ıbul 2004 
orest, 
takes 
es at 
where 
and 
aling, 
| Case 
cover 
SAR 
nones 
me is 
y, the 
ently 
r map 
order 
cover 
X the 
caling 
ire. to 
ERS- 
es of 
- land 
, only 
pixel 
to be 
t with 
y are 
lution 
terest. 
and a 
pixel 
reases 
meter 
f6m 
| "The 
t. The 
els of 
RSAR 
scaled 
earest 
es the 
mpled 
nto à 
ution 
hat in 
e with 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
continuing upscaling to lower spatial resolution while the 
proportion of unclassified pixels will increase (Gupta et al., 
2000). An aggregation of classes will occur. 
  
  
Primary Secondary Pastures Recently 
forest forest fai ___Cut areas 
AIRSAR map 68.41% 17.98% 6.51% 7.11% 
Approach 1 68.39% 17.94% 6.70% 6.97% 
Approach 2 68.08% 18.17% 6.63% 7.11% 
  
Table 2. Comparison of upscaling approaches based on the 
proportion of land cover classes 
In this research, upscaling was done by nearest neighbour 
resampling of classified data, so there was neither a problem of 
class aggregation nor of an increase of unclassified pixels. 
As shown in Table 2, both in the first and in the second 
upscaling approach barely any change in the proportion of land 
cover classes could be discovered. By comparing the 
percentages of the land cover classes, it can be seen that the 
proportions stayed almost the same. 
4.1.2 Changes in number and size of patches 
As another possibility to detect changes caused by the process 
of upscaling of the AIRSAR land cover map, unique identifiers 
were assigned to the pixels with the same class names that are 
horizontally, vertically and diagonally connected. This was 
applied for each upscaling step. These connected areas are 
called patches, since a patch is a set of neighbouring pixels of 
the same class. The output was a map in which the connected 
areas are coded. Furthermore, an attribute table was created for 
the output map containing the size of the unique output units. 
A 3 x 3 filter was moved over the map and a value was assigned 
to the centre pixel of the filter in the output map depending on 
the values of the centre pixel itself and its eight neighbouring 
pixels in the input map. 
  
8 
Number of patches 
a a 
8 8 
8 
  
Le] 
6 7 8 9 10 11 12 
  
Figure 3. Comparison of upscaling approaches based on the 
number of patches 
By comparing the two upscaling approaches, a strong decrease 
in the number of patches was found in both. In the first 
approach, stepwise upscaling with interim results, 763 out of 
2638 patches remained at a resolution of 12.5 m. In the second 
approach, direct upscaling to different levels from the same 
basis, 913 patches out of 2638 remained. As shown in Figure 3, 
both curves have a similar course with a small difference along 
the y-axis. Nevertheless, the direct upscaling approach gave a 
larger number of remaining patches and better visual results. 
959 
Furthermore, the size and the shape of the patches were kept 
longer during the direct upscaling process. 
With regard to the size of the patches, it can be stated that in 
both upscaling processes small patches loose their shape or 
disappear during upscaling to lower spatial resolution and 
bigger patches remain longer and keep their recognizable shape 
as well. 
4.2 Conformity of land cover maps 
The conformity of the land cover maps was determined with the 
help of the so-called cross operation. This operation performs 
an overlay of two land cover maps and compares the class 
values on the same positions in both maps. The combinations of 
class values that occur are stored. The output is a cross map and 
a cross table. The cross table includes all combinations of the 
input classes of both maps and the number of pixels for each 
combination. With the help of the cross table a cross matrix is 
calculated to compare two land cover maps by evaluating the 
number of matching pixels. 
  
  
  
  
  
  
Land cover class | Conformity 
AIRSAR and ERS-1 (until 28-09-1993) 
Pasture 59% 
Pasture and secondary forest 0% 
Primary forest 91% 
Recently cut areas 0% 
Secondary forest 16% 
AIRSAR and ERS-1 (until 05-09-1994) 
Pasture 63% 
Pasture and secondary forest 0% 
Primary forest 86% 
Recently cut areas 0% 
Secondary forest 24% 
  
  
  
Table 4. Conformity of cross maps 
When assessing the conformity of two land cover maps, the 
same logic as in an ordinary confusion matrix is used. Not the 
single classification but rather the difference between the two 
classifications is considered. Nevertheless, the outcome of a 
cross map is strongly influenced by the accuracies of the two 
independent classifications used for the cross matrix. 
Classification errors in either of the classifications could result 
in non-conformity of classes. 
According to Quifiones (1995), the results of the classification 
of the original AIRSAR image indicate that 89% of the 
secondary forest, 100% of the pastures, 97% of the primary 
forest and 92% of the recently cut areas were classified 
correctly. Consequently, the overall accuracy is 95%. The 
overall accuracy of the ERS-1 land cover maps varies with time 
from 65% to 70%, but they contain a mixed class of pasture and 
secondary vegetation. The pixels of this mixed class were 
included into the other classes, which influences the result of 
the conformity. The mismatches occurred mainly between this 
mixed class of the ERS-1 classification and the AIRSAR land 
cover classes secondary forest and pastures. 
First, the ERS-1 land cover map of 28-09-1993 was chosen 
because it has the same year of acquisition as the AIRSAR 
image. The low value for the conformity of the secondary forest 
is influenced by the mixed class pasture and secondary 
vegetation of the ERS-1 land cover map, but also by the fact 
that the ERS-1 sensor has difficulties in separating the 
secondary forest from primary forest and pastures. The cross