SUBSPACE MATCHING PURSUIT WITH DICE COEFFICIENT FOR SPARSE
UNMIXING OF HYPERSPECTRAL DATA
Dan Li, Chunmei Zhang, Qianqi Zhou, Junyan Wang, Guodong Xu
School of Computer Science and Engineering, Beifang University for Nationalities,
Ningxia 750021, China
e-mail:chunmei66@hotmail.com
ABSTRACT
Sparse unmixing is a popular linear spectral unmixing tool
in remotely sensed hyperspectral data interpretation. It can
be worked out in semisupervised fashion by taking the
advantage of the spectral library known in advance.
Most sparse regressions methods are based on convex
relaxation methods which try to obtain the global solution of
a well-defined optimization problem. Recently, more and
more attention has been paid to greedy algorithms for sparse
unmixing because of their low complexity. Among of them,
subspace matching pursuit (SMP) is a preferable one to
recover the optimal endmembers according to the referred
subspace of the original image. In this paper, we study the
linear spectral unmixing problem under the method of SMP.
Furthermore, in order to increase the robust in the case of
low SNR, we provide a novel algorithm to replace the inner
product of the matching measurement criteria of sparse
representation with the Dice coefficient to improve the
unmixing accuracy.
Index Terms—Hyperspectral unmixing, sparse
unmixing, subspace matching pursuit, Dice coefficient
1. INTRODUCTION
Hyperspectral remote sensing measures radiance from
earth’s surface materials at hundreds of narrow and
contiguous wavelength bands. However, due to the low
spatial resolution of a sensor as well as the combination of
distinct materials into a mixture, each pixel in the
hyperspectral image often contains more than one pure
substance. Unmixing aims at decomposing the measured
spectrum of each mixed pixel into a collection of constituent
spectra (endmembers) and a set of corresponding fractions
(abundances)[1]. Using unmixing to process hyperspectral
data can get a more accurate measurement in the pixel level
than classifying to satisfy higher quality requirements.
Depending on the mixing scales at each pixel and on the
geometry of the scene, the observed mixture is either linear
or nonlinear. Among them, linear unmixing model considers
each mixed pixel a linear combination of endmembers
weighted by their corresponding abundance fractions. This
linear model has been widely used to solve the unmixing
problem. Under this model, a group of unmixing approaches
based on geometry[2], statistics[3] and sparse regression[4]
have been proposed.
Sparse regression has been used to model each mixed
pixel in the hyperspectral image using only a few spectra in
a spectral library which is known in advance. So it is a semi-
supervised unmixing approach with the potential advantage
to the applications in which the spectral library can be
obtained. Unmixing then amounts to finding the optimal
subset of signatures in a (potentially very large) spectral
library that can best model each mixed pixel in the scene[5].
Because the number of spectral signatures contained in the
spectral library is much larger than the number of
endmembers presented in the hyperspectral image, this
approach often leads to a sparse solution.
Most sparse unmixing methods are based on convex
relaxation methods. However, convex relaxation methods
are far more sophisticated than the greedy algorithms as
they obtain the global solution of a well-defined
optimization problem. Therefore, the greedy algorithms
attract more and more attention in sparse unmixing. Several
such techniques ,e.g. orthogonal matching pursuit (OMP),
basis pursuit (BP), subspace matching pursuit (SMP) and
iterative spectral mixture analysis (ISMA)[6] are emerged in
sparse unmixing. Among of them, SMP, a preferable greedy
algorithm with the split subspaces, is used to solve the
unmixing problem. In each iteration, SMP looks for some
low-degree mixed pixels from the residual and extracts
several corresponding endmembers. By utilizing the low-
degree mixed pixels in the hyperspectral data, it can find the
actual endmembers more accurately than the other sparse
regression algorithm. However, it ignores the relationship of
the inner spectrum blocks which can be taken advantage to
increase the robust in the case of noised data.
To exploit the spectral information, a novel algorithm in
this paper is proposed to increase the anti-noise ability. In
this method, the inner product of the matching measurement
criteria of sparse representation is replaced by the Dice
coefficient [7]. The emulation experiment is presented to
show the inversion results.
6585978-1-5090-3332-4/16/$31.00 ©2016 IEEE IGARSS 2016