Hyperspectral Compressive Sensing Imaging Via Spectral Sparse
Constraint
Qi Wang
1
, Hong Xu
2
, Lingling Ma
1,*
, Chuanrong Li
1
, Yongsheng Zhou
1
and Lingli Tang
1
1
Key Laboratory of Quantitative Remote Sensing Information Technology, Academy of Opto-Electronics,
Chinese Academy of Sciences, Beijing, 100094, China
2
High Technology Research and Development Center, Ministry of Science and Technology, Beijing 100862, China
Keywords: Compressive Sensing, Hyperspectral Imaging, Sparse Representation, Dictionary Learning, Reconstruction
Algorithm.
Abstract: The existing algorithms to reconstruct hyperspectral compressive sensing images mainly use the sparse
property of spatial information and some simple non-adaptive spectral constraint such as the low-rank
property. However, these strategies cannot remove the spectral redundancy efficiently and a new method to
make full use of the abundant redundancy of spectral information and improve the quality for hyperspectral
CS reconstruction is necessary. A new CS sampling and reconstruction model based on spectral sparse
representation was proposed in this paper. The spectral sparse dictionary was constructed from training
samples to enhance the effect of sparse representation and the total variation constraint of spatial images was
also considered to further enhance the precision during the reconstruction. The experiment to reconstruct
AVIRIS hyperspectral images of 200 bands show that the hyperspectral image was almost perfectly
reconstructed at 25% sampling rate and the spatial and spectral precision was higher than traditional methods
which only adopt the spatial sparsity and simple non-adaptive spectral constraint in the same condition.
1 INTRODUCTION
Hyperspectral remote sensing technique has the
ability to acquire and analyze the physical and
chemical properties of land surface objects. Over the
past 30 years it has witnessed a great progress in
various fields such as atmosphere monitoring, ocean
monitoring, mineral exploration, precision agricul-
ture and target detection. However, hyperspectral
images have much larger amount of data which makes
data transmission, storage and onboard processing
more difficult. Furthermore, the optics system will
also become more complicated and costly with the
improvement of resolution. As far as the traditional
imaging system based on Nyquist sampling theory is
concerned, it is almost impossible to conquer the
contradiction between high efficiency and high
resolution demand. Therefore, new theory is expected
to be developed so as to promote more efficient
application of hyperspectral remote sensing.
As a novel theory in signal processing,
compressive sensing (CS) theory which integrates
compression and sampling process has drawn much
attention in many fields including remote sensing
imaging (Donoho, 2006). By concerning the sparse
characteristic of the object, images can be
reconstructed from very few measurements.
Therefore the size and complexity, as well as the on-
orbit computational cost of the CS imaging system is
much lower than conventional systems, making it a
very promising application in the remote sensing (Li,
2014). A. Wagadarikar (2008), Thomas A. Russell
(2012), and J. Wu (2014) have developed various CS
spectral imaging systems respectively.
The images of CS imaging system are
reconstructed from compressive measurements by
solving specific optimization problem. The sparse
characteristics of target signal is the foundation of CS
reconstruction: the better the signal is sparsely
represented, the higher reconstruction precission
would be obtained. The total variation (TV) model
based on the sparsity of two dimensional discrete
gradient was widely used to reconstruct spatial
images (Combettes, 2014). However, in the
hyperspectral remote sensing imaging which
demands for higher compression rate, the
Wang, Q., Xu, H., Ma, L., Li, C., Zhou, Y. and Tang, L.
Hyperspectral Compressive Sensing Imaging via Spectral Sparse Constraint.
DOI: 10.5220/0006664202730278
In Proceedings of the 6th International Conference on Photonics, Optics and Laser Technology (PHOTOPTICS 2018), pages 273-278
ISBN: 978-989-758-286-8
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
273
reconstruction effect of non-adaptive total variation
model is usually dissatisfied. There is strong
correlation among the bands of hyperspectral image.
Regarding to this characteristic, several constraint
models using spectral correlation had been proposed.
