IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO.
7, JULY 2016
Wishart Deep Stacking Network for
Fast POLSAR Image Classification
Licheng Jiao, Senior Member, IEEE, and Fang Liu
Abstract— Inspired by the popular deep learning architecture,
deep stacking network (DSN), a specific deep model for polari-
metric synthetic aperture radar (POLSAR) image classification is
proposed in this paper, which is named Wishart DSN (W-DSN).
First of all, a fast implementation of Wishart distance is achieved
by a special linear transformation, which speeds up the classi-
fication of POLSAR image and makes it possible to use this
polarimetric information in the following neural network (NN).
Then, a single-hidden-layer NN based on the fast Wishart
distance is defined for POLSAR image classification, which is
named Wishart network (WN) and improves the classification
accuracy. Finally, a multi-layer NN is formed by stacking WNs,
which is in fact the proposed deep learning architecture W-DSN
for POLSAR image classification and improves the classification
accuracy further. In addition, the structure of WN can be
expanded in a straightforward way by adding hidden units if nec-
essary, as well as the structure of the W-DSN. As a preliminary
exploration on formulating specific deep learning architecture
for POLSAR image classification, the proposed methods may
establish a simple but clever connection between POLSAR image
interpretation and deep learning. The experiment results tested
on real POLSAR image show that the fast implementation
of Wishart distance is very efficient (a POLSAR image with
768 000 pixels can be classified in 0.53 s), and both the single-
hidden-layer architecture WN and the deep learning architecture
W-DSN for POLSAR image classification perform well and work
efficiently.
Index Terms— Deep stacking network (DSN), POLSAR image
classification, Wishart network (WN), Wishart deep stacking
network (W-DSN).
I. I NTRODUCTION
W ITH A booming development of deep learning in recent
years, many deep learning architectures have been well
known in various fields, such as image processing [1]–[5],
speech recognition [6], [7], natural language processing [8]
and so on. These architectures include, but are not limited
to, Convolutional Neural Network (CNN) [2], [4], [9], [10],
Deep Belief Network (DBN) [3], [6], [8] and Deep Stacking
Network (DSN) [1], [7], [11], [12]. CNN is aimed at image
Manuscript received October 23, 2015; revised March 30, 2016; accepted
April 28, 2016. Date of publication May 11, 2016; date of current version
May 24, 2016. This work was supported in part by the National Basic
Research Program (973 Program) of China under Grant 2013CB329402 and
in part by the National Natural Science Foundation of China under Grant
61271302, Grant 61272282, Grant 61572383, and Grant 61573267. The
associate editor coordinating the review of this manuscript and approving
it for publication was Prof. David Clausi.
The authors are with the Key Laboratory of Intelligent Perception and Image
Understanding of Ministry of Education, International Research Center for
Intelligent Perception and Computation, School of Electronic Engineering,
Xidian University, Xi’an 710071, China (e-mail: lchjiao@mail.xidian.edu.cn;
fayliu77@163.com).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIP.2016.2567069
1057-7149 © 2016 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.
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
3273
processing, including natural image classification [2], [9],
object detection [4] and scene labeling [5], [10]. DBN is a
deep network as a probabilistic generative structure, which
performs good in many tasks, including image recognition [3],
speech recognition [6] and natural language processing [8].
With successful application in many tasks (such as speech
classification [7], [11], information retrieval [12], and image
classification [1], [7]), the deep learning architecture DSN
has received increasing attention as well. Despite many fields
have been studied, this paper focuses on issues in the field of
remote sensing, where deep learning is not fully developed.
To be exact, an exploration of polarimetric synthetic aperture
radar (POLSAR) image classification by DSN is made in this
paper.
With the fact that more than one polarization is used,
POLSAR acquires a much richer characterization of the
observed land and plays an important role in many areas,
such as military, agriculture and geology [13], [14]. As a
result, it is of great significance to interpret POLSAR image
effectively. POLSAR image classification is one of the most
fundamental issues in the process of interpretation, where
each pixel in POLSAR image is assigned to one class (such
as urban, water, and grass). In fact, the task of POLSAR
image classification corresponds to the task of scene label-
ing in nature image. Nevertheless, POLSAR data is always
represented by a coherency/covariance matrix, which contains
fully polarimetric information, rather than by a real-valued
scalar/vector (in gray/color image).
Methods for POLSAR image classification have been stud-
ied for decades. Taking the specificity of POLSAR data
into consideration, both scattering mechanism and statistical
property of POLSAR data are widely used in many classical
methods, such as Pauli decomposition [15], Entroy/Alpha
(H − α) decomposition [16] and Wishart distance [17], [18].
In the usage of scattering mechanism, a professional and
thorough analysis of POLSAR data is of great importance
to design a proper classifier [19]–[21], which is challenging
for scholar who does not specialize in POLSAR and still
intends to settle the task by machine learning. As for sta-
tistical property, the popular Whishart distance was proposed
based on the Wishart distribution of coherency matrix and
covariance matrix [17], [22], [23], which is a maximum
likelihood classifier actually [17] and has been used in both
unsupervised and supervised POLSAR image classification
[18], [24], [25]. However, the calculation of Wishart distance is
time-consuming and the accuracy of classifier Wishart distance
is low. Methods based on more complicated distribution, take
too much time in estimating parameters [26], [27]. Besides the
3274
property of POLSAR data, lots of machine learning methods
are also used in the task of POLSAR image classification,
including Neural Network (NN) [28], [29], Support Vector
Machine (SVM) [30], [31] and so on. Even though these
machine learning methods perform well, few of them are
specially designed for POLSAR data. They are only used as
classifiers, with features extracted from POLSAR data. What’s
more, the speed of POLSAR image classification is limited
by the process of training or testing, which is necessary in
machine learning.
Yet deep learning architectures are less popular in remote
sensing. Three available works about POLSAR image classi-
fication by deep models were revealed in [32], [33], and [34]
respectively, where all of them employed the deep
model DBN, which is staked by Restricted Boltzmann
Machines (RBMs). The first used the traditional RBM without
any modification, while the second revised it into Wishart
RBM (WRBM) according to the Wishart distribution of
POLSAR data, and the third stepped further by providing
a more mathematic stricter and stable version, i.e., Wishart-
Bernoulli RBM (WBRBM). Unfortunately, the generative
structure DBN with RBMs is not a task-oriented model,
where DBN is trained by unsupervised training to model input
data at first, and then it is modified to fit particular task
(e.g., classification) by supervised training. Hence the process
of training is time-consuming in [32]–[34]. By contrast, DSN
is task-oriented, which is stacked by modules and trained mod-
ule by module. Based on DSN, Wishart Deep Stacking Net-
work (W-DSN) is proposed in this paper, which is a specific
deep learning architecture for POLSAR image classification.
At the very beginning, two primary missions should be
finished to reach this point, including a linear implementation
of Wishart distance and a single-hidden-layer network named
Wishart Network (WN). The first provides a fast implemen-
tation of Wishart distance, where the equation of Wishart
distance is re-organized in a linear form. The re-organization
makes it possible to use this polarimetric information in
the proposed network. Both the computation analysis and
experiment results show that the proposed fast implementation
of Wishart distance works very well (reducing the cost from
336s to 0.53s for a POLSAR image with 768000 pixels
and of 15 classes). Based on Wishart distance, the second
(i.e., WN) is a simple but valuable usage of prior information
of POLSAR data in NN, where the column-vector form of
coherency matrix for a pixel is used as an input data and
the corresponding output of WN is an estimated label vector.
These two researches establish a subtle connection between
POLSAR image data and NN, and then W-DSN is achieved
by stacking WNs. Experiment results show that the proposed
W-DSN performs well in both accuracy and efficiency. More-
over, the structure of WN can be expanded in a straightforward
way by adding hidden units if necessary, as well as the
structure of the W-DSN.
The most important contribution of our work is the
proposed specific deep learning architecture W-DSN for
POLSAR image classification, which establishes a simple
but clever bridge between POLSAR image classification and
deep learning. In addition, two byproducts are also obtained,
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO. 7, JULY 2016
Fig. 1. DSN with three modules.
including a fast implementation of Wishart distance and a
single-hidden-layer network WN, where the first makes it
feasible to directly speed up the classification methods based
on Wishart distance and the second improves the classification
accuracy by supervised training. Experiment results show that
these methods work very well. Specially, a separate experiment
comparison with DSN is made to show the necessary of
W-DSN in the task of POLSAR image classification.
The remainder of this paper is organized as below.
Section II states the problem to be solved, which introduces
the traditional DSN and POLSAR image classification. The
connection between the prior knowledge of POLSAR data
and NN is described at large in the next section, including the
fast implementation of Wishart distance and the single-hidden-
layer neural network WN. In section IV, the specific deep
leaning architecture W-DSN for POLSAR image classification
is formed by stacking WNs, which is the ultimate goal in this
paper. Experiment results are demonstrated in section V to
verify the effectiveness of the proposed methods. At last, a
brief conclusion and the future work are discussed. Bold-faced
letters in this paper denote vector or matrix, where lower-
case letter in bold denotes vector and upper-case letter in bold
denotes matrix.
II. P ROBLEM S TATEMENT
DSN comprises a series of layered modules and the structure
of DSN with three modules is illustrated in Fig. 1. Each
module includes three layers, i.e., a linear input layer, a
nonlinear hidden layer and a linear output layer again [11],
referring to the sub-structure T − H − Y in Fig. 1. In the
lowest module, the input units contain only the raw training
data. While in each higher module, the input units contain raw
training data as well as the output units of lower modules.
In scene labeling, the input of DSN in this task should be a
pixel block, where the pixel block contains the pixel to be
classified and is used to represent that pixel. The output layer
of each module output the estimated label vector for the input
data. Thus, the estimated label vector of higher module can
be regarded as a rectification of the adjacent lower one, since
the output of the lower module is used as a portion of input
data for the higher one. What’s more, the DSN is trained in a
JIAO AND LIU: W-DSN FOR FAST POLSAR IMAGE CLASSIFICATION
module-by-module and supervised way, without propagation
over all modules.
As stated in the previous section, the classification of
POLSAR image is consistent with the scene labeling of
natural image, i.e., each pixel in POLSAR image corresponds
to one label. Fortunately, benefiting from the polarimetric
characteristics of POLSAR data, each pixel can be recognized
without considering its neighborhoods in the task of POLSAR
image classification. Then a pixel is represented by itself here,
rather than by a pixel block, and used as the input data.
Furthermore, each POLSAR pixel is denoted by a complex-
valued matrix, which is different from the real-valued pixel
in natural image. Specifically the 3 ∗ 3 complex coherency
matrix T is used to represent a POLSAR data for its Wishart
distribution, where the corresponding Wishart distance is fully
used.
Based on DSN, this paper reveals a specific deep model
named W-DSN, which is designed for POLSAR data and
classifies each POLSAR pixel T effectively and efficiently.
III. F ROM POLSAR DATA TO NN
To exploit a specific deep model for POLSAR image
classification, unique characteristics of POLSAR data should
be considered. In what follows, it is fully discussed that the
Wishart distance serves for a guide to transfer the information
of POLSAR data to NN.
A. Fast Implementation of Wishart Distance
Having a complex Wishart distribution [22], coherency
matrix and covariance matrix have been widely used in the
analysis of POLSAR data [17], [18]. J. S. Lee proposed a
maximum likelihood classifier based on Wishart distribution
using covariance matrix [17], which is also suitable for
there are two items in the Wishart
coherency matrix (coherency matrix and covariance matrix
T r ace  m and ln |Tm |. The first
can be transformed into each other by a linear operator).
Complex coherency matrix T is chosen in this paper, which
is conjugate symmetric, i.e.,
⎡ ⎤
T11 T12 T13
T = ⎣T12 T22 T23 ⎦
T13 T23 T33
where T11 , T22 and T33 are real-valued, the remaining elements
are complex-valued and • is a conjugate of an element.
1) Wishart Distance: With the maximum likelihood clas-
sifier [17], a multi-look POLSAR pixel T is classified
according to a so-called Wishart distance d (T |Tm  ), as
shown in eq. (1),
 
