978-1-4673-0024-7/10/$26.00 ©2012 IEEE 1707
2012 9th International Conference on Fuzzy Systems and Knowledge Discovery (FSKD 2012)
An endmember extraction algorithm for hyperspectral imagery based on kernel
orthogonal subspace projection
Liaoying Zhao, Fujie Li
Institute of Computer Application Technology, HangZhou
Dianzi University
Hang zhou, China
Jiantao Cui
College of Electrical Engineering
Zhejiang University
Hang zhou,,China
Abstract—Endmember extraction is a key step of spectral
unmixing. In order to extract endmembers more precisely
from nonlinear mixed hyperspcetral imagery, an unsupervised
kernel-based orthogonal subspace projection (UKOSP)
technique is proposed in this paper. Without considering the
noise, the maximal pixel vector in the imagery would be
regarded as an endmember, then was removed the effect of it
by kernel orthogonal subspace projection method to get
another orthogonal imagery. Experimental results of simulated
and real data prove that the proposed UKOSP approach
outperforms the linear endmember extraction algorithms such
as vertex component analysis and unsupervised kernel-based
orthogonal subspace projection.
Keywords-Endmember extraction, unsupervised, kernel
subspace projection, hyperspectral imagery
I. INTRODUCTION
Endmember extraction from hyperspectral imagery is the
precondition of understanding hyperspectral data and doing
further analysis (for instance, mixed pixel decomposition,
feature mapping and target detection) [1].How to extract
endmember from hyperspectral imagery is usually payed
high interest and lots of mature methods have been proposed.
Typical endmember extraction algorithms include pixel
purity index (PPI) [2], the N-FINDR[3], vertex component
analysis(VCA) [4], endmember extraction algorithm based
on RMS error analysis [5], maximal distance method,
simplex volume algorithm [6], unsupervised orthogonal
subspace projection (UOSP) [7] and so on.
All the above methods are proposed based on linear
spectrum mixture model, the nonlinear mixture model is
more reasonable for the fine spectral analysis in microscopic
scale or for the detection of object with small probability,
while the nonlinear relationship is often hard to define, and
the computation of the nonlinear model is usually complex.
Kernel method [8] is a new method established on the
statistical learning theory, used in nonlinear pattern analysis
and recognition. Kernel method has been widely used in
hyperspectral imagery classification and target detection.
Particularly, Kwon [9] who worked in the America Army
Research Lab has extended some existed linear methods, and
proposed the kernel-matched subspace detector (KMSD), the
kernel spectral-matched filter (KSMF), and the kernel
adaptive subspace detectors (KASD) and kernel orthogonal
subspace projection (KOSP). Based on the idea of UOSP [7],
a nonlinear unsupervised kernel orthogonal subspace
projection (UKOSP) is proposed in this paper, first analyze
the nonlinear orthogonal subspace projection in theory, then
deduce the computation of the vector value after projection
based on kernel theory, and make the analysis that the
maximal pixel vector in the image cube is rightly an
endmember when the noise is ignored. At last, through
computing the largest value of vectors and kernel subspace
projection alternate, each endmember is extracted during the
process. Experimental results of simulated and real data
prove that the proposed UKOSP approach is reasonable and
effective.
II. N
ONLINEAR SPECTRUM MIXTURE MODELS
According to the mixture of materials and the spatial
scale of physical distribution, spectrum mixture may be
linear or nonlinear mixture [1].Currently, there are few
research fruits for nonlinear spectrum mixture model, the
typical mathematical model is called bilinear spectrum
mixture model [10,11], the bilinear model considers second
order interactions between endmembers, the model is written
as
1
,
11
ss
ij i j i j
iji
−
==+
=+ ⊗+
∑∑
rMa m m n
βαα
(1)
where
i
α
is the abundance vector,
i
m is the endmember,
,ij
is a coefficient that controls the interactions between
endmembers. ⊗ is the Hadamard (term-by-term) product
operation. The abundances have to satisfy the following
positivity and sum-to-one constraints:
01
k
a≤≤
,
01
ij
r≤≤ (2)
1
1
S
k
k
a
=
=
∑
(3)
Note that higher order interaction terms are also received
by the hyperspectral sensor. However, experiments
conducted in [10] have shown that these higher order terms
can be neglected.
Linear mixture is the special case under the condition that
multiple reflections are neglected in the nonlinear mixture. If