Table 2 Confusion matrix
( AVE: Leaves, F: Soil, G: Stone, H: Concrete )
ures used,
about 19% (a) Classification using first feature 6 +
jon matrix (%)
di Mont "Dre aot fall gf db 3 4
A CASS 0 20, 112. 16. iT Booed S coi
B 0 46.4 25.5 0 19.4... 8.7 0 0 9 NN
C 0 1.5 . 98.5 0 0 0 0 0 S 0
Dl 27T0::26 :0(nad635204227.0..19.30. 5.1 "d / Se gap
m El.O0. .10...0,4, 052388. 429..2.6. 14.2 5 21
E 1. 0. bin: À 0 168 49.5 7.1 25.5 à at
G 1.0 0 0 6.6. 0.5: 321.347. 250
H 0 0 0 1.0 ,.16.3, 52.0. 143 18.3 -6 1 1 1 1 i
500 550 600 650 700 750
(b) Classification using first two features Wavelength [nm]
] (96) (a) Feature 1
A B C D E F G H
1 AES 10.75: 0.5 TT 0 CDI 120 6 à
B 0 97.5 512.6 0 0 0 0 0
750 C 0 0 100 0 0 0 0 0 3 4 Lr
D 0 0 0 903 3.1 1.0 4.1 1.5 ©
E gio gage 50 1 ROC TRAINS 000 ad Ths, /^
Fio odere Quran 69400999/0:64506:9 3 0 P
G | © OHV QUE FL -1505 13065446 415 199 > 3 tol ird
ne Hoi spage 0005930009 5/17,9 282^ 05199 c
©
S 4}
Next, we applied our method to the case where two most
significant classes A and B are appointed. The distance -6 : : i
of each class from class A and B is shown in Table 3 (a). 500 550 600 à 650 700 750
Pv e" In this case we set the first feature between A and B. The Wavelength [nm]
table of distance in this feature space is (b). In the same (b) Feature 2
way above we successively set two features. The distances
in 2 dimension are shown in (c). Fig.6 Weighting factors for the significant classes A and B
750 Table 3 Distance from A and B (Relative distance)
( A~E: Leaves, F: Soil, G: Stone, H: Concrete ) => 100 7 1
TS ner = 90 4E |
(a) Distance in 7 dimension 9 3
class A A TOD TE ECTS OUR 2:80 y
A — 41 197.183 155 161 180 171 à 70 Zt
B'l.44..— 7.0. 185 152..16.0...16.1... 17.1 5 60 7
© / edi
-—— T] (b) Distance in 1 dimension i 50 i crono memor tete
A B C D E F G H 8 40 enc s ripcuuup ut c E Are BEE rer EEE ben actes En
Al — 44 T4 1.6 3.4 2.7 1.9 2.5 o 30
B.4.4. ++ 3,0 2.8 0.9 1.7 2.5 1.9 1 2 3
ae 1 Number of Features
"Temper (c) Distance in 2 dimension Fig.7 Classification accuracy of calsses A and B
M versus number of features
RET A B C D E F G H
ia A|-- 44 92 145 155 15.0 140 146
B|44 — 62 147 152 148 1441-145
4. APPLICATION FOR SENSOR EVALUATION
ris It is seen that the two features are sufficient. The features We applied the result of feature extraction to the eval-
are shown in the form of the weighting factors in Fig. 6. uation of performance of sensors. The performance of a
The classification accuracy for A and B is about 16% (one sensor is evaluated by classification accuracy of particular
feature) and 6% (2 features) higher than that by canonical classes which are defined to be significant for a specific
analysis (Fig. 7). purpose.
237
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