3275
classification as well [18]. In unsupervised classification [18],
J. S. Lee et al. initially clustered the POLSAR image by
polarimetric decomposition, and then estimated the cluster
mean Tm  by averaging all pixels from class m, i.e.,

Tm  = E [T |T ∈ m ] = |1m | T, (2)
T∈m
where m is the set of pixels from class m, and |m | is
the number of pixels in m . In supervised classification,
the
labeled training data set  = 1 2 · · ·  M is given,
where M is the number of classes. Then the cluster mean of
each class can be calculated directly by eq. (2) with the labeled
training data set.
2) A Linear Implementation of Wishart Distance: Although
Wishart distance has been widely used in a great many of
literatures for POLSAR image classification [18], [24], [35],
little attention is paid to the high-computation in calculating
the Wishart distance. Traditionally, a two-level loop (where
one level corresponds to the number of classes and the
other one corresponds to the number of pixels) is needed
to compute the Wishart distance of each pixel from each
cluster mean, since the equation of Wishart distance contains
matrix operations (as shown in eq. (1)). It is a very important
issue in POLSAR image classification, since that there are
always millions of pixels in a POLSAR image and the Wishart
distance of every pixel from each class should be calculated.
In 2006, W. Wang el al. explored a fast implementation of
H-Alpha Wishart classification [36]. But, it just reduces the
number of pixels in computing new cluster means to speed up,
without realizing that what matters most is the computation of
Wishart distance. A fast implementation of Wishart distance
is proposed in this paper, which exactly computes the Wishart
distance by a special linear transformation.
As shown in eq. (1),
T  −1 T
distance, i.e.,
item T r ace Tm −1 T is usually calculated by multiplying
Tm −1 by T to obtain the matrix Tm −1 T firstly and
then the trace of the obtained matrix is computed. It needs
27 multiplication operations and 20 addition operations, with
the obtained Tm −1 . Let  = Tm −1 T. Both Tm −1
and T are 3*3 complex matrices, so is . It is noticed
that the trace of a matrix is just a summation of diagonal
elements of the matrix. Namely, the computation of getting 
is redundant and only the diagonal elements are necessary. The
fast implementation of Wishart distance is based on leaving
out the redundant computation.
Let σ = f () be a function which arranges
all elements in the matrix  into column and out-
puts the column-vector form of  (e.g., f (T) =
d (T |Tm  ) = T r ace Tm −1 T + ln |Tm | (1)

where tr ace (•) is the trace of a matrix, •-1 is the inverse of a
matrix, |•| is the determinant of a matrix, Tm  for class m is
estimated by training samples from class m and it is regarded
as the cluster mean of the m-th class. With calculated distance
for each class, a pixel is assigned to the class with minimum
What’s
distance.
Even through Wishart distance was originally proposed for
supervised classification [17], it has been used in unsupervised
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.

