International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
smaller than 10, the LDB methods which take into account the 
discriminant information from the training data have better 
results. Furthermore, the matching pursuit method has the best 
accuracies when the number of features is larger than 20. One 
may notice that the best classification accuracy of some 
nonlinear wavelet-based methods is occurred when the number 
of feature is 10. This corresponded with the conclusion of ideal 
features that the L-1 features are the smallest set needed to 
classify L classes where L = 11 in this experiment (FuKunaga, 
1990). 
  
0.98 : ET 
| 
0.96 R ; 
i + ; 
opt ot bl OR 
= ur B | 
8 0.92 x i x. * 5 i Án 
3 : ¥ x 
o x 
< 
< 09 > 
9 x 
8 ; 
= 0.88 ; © Linear WFE__A 
2 {2 Non-Linear WFE 
o ® Best-Basis Algorithm (Entropy) 
0:86 <7 LDB (J-Divergence) 
PCT 
0.84 ». Matching Pursuit 
0.82 4 - 
5 10 15 20 25 30 
Number of Features 
Figure 7. Classification results using wavelet-based feature 
extraction methods 
5. CONCLUSIONS 
In this study, several feature extraction methods based on WT 
and matching pursuit algorithm are used to reduce the 
dimensionality of hyperspectral data. The experiment results 
show that the WT is exactly an effective tool for feature 
extraction. Although some wavelet-based methods such as the 
nonlinear WFE, best basis algorithm and matching pursuit are 
based on the best approximation for data representation, they 
are still effective for classification. Especially, the nonlinear 
wavelet-based methods are more effective for classification 
than linear methods. In some circumstances, the matching 
pursuit basis has better results than the best wavelet packet 
basis. In the LDB methods, the resulted features are selected 
within the subspace of wavelet packets, thus the problem of 
limited training sample size is avoided. In the future, the 
matching pursuit methods based on the discriminant 
information between different classes derived from the training 
data set will be studied for feature extraction. Furthermore, 
because the results of wavelet-based feature extraction methods 
are strongly depend on the choice of wavelet basis, the 
classification accuracies of wavelet-based features using 
different wavelets function will be tested in the future. 
6. REFERENCES 
Bellman, R., 1961. Adaptive Control Processes: A Guided Tour, 
Princeton University Press. 
Benediktsson, J. A., J. R. Sveinsson, and K. Arnason, 1995. 
Classification and feature extraction of AVIRIS data. IEEE 
Trans. Geoscience and Remote Sensing, 33(5), pp. 1194-1205. 
Coifman, R. R., Y. Meyer, and M. V. Wickerhauser, 1992. 
Wavelet analysis and signal processing. In: Wavelets and Their 
Applications, Jones and Barlett, Boston. 
Fukunaga, K., 1990. /ntroduction to Statistical Pattern 
Recognition. Academic Press, San Diego. 
Firedman, J. D. and W. Stuetzle, 1981. Projection pursuit 
regression. J. of Amer. Stat. Assoc., 76, pp. 817-823. 
Goetz, A. F. H., G. Vane, J. E. Solomon, and B. N. Rock, 1985. 
Imaging Spectrometry for earth remote sensing. Science, 
228(4704), pp. 1147-1153. 
Hughes, G. F., 1968. On the mean accuracy of statistical pattern 
recognizers. [EEE Trans. Information Theory, YT-14(1), pp. 55- 
63. 
Hsu, P. H., Y. H. Tseng and P. Gong, 2002. Dimension 
reduction of hyperspectral images for classification applications. 
Geographic Information Sciences, 8(1), pp. 1-8. 
Hsu, P.H., 2003. Spectral Feature Extraction of Hyperspectral 
Images Using Wavelet Transform. Ph.D. Thesis, Department of 
Surveying Engineering, National Cheng Kung University, 
Tainan, Taiwan, R.O.C. 
Jain, A. K., R. P. W. Duin, and J. Mao, 2000. Statistical pattern 
recognition: A Review. [EEE Trans. Pattern Analysis and 
Machine Intelligence, 22(1). pp. 4-37. 
Jimenez, L. O. and D. A. Landgrebe, 1995. High Dimensional 
Feature Reduction Via Projection Pursuit. Ph.D. thesis, School 
of Electrical & Computer Engineering, Purdue University, West 
Lafayette. 
Landgrebe, D. A., 2001. Analysis of multispectral and 
hyperspectral image data. In: Introduction to Modern 
Photogrammetry, John Wiley & Sons, Inc. 
Lee, C. and D. A. Landgrebe, 1993. Analyzing high- 
dimensional multisepctral data. IEEE Trans. Geoscience and 
Remote Sensing, 31(4), pp. 792-800. 
Lillesand, T. M. and R. W. Kiefer, 2000. Remote Sensing and 
Image Interpretation. John Wiley & Sons, New York. 
Mallat, S., 1989. A theory for multiresolution signal 
decomposition: the wavelet representation. /EEE Trans. Pattern 
Analysis and Machine Intelligence, 11(7), pp. 674-693. 
Mallat, S., Z. Zhang, 1993. Matching pursuit with time- 
frequency dictionaries. /EEE Trans. Signal Processing, 41( 12), 
pp. 3397-3415. 
Mallat, S., 1999. A Wavelet Tour of Signal Processing. 
Academic Press, San Diego. 
Pittner, S. and S. V. Kamarthi, 1999. Feature extraction from 
wavelet coefficients for pattern recognition tasks. [EEE Trans. 
Pattern Analysis and Machine Intelligence, 21(1), pp. 83-88. 
Saito, N. and R. R. Coifman, 1994. Local discriminant bases. In: 
Proceeding of SPIE. 
Tadjudin, S. and. D. A. Landgrebe, 1998. Classification of high 
dimensional data with limited training samples. Ph.D. thesis, 
School of Electrical & Computer Engineering, Purdue 
University, West Lafayette. 
888 
Re 
pr 
foi 
na 
th: 
the 
riv 
the 
reg 
Re 
de: 
ser 
dis 
im 
riv 
an 
big 
of 
fou 
she 
che 
wa 
Spe 
me 
len 
dat 
est 
SA 
2.1 
JEI 
to | 
Bes 
bra 
Rio