Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

POLSAR IMAGE CLASSIFICATION VIA COMPLEX-VALUED CONVOLUTIONAL

NEURAL NETWORK COMBINING MEASURED DATA AND ARTIFICIAL FEATURES

Xianxiang Qin1, Tao Hu1, Huanxin Zou2, Wangsheng Yu1, Peng Wang1

Information and Navigation College, Air Force Engineering University, Xi’an 710077, China
1
2
College of Electronic Science, National University of Defense Technology, Changsha 410073, China

ABSTRACT images [1, 2]. However, since these features are designed
artificially, they are often only available to specific images.
Recently, many deep convolutional neural networks (CNNs) Recently, deep learning techniques especially the
have been developed for the polarimetric synthetic aperture convolutional neural networks (CNNs) have become
radar (PolSAR) image classification. For them, it is often a powerful tools for image processing. Referring to the
hard task to obtain good results with limited training PolSAR image classification, some meaningful work with
samples. To release this problem, some strategies, such as CNNs has been performed. For examples, in [4], a four-
the feature-driven method that takes artificial features as the layer CNN for PolSAR image classification is designed and
input of CNN, have been proposed. However, since provides satisfactory results. Besides, to satisfy the complex
artificial features are usually difficult to be universal, their form of PolSAR data, a complex-valued CNN (CV-CNN)
applicability is limited to a certain extent. For this issue, this for PolSAR image classification is further developed and
paper proposes a scheme of combining measured data and adopted in [5], which can well utilize phase information to
artificial features with a complex-valued CNN (CV-CNN). improve the performance. Aiming to the limitation of
In our algorithm, not only the measured PolSAR data but labeled samples, in [6], a polarimetric-feature-driven CNN
also some discriminative artificial features are employed as is proposed by taking some discriminative features as the
the input of CV-CNN. The basic idea is that the measured input of CNN. Though fine classification results were
data contains the fully acquired information of targets, while obtained, its application is still somewhat limited since the
the artificial features include expert knowledge. Therefore, artificial features are usually difficult to be universal.
by fusing them, better and more stable performance may be Focusing on this problem, in this paper, we propose a
obtained. The experiments performed on both actual and scheme of combining both measured data and artificial
simulated PolSAR images have validated the effectiveness features by using the CV-CNN. The basic idea is that the
of the proposed algorithm. measured data contains original fully information of terrains,
while the artificial features include the expert knowledge.
Index Terms—Classification, polarimetric synthetic So, by fusing them, better and more stable performance may
aperture radar (PolSAR), complex-valued convolutional be obtained.
neural network (CV-CNN), artificial features The paper is organized as follows. Section 2 describes
the PolSAR data and some typical discriminative artificial
1. INTRODUCTION polarimetric features. The proposed classification scheme is
given in Section 3. Section 4 shows the experiment results
Polarimetric synthetic aperture radar (PolSAR) is an and we give our conclusion in Section 5.
advanced microwave imaging sensor that can acquire rich
information of the earth’s surface. This makes the PolSAR 2. MEASURED POLSAR DATA AND TYPICAL
image classification an interesting and important application ARTIFICIAL FEATURES
of PolSAR data [1-6].
In the past decades, plenty of algorithms for PolSAR 2.1. Measured PolSAR Data
image classification have been developed. In general, they For the general monostatic PolSAR, the measured data of
contain two main steps, i.e., feature extraction and classifier each pixel in a PolSAR image can be represented by a
designing [6]. Therefore, a large part of such algorithms are coherency matrix as [1]
devoted to extracting discriminative features of PolSAR T11 T12 T13 
T= k Pk P*T = T12* T22 T23  , (1)
___________________________________________________

This work is supported by National Science Foundation of China


T13 T23 T33 
* *
under Grant 41601436, 61403414 and 61703423, and in part by
the Natural Science Basic Research Plan in Shaanxi Province of
China under Grant 2018JM4029.  
where k P  1 / 2  S HH  S VV , SHH  S VV , 2S HV 
T
is the Pauli