T11 , T12 , T13 , T12 , T22 , T23 , T13 , T23 , T33 ).  = f −1 (σ ) is
the inverse function of σ = f (). In the following part of
this paper, the n-th pixel in POLSAR image is denoted by tn ,
n = 1, 2, ..., N, where N is the total number of POLSAR
pixels. Then let T = [t1 , t2 , ..., t N ], where tn = f (Tn )
is the column-vector form of matrix Tn  and Tn  is the
coherency matrix of the n-th pixel in POLSAR image. 
 
more, let W = [w1 , w2 , ..., w M ], where wm = f Tm −1 ,
m = 1, 2, ..., M. Thus, it is easy to notice the fact that
3276 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO. 7, JULY 2016

Fig. 2. Comparison of traditional implementation and the fast implementation. (a) Traditional implementation. (b) Fast implementation.

  
T r ace Tm −1 Ti = (wm ) ti , where both wm and ti are form of Wishart distance is the key point to establish the

9-dim complex column vectors and (wm ) is merely the following WN.
transposition of wm without conjugation. The computation Experiment results listed in section V verify this improve-

of (wm ) ti includes 9 multiplication operations and 8 addi- ment: a POLSAR image with 768000 pixels and of 15 classes
tion operations, can be classified in 0.53s, while it needs over 330s with

 which
 are almost just one-third of what
T r ace Tm −1 Ti needs in traditional way. traditional implementation by eq. (1).
Furthermore, let

b = [ln (|T1 |) , ln (|T2 |) , . . . , ln (|T M |)] B. Wishart Network
be a column vector. As a result, the Wishart distances of a Despite the fact that it is highly efficient to compute the
pixel t from different cluster means are calculated by the linear Wishart distance by the proposed method as discussed in

transformation W t + b, where t = f (T). Thus the Wishart section III-A, it does not improve the classification accuracy
distance matrix is calculated by eq. (3), at all. It is because the proposed eq. (3) is just a fast but

exact implementation of Wishart distance. As a matter of
D=WT+B (3) fact, the classification accuracy of Wishart classifier depends
where B = [b, b, ..., b] is formed by repeating b N times, on the cluster mean of each class, which is not changed in
⎡
  
  
  ⎤ the fast implementation of Wishart distance. In supervised
d T1 |T1  d T2 |T1  · · · d T N |T1  POLSAR image classification, the cluster mean of class m is
⎢
 1  
 2  
 N  ⎥
⎢ d T |T2  d T |T2  · · · d T |T2  ⎥ simply estimated by averaging pixels from class m, as shown
D=⎢ ⎢ .. .. ..
⎥
⎥ in eq. (2). Inspired by the method of learning in machine
⎣ . . . ⎦

 1  
 2  
 N   learning, an effort to explore higher accuracy is made by a
d T |T M  d T |T M  · · · d T |T M  learning method named WN, for the task of POLSAR image
is the matrix of Wishart distance, and D (m, n) denotes the classification.
Wishart distance of the n-th pixel Tn  from the m-th cluster 1) The Definition of Wishart Network (WN): The Wishart
mean Tm . In this way, the Wishart distance of each pixel distance matrix D shown in eq. (3) is a linear transformation
from each cluster mean is calculated by eq. (3) without of T, with W and B as weight parameter and bias parameter.
looping, but a relative efficiency of compiled array-handling The n-th pixel is assigned to the class with minimum distance
code, while it needs a two-level loop by eq. (1). The omission in the n-th column of D. In supervised POLSAR image
of loops make it convenient to execute on computer, then the classification by Wishart classifier, labeled training pixels tend
method shown in eq. (3) is a fast and linear implementation to be classified correctly with 
proper cluster

means.
 Let

labeled
of Wishart distance matrix. training pixel set be  = t1l , yl1 , t2l , yl2 , ..., tlK , ylK ,
A clear comparison is demonstrated in Fig. 2. The fast where tkl , k = 1, 2, ..., K denotes the k-th labeled pixel, ylk is
implementation leaves out the redundant computation of , the corresponding label vector and ylk is a unit column vector
and it is further organized in a way so that the Wishart distance where the index of the only non-zero element indicates the
matrix is calculated by a special linear transformation. As a label of tkl , and K is the number of labeled training pixels
result, the efficiency of calculating Wishart distance is highly in .
improved. In addition, the more classes and pixels, the higher Let W = [w1 , w2 , ..., w M ], where each cluster
the acceleration, since the two-level loop is omitted. The linear mean is associated with a column of W, as discussed

Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
JIAO AND LIU: W-DSN FOR FAST POLSAR IMAGE CLASSIFICATION 3277

It should be noticed that the parameters of WN include

both complex and real values. As discussed in the
previous subsection, pixel t is the complex input data
(i.e., a complex column vector t = f (T)), W is a
complex matrix containing cluster means as columns, b =

[ln (|T1 |) , ln (|T2 |) , ..., ln (|T M |)] is a real column vec-

tor, and W t + b is the real value which corresponds to Wishart
distance (i.e., the input of hidden layer is the Wishart distance
of a pixel from each cluster mean). U and c are real  matrix

and real bias vector respectively. Let   = W; b ,
t = [t; 1], W
 

Fig. 3. The structure of WN. 
h = [h; 1] and  U = U; c . The output of WN can be denoted

  
in section III-A2. Given a POLSAR pixel t, the proposed by y = g (t) =  U sigm W  t , then the optimization is
learning method classifies it by the following function, as

K K 
 2 K  
shown in eq. (4),  l   l  2
   mi n E = ek = yk − g (tk ) = yk − U hk 

y = g (t) = U sigm W t + b + c (4) k=1 k=1 k=1
(6)
where sigm (α) = 1+exp(−α)
1
is the sigmoid function, U and c    

are weight parameter and bias parameter respectively, and y is where  
hk = [hk ; 1] = sigm W tk ; 1 and 
tk = [tk ; 1].
the estimated label vector where the index of the maximal ele- 2) Supervised Training: With the training set , the initial-

ment associates with the predicted label. Moreover, W t + b is ization of W is associated with the cluster mean of each class
just the Wishart distance of pixel t from cluster mean of each (which iscalculated by eq. (2)), i.e., each column of W is


class, if there is no update for W. The motivation of eq. (4) wm = f Tm −1 as discussed in section III-A2. Vector
is to learn proper parameters to improve the performance b is initialized by the natural logarithm of each cluster mean

of classification, including W, U, b and c. In supervised [ln (|T1 |) , ln (|T2 |) , ..., ln (|T M |)] .
classification, the predicted label of a training pixel should be  b=
as element, i.e., 
Then W  = W; b . In fact, the initialization of W and b
as close
 as possible
2 to the true label. With the training set ,
ek = ylk − yk  denotes the k-th error between the estimated

agrees with the discussion in section III-A2, which is much
label vector and the true label vector, where yk = g tkl . better than random initialization. This initialization is very
As a result, the proposed method should be trained to minimize important in the training process, because it makes the super-
the total error, as shown in eq. (5). vised training proceeds based on the existing Wishart distance
classifier rather than a fully random initialization, which

