IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 9, NO.

2, FEBRUARY 2016 655

Adaptive Morphological Filtering Method

for Structural Fusion Restoration of
Hyperspectral Images
Yidan Teng, Student Member, IEEE, Ye Zhang, Yushi Chen, Member, IEEE, and Chunli Ti

Abstract—Recovering hyperspectral image (HSI) from mixed applications including target detection, classification, and map-
noise degradation is a challenging and promising theme in remote ping. In order to mitigate these influences, HSI restoration
sensing, particularly when stripes and deadlines exist in sev- is an active and challenging issue in both preprocessing and
eral contiguous bands. This paper proposes a HSI’s restoration
method making use of adaptive morphological filtering (AMF) postprocessing areas.
and fusing structure information of an auxiliary color image. An In the past decades, many different preprocessing restoration
adaptive structuring element (ASE) indicating morphological fea- methods for HSI have been proposed and achieved good results.
tures of each pixel is generated through information fusion, to The existing restoration methods can be mainly classified into
simultaneously remove the mixed noise and preserve fine spatial three types: 1) filter or morphology-based methods; 2) spatial–
structures. This key technology contains three main steps. First,
edges are extracted from the auxiliary image exploiting its color spectral methods; and 3) the sparse representation and low-rank
information; then, an edge-constraint growing algorithm is used (SR–LR)-based methods.
to generate the clustering kernel; finally, the ASE is obtained via Filter- or morphology-based methods are a type of the most
goal-guided k-means clustering. The ASE has extensive applica- traditional and practical HSI restoration methods. Generally
tion value, for it can be an enhancing module for most filters-based speaking, these approaches pay most attention to the spatial
restoration methods, to mitigate the structural damage due to the
fixed mask. Among these methods, Gaussian filter for prepro- information, and separately restore each band of the HSI as
cessing and majority voting for postprocessing are introduced in a gray-level image. They are explicit and efficient for esti-
this paper as representatives. In addition, the auxiliary image can mating target via immediate calculations on its neighboring
be both visible image of multisensor and false RGB component pixels, and their defects are also evident: the structure infor-
of the undamaged bands of the HSI, so it is relatively available. mation will be increasingly damaged with the growth of the
Experiments on simulated and real data sets show obvious effects
on denoising and destriping both subjectively and objectively. The filter window size. Aim at this problem, multifarious improve-
advantage of ASE on structure details preserving, compared to ments are implemented constantly and have made significant
conventional approaches, is clearly demonstrated. The application progress. Wang and Niu [3], [4] proposed denoising approach
value of the proposed restoration frame and ASE is further proved using alternative hyperspectral anisotropic diffusion to preserve
through the decision-level postprocessing experiments. the edges of objects. Mathematical morphology (MM) filtering
Index Terms—Adaptive morphological filtering (AMF), hyper- methods try to highlight structure feature by using operational
spectral image (HSI), information fusion, postprocessing restora- windows with specific shape, which is generally called struc-
tion, preprocessing restoration.
turing element (SE). To allow more freedom in forming SE
that can adapt their shapes according to the local features of
I. I NTRODUCTION
the processed images, adaptive morphological operators are

D UE TO THE rapid development of hyperspectral sensors,

hyperspectral image (HSI) has the ability to capture a
continuous spectral curve and obtain detailed spatial–spectral
further developed in [5]. For HSI processing, Song et al. [6]
improve the morphological filtering via example-based learn-
ing and Markov random field (MRF). In [7], spatial structures
information. Unfortunately, the actual acquired HSIs are often in HSI are detected using a supervised framework according to
degraded by kinds of annoying degradation, e.g., noise contam- their spectral nature.
ination, stripe corruption, and missing data, despite the advance In recent years, more attention has been paid to the restora-
in hyperspectral sensors [1], [2]. These noises have a negative tion methods that use both spatial and spectral information, and
impact on not only the visual perception but also the intelligent these procedures take better advantage of relevance between
Manuscript received September 15, 2014; revised June 17, 2015; accepted different bands. The hybrid spatial–spectral derivative-domain
August 03, 2015. Date of publication August 25, 2015; date of current version wavelet shrinkage noise reduction (HSSNR) approach pro-
February 09, 2016. This work was supported by the National Natural Science posed by Othman and Qian [8] resorts to the spectral derivative
Foundation of China under Grant 61471148 and Grant 61301206.
domain and benefits from the dissimilarity of the signal regular-
The authors are with the Department of Information Engineering, School
of Electronics and Information Engineering, Harbin Institute of Technology, ity in the spatial and the spectral dimensions. Then, Chen and
Harbin 150001, China (e-mail: [email protected]; [email protected]; Qian [9] further improve the above method by adding a PCA
[email protected]; [email protected]). transform before spatial–spectral wavelet shrinkage. Karami
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
et al. [10] evaluated a new HSI denoising method based on
Digital Object Identifier 10.1109/JSTARS.2015.2468593 genetic algorithm and kernel hierarchical nonnegative Tucker

1939-1404 © 2015 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.
656 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 9, NO. 2, FEBRUARY 2016

