ul 2004
inds for
natures
of leaf
features
rmining
g NIR
spect to
3
assess
ectance.
methods
ist 153,
pigment
species,
ellesley-
ion and
n. Crop
'chnique
ctral &
: Stress
hanisms.
nversion
orophyll
FEATURE EXTRACTION OF HYPERSPECTRAL IMAGES
USING MATCHING PURSUIT
P.H. Hsu
Associate Research Fellow, National Science and Technology Center for Disaster Reduction,
106, No. 200, Sec. 3, Hsinhai Rd., Taipei, Taiwan, Republic of China - paihui(g)naphm.ntu.edu.tw
TS Ths 5
KEY WORDS: Hyper-Spectral Sensing, Wavelet Transform, Matching Pursuit, Feature Extraction, Classification
ABSTRACT:
Hyperspectral images contain rich and fine spectral information, an improvement of land use/cover classification accuracy is
expected from the use of such images. However, the classification methods that have been successfully applied to multispectral data
in the past are not as effective as to hyperspectral data. The major cause is that the size of training data set does not correspond to the
increase of dimensionality of hyperspectral data. Actually, the problem of the “curse of dimensionality” emerges when a statistic-
based classification method is applied to the hyperspectral data. A simpler, but sometimes very effective way of dealing with
hyperspectral data is to reduce the number of dimensionality. This can be done by feature extraction that a small number of salient
features are extracted from the hyperspectral data when confronted with a limited set of training samples. In this paper, we tested
some proposed feature extraction methods based on the wavelet transform to reduce the high dimensionality without losing much
discriminating power in the new feature space. In addition, a new feature extraction method based on the matching pursuit with
wavelet packet is used to extract useful features for classification. An AVIRIS data set was tested to illustrate the classification
performance of the new method and be compared with the existing wavelet-based methods of feature extraction.
I. INTRODUCTION
Since the mid 1980s, the new technology of imaging
spectrometer with two-dimensional area arrays of detector
elements was developed to collect spectral data with a large
number of bands simultaneously (Goetz ef al., 1985). The value
of this technique lies in the ability to construct an effectively
continuous reflectance spectrum for each pixel of the sense.
Because of the large number of spectral bands, the images
acquired with imaging spectrometers are also referred to as
hyperspectral images which are distinguished from the
multispectral images with only three to ten bands. The rich and
detailed spectral information provided by hyperspectral images
can be used to identify and quantify a large range of surface
materials which cannot be identified by multispectral images.
By means of the solar reflected spectrum measured by imaging
spectrometers, a wide range of scientific researches and
applications have being proposed based on the spectral analysis
(Lillesand and Kiffer, 2000).
1.1 Curse of Dimensionality
Seemingly the high dimensionality of hyperspectral data should
increase the abilities and effectiveness in classifying land
use/cover types. However, the classification methods that have
been successfully applied to multispectral data in the past are
not as effective as to hyperspectral data. The major cause is that
the size of training data set does not adapt to the increasing
dimensionality of hyperspectral data. If the training samples are
insufficient for the needs, which is common for the
hyperspectral case, the estimation of statistical parameters
becomes inaccurate and unreliable. As the dimensionality
increases with the number of bands, the number of training
samples needed for training a specific classifier should be
increased exponentially as well. The rapid increase in training
883
samples size for density estimation has been termed the “curse
of dimensionality” by Bellman (1961), which leads to the
“peaking phenomenon” or “Hughes phenomenon” in classifier
design (Hughes, 1968). The consequence is that the
classification accuracy first grows and then declines as the
number of spectral bands increases while training samples are
kept the same. For a given classifier, the "curse of
dimensionality” can only be avoided by providing a sufficiently
large sample size. The more complex the classifier, the larger
should the ratio of sample size to dimensionality be to avoid the
curse of dimensionality. However, in practice, the number of
training samples is limited in most of the hyperspectral
applications. Furthermore, the high dimensionality of
hyperspectral data makes it necessary to seek new analytic
methods to avoid a vast increase in the computational time. A
simpler, but sometimes very effective way of dealing with high-
dimensional data is to reduce the number of dimensions (Lee
and Landgrebe, 1993; Benediktsson er a/., 1995; Landgrebe,
2001). This can be done by feature selection or extraction that a
small number of salient features are extracted from the
hyperspectral data when confronted with a limited set of
training samples.
1.2 Spectral Feature Extraction
Feature extraction is generally considered a data mapping
procedure which determines an appropriate subspace of
dimensionality M from the original feature space of
dimensionality N ( M<N ) (Fukunaga, 1990; Lee and
Landgrebe, 1993; Jain et al, 2000). The way of feature
extraction can be a linear or nonlinear data transformation.
Regardless of how the data transformation is implemented, the
feature extraction algorithm must be designed to preserve the
information of interest for a special problem such as
compression, denoising, or classification. For example, in