K K 
 2
 l  reduces uncertainty to a great extent. Given the initialized
mi n E = ek = yk − yk  (5)
W and b, the optimization in eq. (6) is convex and it is
k=1 k=1 K  
 −1 
With the structure shown in Fig. 3, the WN is defined as easy to obtain that  U = hk
hk   is
hk ylk . Then W
k=1
below, to realize the proposed supervised method shown in  =W  − λ ∂ E , where λ is the step.
updated by the equation W ∂W
eq. (4) and (5) by neural network. The input data of the
The updating equations here are obtained just by the simple
network is POLSAR pixel t, the input and output of hidden
 derivation operator.
layer are W t + b and h = sigm (Wt + b) respectively, and
An explanation is provided here in detail that Tm  is
the output of the output layer is the
 corresponding
 estimated
  associated with the first 9 elements in the m-th 
column of

label vector y, where y = U sigm W t + b +c. The number  i.e., wm = W  (1 : 9, m) = W (:, m) = f Tm −1 .
W,
of input units is 9, which is equal to the dimension of
The m-th element of b corresponds to the 10-th element in
input data t (i.e., the number of elements in t); the number  i.e., bm = W  (10, m). However, the
the m-th column of W,
of hidden units is equal to the number of cluster means
bias parameter b is updated in terms 
of the updated W, i.e.,
(i.e., the number of columns in W); the number of output
−1
 −1
units is equal to the number of classes (i.e., the dimension bm = ln (|Tm |), where Tm  = f (wm ) and the
of label vector y). The weight matrix connecting input layer  (10, m) is not used. It is because b is totally determined
W
and hidden layer is W, the bias parameter in hidden layer by the coherency matrix in Wishart distance, which should be
is b and the activation function of hidden layer is the sigmoid preserved anyhow. The way of updating b is a key point in
function. U is the weight matrix connecting hidden layer and the process of WN training, which has a close relationship

output layer and c is the bias parameter in output layer. to the specific design of WN. Let Tl = t1l , t2l , ..., tlK and

  WN defines a function y = g (t),
Generally, the proposed
i.e., g (t) = U sigm W t + b + c. The network WN should
be trained to learn proper parameters so that it can classify
labeled training pixels correctly. Therefore, the training of WN
is aimed at the same optimization with eq. (5).
Yl = yl1 , yl2 , ..., ylK . The training process of WN is summa-
rized as Algorithm 1, where ’Require’ is meant to indicate a
list of inputs and ’Ensure’ a list of outputs, as listed below.
Although the structure of WN illustrated in Fig. 3 is
similar to the general single-hidden-layer network, it should

Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
3278
Algorithm 1 Training of WN
be emphasized that difference between WN and general one
is mainly that the input data is the complex-valued vector t,
the initialization of W is explicit and the update of bias b is
unique (i.e., determined by the updated W, as discussed in the
previous paragraph).
With learned parameters W  and  U, a POLSAR pixel t is
classified by the estimated label vector, i.e., y = g (t) =
 
 
U sigm W t , where t = [t; 1]. The index of maximal value
in the estimated label vector y corresponds to the predicted
label. Experiments shown in section V demonstrate that the
proposed WN does improve the classification accuracy.
3) Expanded WN: The proposed WN mentioned above
contains 9 input units (corresponding to the dimension of input
data), M hidden units (corresponding to the number of cluster
means) and M output units (corresponding to the number of
classes). From the point of view of function learning, the
training of WN listed in Algorithm 1 is aimed to learn a
function (i.e., y = g (t)) so that it can classify training pixels
correctly. WN with more hidden units has a higher ability
to represent complicated function, because more hidden units
bring about more parameters. Hence some hidden units should
be added if the WN with M hidden units does not meet the
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO. 7, JULY 2016
error condition. That is to say, if the total error E of WN
cannot be reduced to a certain threshold, then add some hidden
units in WN. Weights associated with the added hidden units
are initialized by rearranging the picked pixels with f (•),
where the picked pixels are randomly selected from different
classes. WN with M hidden units is called basic WN and WN
with more hidden units is called expanded WN. Experiment
results in section V show that the addition of hidden units can
improve the classification accuracy effectively.
IV. A D EEP M ODEL OF WN-W ISHART D EEP
S TACKING N ETWORK (W-DSN)
As a single-hidden-layer network, WN achieves higher
accuracy than the basic Wishart distance. In the following
discussion, a deep model of WN is formed to improve the
performance further, which is based on the popular DSN and
named as Wishart Deep Stacking Network (W-DSN). It is
precisely the specific deep learning architecture for POLSAR
image classification proposed in this paper.
A. Wishart Deep Stacking Network (W-DSN)
It is obvious that the proposed WN is directly a single
module of DSN, where the main differences are that the raw
input data is complex and the weights connecting input layer
and hidden layer are complex and initialized by cluster means.
The structure of W-DSN with three modules is same to that of
DSN, as demonstrated in Fig. 1. W-DSN is different from the
traditional DSN in the fact that the modules stacked in W-DSN
are WNs and the parameters in WNs are unique. Specially,
W-DSN with only one module is exactly WN.
Suppose that the parameter  in the
 i -th WN module
 is
   
   
Wi , Ui , where Wi = Wi ; bi and Ui = Ui ; ci .
The inputs of three modules are T, [T; Y1 ] and [T; Y1 ; Y2 ]
respectively and the corresponding outputs are Y1 , Y2 and Y3
respectively,
 as listed in eq. (7)∼(9). The lowest module with
parameter W  1, 
U1 is exactly the WN, so it is trained as
described in Algorithm 1. However when the second module
adjacent to the lowest one is trained, W 2 is initialized by com-
bining the trained W  1 with several random real values, where
these real values correspond to the weights connecting Y1
and hidden units, because Y1 is a portion of input data for
the second module. The initialization of W  3 is similar to that

of W2 . Thus each module is trained as Algorithm 1 one by
one, and the whole training process of W-DSN is summarized
as Algorithm 2. The number of modules in W-DSN is needed
by Algorithm 2 besides what is needed by Algorithm 1.
  

Y1 =  U1 sigm W  1T (7)

  
Y2 =  U2 sigm W  2 [T; Y1 ] (8)

 

Y3 =  U3 sigm W  3 [T; Y1 ; Y2 ] (9)
With the fact that the estimated label vector of lower module
is used as a portion of input data for the adjacent higher
one, the estimated label vector of the higher module can
be regarded as a rectification of that of the lower one. In
training process, when the output of lower module (e.g., Y1 ) is
JIAO AND LIU: W-DSN FOR FAST POLSAR IMAGE CLASSIFICATION 3279