decomposition, which also utilizes the spectral and the spatial refining the raw classification; the accuracy of image segmen-
information concurrently. In [11], Chen et al. proposed a novel tation affects a lot to the final performance. MRFs model is
spatial–spectral domain mixing prior taking advantage of dif- an effective probabilistic postprocessing strategy; it is used to
ferent properties of HSI in the spatial and spectral domain. integrate spatial information into image classification problems
Qian and Ye [12] introduced nonlocal similarity and spectral– [26]–[28]. In [29], by using MRF regularization, spatial con-
spatial structure of HSI into sparse representation and improved text information is added to refine classification map obtained
the quality of HSI for object detection and classification. Yuan by support vector machine. In [30], a new supervised approach
et al. [13] proposed a HSI denoising algorithm employing a is proposed by combining spectral and spatial information in
spectral–spatial adaptive total variation (SSAHTV) model. In HSI segmentation. Spatial information is included using a mul-
the SSAHTV algorithm, both the spectral noise differences and tilevel logistic Markov–Gibbs MRF prior. Moreover, in [31],
the spatial information differences are considered in the process the framework further improves the classification map taking
of the noise reduction. They [14] further modified the above full advantage of the spatial and contextual information present
algorithm with a spatial–spectral view fusion strategy. in HSI through a spatially adaptive total variation (SpATV)
SR–LR matrix approximation has emerged as a class of regularization.
powerful tool for HSI analysis and recovery. Unlike the tradi- Generally, the existing methods have made remarkable
tional image restoration methods, SR–LR-based methods take achievements in HSI restoration under certain circumstance
advantage of the assumption that components of the clean or for specific application. However, they lack the stability
HSI have the character of LR in matrix analysis, while the and adaptability when mixed noises exist, especially in several
stripes and deadlines are sparse. A large number of approaches pixels or bands contiguous stripes removing. For the spatial–
have been developed for solving ill-conditioned problem of spectral and SP–LR methods, the correlation of noise among
separating clean HSI and noises, and many of them use the spectrum is intolerable; and these procedures may bring non-
sparsity characteristic in wavelet domain. In [15], sparse syn- linear impact to the original spectrum curves. Besides, it is too
thesis regularization was proposed for HSI denoising using an complex and inefficient for HSI with large scenes and so many
orthogonal wavelet transform and a group Lasso penalty. In bands, for the need of iterative or global optimization. For the
[16], three-dimensional (3-D) nonlocal means filtering from filter- or morphology-based methods, bigger masks need to be
[17] are applied for noise reduction for HSI. A 3-D (blockwise) utilized for suspected wider gaps, which would lead to more
nonlocal sparse denoising method [18] is given in [19]. On the serious damage to the local structure. Nevertheless, this degra-
other hand, Golbabaee and Vandergheynst [20] and Zhang et al. dation is usually linear and can be restrained by exploiting prior
[21] solved the problem utilizing the LR property of HSI and knowledge. Consequently, there is strong potential to develop
expressed the HSI by a linear combination of a limited num- filters- and morphology-based methods into broader applicable
ber of endmembers. Therein, 3-D LR models express better and more efficient HSI restoration system.
potential. In [21], an LR matrix recovery model was devel- In this paper, we investigated and developed a novel HSI
oped to simultaneously remove several types of noises, such as restoration system which combines the efficient filtering meth-
Gaussian noises, impulse noises, stripes, and deadlines. In [22], ods with the adaptive SE (ASE) containing the local morpho-
Zhu et al. introduced spectral nonlocal method (SNL) into the logical details to preserve the spatial structure better. With the
LRMR to further preserve the fine structures. These methods extensive application of the hyperspectral and multisource sen-
can operate effectively in general; however, when stripes and sors, structural information becomes easier to obtain by making
deadlines exist in the same location of some bands, the noisy use of the visible image of multisensor or RGB component of
components also have LR property in the 3-D HSI model, so the undamaged bands of the HSI. Another contribution of this
that they can be hardly separated. Besides, this kind of cir- paper is presenting an ASE generation method with three key
cumstance is not particularly uncommon, e.g., more than 20 steps. First, the edges are extracted from the auxiliary image
continuous bands have deadlines in the same position in EO-1 exploiting its color information. From one hand, edges general-
Hyperion image data sets. ize the geometry structure information and describe distribution
Without doubt that the preprocessing restoration for HSI is of high-frequency information; from the other, though the spe-
indispensable in most cases, in some specific application such cific information is different for all HSI bands or multisource
as classification, the preprocessing algorithms are excessively images, edges are the common information and demonstrate the
time-consuming when applied to HSI enormous data process- morphological features of the boundaries. Then, a small ker-
ing compared directly to postprocessing two-dimensional (2-D) nel is created through region growing constraint by the edges.
class map. Nowadays, researchers have switched their attention Finally, the ASE is generated utilizing the local goal-guided
to classification postprocessing (CPP). The CPP strategies are k-means clustering.
mainly summarized to three aspects [23]: filtering, object-based The ASE can distinguish influential pixels from the irrele-
voting, and random field. Filtering is the most widely used CPP vant ones in a relatively large local area which can cover the
algorithms, here a sliding window is centered at each pixel and widest gaps. The main common characteristic between the ASE
the class is assigned by considering values, labels, or proba- and the conventional SE in MM is that they both indicate the
bilities. Majority, Gaussian, edge-aware filters are conducted local morphological feature using prior knowledge. Classical
with postprocess classification results [24]. Majority voting is morphological filtering can only preserve limited types of geo-
a commonly used algorithm in object-based voting; in [25], metrical details which depend on the shape of the SEs, although
majority voting and object-based segmentation is combined in a series of different SEs can be simultaneously utilized to
TENG et al.: AMF METHOD FOR STRUCTURAL FUSION RESTORATION OF HSIs 657

