Collaborative Sparse Coding with Smoothness
Regularization for Hyperspectral-Image
Classification
Yang Liu, Ruisheng Wang, Xiaofei Ji, Yangyang Wang
Abstract
Hyperspectral-image (
HSI
) classification plays a key role in
numerous applications in remote sensing, urban planning,
and environmental monitoring. Sparse-representation models
have been increasingly explored for
of their compactness, flexibility, an
this article, a novel
HSI
scheme is pr
ative sparse coding combined with
(
CSCSR
). First, based on the spatial
HSIs, spectral-spatial preprocessing is performed to improve
classification accuracy. Second, a collaborative sparse-coding
model with smoothness regularization is proposed and ap-
plied for
HSI
classification. In our model, the sparsity level
and smoothness regularization are tuned to improve clas-
sification performance. In addition, weighted pixel similari-
ties are computed in pixel neighborhoods and then used
to incorporate spatial information in the
HSI
classification
scheme. A feature-sign search algorithm is used for sparse
coding of feature descriptors. Experimental results on real
HSI
data sets demonstrate that the proposed
CSCSR
method
effectively outperforms the current state-of-the-art
HSI
clas-
sifiers in terms of both qualitative and quantitative metrics.
Introduction
Classification of hyperspectral images (
HSIs
) has been a central
research focus in image analysis for remote sensing, with a
wide range of applications including process monitoring,
quality control, environmental surveillance, precision agricul-
ture, and urban mapping (Bioucas-Dias
et al.
2012; Bioucas-
Dias
et al.
2013; Dalponte
et al.
2013; Hu
et al.
2018).
HSIs
can
have hundreds of adjacent or close narrow bands, and hence
these images offer rich and versatile sources of information.
The goal of
HSI
classification is to correctly assign a pixel with
multispectral information to one of a certain set of classes (He
et al.
2018). Some of the main challenges in accomplishing
that goal are the limited numbers of training samples and the
high dimensionality of the spectral patterns associated with
each pixel. Several classifiers have been proposed for
HSI
classification—for example,
K
-nearest neighbor (Ma, Crawford
and Tian 2010; Samaniego, Bárdossy and Schulz 2008), artifi-
cial neural networks (Stathakis and Vasilakos 2006), support
vector machines (Camps-Valls and Bruzzone 2005; Melgani
and Bruzzone 2004), extreme learning machines (W. Li
et al.
2015; Su, Cai and Du 2017), and multinomial logistic regres-
sion (Y. Chen, Nasrabadi and Tran 2013).
Recently, methods based on the theory of sparse signals have
been applied in the analysis of remote sensing imagery, such as
unmixing for
HSI
(Feng
et al.
2017; Iordache, Bioucas-Dias and
Plaza 2012; Zhong, Feng and Zhang 2014) and
HSI
classification
ng
et al.
2014). Sparse-signal models
g both spectral and spatial
HSI
infor-
hods have been successfully applied
t al.
2014; Jia, Zhang and Li 2015;
t al.
2014; X. Zhang
et al.
2015).
Collaborative sparse-coding (
SC
) methods (Li
et al.
2014; W. Li
and Du 2014) seek to exploit the spatial pixel relationships to
achieve better
HSI
-classification accuracy. For these methods,
the class of a testing hyperspectral pixel is determined by the
features of this pixel as well as its neighboring pixels, which
are assumed to represent materials and features similar to those
of the central pixel. Indeed, collaborative classification meth-
ods have demonstrated clear performance improvements over
classical ones (W. Li
et al.
2015; L. Zhang, Yang and Feng 2011).
This article proposes a novel collaborative sparse-coding
scheme with smoothness regularization (
CSCSR
) to improve
HSI
-classification performance. The proposed scheme com-
bines and extends methods of collaborative filtering, sparsity
modeling, and multihypothesis (
MH
) prediction (Sullivan
1993). The contributions of our article are as follows:
1. The proposed
CSCSR
scheme adopts sparse signal modeling
and smoothness regularization to exploit the rich spatial
HSI
information and improve
HSI
-classification accuracy.
Smoothness regularization can show different effects of
neighboring pixels on
HSI
classification.
2. The feature-sign search algorithm (Lee
et al.
2007) is im-
proved and used to solve our
CSCSR
optimization problem.
The improvement lies in the assumption that a testing
pixel can be completely reconstructed by the model dic-
tionary. Hence, the mean of the sparse coefficient vectors
can be iteratively updated.
3.
MH
-based preprocessing is combined with the
CSCSR
meth-
od to achieve state-of-the-art
HSI
-classification performance.
Related Work
One of the early successful applications of sparse coding was
face recognition (Wright
et al.
2009). Since then, sparse coding
has been widely applied in the fields of pattern recognition
and image processing, especially for signal classification. Next,
a brief overview of sparse coding is given. Let
y
denote a test
sample to be classified and
D
= [
D
1
,
D
2
, …,
D
L
] be a dictionary
whose columns are training samples, where
L
is the number of
Yang Liu, Xiaofei Ji, and Yangyang Wang are with the School
of Automation, Shenyang Aerospace University, Shenyang,
China (
,
,
).
Ruisheng Wang is with the Department of Geomatics
Engineering, University of Calgary, Calgary, Alberta, Canada
(
).
Photogrammetric Engineering & Remote Sensing
Vol. 85, No. 9, September 2019, pp. 659–672.
0099-1112/19/659–672
© 2019 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.85.9.659
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2019
659
