Sparse hyperspectral unmixing based on smoothed ‘
0
regularization
Chengzhi Deng
⇑
, Shaoquan Zhang, Shengqian Wang, Wei Tian, Zhaoming Wu
Department of Information Engineering, Nanchang Institute of Technology, Nanchang 330099, PR China
highlights
This paper present a new sparse hyperspectral unmixing method based on smoothed ‘
0
norm.
The smoothed ‘
0
norm is a continuous function, which provides a smooth measure of ‘
0
sparsity and better accuracy than ‘
1
norm and also tolerates noise
to some extent.
The experimental results show effectiveness and accuracy of the proposed method.
article info
Article history:
Received 8 January 2014
Available online 21 August 2014
Keywords:
Sparse unmixing
Smoothed ‘
0
norm
Spectral unmixing
Hyperspectral imaging
abstract
Sparse based approach has recently received much attention in hyperspectral unmixing area. Sparse
unmixing is based on the assumption that each measured pixel in the hyperspectral image can be
expressed by a number of pure spectra linear combination from a spectral library known in advance.
Despite the success of sparse unmixing based on the ‘
0
or ‘
1
regularizer, the limitation of this approach
on its computational complexity or sparsity affects the efficiency or accuracy. As the smoothed ‘
0
regularizer is much easier to solve than the ‘
0
regularizer and has stronger sparsity than the ‘
1
regularizer, in this paper, we choose the smoothed ‘
0
norm as an alternative regularizer and model the
hyperspectral unmixing as a constrained smoothed ‘
0
‘
2
optimization problem, namely SL
0
SU
algorithm. We then use the variable splitting augmented Lagrangian algorithm to solve it. Experimental
results on both simulated and real hyperspectral data demonstrate that the proposed SL
0
SU is much more
effective and accurate on hyperspectral unmixing than the state-of-the-art SUnSAL method.
Ó 2014 Elsevier B.V. All rights reserved.
1. Introduction
Hyperspectral remote sensing collects and processes electro-
magnetic spectrum information from the Earth’s surface at hun-
dreds of narrow and contiguous wavelength bands. It has been
widely applied in various fields, such as target detection, material
mapping, and material identification. However, due to the low spa-
tial resolution and the complexity of natural diversity of surface
features, which leads to homogeneous mixture, hyperspectral
remote sensing imagery of a pixel contains may not be a single fea-
ture. The spectral unmixing technique was proposed to deal with
the problem, which estimates the number of endmembers and cor-
responding fractional abundances in each mixed pixel [1].
There are two models used to analyze the mixed pixel problem
which can be divided into linear and nonlinear mixture models [2].
Compared with the nonlinear mixture model, the linear mixture
model is computational tractability and flexibility and has been
widely used to solve the unmixing problem. In recent years, many
linear unmixing methods have been proposed, including the pixel
purity index (PPI) [3], N-FINDR [4], the iterative error analysis (IEA)
[5] and iterative constrained endmembers (ICE) [6]. However, all
these methods under the framework of linear spectral unmixing
are very likely to fail in highly mixed scenarios.
The advantage that they do not require the estimation of the
endmembers makes the sparse regression to be a new direction
for hyperspectral unmixing proposed by Iordache firstly [7].In
recent years, sparse unmixing becomes a new hotspot [8–14],
which is based on the assumption that the pixel’s spectrum can
be expressed in the form of linear combinations of a number of
pure spectral signatures from a large spectral library that is known
in advance [1]. Because the size of the spectral library is often large,
the number of endmembers in the spectral library will be much
greater than the number of spectral bands. The sparse unmixing
model is a typical underdetermined linear system, which is
difficult to find a unique, stable, and optimal solution [14].
http://dx.doi.org/10.1016/j.infrared.2014.08.004
1350-4495/Ó 2014 Elsevier B.V. All rights reserved.
⇑
Corresponding author.
E-mail addresses: dengchengzhi@126.com (C. Deng), zhangshaoquan1@163.com
(S. Zhang), sqwang113@263.net (S. Wang), tw_0930@163.com (W. Tian), zmwunit@
foxmail.com (Z. Wu).
Infrared Physics & Technology 67 (2014) 306–314
Contents lists available at ScienceDirect
Infrared Physics & Technology
journal homepage: www.elsevier.com/locate/infrared