Zhang et. al. extended the traditional two-
dimensional TV model to 3 dimensions in order to
reconstruct hyperspectral images (Zhang, 2014).
Feng et. al. (2012) used the current reconstructed
band as the reference to predict the following band.
Liu et. al. (2011) presented the hyperspectral
reconstruction algorithm based on the prediction of
the residual vector. Golbabaee et. al. (2012) applied
the low-rank property of hyperspectral data bands and
reconstructed the image by adding the nuclear norm
to the constraint model. Jia et. al. (2014) put forward
a structural-relation-based model via researching the
spectral statistical correlation of hyperspectral image.
These spectral correlation models do enhance the
reconstruction precision to some extent, but the
spectral constraint models are relatively simple and
non-adaptive, thus the detailed spectral information is
difficult to be reconstructed correctly at a low
sampling rate. This shortage is more notable for the
hyperspectral scenes with a wider spectral range and
more bands of data.
Actually, most of the earth surface spectrum
possesses piecewise smooth property with a large
amount of redundant information. The spectral
information redundancy is usually more abundant
than that of the complex spatial information and is not
well removed using the available methods such as
neighboring band prediction or low-rank
minimization of the HSI data. On the other hand, the
sparse representation theory suggests that, the signals
can be precisely represented by very few atoms from
some kinds of dictionaries that can highly reduce the
information redundancy. Recently some researchers
applied the sparse representation theory in the
hyperspectral fields such as spectral classification,
unmixing, and reconstruction (Zare, 2012, Charles,
2011, Wang, 2013). By exploiting the spectral sparse
property, researchers obtained better results than
conventional methods, making it an effective way to
process HSI data.
On such basis, the sparse representation theory is
introduced to reconstruct compressive sensing
hyperspectral images in this paper and a new
hyperspectral sampling model is proposed to
precisely reconstruct the images in the CS scheme. In
Section II, the sparse representation theory and the
method of spectral sparse dictionary learning are
discussed. In Section III, the principles of spectral
compressive sensing model and the hyperspectral
image reconstruction algorithm based on the spectral
sparse dictionary are presented. In Section IV, the
experiment of sampling and reconstructing the
AVIRIS scene is conducted, and our method as well
as three different hyperspectral reconstruction
methods is applied in comparison. In Section V, the
content of this paper is concluded and the problems
and prospects are analyzed.
2 SPECTRAL SPARSE
DICTIONARY LEARNING
Suppose x is a signal of length N and D=[d
1
,d
2
,…d
K
]
is a set of basis (also called as dictionary) in the N-
dimensional Euclidean space. If x can be linearly
represented by the L atoms in the dictionary (usually
L<<K, N), then we declaim that x has a sparse
representation in the dictionary D, the sparsity level
is L. That is
1
, 1,2...
i
L
ii
i
x d K
(1)
where ε is the residual error of the sparse
representation. (1) can be rewritten in the matrix form:
x Ds

(2)
In which s, called as the sparse coefficient, is the
coordinate vector with only L non-zero elements. The
process to calculate s is called as sparse coding and it
solves the minimization problem:
0
min , . .s st x Ds
(3)
where || ||
0
denotes the number of non-zero
elements and (3) can be solved by several algorithms.
A more important question of the sparse
representation is the way to construct the dictionary.
In the area of signal and image processing, DCT
dictionary, wavelet dictionary and Gabor dictionary
are usually adopted. The atoms in these dictionaries
are fixed and hard to be adjusted adaptively, which
results in the limitation that more efficient and
accurate decomposition is needed for hyperspectral
CS reconstruction. In recent years the adaptive
dictionary training method from the characterized
sample set is developed and used in the sparse
dictionary construction by many researchers. The
dictionary learning method is an adaptive approach
for complex signals and usually achieves a better
result than conventional dictionaries such as DCT and
wavelet. The main idea of dictionary training is to
find a matrix that ensures each sample in the training
PHOTOPTICS 2018 - 6th International Conference on Photonics, Optics and Laser Technology
274
set a sparse representation by such matrix. The
sparsity level and the residual error in the
representation are the optimization goal in the
calculation. For the hyperspectral dictionary
construction in this paper, the spectral sparse
dictionary model is presented as
20
,
min . .