improve its flexibility. The proposed ASE can actively reshape information of its neighborhood. In this paper, the ASE con-
itself to simulate the local morphological characteristics of structing procedure is realized via fusing intact structure infor-
the auxiliary image adaptively; thus, various kinds of spatial mation of an auxiliary color image. Three main steps are
structures are preserved. Moreover, the ASE has a wider appli- included as follows:
cability. The performance of most filters will be improved 1) edge extraction based on color information;
when the fixed mask is replaced by the proposed ASE, for 2) cluster-kernel growth constraint by edge;
the local details can be protected by ignoring irrelevant pixels. 3) local k-means cluster guided by target.
Undoubtedly, the ASE can enhance the effect of MM operators By replacing the fixed window with the ASE containing the
on dealing with the stripes and deadlines. However, restora- variable local morphological feature, most traditional effec-
tion results of MM filtering are unstable when Gaussian noise tive window-based filtering methods will renew themselves on
exists, so the corresponding description of procedure and exper- structural information preserving in the HSI restoration field.
iments are not given here. Instead, this paper introduces a
Gaussian filter combined preprocessing recovery method and
a majority voting combined method for the class map postpro- B. Acquisition of ASE Based on Fusion
cessing restoration as representatives. In Section II, the various 1) Edge Extraction Based on Color Information: First, the
steps of the proposed approach are explained in more depth. auxiliary color image, which can be obtained from either multi-
In Section III, experiments are conducted on simulated and source sensors or the false color synthetic of three intact bands
real hyperspectral data set with mixed noise to demonstrate of the HSI, is converted from RGB space to LUV space. LUV
the performance of the methodology. Section IV concludes this color space is recommended by the International Commission
paper. on Illumination (CIE), where L represents brightness, u and
v donate chromaticity. In the LUV space, the nature color
II. A DAPTIVE M ORPHOLOGICAL F ILTERING information and the relationship of refractive index at different
R ESTORATION M ETHOD FOR HSI wavelengths obtain more direct expression, which have more
robust correspondence with the ground truth.
A. Window-Based Filtering
Edge extraction first, respectively, calculates gradients
(x,y) (x,y) (x,y)
Window-based filtering is one of the most widely used GL , Gu , and Gv of the three channel L, u, and v uti-
image restoration methods. When used in HSI restoration, the lizing gradient operator, then we use weighted summation of
window-based filtering method treats each band as a gray- three gradient vector to express the color gradient
level image and restores them separately. Although it mainly
(x,y) (x,y)
takes advantage of the spatial information rather than spectral GColor = αL GL + αu G(x,y)
u + αv G(x,y)
v (1)
correlation, and precisely because of this, it has some irre-
placeable superiority-like parallel computing, bare nonlinear where αL , αu , and αv are the separate weights. In order to avoid
degenerates in the spectral dimension, and adaptability of data oversegmentation caused by multitime phases, atmospheric or
missing in continuous bands. These advantages make window- shadow conditions, and accidental error caused by similar lumi-
based filtering a worth noting and improving method in HSI nance of different objects, more attention should be paid to the
processing. chromaticity information. Therefore, αu and αv are obviously
Generally, the basic idea of window-based filtering method greater than αL . As a rule of thumb, the weights can be set
is to build a pointwise estimate of the image where each as αu = αv = 3 ∗ αL for the visible image. As for the false
pixel is obtained via arithmetic of pixels in a specific filter. color image, the weights should balance the degree of variance
Typical algorithms include weighted average, morphology, and among the selected bands. The more similar the selected bands
median filtering in preprocessing as well as majority voting in are, the lower the energy of chromaticity components is, and
postprocessing. Relatively speaking, the influence of the filter the greater their weights should be. They can thus be set as
window is also important, especially when large area data miss- EL EL
ing exists. With the growing of the window size, the ability αu = αe , αv = αe (2)
Eu Ev
of gap repairing and noise suppression enhances but the blur
effect gets worse at the same time. For HSI restoration, the where EL , Eu , and Ev are the energy of L, u, and v, and αe is
contradiction appears more fierce: on one hand, the mutually the empirical coefficient which is between 2 and 3.
exclusive relationship between spectral and spatial resolution In this paper, Canny operator is chosen to calculate gradi-
leads to more abundant geometry information of HSI within ents; through nonmaximum suppression (NMS) and hysteresis
the scope of the smaller image coordinates; on the other hand, thresholding on the basis of conventional method, we get the
HSI always faces more serious degradation including Gaussian final continuous boundaries.
noise, stripes, deadlines, and so on. All these characteristics 2) Kernel Growth Constraint by Edge: After the edge
cause a dilemma in window size selection that small window detection, the edge-constraint region growth algorithm is
cannot restore wide gap and suppress strong noise, but it will applied to create the clustering kernel Kp , which indicates a
affect the space and spectrum detail seriously if the window is small local homogeneous area with the center of pixel p. Kp
too large. is initialized as a square binary matrix with the radius of w, of
To solve this problem, an ASE is generated for each which the center element value is 1 and the remaining elements
pixel which indicates the morphological feature and structure are set to 0. Ep is the corresponding area to Kp in the edge
658 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 9, NO. 2, FEBRUARY 2016

when the target pixel belongs to a small enclosed scope, the

similar but not continuous regions cannot be covered into the
ASE. To deal with this problem, the growth algorithm is only
used to generate the clustering kernel in a relatively small area
according to the resolution of the image.
3) Goal-Guided K-Means Clustering: To obtain the final
ASE, a simple and efficient partitional clustering algorithm is
applied. K-means clustering algorithm is one of the most widely
used partitional clustering algorithms. In this algorithm, k clus-
tering centroids are assigned to the image for segmenting, then
the clustering result is concluded by iteration with the principle
of maximizing the between-class distance and minimizing the
in-class distance, according to some distance measure.
K-means clustering algorithm has been proved effective in
Fig. 1. EL and corresponding clustering kernels. (a) EL1. (b) EL2. (c) EL3. revealing the contextual information [28]. However, it still has
(d) Kernel 1. (e) Kernel 2. (f) Kernel 3.
some problems when applied to the whole image segmentation:
one of them is that it is difficult to determine the parameter
extraction result. We expect that the elements 1 in final Kp rep- k, especially when the scene is complicated and unfortunately
resent the area connected to the center p if p is nonedge pixel the value of k has a great influence on clustering results; on
or the continuous edge if p lies on the edge. In this paper, the the other hand, due to the random initialization, the cluster-
edge is thin as a result of NMS; thus, the four-neighborhood ing results of each independent experiment will be instable,
center growth algorithm is applied to the nonedge pixels to which increases the risk of the wrong classification. With the
avoid growth leakage and eight-neighborhood of edge pixels to analysis of the above problems, this paper proposes a par-
avoid edge breakage. Some examples shown in Fig. 1(a)–(c) are tial Goal-directed clustering method. Compared with traditional
the edge labels (EL) of target region and (d)–(f) are the corre- methods, the main distinctions lie in the following.
sponding results of clustering kernels. The centers of regions 1) The clustering algorithm is executed in the adjacent area
are marked with red points. The centers of (a) and (b) are of each pixel, but not the whole image. Our main pur-
nonedge, and the third one is exactly on the edge. The proposed pose is to obtain the homogeneous region of the target
edge-constraint kernel growth algorithm is detailed below. pixel, but ignoring the segmentation accuracy of other cat-
egories. Therefore, k can be simply set as a small fixed
Algorithm 1. Edge-constraint Kernel Growth Algorithm value (like 3 or 4) to increase the speed and stability of
clustering.
1. Kp = 0; Kp (w + 1, w + 1) = 1//initialize Kp
2) Considering the similar purpose above, the Goal-guided
2. if Ep (w + 1, w + 1) = 0 //central pixel is nonedge
improvement is introduced in the initialization phase.
3. for n = 1:w
Centroid of target class is set as the mean value of the
4. for euclidean distance between p(x, y) and
acquired clustering kernel and won’t change in the itera-
Kp (w + 1, w + 1) is n
tive procedure. By this method, the accuracy and stability
5. if Ep (x, y) is nonedge and at least one of
of target class can be further guaranteed.
4-neighborhood of Kp (x, y) equals to 1
3) Color information is better used, and the distance of
6. Kp (x, y) = 1
two pixels (L1 , u1 , v1 ) and (L2 , u2 , v2 ) in LUV space is
7. end
calculated as
8. end

