International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
Maximum  Likelihood classifications. Principal component 
analysis was used to reduce the dimension of the data. Four 
principal components were used in the Minimum Distance and 
Maximum  Likelihood classification. Other classification 
methods were carried out fast but Maximum Likelihood 
algorithm needed more computation time. 
  
Figure 7: Classification results of Spectral Angle Mapper 
(upper left), Spectral Correlation Mapper (upper 
right), Minimum Distance (lower left) and 
Maximum Likelihood (lower right). 
Average and overall accuracies were calculated using test areas 
(Figure 8). The average accuracy is the average of the 
accuracies for each class. Overall accuracy is a similar average 
with the accuracy weighted by the proportion of test samples. 
The Maximum Likelihood algorithm led to the best results. 
Spectral Angle Mapper and Spectral Correlatoin Mapper led to 
good results as well but the results of Minimum Distance were 
worse. 
Blaverage accuracy 
Bloverall accuracy 
- 
| 
MM 
  
     
SCM [onn 
I 
MINDIS Mc 
Figure 8: The average and overall accuracies of the Spectral 
Angle Mapper, Spectral Correlation Mapper, 
Minimum Distance and Maximum Likelihood 
classification. 
86 
  
3.2.1 Variation in illumination: Variation in illumination 
affected more strongly when the Maximum Likelihood and 
Minimum Distance were used. The results of the classification 
deteriorated fast while the Spectral Angle Mapper and the 
Spectral Correlation Mapper were better in that case. 
  
  
Figure 9: The image of the test area with variance in 
illumination (up) and the classification results of 
Spectral Angle Mapper (middle) and Maximum 
Likelihood (down). 
3.2.2 Spectral Unmixing: The original pixel size of the 
hyperspectral image was enlarged ten times its original size 
because a square meter size pixel is usually represented by only 
one manmade or vegetation class. Finding suitable reference 
spectra was the hardest part in Spectral Unmixing classification. 
The proportions of different materials were approximated from 
each image pixel and algorithm generated images where pixel 
values represented the proportions of different materials (Figure 
10). Proportions were calculated from O to 100 per cent and 
bright pixels meant larger proportion. 
coniferous forest deciduous forest buildinas 
       
      
    
threshed cornfield 
Figure 10: The proportion images of the Spectral Unmixing 
Classification. 
The proportion images were compared to the original image 
(Figure 6) and there was good association between the original 
International . 
TECH eee eo 
image and | 
deciduous foi 
brighter the pi 
the proportion 
buildings hav 
and road prop 
Generally sp 
example, Ma 
(overall accu 
computation 
Correlation 
classification 
Minimum Di 
algorithm wc 
images for th 
buildings bu 
suitable refer 
There was to 
class. This s 
"attention to « 
calculated. 
Boardman, J 
spectrometry 
Sensing Sym 
remote sensi 
California, U 
Carvalho, O. 
(SCM): An 
Airborne E 
Publication C 
Kallio, K., K 
A, Vepsilài 
quality from 
types in difi 
268, pp. 59- 
Kruse, F., L 
A., Barloor 
processing 
analysis of 
Environmen 
Mäkisara, 
spectromete 
Geoscience 
Remote sens 
II, pp. 1010- 
Ruohomäki, 
M. 2002. S 
test area. Gc