978-1-5386-9154-0/19/$31.00 ©2019 IEEE 3209 IGARSS 2019


scattering vector, the superscript “*” and “H” are the consists of an input layer and an output layer, between
conjugate and transpose operations, respectively, and  which are several alternations of convolutional and pooling
denotes temporal or spatial ensemble averaging. It should be layers, fully connected layers. Different to the general real-
pointed out that the non-diagonal elements of the coherency valued CNNs, all the elements of the CV-CNN, including
matrix are usually complex that contain important phase the input and output of all layers, the weights and biases of
information of PolSAR data. the network, are fully defined in the complex domain.

2.2. Typical Artificial Polarimetric Features


In [6], some typical roll-invariant polarimetric features and
two hidden polarimetric features in the rotation domain are
employed and show good performance. Fig. 1. Architecture of the CV-CNN [5]

2.2.1. Roll-Invariant Polarimetric Features Since the CV-CNN is defined in the complex domain, it
In the H/A/  decomposition model, the coherency matrix is especially suitable to the complex PolSAR data. In [5], it
T is represented by the sum of the scattering matrices of shows superior performance to its real-valued counterpart
three uncorrelated targets, that is [7] for the PolSAR image classification.
T   i 1 i Ti   i 1 i ui u*T
3 3
i , (2)
3.2. Framework of the Proposed Algorithm
where i and ui are the ith eigenvalue and eigenvector of
To enhance the performance of CV-CNN for PolSAR image
T , respectively. The modulus of the first element of ui is classification, in this paper, a novel scheme by combining
represented by both the measured data and artificial features is proposed. It
ui 1  cos i , (3) has four steps:
(1) Extract artificial features and form combined data
where  i  [0 ,90 ] is called the ith alpha angle.
 
For a given PolSAR image, various artificial features can
Then, three roll-invariant polarimetric features including be extracted. In this paper, those used in [6] including four
the entropy H, mean alpha angle  and scattering roll-invariant features and two null angle described in
anisotropy A can be obtained as Section 2 are employed. Then a 6-dimension vector of
  3 artificial features for each pixel can be represented by
H   i 1 Pi log 3 Pi ,    i 1 Pi i , A  2
3 3
, (4)
2  3 D F =  H , A,  , SPAN ,  null _ Re T12  ,  null _ Im T12  ,
T
(8)
where Pi  i /  i 1 i .
3
Besides, the measured PolSAR data is described by the
six elements in the coherency matrix namely
Besides, another classical roll-invariant polarimetric
DT = T11 , T12 , T13 , T22 , T23 , T33  .
T
feature is the total backscattering power SPAN given by (9)
SPAN  T11  T22  T33 . (5) Then, a 12-dimension vector of combined data
combining the measured PolSAR data and artificial features
for each pixel is formed as DC   DT , D F  .
T
2.2.2. Hidden Polarimetric Features in Rotation Domain
The null angles of Re T12  and Im T12  are defined by [6, 8] (2) Determine a CV-CNN.
1 As mentioned before, the measured PolSAR data is
 null _ Re T12    Angle Re T13   j Re T12  ,
(6) represented by complex values, while the artificial features
2
are generally real-valued. Since real values are special cases
1
 null _ Im T12    Angle Im T13   j Im T12  , (7) of complex values, we choose the CV-CNN as the
2 framework to fuse the measured data and artificial features.
where Angle   is to get the phase. Re  and Im  are to (3) Train the CV-CNN.
obtain the real and imaginary parts, respectively. These two To train the CV-CNN, the training samples are randomly
null angles are sensitive to plenty of terrains [6] and then are selected, and the complex backpropagation algorithm using
useful for the PolSAR image classification. stochastic gradient descent (SGD) [5] are employed.
(4) Classify PolSAR image with the trained CV-CNN.
3. PROPOSED CLASSFICATION SCHEME For each pixel of the PolSAR image to be classified, a
neighboring window with the size fitting the input of the
given CV-CNN is firstly determined. Then, by feeding the
3.1. CV-CNN
combined data DC within the window into the trained CV-
The architecture of CV-CNN [5] is shown in Fig. 1. It CNN, the class label of the pixel can be obtained.