i
DX
W DS st i s L
(4)
where the matrix W is composed by the column
vectors of spectral sample set that are extracted from
hyperspectral scenes, D and S are the sparse
dictionary and the coefficients to be solved
respectively, and L is the sparsity level control
parameter. The solving process of (4) consists of two
key steps called as sparse coding and dictionary
updating. The sparse coding step is to solve the sparse
coefficients S under the dictionary D and the updating
step is to adjust D according to the new coefficients.
The two steps are executed alternatively until
convergence. Based on the different approach to
realize dictionary updating, a series of dictionary
learning algorithms such as KSVD, MOD and RLS
are presented (Aharon,2006, Mairal, 2009, Skretting,
2010). Here we choose the KSVD algorithm as the
dictionary learning algorithm and fast OMP as the
sparse coding algorithm for their relatively high
accuracy and efficiency (Azimi, 2014).
3 HYPERSPECTRAL SAMPLING
AND RECONSTRUCTION
Suppose the matrix X represents a hyperspectral data
cube, with N=n
1
*n
2
spatial pixels and B spectral
bands. The compressive sampling of hyperspectral
object is conducted in the spatial region usually. x
i
denotes for the column vector with length N expand-
ing from the i-th band image, and y
i
denotes for the
corresponding measurements. To sample the scene by
the linear mixing matrix P in each band, we have
1 2 1 2
[ , ,..., ] [ , ,..., ]
BB
y y y P x x x
(5)
The usual way to solve (5) is to reconstruct each
spatial image in the constraint of total variation and
adjust the solution based on the spectral correlation
model. As pointed before, the spectral sparse property
with a large amount of information redundancy is not
exploited sufficiently in this model. Therefore, we
propose a new sampling scheme in the spectral region.
λ
i
denotes the spectral data vector with length B of the
i-th spatial pixel and y
i
denotes the corresponding
compressive measurements. The sampling process is
applied in the whole spatial region, that is
1 2 1 2
[ , ,..., ] [ , ,..., ]
NN
y y y P
 
(6)
And the hyperspctral scene is reconstructed based
on (6).
In the spectral sparse model of Section II, the
spectral signal in (6) can be decomposed as the
product of sparse dictionary and coefficients, that is
, 1,2,...,
ii
Ds i N

(7)
Define matrix A=PD and put (7) into (6):
i i i
y PDs As
(8)
And the sparse coefficients s
i
is determined from
solving the optimization problem
(9)
As the sparse representation error is unavoidable,
(9) is presented as an unconstrained optimization
problem with the regularizer β
0
:
2
0
20
1
min
2
i
i i i
s
y As s

(10)
The optimization of the l
0
norm is a NP hard
problem. Therefore we use smooth Gaussian function
to approximate the l
0
norm and convert the problem
to classical convex optimization problem which can
be solved via gradient descent pursuit algorithm. The
detailed steps are presented in (Mohimani, 2007).
