cterize different 
the Cabor and 
to characterize 
sensitivety and 
by texture sig- 
subtle discrimi- 
res. In this pa- 
| to characterize 
multiple-scales 
Performance of 
r the classifica- 
ng texture im- 
enty images in 
with few errors 
jer. The classi- 
resalutions had 
'ICATION 
Wavelet Trans- 
1, we mean the 
multiple levels 
rtion represent 
uency and spa- 
composition, in 
iformation, dl 
. horizonal edge 
e information. 
omposition and 
de original im- 
or Daubechics 
  
  
  
wavelet [6] with filter length 4. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Fig. 1 Fig. 2 
2. 2 Images Selection and Sampling 
T wenty distinct aerial geomorphy texture im- 
ages were selected from the albums of photogram- 
metry and remote sensing images. These images in- 
clude desert, dune, loess dome, mountain, forest 
and so on. Each image was digitized on a scanner at 
180dp. Some images were digitized at 150dp,180dp 
and 230dp. Each image was stored as a 512 X 512 
8bit/pixed digital image. Each image was broken 
down into forty sub-samples of size 256 X 256, in 
whcih twenty sub-samples for training and another 
twenty for testing. 
2. 3 Features Selection and Classification 
Each sub-sample was decomposed into seven- 
teen sub-images by a 4-levels orthonoral wavelet 
transform. The information entropy H (x) of each 
sub-image was used as a feature of the sub-sample. 
Here H (x) is defined by 
Hm S^? 6,5) |1og | p €, 32 | 
5j 
where 
(CG: 
/S eo. 
ij 
and C (i,j) is the number of image. Then each sub- 
p (2,5) € 
sample was represented by a vector of seventeen 
features which are used for classification. 
To decide the efficacy of wavelet transform for 
texture classification, the performance of a simple 
minimun-distance classificatier was evaluated. 
3. Results and Discussion 
3. 1 Classification on the Same of Resolution 
Supposing that decomposition levels of sub- 
sample are 1o, lj, l;, lgandl,, we experimented 
for 400 training sub-samples and 400 testing sub- 
samples. In Table 1. we show the classification re- 
  
  
sult. 
Table 1 
Numble of | Numble of 
Features correct % 
features errors 
Loslyslaslssly 17 4 99. 00 
P slookast, 16 5 98. 75 
lo ,U lla 13 8 98. 00 
lh, 1l. 12 9 97.75 
19,1, 1, 9 11 97. 25 
Liss 8 13 96. 75 
  
  
  
  
3. 2 Classification on Different Resolutions 
Six images digitized at 150dp, 180dp, 230dp 
were used for classfication under the condition of 
different resolutions. We- experimented for 360- 
. training sub-samples and 360 testing sub-samples. 
In Table 2, we show the result in our study. 
  
  
Table 2 
Numble of | Numble of 
Features correct 
features errors 
lo 50 5 la 5l3 514 17 3 99. 1667 
lg 0,0, ls 13 8 97.7777 
D 5l2 5t, 12 il 96. 9444 
  
  
  
  
3. 3 Disscussion 
Table 1 and Table 2 show that remote sensing 
texture images could be classified by wavelet trans- 
form with high correct. And the highest correct is 
hased on 17 features. Our method of classification 
can be used for many regions such as imagery 
recognition, the application of remote sensing and 
computer vision. 
REFERENCES 
[1] Haralick, R. M. , 1979. Statistical and struc- 
tural approaches to texture. Proc. IEEE, 67, 
PP. 786— 804. 
[2]Chen, C. H. , 1982. A study of texture classifi- 
cation using spectral features. in Proc. IEEE 
6th Int. Conf. Pattern Recongnit, Munich, 
1037 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996