IX-B8, 2012 
(D 
  
Q) 
1ain co-occurrence 
ata set for a pixel- 
-series data set for 
MENTS 
1ich are same with 
ories were used in 
P legend). 
HH 
  
  
BP 17 categories, 
ssification classes. 
h class and about 
| as training data. 
training data for 
arsely vegetated", 
ined time-domain 
eaf forest" and " 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
(a) ground view (b) training area 
  
(c)time series reflectance of = (d)time-domain 
one pixel in the training co-occurrence matrix of 
area the training area 
Figure 4. Training data example for "evergreen needleleaf 
forest". 
  
  
(a) ground view (b) training area 
   
(c) time series reflectance of (d)time-domain 
one pixel in the training area co-occurrence matrix of 
the training area 
Figure 5. Training data example for  "barren/sparsely 
vegetated". 
    
(a) "deciduous needleaf (b) " savannas" 
forest" 
Figure 6. Examples of obtained time-domain co-occurrence 
matrix. 
Classification accuracies were measured by using test samples 
of 300 pixels that were sampled randomly from training area of 
each individual class. 
4.3 Classification Results 
Because the cosine distance classifier has always produced 
several percent higher mean producer's classification accuracy 
than that of the Euclidean distance classifier regardless of the 
time-separation delta-t, only the results when the cosine 
distance is used is shown in the following. When time-domain 
co-occurrence matrix is defined with reflectance, the highest 
accuracy of 9596-9696 has been obtained for SR and NBAR 
products when time-separation delta-t is about 3 months as 
shown in Figure 7. 
  
  
e SR —9— NBAR 
  
  
Classification Accuracy [96] 
90 T T T T T 1 
1 2 3 4 5 6 
Time Separation Delta-t [month] 
  
  
  
  
Figure 7. Classification accuracies obtained by 
time-domain co-occurrence matrix which is 
defined with reflectance. 
When time domain co-occurrence matrix is defined with 
spectral cluster, classification accuracies are increased 
according to the number of clusters as shown in Figure 8. The 
highest accuracy of about 99% has been obtained for SR and 
NBAR products when time-separation delta-t is about 4 months 
as shown in Figure 9. 
  
  
  
  
  
  
  
  
  
** SR (delta-t- 5 months) 
77$ NBAR (delta-t- 4 months) 
S 100 
S 
95 
B 
o 
< 
Xm 
5 
= 85 
3 
o 80 T T T T T T T T TTA 
30 60 120 250 500 1000 2000 4000 8000 
The Number of Clusters 
  
  
  
Figure 8. The relationship between the number of clusters 
and classification accuracies obtained by time-domain 
co-occurrence matrix which is defined with spectral cluster.