A NOVEL BAND SELECTION METHOD FOR HYPERSPECTRAL DATA ANALYSIS
B.Mojaradi 3 * , H.Emami b ., M.Varshosaz c , S.Jamali d
a Department of Geodesy & Geomatics Shahid Rajaee Teacher Training University ,P.O.Box 16785/163, Tehran , Iran
bmoj aradi@gmai 1. com
b University of Tabriz, Eng. College of Marand, Tabriz Iran, H_emami@tabrizu.ac.ir
c Department of Geodesy & Geomatics K.N.Toosi University of Technology ,Mirdamad Cross, Tehran, Iran
Mvarshosaz@yahoo.com
d Department of Geodesy & Geomatics, Faculty of Engineering, Zanjan University, Zanjan, Iran, jamali@znu.ac.ir
Commission VII, WG VII/3
KEY WORDS: Hyperspectral data, Band Selection, Search Strategy, Data Representation,
ABSTRACT:
This paper proposes an innovative band selection (BS) method called prototype space band selection (PSBS) based only on class
spectra. The main novelty of the proposed BS lies in band representation in a new space called prototype space, where bands are
characterized in terms of class reflectivity to pose reflection properties of classes to bands. Having clustered the bands by K-means
in the prototype space, highly correlated bands are trapped in a cluster. In each cluster a band that is close to cluster center identified
as representative of clustered bands. In contrast to the previous BS methods, PSBS substitutes the search strategies with K-means
clustering to find relevant bands. Moreover, instead of optimizing separability criteria, the accuracy of classification over a subset of
training data is used to decide which band subset yield maximum accuracy. Experimental results demonstrated higher overall
accuracy of PSBS compared to its conventional counterparts with limited sample sizes.
1. INTRUDUCTION
Hyperspectral technology, compared to the multi-spectral, is
capable of reconstructing spectral signatures of phenomena, as
well as producing a spectral library for earth observation. The
NASA Jet Propulsion Laboratory (NASA/JPL) airborne
visible/infrared imaging spectrometer (AVIRIS) and the Naval
Research Laboratory HYperspectral Digital Imagery Collection
Experiment (HYDICE) are two types of such sensors that
collect image data with hundreds of spectral bands. Since the
signatures of phenomena are sampled systematically in narrow
band width with fixed sample intervals, not all bands are
essentially useful for information extraction. Hence, in the
context of hyperspectral data analysis, determination of
informative bands needs to be considered for efficient
representation of phenomena.
Recent Support Vector Machine (SVM) algorithms have
demonstrated good performance in dealing with high
dimensional data (Mao, 2004). In the relevant, dimensionality
reduction, however, is still required (due to the intrinsic
dimensionality of the studied phenomena) to improve the
generalization ability of the classification algorithm and to
reduce the computational overhead (Mao, 2004; Pal, 2006). In
the relevant literature, different band selection methods for
dimensionality reduction are reported (Kudo and Sklansky,
2000; Guan and et al, 2006; Martinez-Uso and et al, 2007).
Band selection (BS) algorithms are categorized into two main
approaches: supervised and unsupervised BS. In general,
supervised BS methods, regardless of their search strategy stage,
can be classified into two approaches based on their predefined
criteria. The first one, called the filter approach, is based on
optimising a discrimination measure, such as the Mahalanobis
distance, Bhattacharyya distance, etc. The filter approach
operates independently from any classification algorithm, so
undesirable features are omitted before the classification
process begins. The second one, called the wrapper approach,
tries to optimize the classification accuracy of the desired
classifier by selecting feature subsets. Feature shaving
(Verzakov and et al, 2004) is an example of this category.
Recently, also unsupervised feature selection procedures based
on feature similarity (mitraand et al, 2002) and mutual
information (Mart'mez-Us'o et al, 2006) have been proposed.
In particular, the profit of BS depends on many parameters: an
effective search strategy for exploring all possible subsets,
definition of a criterion for evaluation of subsets and a
classification algorithm for assessment of the accuracy of
selected bands as the final subspace (feature size).
Since these methods depend on the search strategy and in filter
methods the criteria are based on predefined pair-wise class
discriminant measure (like Mahalanobis and Bhattacharyya
distances), they result in suboptimal solutions. Sometimes they
suffer from shortcomings, such as a high correlation of
neighbour bands for computing a separability measure, and
classification with limited training samples in high dimensional
space at the beginning of backward search or at the final steps
of forward search algorithms.
In this article, we attempt to introduce an innovative method for
BS that makes use of only classes’ spectra. A natural question
that motivated us was whether one can perform BS based on
first order statistic parameters or class spectra obtained from a
spectral library, while discarding separability criteria based on
distribution of classes in high dimension. For this study, we
propose a new space called the prototype space for band
representation. Feature vectors in this space describe the band
behaviour in terms of their reflectance in dealing with imaging
scene phenomena. Conventional BS analyse the bands in terms
of feature vectors, which are defined based on pixels. In