bul 2004 
imaging 
ind. The 
elds and 
as about 
1999 and 
ely 250 
flection, 
in field 
er 1999, 
he actual 
rformed. 
, 2002). 
tained 17 
1 turn for 
w in last 
pectra as 
ired from 
eas. The 
it defines 
yds were 
flectance 
‘ed under 
s may be 
ere used 
ice of the 
ed by the 
iside one 
ctra from 
s of pixel 
choosing 
letermine 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B 7. Istanbul 2004 
reference spectra. The third method used mode values as a 
reference. It means most frequently occurring value. 
Smaller test area (Figure 3) with 100 pixels was chosen for a 
more detailed analysis. Reference spectra were derived from 
this area (Figure 4). The area represents a dense spruce forest 
and it includes spruce trees, shadows and bedrock of the forest. 
  
Figure 3: 10 x 10 pixel test area of hyperspectral image. Bright 
pixels are sunny spruce crowns. Dark pixels are 
shadows between spruces. 
  
  
   
       
  
    
68 —1]— mean 
-median 
been 
mode 
48| —3-—pure 
4 ^ g af eei 
NE 3€ —á pere 
Reflectance value 
w 
co 
  
  
  
à: 2-13 4::51/6 7:48" 9110/11/12, 13, 4 15.16 7 
Channel 
Figure 4: Reference spectra calculated by mean, median and 
mode method and pure spruce spectrum. 
Besides, bright and pure pixels were chosen to represent the 
reflectance spectra of spruce. Spectra were compared with each 
other and the most descriptive spectrum was selected. Values 
were not calculated from larger group of pixels because it tends 
to generalize or flatten the shape of the spectrum. Most 
descriptive spectrum was considered as pure spruce spectrum. 
Pure spectrum and spectra calculated by different methods were 
used as reference spectra and spectral angles (Figure 1) between 
the each pixel of the area (Figure 3) and reference spectra 
(Figure 4) were calculated. 
3. RESULTS AND DISCUSSION 
3.1 Reference spectra analysis 
Comparison of different methods in gathering the reference 
spectra were analysed using the spectral angles. Spectral angle 
images (Figure 5) were stored. Image colours were inverted to 
help the interpretation of the spectral angle images. Therefore 
bright image pixels represent small values of spectral angle and 
dark pixels represent wide spectral angles. Small spectral angles 
mean that the reference spectrum describes the study area well. 
The more there are bright pixels in the spectral angle images the 
better choice that reference spectrum is for describing the test 
area. These results are significant when choosing reference 
spectra for example Spectral Angle Mapper classification. 
  
   
MecBan Mode Pure 
  
  
  
Figure 5: The spectral angles between the small test area pixels 
and the reference spectra. 
Figure 5 shows that calculating reference spectrum by using 
mean values gives slightly better results than median. By 
comparing the test area image and the spectral angle images we 
can see that reference spectra of average and median method 
cannot define dark pixels of the test area. The smallest values of 
the spectral angles were found from the image pixels between 
sunny crows and shadows. 
The spectral angle values of the reference spectrum calculated 
with the mode method were high and especially mode method 
gave bad results in the case of sunny crown pixels. Pure pixel 
method defined reference spectrum that was outstandingly good 
in case of bright crown pixels. Anyway, the results were poor in 
the dark pixel cases as expected. 
3.2 Classification 
Smaller area was selected from the AISA data and classified 
using several algorithms. The area was classified into the 
following different vegetation and soil types: roads, buildings, 
cornfield, threshed cornfield, sugar beet, deciduous and 
coniferous forests. Results of different classification algorithms 
were compared. Besides, effects of different. training areas and 
variation in illumination were investigated. 
   
   
  
Coniferous forests 
| | | Deciduous forests 
| Threshed cornfield 
| | Cornfield 
| Buildings 
Roads 
Sugar beet 
  
Figure 6: The image of the test area that was used in the 
classifications and the colour codes for the 
classification results. 
Spatial resolution of the data was good and there was some 
variation in the values of the training site pixels. Therefore 
several reference spectra per each training site were used in the 
Spectral Angle Mapper and Spectral Correlation Mapper 
classifications. The reference spectra were gathered from the 
same training sites used in the Minimum Distance and 
85