KEY WORDS: Agriculture, Hyper spectral, Extraction, Classification, Artificial lntelligence, Algorithms
ABSTRACT
A new dynamical dimensional reduction model (HDRM) for hyperspectral images is proposed based on clone selection algorithm
which is inspired from nature immune system in this paper. In existing dimensional reduction method, feature selection is most
inefficient. To improve the efficiency, the feature selection problem in hyperspectral images is taken as a multi-objective
optimization problem. The feasible band sets are regarded as antibodies and the evaluation criteria in feature selection are regarded
as the antigens in HDRM. Go through generation after generation, the sets keep on evolution under the guidance and constrain of
evaluation criteria, ultimately, the optimization sets can be found. The model is trained with a hyperion image data, and the result of
feature selection is used in classification to test its effect. It is proved that the time cost in feature selection is 100s in the experiment
data, and the iterate time is just 10 when /? is 0.631, a> is 0.873.
1. INTRODUCTION
Resent Years, as the sensor technology’s advance,
hyperspectral image has got unprecedented development. The
applications of hyperspectral image data have extended to
agriculture, environment, mine, and so on. Because
hyperspectral image data could provide much higher resolution
than normal remote data. And their resolution can reach
nanometer. The ground objects and their characteristics can be
discriminated more accurately by hyperspectral image data.
combinational optimization new ideas. Castro and Timmis (De
Castro, 2002) have built an Artificial Immune Network for
multimodal function optimization. Li investigates AIS in global
optimization, too (Li, 2005). Based on these researches, the
concept of AIS is introduced to hyperspectral image data. To
improve the efficiency of feature selection, this paper derives
some important algorithms from AIS for dimensional reduction,
and proposes a hyperspectral dimensional reduction model
(HDRM).
The quantity of the bands of the hyperspectral image data is
very large, and the correlation of the bands is quite strong, too.
On one hand, the hugeness of the data brings difficulties to not
only data storage, but also data processing. It holds back the
applying of the hyperspectral image data in some degrees. On
the other hands, traditional methods which have been designed
for multi-spectral image data can not be easily applied to
hyperspectral image data. So, dimensional reduction in
hyperspectral image data without losing important information
about objects of interest has become a topic of a substantial
amount of research in recent years.
In the process of hyperspectral data dimensional reduction,
some prior information could be excavated. When prior
information is useful, we get some structural characters of the
hyperspectral data, with which as the guiding, the dimension
can be reduced effectively and gradually in hyperspectral data.
And the threshold value could be set to and the process. All of
this can be looked as a process of recognition. We can take the
prior information as the primary immune response, and the
process of the hyperspectral data reduction can be took as the
secondary immune response, during which the self- tissue, self
cognition, self-memory and other abilities can be inspired.
In many objects, their reflectance or absorption characteristics
only appear at a very narrow spectral range. The correlation
between the bands is quite strong, too. So, dimensional
reduction in hyperspectral image data is theoretically feasible.
This bas been investigated by Jimenez (Zhao, 2004) in detail.
Roughly speaking, the common ways for dimensional reduction
fall into two categories: One is feature extraction; and the other
is feature selection (Zhao, 2004). The process of feature
extraction is usually fast, but the original spectral properties of
the image are lost after the extraction. Feature selection always
costs longer time, but the spectral properties of the image are
remained. When the evaluation criteria for the optimized band
set are determined, feature selection could be regarded as a
combinational optimization problem. The immune system
contains the abilities of self-tissue, self-cognition, self-memory,
clone mutation, and so on. These abilities provide
This paper is organized as follows. Section 2 is a brief
introduction to clonal selection theory which is derived from
AIS, and it gives a short description of multi-objective
optimization, too. Section 3 presents a model (HDRM) to
reduce the dimension of hyperspectral data. In Section 4, an
experiment based on HDRM is carried out in a hyperion image.
Then the results are provided and discussed. Section 5
concludes the entire work and presents prospects of AIS in
hyperspectral image data process and the future works needed
to do.
THE CLONAL SELECTION IN MULTI-OBJECTIVE
OPTIMIZATION PROBLEM
AIS is a calculation model AIS inspired from theoretical
biology, it makes use of some functions, principles and models
Corresponding author: Liuxncugb@163.com
■ggtWHM