stanbul 2004 
natic image 
I2 algorithm 
llowed by a 
ated pixels. 
ising Kappa 
and use/land 
solved using 
'y ERSDAC 
n 30 to 15 m 
ontrol points 
d resampling 
During the 
reference the 
d. This result 
resolution of 
ints used for 
olution of the 
of 0.05 was 
surface, an 
ACORN 4.0 
ence model 
s of ASTER 
T Vapor were 
> same Terra 
inment from 
he method of 
used (Hill & 
ninate those 
the image 
1ese pseudo- 
e adjusted by 
rence image 
4). 
e/land cover 
iven by the 
tral range of 
9 bands of 
eflectance of 
e area under 
2 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
  
    
Forest/Bamboo, Bare Soil, Straw covered Soil, "Clean" Pasture, 
"Overgrown" Pasture, Water. 
The Kappa statistics was used to evaluate the accuracy of 
classification generated, which allowed the generation of error 
matrices to compare images classified by the ART2 algorithm 
mentioned and ground truth data. Based on these error matrices, 
at Table 2 the different values of global exactness and the 
resulting Kappa coefficients are presented for the best band 
combinations [2,3,4] and [2,3,4,8] from images of both dates. 
  
  
  
  
  
  
  
Bands combinations Global accuracy (%) | Kappa 
[2,3,4] image 2003 68.7260 0.6429 
[2,3,4,8] image 2002 69.2790 0.6455 
  
  
  
  
  
  
- e 
11 
= 
i Not 
gout PY reer 7] 
à | Brash millet 
M. ien. 
r Primary fares 
{ i Dugnaded format 
A a 
| 
Mu 4 
TI 
à = 1$o0 
Werelength (nm) 
dm Ve m N kr Shert Wave Infrared 
  
Figure 2. Spectral signature for the targets corresponding to the 
defined thematic classes. 
At Figure 2 one observes that, in a general sense, the classes 
Primary Forest, Logged Forest, Degraded Forest, Bamboo 
formations, Secondary Forest, Pasture and Overgrown Pasture 
present a typical spectral behavior of vegetation, i.e. a low 
reflectance at bands 1 and 2, corresponding to the visible 
spectrum (VIS), high reflectance at band 3 (near infrared), and 
again a low reflectance at bands 4 to 9, corresponding to middle 
infrared, specially due to presence of water inside the leaves. 
The classes Overgrown Pasture and Pasture have a slight 
increase of reflectance in the SWIR, as related to classes 
Primary Forest, Bamboo and Secondary Forest, due to a lower 
water content in the leaves. This similar high reflectance in the 
SWIR occurs at Bare Soil and Straw covered soil. 
Classes referring to cultures, such as Cotton (Gossypium 
hirsutum), Pearl millet (Pennisetum glaucum), “green” and 
"dried" corn (Zea mays), have a variable spectral behavior 
depending on its' maturation state. As for the class Water, the 
increase of its' reflectance, especially in the near infrared (band 
3), is caused by suspended materials and by the low depth of it, 
because the scene was captured during the dry season. At the 
former figure one can observe that the best discrimination 
between the different targets found in the field is possible with 
band 4 (1600-1700 nm), i.e. at the middle infrared. 
On the other hand, there is an increase of reflectance at band 9 
for all classes, which is not a typical behavior of the targets, 
because there should be a lower or similar reflectance at all 
other SWIR bands. This behavior can be explained by the cross- 
talk phenomenon. 
The digital image classification was done using the 
unsupervised artificial neural network ART2, inserted in the 
program Genetic Synthesis of Artificial Neural Networks 
(SGRNA) prepared by Silva (2003). As entries into the neural 
network, different band combinations were selected: [2,3,4] 
[23.4.5], [2,3,4,6], [2,3,4,7],.[2,3,4,8]; [2.3.4.9], [1.2.3], [1,3,8], 
[3,4,6] and finally all bands [1,2,3,4,5,6,7,8,9]. To start the 
classification process using three or more bands as entries into 
the network, the initial parameters of the artificial neural 
network were used, for the combination of classified bands. 
Besides that, the neural network was trained to recognize 
patterns in the image, taking into account the amount of nine 
thematic classes defined in the field, namely: Crops, Primary 
Forest/ recently logged Forest, Degraded Forest, Secondary 
125 
Table 2 — Global accuracy and Kappa coefficients for 
classifications from the image 2002 and 2003. 
The Figure 3 and 4 presents the results of thematic 
classification with the highest Kappa value for the band 
combinations [2,3,4,8] of year 2002 and [2,3,4] of 2003. 
MAPFADEUSDY 
CORERTURA DA TIERRA 
Ragtha Nine Clés Dont (MT) 
mud rame CE 
a 
| B. 
5 i 
Ë eva ie 
a: * 
Lan Am casses 
  
    
Figure 3. Land use/land cover thematic map of 2002 obtained 
from ASTER/Terra image.