144 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 53, NO. 1, JANUARY 2015
Extended Random Walker-Based Classification
of Hyperspectral Images
Xudong Kang, Student Member, IEEE, Shutao Li, Member, IEEE, Leyuan Fang, Student Member, IEEE,
Meixiu Li, and Jón Atli Benediktsson, Fellow, IEEE
Abstract—This paper introduces a novel spectral–spatial clas-
sification method for hyperspectral images based on extended
random walkers (ERWs), which consists of two main steps. First,
a widely used pixelwise classifier, i.e., the support vector machine
(SVM), is adopted to obtain classification probability maps for
a hyperspectral image, which reflect the probabilities that each
hyperspectral pixel belongs to different classes. Then, the obtained
pixelwise probability maps are optimized with the ERW algo-
rithm that encodes the spatial information of the hyperspectral
image in a weighted graph. Specifically, the class of a test pixel
is determined based on three factors, i.e., the pixelwise statistics
information learned by a SVM classifier, the spatial correlation
among adjacent pixels modeled by the weights of graph edges, and
the connectedness between the training and test samples modeled
by random walkers. Since the three factors are all well considered
in the ERW-based global optimization framework, the proposed
method shows very good classification performances for three
widely used real hyperspectral data sets even when the number
of training samples is relatively small.
Index Terms—Extended random walkers (ERWs), graph,
hyperspectral image, optimization, spectral–spatial image
classification.
I. I NTRODUCTION
H
YPERSPECTRAL image classification gives a high-level
understanding of remotely sensed scenes and is therefore
now widely used in different application domains such as envi-
ronment monitoring [1], precision agriculture [2], and national
defense [3]. However, because of the special characteristics of
hyperspectral data sets, image classification in the hyperspectral
domain still has many unresolved problems [4].
For instance, the high dimensionality of hyperspectral data
sets involves the “Hughes” phenomenon in classification [5].
The “Hughes” phenomenon refers to the fact that, if the number
of training samples is fixed, the classification accuracy may de-
Manuscript received January 28, 2014; revised March 24, 2014; accepted
April 18, 2014. Date of publication May 14, 2014; date of current version
August 4, 2014. This paper was supported in part by the National Natural
Science Foundation for Distinguished Young Scholars of China under Grant
61325007, by the National Natural Science Foundation of China under Grant
61172161, by the Fundamental Research Funds for the Central Universities,
by Hunan University, and by the Chinese Scholarship Award for Excellent
Doctoral Student.
X. Kang, S. Li, L. Fang, and M. Li are with the College of Electrical and
Information Engineering, Hunan University, Changsha 410082, China (e-mail:
xudong_kang@hnu.edu.cn; shutao_li@hnu.edu.cn; fangleyuan@gmail.com).
J. A. Benediktsson is with the Faculty of Electrical and Computer Engineer-
ing, University of Iceland, 107 Reykjavik, Iceland (e-mail: benedikt@hi.is).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TGRS.2014.2319373
crease significantly for some supervised classification methods
as the data dimensionality increases beyond a certain number of
features. In order to deal with this difficulty, several solutions
have been developed such as feature extraction [6]–[8] and
discriminative learning [9], [10]. Feature extraction methods
such as principal component analysis (PCA) [7], independent
component analysis [6], and linear discriminant analysis [8],
[11], [12] project the high-dimensional data into a low di-
mensional feature space while preserving the discriminative
information of different classes. Furthermore, discriminative
learning approaches such as support vector machines (SVMs)
[9], multinomial logistic regression [10], and artificial immune
networks [13] learn the class distributions in high-dimensional
spaces by inferring the nonlinear boundaries between classes
in feature space. These methods can effectively tackle the
aforementioned difficulties caused by high dimensionality.
In addition to research on how to overcome the “Hughes”
phenomenon, another active research topic for hyperspectral
image classification is how to make full use of the spatial infor-
mation of the data in order to further improve the classification
accuracy [14]. To achieve this objective, intensive work has
been performed in the last decade to develop spectral–spatial
hyperspectral image classification methods. For example, spa-
tial feature extraction methods [15]–[21] have been proposed to
define an adaptive neighborhood for each pixel by local filtering
operations so that the adaptive local neighborhood information
could be preserved in the resulting features for classification.
Furthermore, an optimal set of the resulting spatial features
can be selected to further improve the performance of feature
extraction [22].
In addition to spatial feature extraction, image segmentation
is a widely used technique for spectral–spatial image classifi-
cation. Specifically, segmentation-based methods perform the
decision fusion of image segmentation and pixelwise classifica-
tion to make full use of the spatial information of hyperspectral
images. For this kind of methods, the automatic segmentation
of hyperspectral images is a challenging task, and thus, many
different hyperspectral image segmentation methods have been
proposed such as watershed [23], partitional clustering [24],
hierarchical segmentation [25], and stochastic minimum span-
ning forest [26], [27].
In recent years, other types of spectral–spatial methods
such as nonlocal joint collaborative representation [28], spatial
kernel-based methods [29], [30], and probabilistic modeling-
based methods [10], [31]–[34] have also been successfully
applied for hyperspectral image classification. For instance,
probabilistic modeling-based methods first estimate the
0196-2892 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.