9. end 2 2 2
D = αL (L1 − L2 ) + αu (u1 − u2 ) + αv (v1 − v2 )
10. else // central pixel is edge
11. for n=1:w (3)
12. for euclidean distance between p(x, y) and
where αL and αu , αv are weighting coefficients of luminance
Kp (w + 1, w + 1) is n
and chromaticity, and they can be set by similar steps of the
13. if Ep (x, y) is edge and at least one of
edge extraction in Section II-B1.
8-neighborhood of Kp (x, y) equals to 1
After clustering, category labels are added to all pixels in
14. Kp (x, y) = 1
adjacent area. Pixels that have the same label as the target clus-
15. end
tering kernel are marked as the influential spatial region, and
16. end
the corresponding values in the ASE are set as one. Regions
17. end
that have different label are considered to have a bare impact
18. end
on target pixel, and their values in the ASE are set at zero. With
this, the ASE is generated.
While the ASE can be obtained directly from the edge- Experiments and analysis show that the proposed method has
constraint region growing, with the enlargement of window higher stability compared with the previous approaches. The
size, the risk of growth leakage increases any way. In addition, obtained ASEs of every implementation are almost the same.
TENG et al.: AMF METHOD FOR STRUCTURAL FUSION RESTORATION OF HSIs 659

is used to detect edge features in a scene; then a line-tracing

process is employed to identify straight lines that run across the
entire scene vertically or horizontally; finally, actual stripe posi-
tions are located accordingly to trace the boundary lines based
on two thresholds. Simultaneously, ASEs are generated making
use of the referenced color image.
After unmissing area and ASEs are determined, the adaptive
morphological filtering (AMF) method is executed in the inter-
section of them. In consideration of the relevance between the
influence of surrounding pixels to the target pixel and their dis-
tance, a 2-D Gaussian nuclear is used as the assistant weight
window. Consequently, the estimated gray level of target pixel
fˆ(x, y) can be obtained by the function

Gw(x,y) (r, c) · X(r, c)
(r,c)∈W (x,y)∩Q
ˆ
f (x, y) =  (4)
Fig. 2. Schematic of ASE generation method. Gw(x,y) (r, c)
(r,c)∈W (x,y)∩Q

Although cluster computing needs to be performed for many where Gw(x,y) is the Gaussian window with the center of pixel
times, the computation and iteration number are far less in (x, y); W (x, y) ∩ Q is the intersection of ASE and unmissing
each performance. Besides, the capability of parallel processing region Q; X(r, c) is the original value of the usable pixels in
makes the execution speed totally acceptable. the intersection.
Fig. 2 illustrates the ASE obtaining procedure more vividly
with a real sample. It is a data patch of Sandiego data set to the
size 31 × 31, with feature classes of architecture, plants, soil, D. CPP Restoration Combined With Majority Voting
and road. The location of red star marks the target pixel. In the
Although the majority voting algorithm has been proposed
schematic, Canny operator is chosen as the gradient operator in
for many years, it is still one of the methods with most applica-
edge extraction, kernel growth region is limited to 5 × 5, and k
tion value in CPP so far. It definitely faces the similar limitation
is set to 4 in clustering algorithm. In Fig. 2, it can be seen that
as Gaussian window: when the size of filter window is small,
the obtained ASE accurately pinpoints the homogeneous region
it only work well on scatter noise, but may be negative for the
of the target pixel.
case of noise gathered in small scale or block misclassified due
to the stripes or deadlines; with the growing of the window
C. Preprocessing Restoration Combined With Gaussian size, the damage to structure information in class map will be
Filtering increasingly serious.
The proposed approach gives a win-win answer for the above
The next part of our restoration procedure is flexible and contradictions utilizing the ASE. It works well even on non-
replaceable. Many existing noniterative window-based filter preprocessed HSI or class map with dense misclassified pixels
can get a qualitative upgrade by making use of the proposed and structures. The flowchart of the postprocessing restoration
ASE. On account of its spatial structure protection property, method is depicted in Fig. 3(b). It is very similar to the prepro-
greater range can be selected to enhance the capacity on cessing restoration method, the main distinctions are stripes,
noise suppression and gaps or stripes inpainting. This paper detection is not necessary, and the window-based filtering in
takes 2-D Gaussian filter as an example to introduce the HSI ASE is replaced by the majority voting algorithm. For the pur-
preprocessing restoration and majority voting for CPP. pose of verifying the improvement produced by the proposed
Gaussian filter is one of the most widely used and effective ASE, support vector machine classifier is directly used to obtain
window-based filtering method. The weighting of each pixel the initial class map.
in the Gaussian window exponentially decays in its distance to
the center. Usually, it is used as a low-pass filter to suppress
high-frequency noise, and also on inpainting the detected miss- III. E XPERIMENTAL R ESULTS
ing regions. Through integration with the proposed ASE, the
A. Preprocessing Restoration Experiments
damage to the high-frequency information of the original image
will be limited within the homogeneous region, and the geo- To demonstrate the effectiveness of the proposed preprocess-
metric details of different object categories can be effectively ing algorithm, experiments are performed on two simulated
preserved. Despite the unavoidable spatial fuzzy on some level, data sets and a real data set. One of the simulated data sets
the subsequent interpretation algorithms generally have good is taken over Airport, San Diego, California by the Airborne
robustness on this kind of faintish linear mixture. Visible Infrared Imaging Spectrometer (AVIRIS). The scene
The flowchart of the proposed preprocessing restoration contains 400 × 400 pixels with the remaining 126 spectral
method is depicted in Fig. 3(a). First, stripes or deadlines are bands and a ground resolution of 3.5 m and spectral resolution
detected using methods proposed by [32]: Canny edge detector of 10 nm. The second data image is collected in the area around
660 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 9, NO. 2, FEBRUARY 2016

Fig. 3. Proposed restoration procedure. (a) Flowchart of the preprocessing restoration method. (b) Flowchart of the postprocessing restoration method.

