. However, the 
consideration of the pixel in its context, more complete 
information might be derived. We can define three kinds of 
context: 1) spectral context, 2) spatial context and 3) temporal 
context (Khedam e: a/.; 2001). The basic philosophy in non 
punctual approaches is that the response and class of two 
spatially adjacent pixels are highly related. For example, if (i, j) 
and (m, n) are two neighbouring pixels and if (i, j) belongs to 
class k, then there is a high possibility that pixel (m, n) also 
belongs to the same class k. Therefore, the decision for a pixel is 
taken based not only on the observation at (i, j) but also on all 
observations at (m, n) where (m, n) is neighbour of (7, j). Non 
punctual approaches can be contextual or global (MAP and 
MPM) approaches. We are interested in this paper on MAP 
approach. 
4. MAP CLASSIFICATION APPROACH 
In term of global approach where the class assigned to a site 
depends not only on the spectral feature of the site itself, but also 
on the spectral feature of all pixels in the image, our goal is to find 
the optimal classified image X : -Ix, rm X Fa based on the 
observed data Y. Each site of the segmented image is to assigned 
Intern 
into oi 
numbe 
classif 
point c 
Wher 
equati 
X MAI 
The n 
prior c 
2001). 
doesn' 
becon 
prior | 
P(Y/X 
Y give 
the fé 
distril 
labelle 
P(X) « 
as ene 
equati 
poster 
AM A1 
Wher 
Wher 
estimz 
regula 
functi 
q fror 
4-cont 
equati 
V 
(i.i