Solve (10) for the spectrum in each pixel and
calculate the reconstructed spectrum via (7) and
arrange all the spectrum into a three dimensional
matrix to construct the raw reconstructed
hyperspectral data cube X
0
. The accuracy of spatial
information of the reconstructed image is hard to
guarantee with the spectral constraint only. Therefore
the raw reconstructed data is revised in the constraint
of total variation to solve the optimization problem
with the regularizer β
1
:
2
1
2
1
min
2
j
j j j
TV
x
x x x


(11)
in which x
j
is the j-th band spatial image of X
0
and
x
j
is the revised image in the constraint of total
variation. The TV norm represents for
1
1i i i n i
TV
i
x x x x x

(12)
Solve (12) for each band in X
0
to construct the
data cube X by applying the algorithm in
(Chambolle,2004) and further revise the spectrum via
Hyperspectral Compressive Sensing Imaging via Spectral Sparse Constraint
275
solving the optimization problem which is similar to
(10):
22
2
2 2 0
11
min
22
i
i i i i i
s
y As s s s
(13)
Calculate (11) and (13) alternatively to revise
spatial and spectral solution of the reconstructed
scene until the convergence is reached:
2
'
2
2
2
0.001
XX
X
(14)
4 EXPERIMENTS AND RESULTS
Spectral Dictionary Learning: The spectral data from
the remote sensing images acquired by AVIRIS
hyperspectral imagery at the range of 0.4-2.5 μm is
selected as the dictionary learning samples. The
sample set contains 5000 spectrums from 38 types of
ground objects in the scene Indian pines, Salinas,
Cuprite and Kennedy Space Center. Since some
bands are influenced by water vapor absorption, 24
bands are discarded and the data of 200 bands are
used eventually. The spectral data is normalized to
avoid the difference of the intensity in the different
scene and some types of training samples are shown
in Fig. 1. In the dictionary training algorithm we set
the sparsity level L=5, the iteration times T=30, and
the number of atoms K=400. The parameters are
selected from many experiments considering both the
efficiency and precision of the algorithm, while the
influence of the parameter changing on the dictionary
and the method to decide the best parameter is yet to
be further studied.
Experiment 1: The experiment scene to be
reconstructed is selected from the area of Indian Pines
containing 128*128 pixels (the test area is not
included in the training set). The compressive
sampling is conducted to the scene by the random
Bernoulli matrix that has good incoherence property
and is sustainable by hardware. The sampling rate is
25%, and Gaussian noise of 40dB SNR is added to
the measurements. The original data of 200 bands are
compressed to 40 bands after sampling.
The algorithm presented in Section III is applied
to reconstruct the hyperspectral scene from the
compressive measurements. The regularizers of β
0
, β
1
and β
2
are set to 0.4, 0.1, and 0.1 respectively. In
contrast, three different hyperspectral CS
reconstruction algorithms are also tested including
TV algorithm in (Combettes,2004), TVSS algorithm
in (Liu,2011) and TVNU algorithm in
(Golbabaee,2012). Fig. 2 shows the reconstructed
image of the 100th band via the four algorithms in the
same experimental condition. The quality of the
reconstructed image by the proposed method is
significantly better that other algorithms and it
effectively avoids the over-smoothing problem
blurring the image details by TV constraint
algorithms.
Figure 1: Some types of spectrum in the training set
extracted from AVIRIS data containing 200 valid bands.
The spectral intensity is normalized.
The spectral curve of corn in the scene is
compared before and after reconstruction in Fig.3 via
different algorithms. The reconstructed spectrum by
the proposed method achieves very high accuracy
with little deviation. In contrast, the TV algorithm
fails at some bands and makes the spectrum
discontinuous due to the lack of constraint in the
spectral region. The TVSS and TVNU algorithm with
spectral constraint maintain the continuity of the
reconstructed spectrum to some extent, but there are
apparent errors in some details especially at the
wavelength 0.5-1.0μm where the spectrum has a wide
fluctuation.
Experiment 2: Change the sampling rate and
reconstruct the hyperspectral scene via different
algorithms. The sampling rate is set from 10% to 50%.
To assess the reconstruction precision, two
parameters represented for space and spectrum
respectively are calculated for each reconstructed
hyperspectral scene. One is the mean value of the root
square error (RMSE) of each spatial band and the
other is the mean value of the spectral angle (MSA)
between the reconstructed spectrum and the original
one. The result is shown in Fig.4. The precision of the
reconstructed scene via the proposed method is better
than other algorithms in both spatial region and
spectral region. The advantage of our method is more
dominant especially at low sampling rate due to the
precise sparse constraint in the spectral region.