3210
4. EXPERIMENTAL RESULTS AND ANALYSIS The training procedure would stop if the total classification
error is smaller than a given threshold that is 0.1 here. Set
4.1. Experiments on Actual PolSAR Image the sample rates as 0.01, 0.05 and 0.1, the proposed scheme
and two contrast ones are performed for each given sample
In this section, a benchmark AIRSAR data set as shown in
rate. The representative classification results with the
Fig. 2(a) is employed. It is acquired in 1991 over Flevoland, sample rate as 0.05 are shown in Fig. 4.
The Netherlands, with the size as 1020×1024 pixels. The
ground-truth map is shown in Fig. 2(b), which indicates 14
categories in this field, including potato, fruit, oats, beet,
barley, onions, wheat, beans, peas, maize, flax, rapeseed,
grass and Lucerne.

(a) (b) (c)

(a) (b)
Fig. 2. (a) Actual PolSAR image and (b) ground-truth map. (d) (e) (f)
Fig. 4. Classification results of actual PolSAR image obtained
The employed architecture of CV-CNN is shown in Fig. by (a) comparison algorithm 1, (b) comparison algorithm 2 and (c)
the proposed algorithm.
3. Similar to that used in [5], it consists of two convolution
layers, one pooling layer, one fully connected layer, besides
It can be observed from Fig. 4 that all the classification
the input and output layers. The size of input of CV-CNN is
results obtained by the three algorithms are visually in good
12×12×d, where d=12 is the dimension of the input
consistent with the ground-truth. To evaluate the
combined data. In the network, there are 6 and 12 kernels of
performance of methods quantitatively, the overall accuracy
the same size as 3×3 for the first and second convolutional
(OA) of classification is used. The classification
layers, respectively. The average pooling [5] is adopted,
experiments are performed 10 times, and the mean OAs
with the size as 2×2. Moreover, the stride is 1 pixel in the
obtained by three algorithms are listed in Table 1.
convolutional layers and 2 pixels in the pooling layer.

 null _ Im T12  12×12×d


Table 1: OAs obtained by three algorithms
 Sample Comparison Comparison Proposed
...

...

A
H 10×10×6
T33 1×108 rate method 1 method 2 method
...

T
...

...

T1213 5×5×6
T11
3×3×12
1×Nc
0.01 0. 9506 0. 9600 0. 9604
...

0.05 0. 9573 0. 9693 0. 9694


...

...
...

...

Output
0.1 0. 9621 0. 9697 0. 9651
Pooling Conv2 layer
Conv1
2×2 12@3×3×6
6@3×3×d Fully connected
Stride=1 Stride=2 Stride=1
It can be seen from Table 1 that the proposed algorithm
Fig. 3. Architecture of the employed CV-CNN and the CV-CNN with the six artificial features obtains the
results somewhat better than that with the pure measured
To evaluate the performance of the proposed scheme, two data, even though the features are extracted from the
algorithms with the same network architecture but different measured data. It can be explained that the six employed
input are used for comparison. For the first one, only the features are discriminative to the terrains in this actual
measured PolSAR data DT given in (9) is taken as the input PolSAR image, while in the comparison method 1 the
of CV-CNN, while the pure artificial features DF given in information of measured data has not been fully extracted.
(8) are applied for the second one. The sizes of input of CV-
CNN for them are both 12×12×6. For the sake of simplicity, 5.2. Experiment on Simulated PolSAR Image
these two algorithms are called comparison algorithm 1 and Due to the limitation of actual PolSAR images, the flexible
comparison algorithm 2, respectively. simulated images are considered here to further evaluate the
For training the CV-CNN, the complex backpropagation performance of the proposed algorithm. Fig. 5(a) shows a
algorithm with SGD are employed, where the learning rate simulated PolSAR image of 140×490 pixels generated by
 is 0.8, and the batchsize is 100 with 2000 training epoch.