the University of Pavia, Italy, by Reflective Optics System columns 159–166 and 250–254. To evaluate the performance
Imaging Spectrometer (ROSIS-3). The remaining 103 spectral and validity of our recovery algorithm, it is compared to other
channels range from 430 to 860 nm with a scene of 610 rows procedures. As a first benchmark, the Gaussian window-based
and 340 columns. An EO-1 Hyperion image data set is used weighted filter, which is widely used in mature application soft-
in our real data experiment (http://eros.usgs.gov/find-data). The ware, is chosen to demonstrate the strong structure protecting
original data set is of size 6702 × 256 × 242. A subset of size ability of our method. And the LR matrix recovery (LRMR)
256 × 256 × 76 is used here after the removal of water absorp- [21] is another benchmark as a new fashioned mixed noise sup-
tion bands. More than half of the remaining bands are polluted pression method. Of particular note is that the LRMR method
by a mixture of deadlines, stripes, and other noises. cannot effectively deal with the stripes in all bands for low-rank
To simultaneously simulate the two kinds of situations that properties of the noise itself; therefore, the second kind of noise
an additional visible image or several undamaged bands are is only added to several continuous bands, and the undamaged
available for reference, the RGB component extracted from the bands can be taken for an analogous supportive condition as the
HSI serves as the reference image in our experiments. Three reference image in proposed method.
bands that correspond to the wavelengths closest to the center of In this paper, the peak signal-to-noise ratio (PSNR) index and
the red, green, and blue wavelength range, respectively. For the the structural similarity (SSIM) index [33] are adopted to give
Sandiego data set, bands 25, 16, and 7 are selected as red, green, an objective evaluation of the restoration results. The definitions
and blue bands, and for the Pavia University data set, bands 53, of these indices are as follows:
31, and 8 are selected. For calculation and display convenience,
2552 MN
the original data sets are converted to an 8-bit scale. PSNRi = 10 ∗ log10 (5)
Mixed noises are added to the HSI: 1) zero means Gaussian 
M 
N
2
[ẑi (x, y) − zi (x, y)]
noise is added to all the bands and mean SNR value is about x=1 y=1
25 dB for both data sets; 2) deadlines and stripes are introduced (2μzi μẑi + C1 )(2σzi ẑi + C2 )
to all the bands. For the Sandiego data, five pixels width dead- SSIMi = (6)
(μzi 2 + μẑi 2 + C1 )(σzi 2 + σẑi 2 + C2 )
lines in columns 60–64 and three 4-pixel width stripes located
in columns 300–303, 310–313, and 320–323 are added. For where M and N represents the width and the height of the
the Pavia data, 8- and 5-pixel width deadlines are added to image of a specific spectral band, ẑi (x, y) is the restored value
TENG et al.: AMF METHOD FOR STRUCTURAL FUSION RESTORATION OF HSIs 661

Fig. 4. Results of the preprocess restoration experiments of Sandiego data set. (a) Reference color image. (b) Original band 70. (c) Noisy band. (d)–(f) LRMR
with 5, 7, and 9 bands striped. (g)–(i) Gaussian filter with window size 5 × 5, 7 × 7, and 11 × 11. (j)–(l) Proposed method with max ASE size 5 × 5, 7 × 7, and
11 × 11.

TABLE I
PSNR, SSIM VALUE , AND AVERAGE C OMPUTATIONAL T IME OF THE P REPROCESS R ESTORATION E XPERIMENTS ON S ANDIEGO DATA S ET

of a pixel (x, y) of the ith band, and zi (x, y) stands for its orig- best in suppressing mixed noises, stable, and effective, at the
inal value. μzi and μẑi denote the average values of image zi same time protecting the original local details. The common
and ẑi , while σzi and σẑi stand for the variances, and σzi ẑi is Gaussian filter can also restrain the noises; however, the local
the covariance between the images. Moreover, the mean SSIM spatial information, especially on the deadlines and stripes, is
of the striped area and deadlines is calculated to express the impaired increasingly seriously with the rising of the window
character of the impact of the geometric structure of the recover size. Comparatively speaking, the proposed method performs
methods. even better using bigger window, which avoids the dilemma
Fig. 4(a)–(c) shows the referenced, initial, and damaged in window size choosing. So, we can simply tradeoff the win-
image for Sandiego data set, respectively, and Fig. 4(d)–(f) dow size with the execution efficiency, but ignoring the damage
shows the restoration results of LRMR with five, seven, to the structural information of HSI. When the size reaches
and nine continuous bands striped, then recovered results of 11 × 11 pixels, the PSNR of the proposed method improves
Gaussian filter and proposed method with window size of 5 × more than 4 dB, and the SSIM improves about 0.11 in the whole
5, 7 × 7, and 11 × 11 are shown in (g)–(i) and (j)–(l). To further area and more than 0.17 in the striped areas. While the LRMR
demonstrate the improvement of structure preserving of our method can effectively remove the Gaussian noise and protect
method, the PSNR, SSIM values, and average computational the local details, it demonstrates a distinct limitation on repair-
time are presented in Table I. ing the deadlines and stripes. Fig. 4(d)–(f) shows that with more
By comparing the restoration results, it can be observed that continuous bands striped, it performs worse on their restoration,
the proposed preprocessing restoration method performs the even fails in inpainting the missing area when the number of
662 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 9, NO. 2, FEBRUARY 2016

Fig. 5. Results of the preprocess restoration experiments of Pavia University data set. (a) Reference color image. (b) Original band 70. (c) Noisy band.
(d)–(f) LRMR with 5, 7, and 9 bands striped. (g)–(i) Gaussian filter with window size 7 × 7, 11 × 11, and 13 × 13. (j)–(l) Proposed method with ASE size
7 × 7, 11 × 11, and 13 × 13.

TABLE II
PSNR, SSIM VALUE , AND AVERAGE C OMPUTATIONAL T IME OF THE P REPROCESS R ESTORATION E XPERIMENTS OF PAVIA U NIVERSITY DATA S ET

damaged bands comes to 9. Just as the previous analysis, the HSI restoration even lots of bands exist missing area in the same
assumption that the clear HSI is LR, but the noises are high- location, if only several undamaged bands or other sources of
rank, starts to crumble when the amount of damaged bands reference image are provided.
arrives at a certain extent, which is quite low. This can also For the Pavia data set, the corresponding results are shown
be deduced through the comparison of the objective parame- in Fig. 5 and Table II. Similar conclusions can be drawn
ters. The PSNR and SSIM values in deadlines obtained by our that the proposed ASE observably improves the capability
method are higher than those of the other methods, and when of traditional window-based filter in mixed noise suppressing
more than nine bands striped the SSIM of whole band obtained and local details protecting. Overall, it performs better than
by LRMR start to become worse than the proposed method. the iterative recover method on the robustness and parameter
Generally speaking, the method in this paper performs well on sensitivity.
TENG et al.: AMF METHOD FOR STRUCTURAL FUSION RESTORATION OF HSIs 663

Fig. 6. Results of the real preprocess restoration experiments of Hyperion data set. (a) Reference false color image. (b) Original noisy band 20. (c) LRMR.
(d) Gaussian filter with window size 7 × 7. (e) Proposed method with ASE size 7 × 7.

