
Classification of Hyperspectral Image Based on K-means and Structured Sparse
Coding
Yang Liu
College of Automation
Shenyang Aerospace University
Shenyang, China
e-mail: yang97_net@163.com
Yangyang Wang
College of Automation
Shenyang Aerospace University
Shenyang, China
e-mail: wyy2004101@163.com
Abstract—The combination of spatial and spectral information
of hyperspectral image benefits the improvement of
classification accuracy. The structured sparse coding is
proposed to reconstruct the pixels of hyperspectral image. The
reconstructed pixels characterize the spatial structure. The K-
means method is used to form the dictionary, which has
stronger representation ability. Finally, the classification is
implemented according to the reconstruction residuals. The
experiments are conducted on AVIRIS and the results show
that the classification accuracy is improved obviously
compared with the other state-of-the-art methods.
Keywords- hyperspectral image; classification; K-means;
sparse coding; reconstruction
I.
I
NTRODUCTION
In recent years, the remote sensing technology develops
rapidly. The classification of remote sensing image can be
used in precision agriculture, environmental management
and social security and other fields[1]. The processing and
analysis of remote sensing image has become the research
focus. The hyperspectral image holds abundant spectral
information and the number of spectral bands is up to several
hundreds, which can help extract quantitative information
effectively and object classification.
The traditional method of classification of hyperspectral
image is to only deal with the spectral information and
ignore the spatial information of pixels and the spatial
continuity. It is that the classification of each pixel is
independent of spectral information. The main classification
methods are KNN method[2][3], maximum likelihood
estimation[4], artificial neural network[5], kernel based
method[6] and so on, in which the support vector machine
demonstrates a superior performance[7]. Due to the
correlation of spectrums, the spectral information satisfies
the sparsity condition. Therefore, some scholars introduce
sparse coding to hyperspectral image analysis[8,9].
However, in addition to the spectral information of
hyperspectral image, spatial information is very important
for accurate analysis. Combining spectral and spatial
information for classification has become a hot research
topic recently. In this paper, each pixel of hyperspectral
image is reconstructed by using sparse encoding, and the
center pixel is represented by a linear combination of pixels
in its neighborhood. Due to the limited number of samples
and considering the computational burden, the L2 sparse
coding is used instead of the traditional L1 sparse coding.
The K-means method is used to construct the dictionary,
which demonstrates better characters than the dictionary
composed of training samples directly. Finally, test samples
are classified according to the reconstruction residuals. The
experimental results show that the proposed method based
on K-means and structured sparse coding can combine the
spectral and spatial information of hyperspectral image
effectively and improve the classification accuracy. The
overall flow chart of the algorithm is as follows.
Figure 1. The overall flow chart of the algorithm is as follows
2016 3rd International Conference on Information Science and Control Engineering
978-1-5090-2534-3 /16 $31.00 © 2016 IEEE
DOI 10.1109/ICISCE.2016.62
247
2016 3rd International Conference on Information Science and Control Engineering
978-1-5090-2534-3 /16 $31.00 © 2016 IEEE
DOI 10.1109/ICISCE.2016.62
248
2016 3rd International Conference on Information Science and Control Engineering
978-1-5090-2535-0/16 $31.00 © 2016 IEEE
DOI 10.1109/ICISCE.2016.62
248