International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
Mo my In» m3 my mg mg m7 Mg Mo 
  
  
  
  
  
  
  
  
  
  
  
50657 | 3809 520 165 | 118 68 63. [614 331 43 
  
  
  
Table 1. The LSE of 10 iterations by UFCLS 
The six endmembers formed an endmember signature matrix. 
Then we plugged the endmember signature matrix into (1) to 
estimate the abundance fractions of the six endmembers using 
FCLS algorithm for all pixel vectors in the subscene image. The 
results are shown in Fig.3, where the larger the abundance 
fractions are, the higher the brightness are shown in the gray 
scale image, and the smaller the abundance fractions are, the 
lower the brightness are shown. We know the six endmembers 
were extracted without prior knowledge from the image using 
UFCLS method, so which materials can they not be identified 
to be without true data of the land cover. However we can 
judge the m, endmember to be vegetation directly from the 
pseudo color image. In order to verify the results, we compare 
the classification image of vegetation in Fig. 4(a), i.e., the 
abundance fractions image of m3, with NDVI in Fig. 4(b). It is 
shown that the pixel classification image of vegetation agrees 
significantly with the NDVI image, but the contrast of the 
former is a little larger than that of the latter, so there is lack of 
the information of details and edges. It is because the values 
shown in the two images have different physical meanings, the 
former denotes the abundance fractions of vegetation, the latter 
denotes NDVI, which are generally larger than 0 even under the 
circumstance of non-vegetation. And it is also shown, in Fig. 
4(a) denoted by the circle, that the shade areas were not 
classified correctly. 
  
   
(a)m,(UFCLS) (b)NDVI 
(c) m; (CSMA) 
Figure 4. The effects of the classification of m;(UFCLS) and 
m; (CSMA) compared to NDVI 
4.2 Experiment 2 
The data considered in this section are TM data with seven 
bands from Landsat 5, which were obtained on 19 July 1991 
located in WRS123/039. Here we selected six bands 
excluding thermal infrared region. Firstly, the scene was 
corrected geometrically and registered. A subscene of size 
512x512 pixels, larger than the area in experiment 1, was 
selected from Chibi County, Hubei Province in China for study. 
A map showing the location of the study area is presented in 
Fig.5. The color composite image of the study area, of raw 
bands 4, 3 and 2 in red, green and blue respectively, is shown in 
Fig.6. 
  
Figure 5. Location of study area in Chibi County, Hubei 
Province 
  
Figure 6. The color composite image of the study area 
We applied the same UFCLS method in experiment 1 to 
process the data of the study area. Ten endmembers were 
extracted from the subscene. The classification image of 
vegetation, or the abundance fractions image of vegetation 
endmember, is presented in Fig.7. In the same way, we 
compare the classification image of vegetation with NDVI, 
shown in Fig.8. There are the same results with experiment |. 
Furthermore, we find that the pixel classification image of 
vegetation agrees much more with the color composite image of 
raw bands 4, 3 and 2 in red, green and blue respectively in Fig.6, 
in which the red area generally denotes vegetation cover, than 
with the NDVI image. For example, the A region of rectangle 
in Fig.7 maybe had less vegetation than its surrounding region, 
which is seen clearly in Fig.6 than in Fig. 8. The B and C 
region of rectangle in Fig.7 are shown nearly in black, that 
indicates there were not vegetations in the regions, but in Fig.8 
we can not find the result. The same regions are shown in close 
to turquoise in Fig.6. We know in this color composite image a 
1184