297
A DIMENSIONALITY REDUCTION ALGORITHM OF HYPER SPECTRAL IMAGE
BASED ON FRACT ANALYSIS
SU Junying a '*, Shu Ning a
a School of Remote Sensing and Information Engineering, Wuhan University, Luoyu Road 129#, Wuhan, Hubei, P.R. China,430079
jysu_sjy@sina.com
Commission VII, WG VII/3
KEY WORDS: Hyper spectral, dimensionality reduction, spectral domain, fractal analysis, feature image
ABSTRACT:
A dimensionality reduction algorithm based on feature extraction of the spectral curve using fractal analysis which considering both
the spatial characteristic and spectral characteristic of hyper spectral remote sensing image is proposed A spectral domain feature
analysis based on fractal measurement technique is designed for hyper spectral images. Fractal characteristic of spectral curve is
discussed. A brief description is given to explain the nonlinear mechanism resulting in fractal of spectral curve. And the spectral
curves of same objects are presented to show the self similar. And some computational results are given to show exponential relation
between the total length of spectral curves and the different measurement unit. A fractal dimension calculation algorithm of hyper
spectral curve is designed. A noise remove algorithm based on wavelet transformation is done before the fractal analysis of spectral
curve. Then a feature analysis procedure in spectral domain based on fractal measurement is also proposed to reduce dimensionality
of hyper spectral images. The fractal dimension value is taken as the feature of spectral curve and the fractal dimension feature
image is proposed to represent the dimensionality reduction result of hyper spectral image. Experiments of fractal dimension value
of different objects spectral curve show fractal can be used to represent the spectral feature to reduce dimensionality of hyper
spectral image. Finally, the application of fractal measurement of spectral domain feature analysis is briefly discussed.
1. INTRODUCTION
The characteristics of hyper spectral remote sensing data are
numerous channels, high spectral resolution and large amounts
of data, which makes it easy to discriminate objects in the scene,
however, the vast amounts of data not only makes it difficult for
transmission and storage, but also feature extraction and
classification. Therefore, it is very important to reduce the
dimension in the hyper spectral image analysis (C. Lee, 1993; A.
Jain, 1997). As hyper spectral sensors acquire images in very
close spectral bands, the resulting high-dimensional feature sets
contain redundant information. Consequently, the number of
features given as input to a classifier can be reduced without a
considerable loss of information. Dimensionality reduction can
general fall into feature extraction and band selection. Feature
selection techniques generally involve both a search algorithm
and a criterion function while feature selection is usually done
in the image spatial space and feature transformation and
feature extraction is used to reduce the image dimension such as
the principal component analysis, absorption features extraction
and spectral statistical analysis. Band selection is usually based
on the spectral curve of hyper spectral image which can convert
hyper spectral vector into low dimensions or one dimension.
Due to their combinatorial complexity, band selection
algorithms cannot be used when the number of features is larger
than a few tens (L. Bruzzone, 1995; L. Bruzzone, 2000; C.
Lee, 1993; C.I Chang,2006; A. Plaza,2005). And dimensionality
reduction algorithms based on feature extraction and band
selection can not combined spectral and spatial characteristic.
All these dimensionality reduction algorithms have the
disadvantages of less considering the spectral information and
low efficiency. Fractal theory is used in many applications for
the advantage to solve the non-linear system of the complex
phenomenon. Thus Fractal is usually used to solve the complex
non-linear system analysis. Fractal theory has been used in the
remote sensing research while it does not obtain the full
application (Qiu H L; Weng Q, 2003). Commonly, fractal
dimension of hyper spectral remote sensing image is calculated
by the spatial characteristic and then the fractal dimension is
used for the band selection. And the fractal dimension value is
calculated with spectral curve to unify the spectral information
to the spatial distribution feature image thus it can be used to
reduce the dimensionality of hyper spectral remote sensing
image.
In this paper, a dimensionality reduction algorithm based on
feature extraction of the spectral curve using fractal analysis
which considering both the spatial characteristic and spectral
characteristic of hyper spectral remote sensing image is
proposed A spectral domain feature analysis based on fractal
measurement technique is designed for hyper spectral images.
Fractal characteristic of spectral curve is discussed. A brief
description is given to explain the nonlinear mechanism
resulting in fractal of spectral curve. And the spectral curves of
same objects are presented to show the self similar. And some
computational results are given to show exponential relation
between the total length of spectral curves and the different
measurement unit. A fractal dimension calculation algorithm of
hyper spectral curve is designed. A noise remove algorithm
based on wavelet transformation is done before the fractal
analysis of spectral curve. Then a feature analysis procedure in
spectral domain based on fractal measurement is also proposed
to reduce dimensionality of hyper spectral images. The fractal
dimension value is taken as the feature of spectral curve and the
fractal dimension feature image is proposed to represent the
dimensionality reduction result of hyper spectral image.
Experiments of fractal dimension value of different objects
* Corresponding author. SU Junying, PhD, School of Remote Sensing and Information Engineering, Wuhan University, P.R.
China, jysu_sjy@sina.com, 86-27-63003060