tanbul 2004 
> 3. Our data 
posite color 
ng is to have 
le to identify 
base for the 
assifications 
| to uncover 
without prior 
iting classes 
ing to these 
5.(a)) which 
lage (figure 
n accuracy. 
's determine 
ically based 
ng in each 
number of 
minimum of 
1). Numbers 
Or assessing 
algorithm, a 
ly, we first 
to product a 
non-optimal 
execute a 
ation result 
rameter B is 
is taken 0,8 
on result is 
assessing is 
ed between 
s matrix, is 
lton, 1991) 
Let be AY 
elements in 
the diagonal 
x and M the 
e following 
(10) 
etween the 
only some 
ssion (10) 
is another 
ah (Shabah 
class i and 
(11) 
perimental 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
"; "ESPAGNE m 
ES pon SER noa p Avi, TUNS 
     
  
fado 
« Gibraltar (GE) 
Oran 
  
  
Tanger»; 
RABAT, Pes ; 
s Moyet À Lagsoust x 2 
^ \ or 1 3 A tans TRIPOLI 
Casablanca s quss ENG JTougsourt 1 = 
SY Figg Ghardaia* NT av 1 VL M ode " 
Marrakech | oo" > Guarat Tas Massai > 
e x / "gechar S Cuargla™ ACS, i kr N 
e 4 IPOHTA 
MAROC ade > cu jv veu À | j TRIPOLITAINE 
em 108 : ’ ie Ex Y oc / 
A / Ben bes © a en aa 
/ . e) je mt Ghadamés 
à : gno ^ a" 
1, Deus oo mimos ot 09 | 
a du : a 
Md Plateau Hamad? Too i LIBYE 
Adrar. du Todemoit te4 , tt | 
»Tindouf le wie AmAmnas qiadm Sera 
e Gra Oetoiet Aue In Salah Erg Isoovan ; 
his di Ree. * 9 ‘ao s 
you! 2 ; FEZZAN 
ALGER RIE + (ou 
€t 9 Ain Adrar " 
veus! . ; 7 } Gnat 
^ d 
SAH ARA Hoggo: vl 
Djanet 777 
Plateau 
í dy » Tamanrasset > “ 
Adron tee»a — lonezrouft Udo ^ Tamers 
MAURITANIE & NIGER 
MALI s “eax 
Figure 3. The position of the study area in the northwestern part 
of Algeria, 
  
Figure 4. Data set study 
  
CLASS Object 
Less dense urban 
  
  
Less dense vegetation 
  
Blida airport 
  
Non cultivate fields 
  
Dense urban (Blida city) 
  
Cultivate fields 
  
  
  
  
  
  
Dense vegetation 
  
Table 1. Classes of study zone 
  
(b) 
Figure 5. (a) Training samples image — (b) Truth ground image 
eo 
— 
  
  
  
  
  
  
  
  
Classes Training pixels | Ground truth pixels 
] 357 220 
2 1052 483 
3 1348 620 
4 773 378 
5 643 290 
6 984 349 
7 1100 519 
  
  
  
  
  
Table 2. Numbers of training pixels and ground truth pixels 
  
Figure 6. (a) Punctual classification (initial configuration) — (b) 
ICM classification (B—0,8 and 8-connexivity) 
6. DISCUSSION 
We have presented the optimisation algorithm we used to obtain 
Maximum a posterior (MAP) classification of remotely sensed 
data. This iterative algorithm is based on Markov Random Field 
(MRF) and exploits spatial class dependencies between 
neighbouring pixels in an image. It is a simpler and faster version 
of Geman's algorithm (Geman et al., 1984). Applied on the data 
set of size 256x256, ICM convergence is reached after 13 
iterations only. For this reason, ICM classification algorithm is 
selected to keep the computational complexity of MAP approach 
at an acceptable level. Performance of the obtained 
classifications is evaluated by calculating kappa parameters 
derived from confusion matrix and given by equation 10 and 11. 
The resulted classified imagery using context is find to reveal 
globally and locally more patch-like and meaningful patterns. This 
visual interpretation is confirmed by statistics given on table 3 and 
by graph of figure 7. It is shown that the incorporation of 
contextual information leads to impressively improved results, up 
to 84% of global accuracy is achieved in comparison with the out 
put derived from traditional punctual maximum likelihood (MLLH) 
classifier where only around 70% of global accuracy is obtained. 
Also, the classification accuracy is improved for each class. 
  
  
  
Approach Kappa (%) 
Punctual (MLLH) 72.6 
MAP (ICM) 84.23 
  
  
  
  
Table 3. Global classification accuracy