0.5 1 1.5 2 2.5
0
0.05
0.1
0.15
0.2
Wavelength(μm)
Normalized intensity
Woods
Corn
Alfalfa
Grass-pasture
Wheat
PHOTOPTICS 2018 - 6th International Conference on Photonics, Optics and Laser Technology
276
(a) (b) (c) (d) (e)
Figure 2: The 100th band reconstructed image from the AVIRIS scene Indian pines via different algorithms from
25%sampling measurements. (a)Original image. (b) Our method. (c)TV algorithm. (d)TVSS algorithm. (e)TVNU algorithm.
(a) (b)
(c) (d)
Figure 3: The corn spectrum in the AVIRIS scene Indian pines before and after reconstruction via different algorithms from
25%sampling measurements. (a) Our Method. (b)TV algorithm. (c)TVSS algorithm. (d)TVNU algorithm.
(a) (b)
Figure 4: The quality assessment parameters of the reconstructed hyperspectral image at different sampling rate varying from
0.1 to 0.5 in the same CS experiment. (a) The RMSE value to assess spatial reconstruction quality. (b)The MSA value to
assess spectral reconstruction quality.
0.5 1 1.5 2 2.5
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Wavelength(μm)
Normalized intensity
True spectrum
Reconstructed spectrum
0.5 1 1.5 2 2.5
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Wavelength(μm)
Normalized intensity
True spectrum
Reconstructed spectrum
0.5 1 1.5 2 2.5
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Wavelength(μm)
Normalized intensity
True spectrum
Reconstructed spectrum
0.5 1 1.5 2 2.5
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Wavelength(μm)
Normalized intensity
True spectrum
Reconstructed spectrum
0.1 0.2 0.3 0.4 0.5
0
0.005
0.01
0.015
0.02
0.025
Sampling Rate
RMSE
Our Method
TV
TVSS
TVNU
0.1 0.2 0.3 0.4 0.5
0
0.1
0.2
0.3
0.4
0.5
Sampling Rate
MSA
Our Method
TV
TVSS
TVNU
Hyperspectral Compressive Sensing Imaging via Spectral Sparse Constraint
277
5 CONCLUSIONS
Aiming at the problem to utilize the spectral sparse
property in the hyperspectral CS remote sensing
imaging, this paper presents a new sampling and
reconstruction method based on the spectral sparse
representation. By learning the spectral sparse
dictionary to constrain the spectral region in the
reconstruction and optimizing the spatial precision
via total variation constraint, the AVIRIS
hyperspectral scene is reconstructed in very high
quality from 25% compressive measurements, which
provides a new idea to enhance the hyperspectral
sampling efficiency. Compared with other presented
hyperspectral CS reconstruction algorithms, the
reconstruction precision in spatial and spectral region
of our method has a significant superiority in the same
experimental condition.
However, there still exist some problems to be
further studied in order to better apply the new theory.
One is the method of the spectral training sample
construction and dictionary learning. In the
experiment it is found that if the training samples are
extracted from the sensor or the type of ground object
with a great difference from that of the reconstructed
area, the effect of the spectral sparse representation is
significantly affected and the reconstruction precision
decreases. The other one is the realization of spectral
random coding in hardware, for the spectral sampling
scheme is more difficult to realize than spatial
sampling.
ACKNOWLEDGEMENTS
This work is supported by the National Key Research
and Development Program of China under Grant
2016YFB0500402, CAS/SAFEA International
Partnership Program for Creative Research Teams
under Grant 2013AA1229 and the Strategic Priority
Research Program of the Chinese Academy of
Sciences under Grant XDA13030402.
REFERENCES
Donoho, David L, 2006. Compressed sensing. IEEE Trans.
Inform. Theory. 52.4, p.1289-1306.
Li, Sheng Liang, et al. 2014, Innovative remote sensing
imaging method based on compressed sensing. Optics
& Laser Technology. 63.4, p.8389.
