Hyper-dimensional data acquired are separated into two 
groups, one is used for feature extraction and the other 
for estimating the classification accuracy. In the first step, 
significance-weighted features are extracted using the first 
half of the data. In the next step, the other half of the data 
are classified by using the extracted features and by using 
the spectral bands of a sensor under consideration. We use 
the accuracy of classification as an index of performance 
for a specific purpose. If the accuracy by a sensor is as 
high as the accuracy by the extracted features, the sensor 
can be considered to have sufficient performance. 
Our spectrometer for experiments does not cover all the 
spectral region of current sensors. In order to confirm 
the validity of this method, we applied it to the subset 
of bands from the Coastal Zone Color Scanner (CZCS). 
Though the CZCS has six spectral bands, only three of 
them (bands 2, 3 and 4 in Fig. 8) are covered by our spec- 
trometer. Figure 9 shows the classification accuracy for 
classes À and B, by the three bands of CZCS, and by the 
extracted features. The former was lower than the latter 
by about 6% when the number of bands is three. We know 
that the performance of the bands 2, 3 and 4 of CZCS is 
not sufficient for classifying the classes A and B. 
  
  
  
6 T T T 
n m 
— 4. Band S band 4 ; | ] 
= band | fea 
= 2 iE 
o b eM. 
3 à A 
S 0 7j NC. 
£ 2}; 1 
v 
E xL J 
-6 A 1 1 
  
  
  
500 550 600 650 700 750 
Wavelength [nm] 
Fig.8 Three bands used for experiments 
(bands 2, 3 and 4 of CZCS) 
100 
90 
80 
70 
60 
50 
40 
30 
Classification Accuracy [%] 
  
2 
Number of Bands 
Fig.9 Classification accuracy by bands 2, 3 and 4 
of CZCS and by extracted features 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
5. CONCLUSIONS 
We have proposed a feature extraction method for 
significance-weighted classification of hyper-dimensional 
data. The method was tested using 411 dimensional hyper- 
spectral data, in which one or two significant classes were 
appointed. By successively extracting features, a sufficient 
number of features to classify the prescribed classes were 
extracted. It was found that classification accuracy of par- 
ticular classes increased by more than several percents, 
compared with classification using the features extracted 
by canonical analysis. To expand this method to the case 
of more than two significant classes or of many classes in 
an image is straight forward. 
We have also presented a method for evaluating the perfor- 
mance of current sensors by comparing classification accu- 
racy with extracted features. It would be shown that the 
spectral bands of current sensors are not always optimal 
for a specific purpose, and can be improved by designing 
them appropriately. 
Extension of this method to designing spectral bands for 
a specific purpose and to extracting quantitative infor- 
mation efficiently and accurately from hyper-dimensional 
data are subjects for a future study. 
References from Journals: 
Ready,P.J. and Wintz,P.A.,1973. Information Extraction, 
SNR Improvement, and Data Compression in Multispec- 
tral Imagery, IEEE Trans. Comm., COM-21, pp.1123-1131. 
References from Books: 
Kullback,S.,1959. Information Theory and Statistics, Wi- 
ley. 
Schowengerdt,R.A.,1983. Techniques for Image Process- 
ing and Classification in Remote Sensing, Academic Press, 
pp.159-167. 
Swain,P.H. and Davis,S.M.(ed.), 1978. Remote Sensing: 
The Quantitative Approach, McGraw-Hill, pp.358-361. 
References from Other Literature: 
Kiyasu,S. and Fujimura,S., 1993. Successive Feature Ex- 
traction from Hyperspectral Data, Proc. Int. Geoscience 
and Remote Sensing Symp. (IGARSS'93), Tokyo, Japan, 
pp.469-471. 
Fujimura,S. et al., 1981. A Comparison of Automatic Clas- 
sification Algorithms for Land Use Map by Remotely Sensed 
Data, Proc. IEEE Southeastcon '81, Huntsville, Alabama, 
pp.299-303. 
Fujimura,S. and Kiyasu,S., 1994. Significance- Weighted Fea- 
ture Extraction from Hyper-Dimensional Data, Proc. SPIE, 
vol.2318, pp.63-68. 
Vane,G., 1988. First results from the Airborne Visible/ In- 
frared Imaging Spectrometer (AVIRIS), Proc. SPIE, vol.834, 
pp.166-174. 
238 
   
KEY W 
ABSTR 
Sub-pb 
cover c 
covers. 
by fuzz 
In satel 
with a 
when n 
tinuous 
inter-gi 
satellit 
types o 
Pathira 
ent cla: 
ing lab 
racy gr 
7596 cc 
pixels ( 
fact th: 
tral res 
compo 
the inh 
schem 
To get 
basical 
and te: 
classe 
Eagles 
cally, | 
cation 
this is | 
the str 
derive 
and th 
This pe 
cation 
classe 
Theap 
premis 
tion frc