Vol. XXXVIII, Part 7B 
In: Wagner W„ Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
images compound by 
îock-out process (table 
the features types and 
portant role in the 
ns that achieved equal 
ccuracy were ranked 
3.1 Distance definition 
A supervised classification based solely on spectral bands of the 
Ikonos image has reached an overall accuracy of 72,5%. This is 
a control value to verify the efficiency of add texture features in 
the classification process. 
re variables of texture 
13, 15 .... 45) were 
tance. We submit data 
acess as previously 
classification to each 
>ined with the spectral 
were used to find the 
After perform 150 classification according to the knock-out 
process, 68 results achieved values higher than 80% of overall 
accuracy (Combinations of lag distances 3 to 7 and windows 
size 15 to 41). The distance 3 was present in most of the results 
(34%) and there was no prevalence of a given window size. The 
texture features compound the results in the following 
percentages: ASM ( 72%), IDM (54%), Ent (50%), Corr (26%) 
and Cont (18%). However, when we consider the results with 
one single feature of texture, Ent. represents 40% of the results 
and Cont 40%. as well. The best result reached was 86,5% with 
a lag distance of 3, a window size of 41 and using IDM and Ent. 
(Table 1) 
'), Inverse Difference 
lation (Corr.). 
Bands 
Distance 
Window 
Overall 
Acurracy 
IDM+Ent+S 
3 
41 
86,5 
Corr 
Spectra 
ASM+IDM+Ent+S 
3 
35 
85 
1 
Ent+S 
3 
41 
84,9 
77,3 
59 
Ent+S 
6 
31 
84,1 
77 8 
60 3 
Ent+Vis 
5 
21 
83,9 
Cont+ASM+IDM+ 
3 
35 
83,9 
79,8 
65 
Ent+Cor+S 
40,1 
ASM+IDM+Ent+S 
3 
21 
83,8 
Cont+S 
4 
41 
83,2 
IDM+Ent+S 
5 
21 
83 
wind+dist (window + 
ASM+IDM+Ent+S 
3 
25 
83 
Table 2. Ten best classification results of Knock - out using all 
5 texture features combined with spectral bands, lag distances 
between 3 and 7 and the following window sizes: 15, 21, 25, 
31,35 and 41. 
difficult and a fully 
ne was impossible, 
fted area, most of it is 
always able to have 
»till, we were able to 
rom the interpretation 
validation data. To 
0 used a micro-light 
1 to acquire over 700 
ligital camera (Nikon 
5-200 mm 1:3.5-5.6). 
gation GPS (Global 
interval was coupled 
>hs to account for the 
a and the GPS were 
of detail on these 
plant families could 
vo botanists and the 
3.2 Window definition 
With a fixed distance of tree, 90 new classifications were 
performed following the knock-out approach and 67 results 
showed an overall accuracy exceeding 80%. Again, there was 
no prevalence of a given window size. We expected this result 
since the Knock-out technique aimed to find the best result for 
each size of window. In this case, 76% of the results were 
related with texture feature ASM, 60% with Ent, 50% with 
IDM, 30% with Cont and 20% with Corr. The results compound 
by one Texture band were analysed and the same result order 
was achieved. (ASM - 47%, Ent. - 29%, IDM - 18%, Cont. 
12% and Corr. 0%). The best result reached was 86,7% with a 
window size of 41 using only contrast. (Table 2) 
rSSION 
•ee blocks: Distance 
laviour of textures 
Bands and features 
Distance 
Window 
Overall 
Acurracy 
Cont+S 
3 
37 
86,7 
IDM+Ent+S 
3 
41 
86,5 
IDM+Ent+S 
3 
43 
86,4 
Ent+S 
3 
45 
86,4 
IDM+Ent+S 
3 
45 
86,2 
Ent+S 
3 
43 
85,7 
ASM+IDM+Ent+S 
3 
35 
85 
Ent+S 
3 
41 
84,9 
Ent+S 
3 
39 
84,8 
IDM+Ent+S 
3 
39 
84,6 
Table 3. Ten best classification results of Knock - out using all 5 
texture features combined with spectral bands, a fixed lag 
distances of 3 and odd values of window sizes between 11 and 45. 
This result was unexpected and led us to evaluate separately the 
performance of each texture features, alone and combined with 
spectral Ikono’s bands. During this process we have conducted 
180 processes, 90 of each combining texture with the spectral 
bands and 90 only with each texture feature. 56 results obtained 
values superior to 80%. Entropy responded by 32% of cases. The 
contrast was surprisingly the second best feature of texture with 
34% followed by IDM (23%) and ASM (18%). The top 3 results 
for image classification using a combination of a unique texture 
with the spectral bands were obtained with the contrast and the 
following window sizes: 37, 41 and 43. 
Window 
Contrast 
Window 
Entropy 
37 
86,7 
45 
86,4 
41 
86,5 
43 
85,7 
43 
86,5 
41 
84,9 
39 
86,4 
39 
84,8 
45 
85,6 
27 
84,2 
Table 4. Ten best results of image classifications using each 
texture feature separately combined with spectral bands, .a fixed 
lag distances of 3 and odd values of window sizes between 11 and 
45. 
When we evaluate the dispersion of the overall accuracy of 
classification by size of the window, we notice a general trend in 
♦ Texture 
Window size 
Figure 3. Graph of general dispersion of classification results 
of texture features according to window size. 
65