Ashwin, Wagadarikar, et al. 2008, Single disperser design
for coded aperture snapshot spectral imaging. Applied
Optics. 47.10, p.44-51.
Russell, Thomas A., et al. 2012, Compressive hyperspectral
sensor for LWIR gas detection. Proceedings of SPIE -
The International Society for Optical Engineering.
8365, p. 83650C-83650C-13.
Wu Jianrong, Shen Xia, Yu Hong, et al. 2014, Snapshot
compressive imaging by phase modulation. Acta Optica
Sinica. 10, p.113-120.
Combettes, Patrick L, and P. Jean-Christophe. 2004, Image
restoration subject to a total variation constraint. IEEE
Transactions on Image Processing a Publication of the
IEEE Signal Processing Society. 13.9, p.1213-22.
Zhang L, Zhang Y, Wei W, 2014, 3D total variation
hyperspectral compressive sensing using unmixing.
Geoscience and Remote Sensing Symposium (IGARSS)
2014, p.2961-2964.
Feng Yan, JiaYingbiao, Cao Yuming, et al. 2012,
Compressed sensing projection and compound
regularizer reconstruction for hyperspectral images.
Acta Aeronautica et Astronautica Sinica, 33(8), p.1466-
1473.
Liu Haiying, Wu Chengke, Lv Pei, et al. 2011, Compressed
hyperspectral image sensing reconstruction based
oninterband prediction and joint optimization. Journal
of Electronics & Information Technology, 33(9), p.
2248-2252.
Golbabaee, Mohammad, and P. Vandergheynst. 2012,
Compressed sensing of simultaneous low-rank and
joint-sparse matrices. IEEE Transactions on
Information Theory .
JiaYingbiao, Feng Yan, Wang Zhongliang, et al. 2014,
Hyperspectral compressive sensing recovery via
spectrum structure similarity. Journal of Electronics
and Information Technology, (6), p.1406-1412.
Zare, A., P. Gader, and K. S. Gurumoorthy. 2012, Directly
measuring material proportions using hyperspectral
compressive sensing. Geoscience & Remote Sensing
Letters IEEE. 9.3, p.323-327.
Charles, A. S., Olshausen, B.A., and C. J. Rozell. 2011,
Learning sparse codes for hyperspectral imagery.
Selected Topics in Signal Processing IEEE Journal
of. 5.5, p.963-978.
Wang Qi, Li Chuanrong, Ma Lingling, et al. 2013,
Compressive sensing spectral sparsification method
based on training dictionary. Remote Sensing
Technology and Application, 28.6, p.59-60.
Aharon, M., M. Elad, and A. Bruckstein. K-SVD: An
algorithm for designing overcomplete dictionaries for
sparse representation. IEEE Transactions on Signal
Processing, 54.11(2006):4311-4322.
Mairal, Julien, et al. 2009, Online dictionary learning for
sparse coding. Proceedings of the 26th Annual
International Conference on Machine Learning ACM,
p. 689-696.
Skretting, Karl, and K. Engan. 2010, Recursive Least
Squares Dictionary Learning Algorithm. IEEE
Transactions on Signal Processing. 58.4, p.2121 - 2130.
Azimi-Sadjadi M R, Kopacz J, Klausner N. 2015, K-SVD
dictionary learning using a fast OMP with applications.
IEEE International Conference on Image Processing. p.
1599-1603.
PHOTOPTICS 2018 - 6th International Conference on Photonics, Optics and Laser Technology
278
Mohimani, G. Hosein, M. Babaiezadeh, and C. Jutten. 2007,
Fast Sparse Representation Based on Smoothed L0
Norm. Proceedings of the 7th international conference
on Independent component analysis and signal
separation, Springer-Verlag, p.389-396.
Chambolle, Antonin. 2004, An Algorithm for Total
Variation Minimization and Applications. Journal of
Mathematical Imaging & Vision. 20.1-2, p.89-97.
Hyperspectral Compressive Sensing Imaging via Spectral Sparse Constraint
279