Bouzidi, Sonia 
  
  
  
Figure 2: Bare soil. On the left, composition computed from the classification image; on the right, proportion obtained 
after the unmixing process of the NOAA pixels. 
  
Figure 3: Grass. On the left, composition computed from the classification image; on the right, proportion obtained after 
the unmixing process of the NOAA pixels. 
26.7996, field 12.796). These euclidian distances are computed from exact numerical values of composition and proportion 
and are not so easy to interpret even if we can see that fields and grass seem to be the land cover types difficult to analyse. 
We then propose to consider a qualitative criterion for comparison of results. So, for each land cover type, for each 
process (counting process to compute composition, unmixing process to compute proportion), and for each pixel, we 
define a label according to table 3. Consequently, for each land cover, we generate two new images of labels values (1...5) 
  
| Label | Range of composition | Range of proportion | 
1 0% < composition < 20% 0% < proportion < 20% 
20% < composition < 40% | 20% < proportion < 40% 
40% < composition < 60% | 40% < proportion < 60% 
60% < composition < 80% | 60% < proportion < 80% 
80% < composition < 100% | 80% < proportion < 100% 
Table 3: Labels and their corresponding values for proportion and composition. 
  
  
  
  
  
  
Un SI N 
  
  
  
  
  
corresponding to the two process (counting for composition, unmixing for proportion) as it is illustrated by figures 4 et 5, 
respectively for bare soil and grass. Then, we compute for each land cover, a (5 x 5) matrix L, where each value l(c, u) 
(c, u — (1...5)) represents the quantity of pixels belonging to label c for composition (computed after the counting process 
of the Landsat classification image) and retrieved with a label u for the proportion (computed after the unmixing process 
of the NOAA data). The normalized trace of this matrix is then a global measure of accuracy for the unmixing process, 
compared with the classification process of the Landsat images. The table 4 displays the results obtained for the different 
land covers. High rates of label recognition are obtained for the land covers urban, forest and bare soil. However, the 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 209