Spectral-Spatial Clustering of Hyperspectral Remote Sensing Image with Sparse Subspace
Clustering Model
Han Zhai
1
, Hongyan Zhang*
1
, Liangpei Zhang
1
, Pingxiang Li
1
and Xiong Xu
2
1: The State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, and Collaborative Innovation
Center for Geospatial Technology, Wuhan University, P.R. China
2: College of Surveying and Geo-Informatics, Tongji University, Shanghai, P.R. China
*Corresponding author: zhanghongyan@whu.edu.cn
ABSTRACT
Clustering for hyperspectral imagery (HSI) is a very
challenging task due to its inherent spectral and spatial
complexity. In this paper, we propose a novel
spectral-spatial sparse subspace clustering (S
4
C) algorithm
for hyperspectral imagery. Firstly, by treating each kind of
ground class as a subspace, we introduce sparse subspace
clustering (SSC) algorithm to HSIs. Then considering the
spectral and spatial property of HSI, the high spectral
correlation and rich spatial information of the HSIs are
taken into consideration in the sparse subspace clustering
model to obtain a more accurate coefficient matrix, which is
used to build the adjacent matrix. Lastly, spectral clustering
is applied to the adjacent matrix to obtain the final image
clustering result. Several experiments were conducted to
illustrate the performance of the proposed algorithm.
Index Terms—hyperspectral image, sparse
representation, subspace cluster, spectral clustering
1. INTRODUCTION
Hyperspectral image clustering can be defined as the
process of segmenting pixels into corresponding sets, which
satisfies that differences between sets are much greater than
differences within sets. As known to us all, HSIs are typical
high-dimensional data with large spectral variability and
complex structure [1], which makes HSI clustering a very
challenging problem.
To date, many clustering methods for HSIs have been
proposed, such as the centroid clustering methods [2], the
density-based methods [3], the biological methods [4], the
spectral methods [5] and so on. However, most of them
suffer from heavy misclassification because of the uniform
feature point distribution caused by the large spectral
variability of HSI. In recent years, sparse subspace
clustering (SSC) algorithm was proposed to group data
points into different subspaces by finding the sparsest
representation for each data point with only selecting the
data points from its own subspace to represent itself, and has
been widely applied in various computer vision fields [6].
For HSIs, although pixels of the same class may have
different spectrums because of the varying illumination,
topography and imaging condition, they would lie in the
same subspace. Based on this fact, it is natural to introduce
the SSC algorithm to the HSI clustering task. However,
directly applying the SSC algorithm to the HSI usually fails
to take advantage of the high spectral correlation and rich
spatial information of the HSI, which is far from fully
exploiting the potential of the SSC algorithm for HSI
clustering.
In view of this, we proposed a novel spectral-spatial
sparse subspace clustering algorithm (S
4
C) for hyperspectral
imagery. The contributions of this paper are summarized as
follows. Firstly, to the best of our knowledge, by
considering that the pixels of the same class lie in one
independent subspace, we are the first to segment HSI pixels
into different clusters with the SSC algorithm. Secondly,
considering the working mechanism of sparse representation,
a spectral weighted sparse subspace clustering model
(SWSSC) is built to ensure that each pixel can be
represented by the corresponding signals with the highest
correlation. Thirdly, with the spectral similarity of a local
neighborhood of the HSI, the rich spatial information is also
incorporated to the SSC model to improve the performance.
The experimental results demonstrate that the proposed S
4
C
algorithm significantly improve the clustering performance
both in visual and quantitative evaluations.
2. HSI CLUSTERING VIA SPARSE SUBSPACE
CLUSTERING MODEL
In the SSC model, let
be an arrangement of
affine subspace of
of dimensions
, where
is the
dimension of the ambient space. Considering a given
collection of N noise-free data points
that lie in the
union of the
subspaces [6]. Denote the matrix containing
all the data points as
11
, , , ,
Nl
yyY Y Y Γ
where
is a rank-
matrix of the
points
that lie in
and
is an unknown permutation
matrix.
For HSI, every hyperspectral pixel can be denoted as a
-dimensional vector, where
refers to the number of band.
By this way, the HSI can be denoted by a 2-D matrix
12
, , , ,
p MN
MN
Y Y Y Y Y
, where
represents the
width of the HSI data, and
stands for the height of the
data. By using the data matrix itself as a dictionary, the
sparse representing model can be built as
, ,
T
diag 0 1 1 Y YC N C C
where
is the self-representation coefficient