Algorithm 2 Training of W-DSN where tk,1 = [tk ; 1], as shown in eq. (6). When the second
lowest module (which is adjacent to the lowest one) is trained,
the optimization is
K 
   2
 l   2 
mi n E 2 = yk − U2 sigm W tk,2  , (11)
k=1


where  tk,2 =  t ; y . The initialization of parame-
combined with raw input data to form a new input in adjacent
higher module (e.g., [T; Y1 ]), the output (e.g., Y1 ) is exactly
the linear combination of hidden units in the lower module,
without estimating the predicted label. That is, only the output
of the last module (e.g., Y3 ) is used to estimate the final
predicted label, in the process of testing.
A detail should be noticed that the module used here can be
the basic WN with M hidden units or the expanded WN with
more hidden units. The more hidden units there are, the higher
ability to express complicated function the WN has, so is the
W-DSN. If the number of classes is big or the training pixels
from different classes are difficult to distinguish from each
other, which lead to the fact that the classification function
should be complicated enough to distinguish these pixels, more
hidden units will be needed.
Some notes are laid out here for a clearer understanding of
our work. W-DSN is different from DSN in that the modules
stacked in W-DSN are WNs, which are defined specialized
for POLSAR data in section III. Thus the initialization and
training of W-DSN are modified accordingly, despite the
similar procedure.
B. Analysis of W-DSN
Although the explanation that estimated label vector of the
higher module can be regarded as a rectification of that of
the lower one is intuitional, a rigorous mathematic analysis is
provided. For simplicity, only the first two lower modules are
considered, since higher modules work in a similar way.
When the lowest module is trained, the optimization is
K 
   2
 l   1  the i -th row and the j -th column, r e (•) and i m (•) denote the
mi n E 1 = yk − U1 sigm W tk,1  ,
k=1
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
k,1 k,1
ter is W 2 = W  1 ; P , where P is a M × M ran-
dom real-valued matrix associated with yk,1 . Thus E 2 =
K 
   2
 l    
yk − U2 sigm W 1 t k,1 + P y k,1  . If P a null matrix,
k=1
these two optimizations shown in eq. (10) and (11) are the
same, i.e., mi n E 1 is equal to mi n E 2 . Yet, P is a variable
in the second optimization, hence the latter optimization is
a generalization of the former one. Consequently, E 2 may
have a smaller minimum than E 1 , and parameters learned by
minimizing E 2 provide a better estimated label vector than that
learned by minimizing E 1 . Namely, the second lowest module
provides a better estimated label vector than the lowest one.
This will continue for higher modules. So W-DSN with more
modules can improve the classification accuracy further.
It has been discussed that both W-DSN (with more than
one WN) and expanded WN have a higher ability to express
complicated classification function than the basic WN. There-
fore, W-DSN with expanded WN will exhibit better classi-
fication performance than both of them. Experiment results
demonstrated in the following section confirm this conclusion
completely.
V. E XPERIMENTS
The POLSAR image which is tested on to verify the
effectiveness of the proposed methods is related to the site of
Flevoland, the Netherland. The size of this POLSAR image is
750*1024 and it has been used in many papers [37], [38].
Fig. 4 illustrates the corresponding Pauli RGB image
and groundtruth respectively. There are 15 classes in this
groundtruth, where each class indicates a type of land covering
and is identified by one color. 167712 pixels are labeled in the
groundtruth and only 5% of them are used as training pixels.
The reported testing accuracies are obtained by testing on
the 95% residual pixels. What’s more, the proposed methods
are compared with six other methods, including Support
Vector Machine (SVM) [30], Radial Basis Function (RBF)
network, Wishart classifier, DBN with RBM [32], DBN with
WRBM [33] and DBN with WBRBM [34]. The importance
of the proposed W-DSN is emphasized by comparing W-DSN
with DSN separately, which is demonstrated in section V-D.
As discussed in previous sections, the input data of WN
and W-DSN is the column-vector form of coherency matrix,
i.e., t = f (T). Wishart classifier classifies pixel using the
original coherency matrix T. While in the last five comparing
methods, POLSAR pixel is denoted by a 9-dim real-valued
vector, i.e., [T11 , T22 , T33 , r e(T12), i m(T12 ), r e(T13 ), i m(T13),
r e(T23), i m(T23 )], where Ti, j denotes the element locating in
(10)
real part and the image part of a complex number respectively.
3280 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO. 7, JULY 2016

Fig. 4. (a) The Pauli RGB of Flevoland. (b) Groundtruth.

TABLE I
C OMPARISON OF T RADITIONAL AND FAST I MPLEMENTATION

This real-valued vector contains all information of T,
so does t.
Experiment results below are reported to confirm three main
contributions of this paper, including the fast implementation
of Wishart distance, the single-hidden-layer network WN and
the deep learning architecture W-DSN for POLSAR image
classification. Finally, a comprehensive evaluation is provided.
On a 3.20 GHz machine with 4.00 GB RAM, all the experi-
ments are conducted using MATLAB.
A. The Comparison of Traditional and Fast
Implementation of Wishart Distance
As discussed in section III, the proposed fast implementa-
tion of Wishart distance is achieved by a special linear trans-
form, without redundant computation and the two-level loop.
In supervised classification, Wishart classifiers with traditional
implementation and with fast implementation respectively are
both tested. A simple preparation (i.e., the construction of
W and b, which corresponds to the first loop in Fig. 2 (b)
and the ’preparation’ part of Algorithm 1) is needed in the
fast implementation and the corresponding time-consumption
is called ’preparation time’. The time for classifying the whole
image of 768000 pixels is called ’classification time’.
Running times and accuracies are listed in TABLE I. The
proposed fast implementation takes 0.0003s to construct W
and b, which is unnecessary in the traditional implementation.
The traditional implementation takes 336.4s to classify the
whole image, while the fast implementation takes just 0.5288s,
where the latter time-consumption is less than 1/600 of the
former one. This is because the redundant computation and
two-level loop embedded in the traditional implementation
are left out of the proposed fast implementation. Even though
an extra step is needed in the fast implementation, it works
very efficiently, noting that the preparation time is negligible
(0.0003s) and the speedup is large (over 600 times). The
accuracy of the implementation of the Wishart distance is
unchanged, since the fast implementation gives exactly the
same results as the traditional method.
B. Effectiveness of WN
The proposed learning method WN, which is initialized by
cluster mean of each class and uses the column-vector form of
original coherency matrix as input data, improves the accuracy
by supervised training. The result of WN is compared with
that of SVM, RBF, Wishart classifier, and DBNs with RBM,
WRBM and WBRBM respectively. For fairness, there are only
one hidden layer in each DBN and linear classifier is used,
which are the same to the basic WN. In the following, these
three DBN methods are called RBM, WRBM and WBRBM
respectively for conciseness.
First, all methods are repeatedly tested 50 times without
changing training samples and the total accuracies are demon-
strated in boxplot in Fig. 5, to explore their robustness to
randomness. It is clear that SVM, RBF, Wishart classifier
and the proposed WN are stable, which are not affected by
random parameters. However, RBM, WBRBM and WBRBM

Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
JIAO AND LIU: W-DSN FOR FAST POLSAR IMAGE CLASSIFICATION 3281

TABLE II
T HE A CCURACIES

Fig. 5. Total accuracy of each method by repeatedly running. Fig. 6. Total accuracy of each method by repeatedly sampling.

