Spatial-Spectral Unsupervised Convolutional
Sparse Auto-Encoder Classifier for
Hyperspectral Imagery
Xiaobing Han, Yanfei Zhong, and Liangpei Zhang
Abstract
The traditional spatial-spectral classification methods ap-
plied to hyperspectral remote sensing imagery are conducted
by combining the spatial information vector and the spectral
information vector in a separate manner, which may cause
information loss and concatenation deficiency between the
spatial and spectral information. In addition, the traditional
morphological-based spatial-spectral classification methods
require the design of handcrafted features according to expe-
rience, which is far from automatic and lacks generalization
ability. To automatically represent the spatial-spectral fea-
tures around the central pixel within a spatial neighborhood
window, a novel spatial-spectral feature classification method
based on the unsupervised convolutional sparse auto-encoder
(
UCSAE
) with a window-in-window strategy is proposed in this
study. The
UCSAE
algorithm features a unique spatial-spectral
feature extraction approach which is executed in two stages.
The first stage represents the spatial-spectral features within a
spatial neighborhood window on the basis of spatial-spectral
feature extraction of sub-windows with a sparse auto-en-
coder (SAE). The second stage exploits the spatial-spectral
feature representation with a convolution mechanism for the
larger outer windows. The
UCSAE
algorithm was validated by
two widely used hyperspectral imagery datasets (the Pavia
University dataset and the Washington DC Mall dataset)
obtaining accuracies of 90.03 percent and 96.88 percent,
respectively, which are better results than those obtained by
the traditional hyperspectral spatial-spectral classification
approaches.
Introduction
Hyperspectral imagery (
HSI
) data have become a valuable tool
in a wide variety of applications (Chang
et al.
, 2013; Fauvel
et
al.
, 2013) such as agriculture, surveillance, astronomy, min-
eralogy, and environmental sciences, due to the rich spectral
and spatial information (Chuvieco
et al.
, 1989; Congalton
et
al.
, 2015; Lu
et al.
, 2013; Qin
et al.
, 2014; Svejkovsky
et al.
,
2012). Among all of the above research fields, the most com-
mon utilization of
HSI
data is for ground feature classification,
which usually means classifying each pixel from the
HSI
into
an accurate land-cover category (Grahn
et al.
, 2007; Camps-
Valls
et al.
, 2014; Landgrebe
et al.
, 2003).
HSIs contain rich spectral information, which increases the
possibility of more accurately discriminating ground features.
Furthermore, with the development of hyperspectral sensors,
the fine spatial resolution of HSIs allows most neighboring pix-
els to contain homogeneous spectral profile signatures, which
enables the small spatial structures in the images to be correct-
ly delineated (Ji
et al.
, 2014; Jimenez
et al.
, 2005; Kang
et al.
,
2014; Yuan
et al.
, 2014; Zhou
et al.
, 2015). Therefore, efficiently
exploiting the combination of the finer spatial and spectral in-
formation from the
HSI
is of significance to further improve the
classification performance of the ground features. Various spa-
tial-spectral feature classification methods have been proposed,
including algorithms based on neighborhood window opening
operations (Chen
et al.
, 2014; Plaza
et al.
, 2009), morpholog-
ical operations (Fauvel
et al.
, 2008; Mauro
et al.
, 2011), and
segmentation approaches. In general, all these methods can be
categorized as spatially constrained approaches, which process
the spatial and spectral features in a discrete manner and adopt
a handcrafted feature design approach, especially the morpho-
logical-based methods. However, there are two significant as-
pects impeding the development of the traditional spatial-spec-
tral classification methods. On the one hand, the handcrafted
feature design approaches need expert human experience to
design the specific parameter settings for different experimental
data, which is far from automatic. On the other hand, the direct
concatenation of the spatial and spectral features within a cer-
tain local spatial neighborhood window in the
HSI
may lead to
a low-efficient feature representation and is unable to mine the
deeper-level features within the spatial window. Whether there
could be an algorithm that can take both the automatic feature
extraction ability and exhaustive feature representation within
a certain spatial window into consideration is an interesting
and deserving research direction.
From the above description, automatically and effectively
representing the spatial and spectral features within a certain
spatial neighborhood window is a critical problem. Along
with the development of deep learning (Hinton and Salakhut-
dinov, 2006; Hinton
et al.
, 2006; Bengio
et al.
, 2007), an auto-
matic feature learning and feature representation framework
for classification tasks has also been constructed (LeCun
et al.
,
2015). Among the deep learning research models, supervised
feature learning models and unsupervised feature learning
models are the two typical feature learning approaches. A typ-
ical supervised feature learning model is the stacked sparse
auto-encoder (
SSAE
). A typical unsupervised feature learn-
ing model is the sparse auto-encoder (
SAE
). In recent deep
learning research, some models have been based on non-con-
volutional models (e.g., the
SSAE
) and some have been based
on the convolution mechanism. Generally, for the non-con-
volutional models, the spatial and spectral information are
processed separately (Chen
et al.
, 2014), and this is usually
undertaken by concatenating the spatial and spectral informa-
tion vectors, which may cause information loss. Meanwhile,
State Key Laboratory of Information Engineering in Sur-
veying, Mapping, and Remote Sensing, Wuhan University,
Wuhan 430079, P.R. China; and the Collaborative Innovation
Center of Geospatial Technology, Wuhan University, Wuhan
430079, P.R. China (
).
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 3, March 2017, pp. 195–206.
0099-1112/17/195–206
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.3.195
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2017
195