Fig. 7. Results of the postprocess restoration experiments of Pavia University data set. (a) Reference color image. (b) Ground truth. (c) Original class map on
ground reference. (d) Original class map. (e)–(h) Majority voting with window size 5 × 5, 7 × 7, 11 × 11, and 15 × 15, (i)–(l) Proposed method with ASE size
5 × 5, 7 × 7, 11 × 11, and 15 × 15.

For the real data experiments on the EO-1 Hyperion image noises and preserve the details as well. The restoration result
data sets, the restoration results of band 20 (the band 27 of orig- in Fig. 6(e) reveals that the traditional filtering method injures
inal HSI) are shown in Fig. 6 to demonstrate the effectiveness the high-frequency spatial information in spite of the inhibitory
of the proposed method, for there exist continuous deadlines effect on mixed noises. Although LRMR preserves more fine
of columns 1 and 141 from bands 1 to 28 and deadlines of structures in unmissing area and patches of the two-band last-
column 47 in bands 20 and 21. For the proposed method, the ing deadlines in column 47, it fails in repairing the deadlines in
undamaged bands 65, 31, and 106 are selected as red, green, columns 1 and 141 existing in 28 continuous bands, as shown
and blue components of the reference color image, as shown in Fig. 6(c).
in Fig. 6(a). For both filtering methods, the window size is set
as 7 × 7. For LRMR, the data set is divided into patches of
size 16 × 16 × 176 and transformed into a 2-D matrix of size B. Postprocessing Restoration Experiments
256 × 176. Fig. 6(b) shows the original noisy image of band In the postprocessing experiments, the first test of the algo-
20. As shown in Fig. 6(e), the proposed method performs the rithm still made use of the Pavia University HSI for simulative
best among the three methods. It can remove almost all the recovery. The training and test sets are collected in the area
664 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 9, NO. 2, FEBRUARY 2016

Fig. 8. Results of the postprocess restoration experiments of damaged Pavia University data set. (a) Original class map of the damaged HSIs. (b)–(e) Majority
voting with window size 5 × 5, 7 × 7, 11 × 11, and 15 × 15. (f)–(i) Proposed method with ASE size 5 × 5, 7 × 7, 11 × 11, and 15 × 15.

Fig. 9. Results of the postprocess restoration experiments of damaged Québec data set. (a) Reference color image. (b) Noisy band 80. (c) Ground truth. (d) Original
class map. (e)–(g) Majority voting with window size 21 × 21, 31 × 31, and 61 × 61. (h)–(j) Proposed method with ASE size 21 × 21, 31 × 31, and 61 × 61.
TENG et al.: AMF METHOD FOR STRUCTURAL FUSION RESTORATION OF HSIs 665

TABLE III TABLE IV

C LASSIFICATION R ESULTS ON PAVIA U NIVERSITY DATA S ET C LASSIFICATION R ESULTS ON DAMAGED PAVIA U NIVERSITY DATA S ET

TABLE V
around the University of Pavia, Italy, and consist of nine C LASSIFICATION R ESULTS ON DAMAGED Q UÉBEC DATA S ET
classes: Trees, asphalt, bitumen, gravel, metal sheets, shadow,
self-blocking bricks, meadows, and bare soil.
In addition, a real multisource data provided by the
2014 IEEE GRSS Data Fusion Contest is used, which is
available at website: http://cucciolo.dibe.unige.it/IPRS/IEEE_
GRSS_IADFTC_2014_Data_Fusion_Contest.htm. In this data
set, a region over Thet-ford Mines in Québec, Canada is col-
lected by Telops’ Hyper-Cam, an airborne long-wave infrared
(LWIR) HSI and a digital color camera. The LWIR HSI consists
of 84 spectral bands with a spatial resolution of approximately Visual inspection of the results demonstrates that, when the
1 m and size of 564 × 795; the fine-resolution visible (VIS) window size is restricted relatively small, the two methods lead
data have a spatial resolution of 0.2 m and 2830 × 3989 pix- to a similar enhancement to the SVM class map; the classifica-
els, 3 channels. For this data set, a benchmark training set is tion accuracy also reveals this. It is mainly due to that, in this
publicly available. Seven classes including road, trees, red roof, situation, both methods act well in wiping off the dispersive
green roof, gray roof, concrete roof, vegetation, and bare soil misclassification pixels, but leave much to be desired in deal-
are marked on a 2830 × 3989 pixels training map. ing with intensive or even regional misclassification, since not
Since the purpose is to demonstrate the performance and enough accurate classification pixels are included in the voting
validity of our postrestoration algorithm, a direct SVM-based region. The morphology modifying effect of the proposed ASE
classifier is used as the original classification and benchmark. is nonsignificant yet.
For both data sets, 15% of the available ground-truth pixels With the increasing of window size, the suppression of
are used for training. Spectral information without preprocess- speckle noise and regional error get stronger for both methods,
ing is used to train and test the SVM classifier for the Pavia especially in the deadline experiments. The main improve-
HSI. For the-Québec data set, a PCA transform is taken to the ment of this method is that the high-frequency geometric shape
LWIR HSI, and the up-sampling images of three main bands, information, which is very important for visual sensibility and
combined with the high-resolution color image, are utilized to machine recognition, is preserved intact even though a sizeable
generate the class map. window is used. However, when the quadrate window with the
To make a comparative analysis in conditions of different same size is used in the majority voting method, the constructed
window sizes, the results of the common majority voting post- particulars are damaged seriously and the small isolated objects
processing are taken as another benchmark. In the experiments, are devoured. The results of classification accuracy can also
the vote-filters are constrained by a series of square windows reflect the differences, although most of the referenced ground
and the propose ASE, and the results can be seen in Figs. 7–9. is in the homogeneous regions. Tables III and IV reveal that
Fig. 7(a) and (b) shows the Pavia color image and the class map the accuracy of majority voting declines with the increasing
of the ground truth, and the initial classification results of the of window size after 7 × 7. But for the proposed method, the
referenced ground and the whole area are shown in (c) and classification accuracy keeps rising and remains stable finally.
(d). It is observed that there exist a mass of misclassifica- When the size reaches 15 × 15, the proposed method shows an
tion pixels in the primitive SVM class maps, and some of improvement of 12% and 16% to the classification of the initial
them have intensive distribution. The corresponding classifi- and striped HSI, respectively, which is about 4% higher than
cation maps obtained using the two different postprocessing majority voting. Simultaneously, the accuracies of the initial
methods are exhibited, respectively, in (e)–(h) and (i)–(l). The and striped HSI processed by this method are approximately
classification accuracy in ground reference available area is same, which reveals that the deadlines in class map are removed
shown in Table III. Furthermore, to verify the recovery capa- almost perfectly.
bility of the proposed method under the condition of different On the Québec data set, Gaussian noise with both PSNR of
kinds of noise, the same deadlines of 5- and 8-pixel widths 20 dB and stripe with 1-pixel width 20-pixel distance are intro-
as Section III-A are added to the HSI, and the corresponding duced to the HSI. Due to the super resolution process, these
results of classification and postprocessing are shown in Fig. 8 noise morphs into 5 × 5 pixel blocks and 5-pixel width stripes
and Table IV. in the SVM class map. Therefore, the larger window is needed
666 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 9, NO. 2, FEBRUARY 2016