3211
the Monte Carlo method [1, 9]. It consists of 14 classes sufficient to discriminate all the classes well, yielding
following different Wishart distributions [1], of which the obvious confusion between classes 6 and 8. By comparison,
coherent matrices are determined by averaging the previous the algorithm with pure measured data gives better results,
actual coherent matrix data of each class according to the while the proposed algorithm provides best ones. This is
ground-truth. Besides, the equivalent number of looks (ENL) because the proposed scheme can make full use of the
is set as 10. Fig. 5(b) illustrates its ground-truth map. advantages of both measured data and artificial features.

6. CONCLUSIONS

This paper proposes a novel classification scheme of


(a) PolSAR images based on the CV-CNN. In our method, both
measured data and artificial features are taken as the input of
CV-CNN to make full use of their advantages. The
experiment results demonstrate that our method is superior
to that with pure measure data or pure artificial features.
(b)
Fig. 5. (a) Simulated PolSAR image and (b) ground-truth map. 7. REFERENCES
In the experiment, five training sample rates are [1] J.-S. Lee, and E. Pottier. Polarimetric radar imaging:
considered, including 0.005, 0.01, 0.05, 0.1 and 0.15. The from basics to applications: CRC press, 2009.
other parameters for training CV-CNN are the same as [2] C. He, S. Li, Z. Liao, and M. Liao, “Texture
before. The OAs of three algorithms are illustrated in Fig. 6. classification of PolSAR data based on sparse coding of
Moreover, the representative classification results with the wavelet polarization textons,” IEEE Trans. Geosci. Remote
sample rate as 0.05 are shown in Fig. 7. Sens., vol. 51, no. 8, pp. 4576-4590, 2013.
[3] B. Hou, H. Kou, and L. Jiao, “Classification of
polarimetric SAR images using multilayer autoencoders and
superpixels,” IEEE J. Sel.Topics Appl. Earth Observ.
Remote Sens., vol. 9, no. 7, pp. 3072-3081, 2016.
[4] Y. Zhou, H. Wang, F. Xu, and Y.-Q. Jin, “Polarimetric
SAR image classification using deep convolutional neural
networks,” IEEE Geosci. Remote Sens. Lett., vol. 13, no. 12,
pp. 1935-1939, 2016.
Fig. 6. OA plots of three algorithms.
[5] Z. Zhang, H. Wang, F. Xu, and Y.-Q. Jin, “Complex-
valued convolutional neural network and its application in
polarimetric SAR image classification.” IEEE Trans. Geosci.
Remote Sens., vol. 55. no.12, pp. 7177-7188, 2017.
[6] S.-W. Chen and C.-S. Tao, “PolSAR image
(a) classification using polarimetric-feature-driven deep
convolutional neural network,” IEEE Geosci. Remote Sens.
Lett., vol. 15, no. 4, pp. 627-631, 2018.
[7] S. R. Cloude and E. Pottier, “An entropy based
(b) classification scheme for land applications of polarimetric
SAR,” IEEE Trans. Geosci. Remote Sens., vol. 35, no. 1, pp.
68-78, 1997.
[8] S.-W. Chen, X.-S. Wang, and M. Sato, “Uniform
(c) polarimetric matrix rotation theory and its applications,”
Fig. 7. Classification results of simulated PolSAR image IEEE Trans. Geosci. Remote Sens., vol. 52, no. 8, pp. 4756-
obtained by (a) comparison algorithm 1, (b) comparison algorithm 4770, 2014.
2 and (c) the proposed algorithm.
[9] X. Qin, H. Zou, S. Zhou, et al. “Simulation of spatially
It can be observed from Fig. 6 and Fig. 7 that, for this correlated PolSAR images using inverse transform method,”
simulated PolSAR image, the artificial features are not J. Appl. Remote Sens., 9(1):095082, 2015.

3212

You might also like