stanbul 2004 
) 
) 
) 
) 
) 
) 
) 
) 
/ 
) 
) 
) 
ments) 
‚91% for (a), 
, respectively. 
improved by 
ma site 
(Unit: 96) 
s) 
6 C7 
0 0 
0) (0) 
0 0 
0) (0) 
0 0 
0) (0) 
0 5.0 
0) (122) 
0 0 
0) (0) 
4 233.6 
5) (1645) 
.6 634 
6) (444) 
  
(0) 
65.0 
(766) 
C6 
0 
(0) 
3.9 
(260) 
0.3 
(14) 
0 
(0) 
0 
(0) 
88.2 
(1039) 
  
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
(b) Maximum likelihood method (six elements) 
  
  
  
  
  
  
  
  
  
  
  
Cp G3 HERE Wags ig CU 
CI 32.0 0- 4312 0 268 0 0 
(215) (0) (277) (0) (180) (0) (0) 
C2 0 9l. 3.0 5.9 0 0 0 
" (0) (509 (17) (33) (0) (0) (0) 
C3 0.5 02: . 97.5 1.8 0 0 0 
e (20) (7) (3792) (70) (0) (0) (0) 
Ca 0.1 12.3 6.6 81.0 0 0 0 
(2) (299) (161) (970) (1) (0) (0) 
CS 7.9 20 33212 02 68.6 0 0 
(39 (10) (104) (0) (337) (0) (0) 
C6 0 0 0. 132 0 155 713 
(0) (0) (0) (646) (0) (761) (3493) 
CI 0 1.6 0 143 0 23 3819 
(037^ (1i) (0) (100) (0 (16 (573) 
(c) Euclid distance method (twelve elements) 
Ci C2 C C4 C5 06,67 
ci 76.8 0 0 023.2 0 0 
(384) (0) (0) (0) .(116).......(0) (0) 
C2 0. 83.5 0 16.5 0: 0 0 
(0) (449) (0). ...(89) (0) (0) (0) 
C3 0 18 952 2.6 0.4 0 0 
m (0) (71) (36813 (100) (15) (0) (0) 
Ca 0 7.2 19 90.8 0.1 0 0 
(0) (171) (45) (2146) (2) (0) (0) 
cs 1.7 1.0 1.2 0 901 0 0 
(32) (4) (5) (0) (375) (0) (0) 
C6 0 0 0 0 0 840 160 
(0) (0) (0) (0) (0) (3858) (737) 
CT 0 0 0 1.5 0 5.8 92.6 
(0) (0) (0) (10) (0) (38) (605) 
(d) Maximum likelihood method (twelve elements) 
FI Fy "Ty Ca ES Ce C7 
CI 63.2 0 0 0 3638 0 0 
(316) 0 © d'in à (9 
C2 0 901] 2.6 72 0 0 0 
(0) (435). (14) . (39) (0) (0) (0) 
C3 0 0 99.9 0 0.1 0 0 
ee (0) (0) (3866) (0) (1) (0) (0) 
CA 0 5.7 69 87.3 0.1 0 0 
: (0) (136) (162) (2064) (2) (0) (0) 
cs 1.0 0 1.7 0 974 0 0 
(4) (0) (7) (0) (405) (0) (0) 
C6 0 0 0 0 0 so]. 409 
(0) (0) (0) (0) (0) (2717) (1878) 
C1 0 0.2 0 0.3 0 54 94.2 
(0) (1) (0) (2) (0) (35) (615) 
  
  
  
  
C1: urban, C2: grass, C3: paddy 1, C4: paddy 2, 
CS: factory, C6: lake, C7: river 
4.2 Classification Results for Pi-SAR Data 
Pi-SAR is an airborne polarimetric SAR that can observe with 
multi-frequency, L-band and X-band, and multi-polarization. 
However, the observation in 1998 for Hitachi site could obtain 
only two kinds of polarization in four polarizations. For that 
data, polarization synthesize and pseudo color synthesize have 
been difficult. Finding Rajski distance proposed in this paper 
for two obtained amplitude images, influence for land-cover 
classification results using feature vector that have increased 
25 
elements was estimated (Yamada, T. and Hoshi, T., 2002b). For 
this data, land-cover classification score matrices were 
calculated by most likelihood method. 
The results are shown in Table 3. In the table, (a) is the case 
that feature vector has two elements, and (b) is the case of three 
elements. In these tables, category names are defined under the 
tables. 
In table 3, average classification accuracy was 74.0% for (a) 
and 85.28% for (b), respectively. As the same as the case of 
SIR-C, classification accuracy was improved by increasing the 
elements of feature vector. 
Table 3. Classification score matrix for Hitachi site 
(Unit: 9 0) 
(a) Two elements 
  
  
  
  
  
  
  
  
C1 F3 Gs C4 C5 
CI 93.1 4.8 2.0 0 0.1 
(1489) (77) (32) (0) (2) 
C2 9.8 49.4 372 0 3.6 
(33) (169) (126) (0) (12) 
C3 1.5 17.0 64.8 0 16.7 
(14) (153) (583) (0) (150) 
CA 0 0 0 87.7 12.3 
(0) (0) (0y (3157) (443) 
CS 1.7 3.3 6.7 13:3 75.0 
(1) (2) (4) (8) (45) 
(b) Three elements 
Ci C2 C3 C4 €3 
Ci 99.3 0.2 0.1 0 0.4 
(1589) (3) (2) (0) (6) 
C2 5:1 59.8 28.7 0 6.4 
1 (17) (203) (98) (0) (22) 
C3 1.9 13.4 79.1 0 5.7 
= (17) (121) (711) (0) (51) 
CA 0 0 0.1 99.9 0 
(0) (0) (4) (3596) (0) 
C5 1.7 5.0 5.0 0 88.3 
(1) (3) (3) (0) (53) 
  
C1: urban, C2: grass, C3: forest, C4: sea, C5: shadow 
5. DISCUSSIONS 
By increasing elements of feature vector for land-cover 
classification, improvement of average classification accuracy 
was confirmed generally for Table 1, 2, and 3. Although 
classification accuracy was improved for most of category by 
increasing. elements of feature vector, that was lower for 
category C2 and C5 in Sarobetsu site by Euclid distance method. 
Sub area extracted to calculate Rajski distance also realizes 
filter effect. So misclassifications are happened in boundary of 
areas. Spatial ranges for both categories that classification 
accuracy was lower are narrow, so that the influences caused by 
misclassification in boundary parts are higher than other areas. 
This problem will be avoided by correcting the size to construct 
GLCM and selecting targets distributing widely. Actually, 
classification accuracy was increased in urban pattern of 
Kashima site distributing widely. 
For Hitachi site Pi-SAR data that suffers a loss of some 
polarization data, classification accuracies for some categories 
were improved remarkably by increasing elements of feature 
vector although peculiar elements were scanty. Feature vector