to improve the noise suppression, and the advantage of our [2] H. Zhang, H. Shen, and L. Zhang, “A super-resolution reconstruction
method in local details protecting is performed more obviously. algorithm for hyperspectral images,” Signal Process., vol. 92, no. 9,
pp. 2082–2096, Sep. 2012.
The results are shown in Fig. 9 and Table V. [3] Y. Wang and R. Niu, “Hyperspectral urban remote sensing image smooth-
ing and enhancement using forward-and-backward diffusion,” in Proc.
Joint Urban Remote Sens. Event, Shanghai, China, May 2009, pp. 1–5.
C. Discussion [4] Y. Wang, R. Niu, and X. Yu, “Anisotropic diffusion for hyperspectral
imagery enhancement,” IEEE Sensors J., vol. 10, no. 3, pp. 469–477, Mar.
In all of the above preprocessing and postprocessing exper- 2010.
iments, the cluster kernel is restricted to 3 × 3 and the [5] F. Cheng and A. N. Venetsanopoulos, “Adaptive morphological operators,
class number of local clustering is selected as k = 4. In fact, fast algorithms and their applications,” Pattern Recognit., vol. 33, no. 6,
pp. 917–933, Jun. 2000.
the class number k does not have a large effect for only the tar- [6] J. Chen, Y. Shao, H. Guo, W. Wang, and B. Zhu, “Shadow detection and
get class is taken into account, k can be set small to enhance the reconstruction in high-resolution satellite images via morphological fil-
rate and stability of the local clustering. The dilemma on win- tering and example-based learning,” IEEE Trans. Geosci. Remote Sens.,
vol. 53, no. 5, pp. 2545–2554, May 2014.
dow size selecting is settled owe to the details protecting the [7] S. Velasco-Forero and J. Angulo, “Spatial structures detection in hyper-
ability of the proposed ASE. Therefore, large enough window spectral images using mathematical morphology,” in Proc. 2nd Workshop
can be selected to fit different kinds of uncertain noise condi- Hyperspectral Image Signal Process. Evol. Remote Sens., 2010,
pp. 1–4.
tions, just considering the computing efficiency requirements [8] H. Othman and S. Qian, “Noise reduction of hyperspectral imagery
of the application. This is not a critical problem for our noniter- using hybrid spatial-spectral derivative-domain wavelet shrinkage,” IEEE
ative and parallelizable procedure. At last, it has to be noticed Trans. Geosci. Remote Sens., vol. 44, no. 2, pp. 397–408, Feb. 2006.
[9] G. Chen and S. Qian, “Denoising of hyperspectral imagery using prin-
that the availability of a reference color image is a severe limita- cipal component analysis and wavelet shrinkage,” IEEE Trans. Geosci.
tion to this method. Fortunately, for HSI, several bands in good Remote Sens., vol. 49, no. 3, pp. 973–980, Mar. 2011.
condition are as competent as multisource data set. [10] A. Karami, M. Yazdi, and A. Zolghadre-Asli, “Noise reduction of hyper-
spectral images using kernel non-negative Tucker decomposition,” IEEE
J. Sel. Topics Signal Process., vol. 5, no. 3, pp. 487–493, Jun. 2011.
IV. C ONCLUSION [11] S. Chen, X. Hu, and S. Peng, “Hyperspectral imagery denoising using a
spatial-spectral domain mixing prior,” J. Comput. Sci. Technol., vol. 27,
In this paper, we have proposed an AMF and fusion-based no. 4, pp. 851–861, Jul. 2012.
HSI restoration method. The key technique of this method is [12] Y. Qian and M. Ye, “Hyperspectral imagery restoration using nonlocal
spectral-spatial structured sparse representation with noise estimation,”
generating an ASE, which contains the local structure details, IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 6, no. 2,
making use of an auxiliary color image as the prior knowledge. pp. 499–515, Apr. 2013.
Edge is first extracted from the LUV format image; then, a clus- [13] Q. Yuan, L. Zhang, and H. Shen, “Hyperspectral image denoising
employing a spectral-spatial adaptive total variation model,” IEEE Trans.
ter kernel is built by edge-constraint region growing; finally, the Geosci. Remote Sens., vol. 50, no. 10, pp. 3660–3677, Oct. 2012.
ASE is determined via goal-guided local k-means clustering. [14] Q. Yuan, L. Zhang, and H. Shen, “Hyperspectral image denoising with a
The obtained ASE can be a flexible enhancement compo- spatial–spectral view fusion strategy,” IEEE Trans. Geosci. Remote Sens.,
vol. 52, no. 5, pp. 2314–2325, May 2014.
nent for most local filters. Therein, two classical combinations [15] A. Zelinski and V. Goyal, “Denoising hyperspectral imagery and recover-
are explained as instances, including the Gaussian filter for ing junk bands using wavelets and sparse approximation,” in Proc. IEEE
the original image and the majority voting method for class Int. Geosci. Remote Sens. Symp. (IGARSS), Aug. 2006, pp. 387–390.
[16] Y. Qian, Y. Shen, M. Ye, and Q. Wang, “3-D nonlocal means filter with
map correction. Experiments on both simulated and real data noise estimation for hyperspectral imagery denoising,” in Proc. IEEE Int.
sets demonstrate the advantage of the proposed approach on Geosci. Remote Sens. Symp. (IGARSS), 2012, pp. 1345–1348.
recovery of HSI. [17] A. Buades, B. Coll, and J. M. Morel, “A review of image denoising
algorithms, with a new one,” Multiscale Model. Simul., vol. 4, no. 2,
pp. 490–530, 2005.
ACKNOWLEDGMENT [18] J. Mairal, F. Bach, J. Ponce, G. Sapiro, and A. Zisserman, “Non-local
sparse models for image restoration,” in Proc. IEEE 12th Int. Conf.
The authors would like to thank the Associate Editor and the Comput. Vis., 2009, pp. 2272–2279.
anonymous reviewers for their careful reading and construc- [19] Y. Qian and M. Ye, “Hyperspectral imagery restoration using nonlocal
spectral–spatial structured sparse representation with noise estimation,”
tive comments that greatly improved the quality of this paper. IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 6, no. 2,
They would also like to thank Telops Inc. (Québec, Canada) for pp. 499–515, Apr. 2013.
acquiring and providing the data used in this study, the IEEE [20] M. Golbabaee and P. Vandergheynst, “Hyperspectral image compressed
sensing via low-rank and joint-sparse matrix recovery,” in Proc. IEEE Int.
GRSS Image Analysis and Data Fusion Technical Committee Conf. Acoust. Speech Signal Process. (ICASSP), 2012, pp. 2741–2744.
and Dr. M. Shimoni (Signal and Image Centre, Royal Military [21] H. Zhang, W. He, L. Zhang, H. Shen, and Q. Yuan, “Hyperspectral image
Academy, Belgium) for organizing the 2014 Data Fusion restoration using low-rank matrix recovery,” IEEE Trans. Geosci. Remote
Sens., vol. 52, no. 8, pp. 4729–4743, Aug. 2014.
Contest, the Centre de Recherche Public Gabriel Lippmann [22] R. Zhu, M. Dong, and J. Xue, “Spectral nonlocal restoration of hyper-
(CRPGL, Luxembourg) and Dr. M. Schlerf (CRPGL) for their spectral images with low-rank property,” IEEE J. Sel. Topics Appl. Earth
contribution of the Hyper-Cam LWIR sensor, and Dr. M. De Observ. Remote Sens., vol. 8, no. 6, pp. 3062–3067, Jun. 2015.
[23] X. Huang, Q. Lu, L. Zhang, and A. Plaza, “New postprocessing methods
Martino (University of Genoa, Italy) for her contribution to data for remote sensing image classification: A systematic study,” IEEE Trans.
preparation. Geosci. Remote Sens., vol. 52, no. 11, pp. 7140–7159, Nov. 2014.
[24] K. Schindler, “An overview and comparison of smooth labeling methods
for land-cover classification,” IEEE Trans. Geosci. Remote Sens., vol. 50,
R EFERENCES no. 11, pp. 4534–4545, Nov. 2012.
[25] Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Spectral-spatial clas-
[1] X. Liu, S. Bourennane, and C. Fossati, “Nonwhite noise reduction in sification of hyperspectral imagery based on partitional clustering tech-
hyperspectral images,” IEEE Geosci. Remote Sens. Lett., vol. 9, no. 3, niques,” IEEE Trans. Geosci. Remote Sens., vol. 47, no. 8, pp. 2973–2987,
pp. 368–372, May 2012. Aug. 2009.
TENG et al.: AMF METHOD FOR STRUCTURAL FUSION RESTORATION OF HSIs 667

