of land covers. Each elements (i, j) of the time-domain co- 
occurrence matrix is defined as probability that two pixels with 
a specified time-separation delta-t in the same spatial position 
have pixel value i and j. Conventional co-occurrence 
matrix(that is spatial domain co-occurrence matrix) represents 
spatial texture while the proposed time-domain co-occurrence 
matrix represents time-series signature. 
Figure 3a shows pixel values of annual time-series data 
conceptual. The time-domain co-occurrence matrices shown in 
Figure 3b are derived from this time-series data in the case of 
one month separation. That is, a time-series changing pattern of 
pixel values produces the corresponding probability distribution 
pattern in the matrix. Time-domain co-occurrence matrix takes 
advantage of robustness against data loss and noise derived 
from cloud and undesirable fluctuation of calculated reflectance 
values. 
* Deciduous Forest * Desert 
E 
a & 
0 
= # 
F (1 6,68 à 6 6 6,8 à 9 
> 8 E 
© 
X i d 
a 
ge 
a 
  
JAN A 
FEB à 
MAR 4 
APR A 
MAY A 
JUN A 
UE = 
AUG A 
SEP i 
SOU dj 
NOV 
DEC 4 
(a) annual time-series 
  
(b) time-domain co-occurrence matrices 
Figure 3. Conceptual examples of time-domain co- 
occurrence matrix. 
In our experiments, two kinds of pixel value were examined. 
The first one is surface reflectance. The second one is spectral 
cluster that is extracted by clustering in seven spectral bands for 
46 scenes data set. It is expected that spectral clusters absorb 
undesirable fluctuation of surface reflectance. And time 
separation delta-t from one to six months were examined in 
order to search proper delta-t. 
3.2 Classifier 
The non-parametric minimum distance classifier was introduced 
for time-domain co-occurrence matrix. Euclidean distance 
dE(x,c) and cosine distance dn(x,c) between a pixel-x and a 
training class-c were examined in this experiments. The 
distance dE(x,c) and dn(x,c) are defined as Eq.(1) and Eq.(2), 
respectively. 
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 
de(x,c)= } | 2) 2. (Maij) -MesCij))2 
E à (D 
S^ 3 Y" Maij) Mes ij) 
dn(x,0)=— LS 
2 12 we , 2 2 1 MG) 
  
bz1 
Q) 
M,»(i.j) is a component (i, j) of the time-domain co-occurrence 
matrix measured from band-b in time-series data set for a pixel- 
X. M;p(i,j) is that measured from band-b time-series data set for 
the training area of a class-c. 
4. CLASSIFICATION EXPERIMENTS 
4.1 Land Cover Category 
Table 1 presents the land cover categories which are same with 
IGBP Land cover categories. These 17 categories were used in 
our classification experiments. 
Table 1. Land cover categories(IGBP legend). 
. Water EN 
. Evergreen Needleleaf Forest 
  
. Evergreen Broadleaf Forest 
. Deciduous Need leaf Forest 
. Deciduous Broadleaf Forest 
. Mixed Forests 
. Closed Shrublands 
. Open Shrublands 
9. Woody Savannas 
© NN ON tA AW ON — 
10. Savannas 
11. Grasslands 
12. Permanent Wetlands 
13. Croplands 
14. Urban and built-up 
15. Cropland/Natural Vegetation Mosaic 
  
16. Permanent snow and ice 
17. Barren/Sparsely vegetated BE 
  
4.2 Training and Accuracy Estimation 
84 classification classes were prepared for IGBP 17 categories, 
because each category consists of several classification classes. 
About 9,000 pixels on the average for each class and about 
400,000 pixels in total have been extracted as training data. 
Figure 4 and Figure 5 show examples of training data for 
"evergreen needleleaf forest" and " barren/sparsely vegetated”, 
respectively. Figure 6 shows examples of obtained time-domain 
co-occurrence matrix for "deciduous needleaf forest" and " 
savannas". 
    
  
  
   
   
  
    
   
    
  
    
    
    
   
  
   
    
  
   
    
     
    
  
   
     
     
   
    
  
  
(c)tim. 
one 
area 
Figure 
  
(c) tin 
one pi 
Figure 
(a)" 
Figure