KERNEL LOW-RANK REPRESENTATION FOR HYPERSPECTRAL IMAGE
CLASSIFICATION
Lu Du
1
, Zebin Wu
1, 2 *
, Yang Xu
1
, Wei Liu
1
, Zhihui Wei
1
1
School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China;
2
Jiangsu High Technology Research Key Laboratory for Wireless Sensor Networks, Nanjing, 210003, China
ABSTRACT
In this paper, a novel kernel low rank representation
(KLRR) method for hyperspectral image classification
is proposed. Firstly, we extract the global structure
characteristics information of the hyperspectral image
based on low rank representation (LRR), then use it as
a prior to constrain the recovery coefficient matrix. In
order to further improve the classification efficiency
and deal with the linearly non-separable problems
directly, we transformed the linear LRR classifier to a
non-linear one by feature space transformation using
the kernel trick. The proposed algorithm is solved by
the ADMM algorithm effectively. Experimental
results on real hyperspectral images demonstrate that
the proposed method outperforms many state-of-the-
art approaches.
Index Terms— Hyperspectral image classification,
Low Rank Representation, kernel trick
1. INTRODUCTION
Hyperspectral image classification(HC) is one of the
most popular tasks in hyperspectral and remote
sensing processing field. In hyperspectral case, the
recorded spectra have fine wavelength resolution and
cover hundreds of narrow and continuous spectrum
from the visible to the longwave infrared. And
different materials usually reflect electromagnetic
energy differently at specific wavelengths, which is
just the basis of classification.
Various techniques have been proposed for HC.
Among the existing classification methods, support
This work was supported in part by the National Natural Science
Foundation of China under Grant No. 61471199, 91538108, 11431015, the
Research Funds of Jiangsu High Technology Research Key Laboratory for
Wireless Sensor Networks under Grant No. WSNLBKF201507, the Jiangsu
Province Six Top Talents project of China under Grant No. WLW-011.
*Corresponding author. Email: Zebin.wu@gmail.com
vector machine (SVM) has been proved to be a
efficient way and achieved good performance. Many
modifications of SVM have also been proposed to
improve the classification accuracy further by
incorporating the spatial information, such as SVM
with composite kernels (CKs) [1].
Recently, sparse representation classification
(SRC) [2] has been applied to HC. Using the classical
sparsity priors or other structured ones [3], the SRC
method greatly improves the classification accuracy
compared with other traditional methods. However,
most existing priors only take into consideration the
local neighborhood correlation, and usually confine
the sparsity of the pixel neighborhood coefficients too
much, which in some cases strictly limit the recovery
coefficients and sometimes even destroy internal
structure of data. For example, the joint sparse prior [4]
use a set of neighboring samples to represent the
centric test pixel, but if test sample is the boundary
pixel, the fact that neighborhood pixels belong to
different categories may lead to misclassification.
Therefore, the intrinsic global structure of
hyperspectral image should be extracted to improve
the classification accuracy.
On the one hand, rank is a reasonable measure of
the matrix’s sparsity. Rank has been used in signal
processing and machine learning, such as image
segmentation[5] and saliency detection[6]. In [5], the
authors apply the low-rank property to improve the
classification results in HC, however, the low-rank
prior is used only in a neighborhood structure.
On the other hand, SRC is a linear classifier,
which is difficult to deal with linearly inseparable
problem. Fortunately, in machine learning, kernel trick
has been successfully used to construct nonlinear
support vector machine, which improves the
classification efficiency by the feature space
transformation. After that, literature [7] relies on
sparsely representing hyperspectral image in a feature
477978-1-5090-3332-4/16/$31.00 ©2016 IEEE IGARSS 2016