[26] G. Moser, S. B. Serpico, and J. A. Benediktsson, “Land-cover mapping Ye Zhang received the B.S. degree in communica-
by Markov modeling of spatial-contextual information in very high reso- tion engineering, and the M.S. and Ph.D. degrees in
lution remote sensing images,” Proc. IEEE, vol. 101, no. 3, pp. 631–651, communication and electronic system from Harbin
Mar. 2013. Institute of Technology, Harbin, China, in 1982,
[27] S. Z. Li, Markov Random Field Modeling in Image Analysis. Berlin, 1985, and 1996, respectively.
Germany: Springer-Verlag, 2010. Currently, he is a Professor and Doctoral
[28] Z. Mahmood, M. A. Akhter, G. Thoonen, and P. Scheunders, “Contextual Supervisor with the Department of Information and
subpixel mapping of hyperspectral images making use of a high resolu- Communication Engineering, Harbin Institute of
tion color image,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., Technology. His research interests include remote
vol. 6, no. 2, pp. 779–791, Apr. 2013. sensing image analysis, image/video compression
[29] Y. Tarabalka, M. Fauvel, J. Chanussot, and J. A. Benediktsson, “SVM-and and transmission, and multisource information
MRF-based method for accurate classification of hyperspectral images,” collaboration processing.
IEEE Geosci. Remote Sens. Lett., vol. 7, no. 4, pp. 736–740, Oct. 2010.
[30] J. Li, J. M. Bioucas-Dias, and A. Plaza, “Spectral–spatial hyperspectral
image segmentation using subspace multinomial logistic regression and
Markov random fields,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 3,
pp. 809–823, Mar. 2012. Yushi Chen (M’11) received the Ph.D. degree in
[31] L. Sun, Z. Wu, J. Liu, L. Xiao, and Z. Wei, “Supervised spectral–spatial electronics and communication engineering from
hyperspectral image classification with weighted Markov random fields,” Harbin Institute of Technology, Harbin, China, in
IEEE Trans. Geosci. Remote Sens., vol. 53, no. 3, pp. 1490–1503, Mar. 2008.
2015. Currently, he is an Associate Professor with the
[32] F. Tsai and W. W. Chen, “Striping noise detection and correction of School of Electrical and Information Engineering,
remote sensing images,” IEEE Trans. Geosci. Remote Sens., vol. 46, Harbin Institute of Technology. He has authored more
no. 12, pp. 4122–4131, Dec. 2008. than 20 peer-reviewed papers. His research interests
[33] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality include hyperspectral data analysis, ensemble learn-
assessment: From error visibility to structural similarity,” IEEE Trans. ing, deep learning, and remote sensing applications.
Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004.

Yidan Teng (S’14) received the B.S. degree in com-

Chunli Ti received the B.S. and M.S. degrees in elec-
munication engineering from Harbin Engineering
tronics and communication engineering from Harbin
University, Harbin, China, in 2011, and the M.S.
degree in electronics and communication engineering Institute of Technology (HIT), Harbin, China, in
2011 and 2013, respectively. He is currently pursuing
from Harbin Institute of Technology, Harbin, China,
the Ph.D. degree in information and communication
in 2013. She is currently pursuing the Ph.D. degree in
engineering at the same university.
electronics and communication engineering at Harbin
Institute of Technology. His research interests include image processing
and video compression.
Her research interests include remote sensing
image processing, hyperspectral information restora-
tion, and application.