the influence of region likelihood would decrease if the region 
is irregular. This work will be done in future. 
practical 
shape 1 
s wavelet 
ns image Pine 
by using 
sition, the 
xel level 
ion level 
x of each 
ally, the 
del. The 
     
> pyramid 
mentation 
feature to CR 
ation(8) d 
'ottom up E a 
scale up 
Ti na Figure3. Classification result of MSRF 
3 t : à i 3 5 SS E 
; Figure 1. Aerial image of size 626*626 
tion field 
transition TER IT et ER S aan PU 10023 lh 
rithm as el ter jj Mm = 
E E el 
A ut 
Se 
udy. The 
as spatial 
nd blue). 
ain many 
to justify 
improve 
shown in 
ze, which 4 
ial multi- : ÿ Lot 
see from D Le. o E HS T 
di f a he ; ; ; 
sponoinsc Veo poe Figure 4. classificaltion result of UMSRF 
Improve Figure 2. Initial ES rud at B 0 
ive the | 
  
   
pectively. 5, CONCLUSION 
sented by 
Compare In this paper, we have introduced a unified multi-scale MRF 
y, we can model for high resolution image classification. This method try 
tent. For to combine the advantage of pixel based and region based MRF. 
mage are By introducing region shape information into MSRF model, the 
pixels is UMSRF model can complement the pixel based information to 
nsidering classify the land cover object. The merit of the USMF is that it 
nstead of not only inherits the advantage of MSRF, but also employs the 
MSRF as information provided by initial segmentation. This is especially 
yrmation, useful to objects that has increased intra-class variation but has 
| spectral regular shape. Initial experimental results demonstrate that the 
esolution proposed approach performs comparably better than MSRF. 
Future research work involves introducing semantic information 
UMSRF to adjust the influence of region based information 
SRF are automatically to improve classification accuracy further. 
increase B 
ymparing 
sification 
nt and | 
ion show 
shape of 
Is us that 
   
: 
| 
| 
| 
; 
| 
Figure5. Ground truth used for precision estimation