h land 
round 
matic 
Then, 
ly, for 
els on 
er the 
ion of 
puted. 
er the 
RE (1) 
"each 
r this 
ed to 
T, the 
erent 
Down 
Q) 
3) 
(4) 
pixel 
) the 
G) 
ratic 
; for 
; the 
type 
able 
rces 
Bouzidi, Sonia 
  
e a temporal sequence of NOAA images covering an important part of southern Africa, 
e a Landsat TM image covering the main part of the catchment, 
e a land cover classification computed from the Landsat TM image and enhanced by ground truthing. It is described 
in table 1. 
  
  
  
  
  
  
  
  
  
  
  
Name of land cover | Relative surface of land cover 
in the thematic classification 
image 
indeterminate 8.69% 
water 18.36% 
urban 6.91% 
field 12.70% 
bare soil 4.47% 
grass 26.79% 
bushland 11.40% 
woodland 1.16% 
forest 9.52% 
  
  
  
  
Table 1: Characteristics of the Mkomazi Landsat thematic classification image. 
The process of validation is explained in figure 1. We consider a test zone covered simultaneously by the Landsat image 
and the NOAA acquisitions. We first establish the pixels’ composition using the classification image obtained from 
Landsat data (only ground pixels are considered, and a mask is applied on pixels containing water). In fact, each NOAA 
pixel contains approximately 37 x 37 Landsat pixels which can be localized by the geometrical registration. From this 
set of Landsat pixels and from the classification image we compute the composition of the NOAA pixels in terms of 
the different land cover percentages. The obtained compositions serve as reference ones and are compared to the pixels 
proportions obtained after unmixing process of the NOAA data. This can be done for each land cover type. We present 
in figure 2 and figure 3 results obtained respectively for bare soil and grass. The lighter a pixel is, the greater the value of 
proportion or composition (in%) is. 
In order to evaluate the difference between the pixels’ compositions from the Landsat classification and the proportions 
obtained after the unmixing process of the NOAA data, we compute an euclidian distance between both results. So, for 
each land cover type j, a distance value is computed using the equation (6): 
(6) 
  
where : 
e p? denotes the proportion of land cover j in pixel i obtained after the unmixing process of the NOAA pixels, 
> Ci denotes the composition of NOAA pixel i in land cover j obtained after the counting process on the thematic 
Landsat classification image, 
e n is the number of NOAA pixels used for the evaluation. 
The table 2 gives the euclidian distance values obtained for the different land cover types. The highest error values 
  
| Land cover | Euclidian Distance | 
  
  
  
  
  
  
urban 0.07 
field 0.16 
bare soil 0.05 
grass 0.16 
bushland 0.09 
forest 0.11 
  
  
  
  
  
Table 2: Euclidian distances for each land cover. 
(measured by the euclidian distance) are observed for land covers occupying important surfaces of the test area (grass 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 207