are sensitive to some extent, where RBM is the most sensitive
one, WRBM performs more stable than RBM due to its
utility of some polarimetric information and WBRBM is the
most stable one in these three methods because of its stricter
mathematic support.
Second, sample 5% labeled pixels repeatedly 50 times as
training samples for each method, and the total accuracies are
revealed in boxplot as well, as shown in Fig. 6. Similarly,
SVM, RBF, Wishar classifier and the proposed WN perform
rather stable with different training samples. However, the
other three comparing methods are unstable, because they are
not task-oriented. Note that, WRBM is less stable than RBM
here, however WBRBM are still more stable than both of them
(the total accuracies of RBM, WRBM and WBRBM are not as
high as listed in [32]–[34], since more hidden layers with more
units were employed there and nonlinear classifiers were used).
Experiment results demonstrated in Fig. 5 and Fig. 6 show
that the proposed WN is robust to both random parameters
and different training samples, with relative high accuracy.
Third, individual class accuracies are listed in TABLE II
for better insight, since many times they are more attractive to
users. In addition, the expanded WNs with 2M, 3M and 4M
hidden units respectively are tested as well, to illustrate that
the addition of hidden units gives rise to the improvement of
accuracy indeed.
To analyze individual class accuracies, it should be declared
that percentages of different classes in the whole training
sample set are 7.89%, 10.77%, 6.08%, 4.20%, 5.72%, 4.52%,
3.04%, 5.98%, 6.65%, 13.27%, 3.77%, 8.27%, 9.78%, 0.43%
and 9.63% respectively, corresponding to the class order listed
in TABLE II. Even though SVM, RBF and Wishart classifier
reach similar total accuracies, SVM and RBF misclassify

Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
3282 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO. 7, JULY 2016

TABLE III
A CCURACY OF W-DSN W ITH D IFFERENT N UMBERS OF M ODULES

‘Buildings’ (with least training samples, i.e., 0.43%) terribly,

which is well recognized by Wishart classifier. It is because
the optimizations of SVM and RBF are sensitive to small
samples, where more resources are distributed to model larger
samples for better optimal solution. Similarly, by unsupervised
training, RBM, WRBM and WBRBM are used to model all
samples at first, and then supervised training is conducted to
complete the classification task, so they are sensitive to small
samples for the same reason. In fact, they cannot recognize
‘Bare soil’ besides ‘Buildings’, where training samples of
‘Bare soil’ (3.04%) are just more than ‘Buildings’ (0.43%) but
less than other classes. However, Wishart classifier depends on
the center of each class, which is less sensitive to the number
Fig. 7. The variation of total error with more hidden units.
of samples from each class, so it performs stable in classifying
small samples, such as ‘Buildings’ and ‘Bare soil’. Another
opposite evidence is that ‘Water’ is not classified very well
WN and W-DSN with two or three modules are formed as
by Wishart classifier, even though the percentage of ‘Water’ is
illustrated in Fig. 1. The initialization and training process
high up to 7.89%. Moreover, WN is based on Wishart distance
of W-DSN are conducted module by module as discussed in
as well, so it is less sensitive to small samples, similar to
section IV, which is summarized as Algorithm 2.
Wishart classifier. Even so, compared with Wishart classifier,
W-DSNs with different numbers of modules are compared
WN is slightly effected by the number of training samples
with each other and the resulting accuracies are demonstrated
in the process of supervised training, which results in an
in TABLE III. Obviously, W-DSN with more modules or with
improvement in ‘Water’ accuracy and a decrease in ‘Bare soil’
WN of more hidden units achieves higher accuracy, including
accuracy. Next, with increasing total accuracy and comparable
most individual class accuracies and the total accuracy. More
individual class accuracies, the expanded WNs with 2M, 3M
specifically, with M hidden units in each module, W-DSNs
and 4M hidden units respectively perform better than the basic
of 1 module, 2 modules and 3 modules respectively result in
WN, confirming the conclusion in section III-B3 directly.
total accuracies 0.8650, 0.8855 and 0.9138. With 4M hidden
In addition, the final total errors of WNs with different
units in each module, W-DSNs of 1 module, 2 modules
numbers of hidden units are demonstrated in Fig. 7, where
and 3 modules respectively result in total accuracies 0.9018,
vertical axis and horizontal axis denote the total error and the
0.9163 and 0.9268. It is concluded that W-DSN with more
number of iterations for training respectively. It is obvious
modules performs better than W-DSN with less modules.
that the total error deceases with the increase of iterations.
Meanwhile, of the same number of modules (e.g., 2 modules),
What’s important is that the total error of WN with more
W-DSN of module with more hidden units results in higher
hidden units is smaller than that of WN with less hidden units,
total accuracy (e.g., W-DSNs of 2 modules with 4M and
for the same iterations. The results show that additional hidden
M hidden units in each module respectively provide the
units enhance WN’s ability of estimating label vector indeed,
accuracies 0.9163 and 0.8855). The results confirm that the
resulting in higher accuracy as shown in TABLE II.
proposed deep learning architecture W-DSN is an efficient new
type of model for the task of POLSAR image classification.
C. W-DSN for POLSAR Image Classification Similar to Fig. 7, the variation of total error of W-DSN
The deep learning architecture named W-DSN, which is with different numbers of modules is demonstrated in Fig. 8.
formed by stacking WN, could improve the accuracy further. The vertical axis and horizontal axis indicate the total error
Concretely, W-DSN with only one module is exactly the and the number of modules respectively. It is clear that the

Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
JIAO AND LIU: W-DSN FOR FAST POLSAR IMAGE CLASSIFICATION
TABLE IV
A CCURACY AND RUNNING T IME
Fig. 8. The variation of total error with the increase of the number of
modules.
total error decreases with more modules, resulting in better
classification results as listed in TABLE III. With the same
number of modules, W-DSN with modules of more hidden
units has a lower total error, leading to higher accuracy. This
observation is consistent with the fact that WN with more
hidden units has a higher ability of improving the classification
accuracy, as shown in TABLE II in detail.
D. The Comparison of W-DSN and DSN
To emphasize the importance of the proposed W-DSN in
the task of POLSAR image classification, W-DSN is com-
pared with the traditional DSN separately. As discussed in
sections III and IV, W-DSN takes the polarization information
of POLSAR data into consideration, which does not happen
in DSN. The DSN used here is stacked by 3 modules (with M
hidden units in each module) like W-DSN (as shown in Fig. 1),
while these 3 modules stacked in DSN are just general one-
hidden-layer networks. That is to say, all units and weights
in DSN are real-valued and DSN classifies POLSAR pixels
by treating them as general data. Thus the input data of DSN
is a 9-dim real vector, which is same to SVM, RBF, RBM,
WBRBM and WBRBM. Total accuracies of W-DSN and DSN
are shown in Fig. 9 to make the comparison clearly.
Similar to the discussed W-DSN, DSN achieves higher
accuracy with more modules or more hidden units in each
module, as demonstrated in Fig. 9. However DSN performs
much worse than W-DSN, where accuracies obtained by DSN
is much lower than that obtained by W-DSN. In fact, WN
(i.e., W-DSN with 1 module) begins with Wishart classifier,
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
3283
Fig. 9. Accuracies of W-DSN and DSN.
since WN is a learning method based on Wishart distance.
Therefore W-DSN with 1 module would achieve an accuracy
as high as Wishart classifier (0.8504) at the very beginning.
This polarimetric information is a very useful guide for
classification task. While DSN with 1 module starts from a
totally random initialization, which results in a rather low
accuracy (1/M ≈ 0.0667). As a result, W-DSN with 1 module
would perform better than the DSN with 1 module even after
supervised training. The performance of W-DSN with more
modules is better than that of DSN with more modules for
the same reason. Furthermore, the most complicated DSN
(of 3 modules and 4M hidden units in each module) works
rather worse than the simplest W-DSN (of 1 module and
M hidden units in each module), where these two methods
achieve accuracies less than 0.70 and more than 0.85 respec-
tively. The results provide a powerful experimental evidence
that the proposed W-DSN is real necessary for POLSAR image
classification than DSN.
E. Comprehensive Evaluation
At last, the performance of each method is evaluated in
both accuracy and time-consumption, as shown in TABLE IV.
It should be noticed that only WNs with M hidden units and
4M hidden units respectively are listed. The W-DSN shown
here are of 3 modules with 4M hidden units in each module.
Moreover, only the total accuracy is listed to make it clear
to exhibit. However both the time for training and the time
for classifying the whole POLSAR image are listed, which
are called ‘training time’ and ‘classifying time’ respectively.
The total time is obtained by summing these two parts.
3284
Fig. 10. (a)∼(i) show the classification results of SVM, RBF, Wishart classifier, RBM, WRBM, WBRBM, WN (with M hidden units), WN (with 4M hidden
units) and W-DSN (of 3 modules with 4M hidden units in each module).
The ‘preparation time’ in Wishart classifier is renamed as
‘training time’, to keep consistent with other methods. For
RBM, WRBM and WBRBM, both the time for unsupervised
training and the time for supervised training are contained in
‘training time’. Besides, DSN is not included here for its rather
poor performance as revealed in section V-D.
According to the results listed in TABLE IV, the fast
implementation of Wishart distance works very efficient and
reduces the time-consumption greatly, compared with the
traditional implementation. WN with M hidden layer is the
original form of the learning method and achieves a higher
accuracy than Wishart classifier, at the cost of 2.2851s for
training. With the same number of hidden units, WN with M
hidden units takes much less time than RBM, WRBM and
WBRBM (3s is much less than 61.34s, 69.56s and 58.77s),
while WN still achieves higher accuracy. With more hidden
units in WN, the accuracy becomes higher (i.e., the accuracy
of WN with 4M hidden units 0.9018 is higher than that of
WN with M hidden units 0.8650), where a little more time
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO. 7, JULY 2016
is needed (17s). W-DSN with more modules results in higher
accuracy as well (i.e., the accuracy of W-DSN of 3 modules
0.9258 is higher than than of W-DSN of 1 module (i.e., WN)
0.9018). Even though both WN with more hidden units and
W-DSN with more modules take much more time for training,
the total time is much less than that of the comparing methods.
Specifically, W-DSN with 3 modules takes 50s to complete
the task of classification, while Wishart classifier (traditional
implementation), SVM and RBF take 336.40s, 295.33s and
988.89s respectively. In short, the proposed WN and W-DSN
achieve higher accuracy with less time.
Meanwhile, the corresponding classification results are also
illustrated in pictures, as shown in Fig. 10. It is clear that
Fig. 10(g) shows better results than Fig. 10(a)∼(f), con-
firming the conclusion that WN does achieve better results,
compared with SVM, RBF, Wishart classifier, RBM, WRBM
and WBRBM. By comparing Fig. 10(g), (h) and (i) with
each other, it is clearly illustrated that WN with more hid-
den units or W-DSN with more modules can improve the
JIAO AND LIU: W-DSN FOR FAST POLSAR IMAGE CLASSIFICATION
classification results, which agrees with the numerical values
listed in TABLE III.
VI. C ONCLUSION AND F UTURE W ORK
A deep learning architecture named W-DSN is constructed
specialized for POLSAR image classification in this paper.
Two other methods about POLSAR image classification are
also proposed for the construction of W-DSN, including a fast
implementation of Wishart distance and a network named WN.
The fast implementation of Wishart distance makes it possible
to directly speed up the methods based on Wishart distance,
and WN is a single-hidden-layer network which improves the
classification accuracy at a high speed. Most importantly, the
W-DSN achieves higher accuracy further, which is proved to
be an effective and efficient deep architecture specialized for
POLSAR image classification.
The spatial information, which is also very important in
POLSAR image classification, will be considered in our future
work, to complete the classification more rapidly and precisely.
R EFERENCES
[1] J. Li, H. Chang, and J. Yang. (2015). “Sparse deep stacking
network for image classification.” [Online]. Available: https://arxiv.
org/abs/1501.00777
[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification
with deep convolutional neural networks,” in Proc. Adv. Neural Inf.
Process. Syst., 2012, pp. 1097–1105.
[3] O. Russakovsky et al., “ImageNet large scale visual recognition chal-
lenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211–252, 2014.
[4] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature
hierarchies for accurate object detection and semantic segmentation,”
in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2014,
pp. 580–587.
[5] P. H. O. Pinheiro and R. Collobert. (2013). “Recurrent convo-
lutional neural networks for scene parsing.” [Online]. Available:
http://arxiv.org/abs/1306.2795
[6] G. E. Dahl, D. Yu, L. Deng, and A. Acero, “Context-dependent pre-
trained deep neural networks for large-vocabulary speech recognition,”
IEEE Trans. Audio, Speech, Lang. Process., vol. 20, no. 1, pp. 30–42,
Jan. 2012.
[7] L. Deng, D. Yu, and J. Platt, “Scalable stacking and learning for building
deep architectures,” in Proc. IEEE Int. Conf. Acoust., Speech Signal
Process. (ICASSP), Mar. 2012, pp. 2133–2136.
[8] R. Collobert, J. Weston, L. Bottou, M. Karlen, K. Kavukcuoglu, and
P. Kuksa, “Natural language processing (almost) from scratch,” J. Mach.
Learn. Res., vol. 12, pp. 2493–2537, Nov. 2011.
[9] K. Simonyan and A. Zisserman. (2014). “Very deep convolutional
networks for large-scale image recognition.” [Online]. Available:
http://arxiv.org/abs/1409.1556
[10] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and
A. L. Yuille. (2014). “Semantic image segmentation with deep
convolutional nets and fully connected CRFs.” [Online]. Available:
https://arxiv.org/abs/1412.7062
[11] L. Deng and D. Yu, “Deep convex net: A scalable architecture for speech
pattern classification,” in Proc. Interspeech, 2011, pp. 2285–2288.
[12] L. Deng, X. He, and J. Gao, “Deep stacking networks for infor-
mation retrieval,” in Proc. IEEE Int. Conf. Acoust., Speech Signal
Process. (ICASSP), May 2013, pp. 3153–3157.
[13] E. Attema et al., “ENVISAT ASAR science and applications,” ESA Publ.
SP, vol. 1225, no. 1, p. 59, Nov. 1998.
[14] F. T. Ulaby and C. Elachi, Eds., Radar Polarimetry for Geoscience
Applications, vol. 1. Norwood, MA, USA: Artech House, 1990, p. 376.
[15] E. Pottier, “Dr. J. R. Huynen’s main contributions in the development
of polarimetric radar techniques and how the ‘radar targets phenomeno-
logical concept’ becomes a theory,” Proc. SPIE, vol. 1748, pp. 72–85,
Feb. 1993.
[16] S. R. Cloude and E. Pottier, “An entropy based classification scheme
for land applications of polarimetric SAR,” IEEE Trans. Geosci. Remote
Sens., vol. 35, no. 1, pp. 68–78, Jan. 1997.
Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.
3285
[17] J. S. Lee, M. R. Grunes, and R. Kwok, “Classification of multi-look
polarimetric SAR imagery based on complex Wishart distribution,” Int.
J. Remote Sens., vol. 15, no. 11, pp. 2299–2311, 1994.
[18] J.-S. Lee, M. R. Grunes, T. L. Ainsworth, L.-J. Du, D. L. Schuler, and
S. R. Cloude, “Unsupervised classification using polarimetric decompo-
sition and the complex Wishart classifier,” IEEE Trans. Geosci. Remote
Sens., vol. 37, no. 5, pp. 2249–2258, Sep. 1999.
[19] F. Shang and A. Hirose, “Use of Poincare sphere parameters for fast
supervised PolSAR land classification,” in Proc. IEEE Int. Geosci.
Remote Sens. Symp. (IGARSS), Jul. 2013, pp. 3175–3178.
[20] G. Singh, Y. Yamaguchi, and S.-E. Park, “General four-component scat-
tering power decomposition with unitary transformation of coherency
matrix,” IEEE Trans. Geosci. Remote Sens., vol. 51, no. 5,
pp. 3014–3022, May 2013.
[21] J.-S. Lee, M. R. Grunes, E. Pottier, and L. Ferro-Famil, “Unsupervised
terrain classification preserving polarimetric scattering characteristics,”
IEEE Trans. Geosci. Remote Sens., vol. 42, no. 4, pp. 722–731,
Apr. 2004.
[22] N. R. Goodman, “Statistical analysis based on a certain multivariate
complex Gaussian distribution (an introduction),” Ann. Math. Statist.,
vol. 34, no. 1, pp. 152–177, 1963.
[23] F. Cao, W. Hong, Y. Wu, and E. Pottier, “An unsupervised segmentation
with an adaptive number of clusters using the SPAN/H/α/A space and the
complex Wishart clustering for fully polarimetric SAR data analysis,”
IEEE Trans. Geosci. Remote Sens., vol. 45, no. 11, pp. 3454–3467,
Nov. 2007.
[24] L. J. Du and J. S. Lee, “Polarimetric SAR image classification based
on target decomposition theorem and complex Wishart distribution,” in
Proc. IEEE Remote Sens. Sustain. Future, Int, Geosci. Remote Sens.
Symp. (IGARSS), vol. 1. May 1996, pp. 439–441.
[25] G. Zhou, Y. Cui, Y. Chen, J. Yin, J. Yang, and Y. Su, “Pol-SAR
images classification using texture features and the complex Wishart
distribution,” in Proc. IEEE Radar Conf., May 2010, pp. 491–494.
[26] M. Liu, H. Zhang, C. Wang, and F. Wu, “Change detection of multilook
polarimetric SAR images using heterogeneous clutter models,” IEEE
Trans. Geosci. Remote Sens., vol. 52, no. 12, pp. 7483–7494, Dec. 2014.
[27] J. S. Lee, D. L. Schuler, R. H. Lang, and K. J. Ranson,
“K-distribution for multi-look processed polarimetric SAR imagery,” in
Proc. Surf. Atmospheric Remote Sens., Technol., Data Anal. Interpre-
tation, Int. Geosci. Remote Sens. Symp. (IGARSS), vol. 4. Aug. 1994,
pp. 2179–2181.
[28] Y. C. Tzeng and K. S. Chen, “A fuzzy neural network to SAR image
classification,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 1,
pp. 301–307, Jan. 1998.
[29] C. Chen, K.-S. Chen, and J.-S. Lee, “The use of fully polarimetric
information for the fuzzy neural classification of SAR images,” IEEE
Trans. Geosci. Remote Sens., vol. 41, no. 9, pp. 2089–2100, Sep. 2003.
[30] C. Lardeux et al., “Support vector machine for multifrequency SAR
polarimetric data classification,” IEEE Trans. Geosci. Remote Sens.,
vol. 47, no. 12, pp. 4143–4152, Dec. 2009.
[31] S. Fukuda and H. Hirosawa, “Support vector machine classification of
land cover: Application to polarimetric SAR data,” in Proc. IEEE Int.
Geosci. Remote Sens. Symp. (IGARSS), vol. 1. Jul. 2001, pp. 187–189.
[32] Q. Lv, Y. Dou, X. Niu, J. Xu, J. Xu, and F. Xia, “Urban land use and
land cover classification using remotely sensed SAR data through deep
belief networks,” J. Sensors, vol. 2015, Jul. 2015, Art. no. 538063.
[33] Y. Guo, S. Wang, C. Gao, D. Shi, D. Zhang, and B. Hou, “Wishart
RBM based DBN for polarimetric synthetic radar data classification,”
in Proc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Jul. 2015,
pp. 1841–1844.
[34] F. Liu, L. Jiao, B. Hou, and S. Yang, “POL-SAR image classification
based on Wishart DBN and local spatial information,” IEEE Trans.
Geosci. Remote Sens., vol. 54, no. 6, pp. 3292–3308, Jun. 2016.
[35] C. Wang, W. Yu, R. Wang, Y. Deng, F. Zhao, and Y. Lu, “Unsuper-
vised classification based on non-negative eigenvalue decomposition
and Wishart classifier,” IET Radar, Sonar Navigat., vol. 8, no. 8,
pp. 957–964, Oct. 2014.
[36] W. Wang, J. Wang, S. Mao, and P. Wu, “Fast implementation of
H/α-Wishart classification to polarimetric SAR images images,” in Proc.
CIE Int. Conf. Radar, Oct. 2006, pp. 1–4.
[37] T. L. Ainsworth, J. P. Kelly, and J.-S. Lee, “Classification comparisons
between dual-pol, compact polarimetric and quad-pol SAR imagery,”
ISPRS J. Photogram. Remote Sens., vol. 64, no. 5, pp. 464–471, 2009.
[38] P. Yu, A. K. Qin, and D. A. Clausi, “Unsupervised polarimetric
SAR image segmentation and classification using region growing with
edge penalty,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 4,
pp. 1302–1317, Apr. 2012.
3286 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 25, NO. 7, JULY 2016

