UNSUPERVISED SPECTRAL MIXTURE ANALYSIS WITH HOPFIELD NEURAL NETWORK
FOR HYPERSPECTRAL IMAGES
Shaohui Mei
1
, Mingyi He
1
, Zhiyong Wang
2
, and Dagan Feng
2
1. School of Electronics and Information, Northwestern Polytechnical University, Xi’an 710129, China
2. School of Information Technologies, The University of Sydney, NSW 2006, Australia
E-mail: meish@nwpu.edu.cn
ABSTRACT
Spectral Mixture Analysis (SMA) has been widely utilized to
address the mixed-pixel problem in the quantitative analysis
of hyperspectral remote sensing images. Recently Nonneg-
ative Matrix Factorization (NMF) has been successfully uti-
lized to simultaneously perform endmember extraction (EE)
and abundance estimation (AE). In this paper, we formulate
the solution of NMF by performing EE and AE iteratively.
Based on our previous Hopfield Neural Network (HNN)
based AE algorithm, an HNN is also constructed for EE to
solve the multiplicative updating problem of NMF for SMA.
As a result, SMA is conducted in an unsupervised manner
and our algorithm is able to extract virtual endmembers with-
out assuming the presence of spectrally pure constituents
in hyperspectral scenes. We further extend such strategy to
solve the constrained NMF (cNMF) models for SMA, where
extra constraints are imposed to better model the mixed-pixel
problem. Experimental results on both synthetic and real
hyperspectral images demonstrate the effectiveness of our
proposed HNN based unsupervised SMA algorithms.
Index Terms— Hyperspectral images, Spectral Mixture
Analysis, Hopfield Neural Network, Nonnegative Matrix Fac-
torization
1. INTRODUCTION
Hyperspectral remote sensing imaging technology, which si-
multaneously acquires tens of hundreds of images covering
the visible areas, the infrared areas, and even the short wave
areas of electromagnetic spectrum, has been more and more
important in photoelectric remote sensing since 1980s. Due
to the limitation of spatial resolution, pixels in a hyperspec-
tral remote sensing image often contain more than one type of
ground objects, resulting in inefficiency in many applications,
such as image classification, target recognition, and quantita-
This work is partially supported by National Natural Science Founda-
tion of China(61171154, 61101188), Scholarship Award for Excellent Doc-
toral Student granted by Ministry of Education, NPU Foundation for Basic
Research, and ARC grants.
tive analysis. Therefore, Spectral Mixture Analysis (SMA) is
proposed to address the mixed-pixel problem.
Many linear unmixing algorithms have been proposed for
SMA in the past decade, such as fully constrained least square
(FCLS) algorithm [1], Hopfield Neural Network (HNN) [2],
to name a few. However, many of these algorithms demand
that the spectral signature of pure ground objects (known as
endmember) must be known in advance, which are known
as supervised SMA. Therefore, many algorithms have been
developed to extract endmember spectra for these SMA algo-
rithms, such as vertex component analysis (VCA) [3], sim-
plex growing algorithm (SGA) [4], and etc. Nevertheless, in
these algorithms, the pure spectra for each ground objects are
assumed to be present in the image. When the pure-pixel as-
sumption is violated in a highly mixed situation, these algo-
rithms show unsatisfactory results. Nonnegative Matrix Fac-
torization (NMF) has been proven to be capable of extracting
endmember and their corresponding fractional abundance si-
multaneously from highly mixed hyperspectral data [5, 6, 7].
The projected gradient methods have been utilized to solve
this NMF problem. However, searching for a suitable step
size is a critical and time-consuming operation for the pro-
jected gradients methods.
In this paper, a Hopfield minimization algorithm is pro-
posed to solve the NMF problems for SMA. Two different
Hopfield Neural Networks (HNNs) are constructed to solve
the multiplicative updating problems of NMF. In addition, the
minimum distance constrained NMF [6] and minimum dis-
persion constrained NMF [7] are adopted to further model
mixture problems, respectively, and the proposed HNN based
algorithm is also utilized to solve these two constrained NMF
(cNMF) models for SMA. Consequently, unsupervised SMA
in highly mixed hyperspectral data can be solved by the pro-
posed HNN based algorithms.
2. UNSUPERVISED SMA WITH HNN
2.1. NMF for SMA with HNN
In the LMM, the photons reflected from different ground ob-
jects contained in one pixel are assumed not to interfere with