International Archives of the Photogramme 
images. Therefore, a specific analysis is required for each of the 
images with a different resolution. 
Coocurrence matrix method: The distance between pixels 
(from 1 to 3) does not seem to effect on the results, so a 
distance of one pixel was used. In general, the increase of 
the window size rises the level of the accuracy in the inner 
part of the texture areas, but produces a progressive 
increase in error due to the border effect. A neighbourhood 
size of 25x25 was used, except for the forest area 3 
(Menorca), where a size of 15x15 optimized the accuracy 
results. 
e Energy filters and edgeness: A common window size of 7 
pixels was used to apply the filters, while for the post- 
processing operation the window size ranged from 7 to 15 
pixels, depending on the area. The optimal distance for the 
edgeness factor was 3 pixels. 
e Gabor filters: The main parameters are the standard 
deviation of the filter, what has an interpretation similar to 
the window size, and the frequency. After the selection 
process, banks of filters with standard deviations of 2.86, 
5.73 and 11.44, and respective frequencies of 0.3536, 
0.1768 and 0.0884 were created. They were defined by the 
six dominant directions and then averaged to eliminate the 
orientation factor. 
e Wavelet based method: Four types of wavelet families 
were tested, Daubechies 4 and 8, and Coiflet 12 and 24, as 
well as 3 different levels of fluctuations and details. The 
best results were obtained using the wavelet Coiflet-24 and 
its reconstructed details form the 3 levels, because cach 
level provides texture information from a different scale 
(figure 2). 
As a result of these preliminary tests, a reduction of the texture 
features to be used in the comparative classification process was 
made for cach of the four methods tried. 
MS+Textures 
GLCM+WV+Gabor+Energy 
GLCM+Energy 
GLCM+Gabor 
GLCM+WV(Coif24) 
Energy 
Gabor 
WV(Coif24) 
GLCM(8) 
MS 
50 55 60 
Figure 3. Overall accuracy percentages obtained for the four test areas using different methods and combinations of texture var! 
    
   
try, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
Overall Accuracy (%) 
83 : - ete ent d 
1+2 1+2+3 1+2+3+4 1+2+3+4+5 
Groups of variables 
—eDaub4 —#— Daub& —— Coif12 Coif24 
Figure 2. Results for the selection of wavelet type and level of 
decomposition used for the urban area. (Groups of variables: 
l:Original image. 2:Textural variables from original image, 
3-Variables from details of 1* level. 4:Variables from details of 
2'4 level. 5:Variables from details of 3 level). 
3.3 Comparison of methods 
The algorithm used in the classification process was the 
maximum likelihhod classifier, and two sets of texture samples 
were defined for each area: a training set and a testing set, both 
independent and chosen to be representative of the different 
classes considered. After the aforementioned selection of 
variables, several combinations of groups of variables were 
tested to compare the texture methods. The results of the 
different classifications, in terms of overall accuracy, are shown 
in figure 3. 
As expected, due to the spectral heterogeneity of most of the 
classes, the lower accuracy levels correspond to the only 
spectral classification that uses the four multispectral bands of 
the QuickBird image (only for the area of Menorca). The 
accuracy increases by combining different groups of texture 
variables. 
; 
i 
| 
i 
| 
| 
| 
} 
| 
| 
n FOREST 1 
GFOREST 2 
E) FOREST 3 
EIURBAN 
70 75 80 85 90 95 
ables. 
1411? 
  
Rega 
accu 
areas 
(34% 
repre 
the sj 
I she 
class 
FF 
[eas] 
  
===]