443 
IMPROVING HYPERSPECTRAL CLASSIFICATION BASED ONWAVELET 
DECOMPOSITION 
O. Almog M. Shoshany a , V. Alchanatis b , F. Qizel a 
a Faculty of Civil and Environmental Engineering, Technion - Israel Institute of Technology, Haifa, Israel - 
almogo@tx.technion.ac.il 
b Institute of Agricultural Engineering, ARO - The Volcani Center, Bet Dagan, Israel. 
KEY WORDS: Remote Sensing, Hyperspectral, Classification, Calibration, Analysis, Data mining. 
ABSTRACT: 
Information extraction from Hyperspectral imagery is highly affected by difficulties in accounting for flux density variation and 
Bidirectional reflectance effects. Calculation of flux density requires digital description of the surface structure at the pixel level, 
which is frequently not available at the accuracy required (if exists). The result of these shortcomings in achieving accurate radio- 
metric image calibration is reduced separability of surface types: limiting the performance of spectral classification schemes. In this 
study an alternative approach is presented: application of features of the spectral signature which mainly represent the shape of the 
spectral curve. This is achieved by applying features calculated based on Wavelet decomposition. 
1. INTRODUCTION 
Hyperspectral remote sensing involves image acquisition and 
analysis of spectral cubes, which are composed of tens and hun 
dreds of narrow spectral bands. This process is used for extract 
ing, identifying and classifying materials and environmental 
phenomena. The main assumption is that there are relations be 
tween the chemical, biological and physical properties of those 
materials and phenomena and the characteristics of their re 
flected radiation distribution. Those relations are the basis of re 
mote sensing analysis (Landgrebe, 2002; Penn, 2002). The use of 
large number of narrow bands is supposed to increase classifica 
tion accuracies. However, it seems that there are some obstacles 
in achieving these analysis improvements like: (1) various ac 
quiring conditions such as: atmospheric conditions, illumination 
and relative sensor position; (2) various materials characteristics; 
(3) Lack of adequate information regarding the surface topogra 
phy and micro-topography ; and (4) high dimensionality of in 
formation including noise added during the acquisition process. 
In this study it is suggested to improve the spectral separation be 
tween surface objects under these conditions by applying fea 
tures of the spectral signature which mainly represent the shape 
of the spectral curve. This is achieved by applying features cal 
culated based on Wavelet decomposition. 
Wavelet analysis is a space localized periodic analysis tool, 
which enables analysis of a signal in both time and frequency 
domains (Bruce et al, 2002; Kaewpijit et al, 2003; Kempeneers 
et al, 2005; Li, 2004). The reflectance signature is decomposed 
into different scale components; each scale component repre 
sents periodical behavior of the reflectance signature at that spe 
cific scale. The periodical behavior preserves the shape of the 
original reflectance signature. In an earlier study, Almog et al, 
(2006) presented that a selection of such scale components by it 
self may improve classification robustness. Here it is hypothe 
sized that applying several relationships between wavelet coeffi 
cients and the original reflectance curve would be less affected 
by illumination intensity while preserving the unique features of 
each of the signatures. For applying those relationships, we 
transformed the reflectance signal into a new domain combined 
both radiometric and geometric information of the reflectance 
signal. 
2. WAVELET TRANSFORME 
Wavelet transform is a signal periodic analysis tool, which en 
ables analyzing a signal in both time and frequency domains and 
consider it as a multi resolution process. The signal is analyzed 
by a mother wavelet function, which is translated relative to the 
signal in various extended scaling factors. 
Wavelet transform output is a pyramidal form of coefficients; 
each of them describes the correlation between the mother wave 
let function and a specific signal segment. The size and location 
are then derived from a corresponding scale and translation fac 
tors. Scaling process is achieved by stretching the mother wave 
let among its spectral axis. Each of this stretching procedure re 
duces the mother wavelet frequency, and hence reduces the 
number of translations among the signal as well. 
Initially at level 1, the mother wavelet scale is set to 1 and 
translated relative to the signal, such that each translation pro 
duces a correlation coefficient. All the coefficients calculated in 
the first step represent high frequencies hidden among the sig 
nal and called Detailed Coefficients at level 1. In the following 
step the mother wavelet is stretched, usually by a power of two, 
the level is ascended and the translation process relative to the 
signal is repeated. Figure 1 describes different mother wavelet 
scales, (la) scale = 1, the mother wavelet is detailed and en 
ables analysis of high frequencies along the signal, (lb) scale = 
2, the mother wavelet is less detailed hence it enables analysis 
of lower frequencies along the signal. The number of points that 
define the mother wavelet at scale = 2 is halved compared to the 
number of points at scale = 1, hence, the number of translation 
is halved accordingly. Finally the coefficient pyramid consists 
of the maximum number of coefficients at first level, depending 
on chosen mother wavelet. Assuming scaling factor of 2, each 
ascending level consists of half the number of coefficients of 
the level above.