Licheng Jiao (SM’89) received the B.S. degree Fang Liu was born in China, in 1990. She received
from Shanghai Jiaotong University, China, in 1982, the B.S. degree in information and computing
and the M.S. and Ph.D. degrees from Xi’an science from Henan University, Kaifeng, China,
Jiaotong University, Xi’an, China, in 1984 and 1990, in 2012. Since then, she has been taking suc-
respectively. Since 1992, he has been a Professor cessive post-graduate and doctoral programs with
with the School of Electronic Engineering, Xidian the Key Laboratory of Intelligent Perception and
University, Xi’an, where he is currently the Direc- Image Understanding of Ministry of Education,
tor of the Key Laboratory of Intelligent Perception Xidian University, Xi’an, China. Her research inter-
and Image Understanding of the Ministry of Edu- ests include deep learning, polarimetric sar image
cation of China, International Research Center of classification, and change detection in polarimetric
Intelligent Perception and Computation. His current sar images.
research interests include intelligent information processing, image processing,
machine learning, and pattern recognition. He is a member of the IEEE Xi’an
Section Execution Committee, the President of Computational Intelligence
Chapter, the IEEE Xi’an Section and IET Xi’an Network, the Chairman of
the awards and Recognition Committee, the Vice Board Chair-Person of the
Chinese Association of Artificial Intelligence, a Councilor of the Chinese
Institute of Electronics, a Committee Member of the Chinese Committee of
Neural Networks, and an Expert of the Academic Degrees Committee of the
State Council.

Authorized licensed use limited to: Chang Gung Univ.. Downloaded on December 28,2020 at 10:18:18 UTC from IEEE Xplore. Restrictions apply.