UNSUPERVISED POLSAR IMAGE CLASSIFICATION USING BOUNDARY-PRESERVING

REGION DIVISION AND REGION-BASED AFFINITY PROPAGATION CLUSTERING

Biao Hou, Yuheng Jiang, Bo Ren, Zaidao Wen, Shuang Wang, and Licheng Jiao

Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China,
Xidian University, China (Email: [email protected])

ABSTRACT classification using wishart MRF, in which the image is

divided into large amounts of rectangular regions, and then a
This paper presents a new method for polarimetric synthetic spatial context model based on MRF is used to classify the
aperture radar (PolSAR) image classification. Firstly, to get regions with training data. Y. H. Wang et al. [5] presented a
a reasonable edge strength map, polarimetric information is segmentation method for fully PolSAR data by coupling the
used in edge strength calculation, and watershed algorithm is cluster analysis in tensor space and MRF framework. In
used to obtain the oversegmentation using the edge strength. addition to studying the contextual information, using
Secondly, a searching table is used to determine the most polarimetric information or feature vectors as the input,
suitable region to be merged. Finally, region-based affinity neural network (NN) [6] and support vector machine (SVM)
propagation clustering is employed to achieve an initial [7, 8] are also used to classify PolSAR data. Other methods
classification map, and the method provides an adjacent for integrating region information can be found in the
Wishart classifier with spatial relations to obtain the final literatures. B. Liu et al. [9] proposed a superpixel-based
classification result. method which uses superpixels instead of individual pixels
and normalized cuts algorithm is emplyed to generate
Index Terms— Polarimetric SAR, Image classification, superpixels which can be regarded as a local region, and
Boundary preservation, Affinity propagation clustering Wishart clustering proposed by J. S. Lee et al. [10] is used
to obtain the final classification result.
1. INTRODUCTION Unsupervised classification may be applied when training
data are not available. Clustering is an important
Polarimetric synthetic aperture radar (PolSAR) data has unsupervised classification technique where data points are
been widely used in many applications as it can provide grouped such that points that are similar in some sense
comprehensive information about the earth’s surface. belong to the same cluster. P. Yu et al. [11] used an
Classification is an important way of image interpretation extension to iterative region growing with semantics (IRGS)
and using polarimetric data to classify land cover [1] and algorithm for PolSAR image classification. But the initial
crop [2] is widely studied. Since there is a large amount of classification does not use the spatial information of image.
polarimetric data from radar platforms launched in recent V. Akbari et al. [12] proposed an unsupervised, non-
decades, automatic method for image interpretation is more Gaussian, and contextual segmentation method that
demanded than manual method. combines an advanced statistical distribution with spatial
Most of PolSAR image classification methods are based contextual information for multilook PolSAR data. The
on pixels [3]. In these methods, each pixel is an individual proposed contextual clustering method uses a specific
element and pixels are assigned with the different labels. Markovian energy function for integrating the K-Wishart
The salt-and-pepper effect due to pixel-based processing is distribution. A. P. Doulgeris et al. [13] presented an
inevitably shown in results. In recent years, region-based automatic image segmentation method for PolSAR data. In
methods are increasingly researched, they regard a region as which the unsupervised clustering algorithm is based upon
an element instead of a pixel, and the region statistics are the expectation maximization (EM) algorithm for finite
used as the pixel value. mixture modeling using the K-Wishart probability density
The region-based classification methods can be divided function for the probabilities and method of moments for
into two categories: supervised and unsupervised. The class parameter updates. F. Cao et al. [14] combined the
supervised methods usually need a priori knowledge. SPAN/H/α/A initialization, the agglomerative hierarchical
Markov random field (MRF) is a common way to include clustering, and the data log-likelihood estimation to perform
spatial relations between pixels, which can model the an unsupervised segmentation for fully PolSAR data.
labeling process and improve the classification performance. In this paper, we propose an unsupervised region-based
Y. H. Wu et al. [4] proposed a region-based image classification method for PolSAR image, which uses an edge

978-1-5090-3332-4/16/$31.00 ©2016 IEEE 5103 IGARSS 2016

measurement considering the scattering characteristics to a class label xv .
segment the image into a number of regions that have
accurate boundary information. Then we utilize an iterative
region growing that preserves the boundary to get more 2.2. Boundary-preserving region division based on
searching table
reasonable number of regions, and a region-based affinity
propagation clustering [15] is used to get an initial
classification. Finally, an adjacent Wishart classifier is used To get a reasonable division result, the proposed method
to obtain the final classification result. use [11, Eq. (15)] and a searching table  of size KK to
merge the regions, K is the number of regions. As the
2. THE PROPOSED METHOD boundary-preserving region division is a region merging
procedure and it is iterative. The merging criterion E (v, w)
2.1. Oversegmentation based on edge strength map [11, Eq. (14)] between the regions and their adjacent regions
is calculated at each iteration. If E (v, w) is the smallest, the
The proposed method uses the oversegmentation of the adjacent region w of v is regarded as the most suitable
image as the first step. A watershed algorithm is used to get region to be merged, and then we set   v, w  1 . If
the oversegmentation using the edge strength of each pixel.
In order to get the accurate segmentation result,   w, v   1 , we merge the two regions w and v as a new
polarimetric information is used in edge strength calculation. region which is labeled as v . At the end of iteration, we
The gradient vector of each pixel s is based on the Pauli relabel all the new regions according to the number of
decomposition of PolSAR Data: regions. The region merging is iterative and stops until the
 S HH  SVV 2 2 S HH  SVV 
2
(1) iteration number achieves the maximum value. The output of
gs  [ R G B]  
Pauli
2 S HV 
 2 2  boundary-preserving region division is an unsupervised
The elements S HH , SVV and S HV are the scattering segmentation of image because there is no use of a priori
information during merging. It has far lower number of
coefficients produced in HH , VV and HV channels. The regions than the oversegmentation and generates a
edge strength map is generated by computing the gradient reasonable result by only merging the regions which meet
 s of each pixel s  S  (i, j ) |1  i  M ,1  j  N  as the conditions mutually. After region merging, a new RAG
follows.  *  (V * ,  * ) is constructed, it has a much less number of
1) Assuming that every element of g sPauli is a parameter vertices than the oversegmentation, each vertex v of  *
*

for Sobel operator, we can get three gradient parameters g sR , denotes the new region generated by region growing
procedure.
g sG and g sB .
2) Computing the gradient of every pixel s as the edge 2.3. Region-based affinity propagation clustering
strength:
s  g sR  g sG  g sB (2) In this paper, an unsupervised region-based affinity
3) Obtaining edge strength   s | s  S of the whole propagation clustering [15] is proposed to get initial
image. classification. The input of affinity propagation clustering is
The proposed method uses a watershed algorithm to a similarity matrix D, and D (i, j) indicates the similarity
produce the oversegmentation of the input image. The image between data point i and j, if j is the cluster center of i. The
is segmented into an amount of irregular regions, and all similarity is usually calculated by Euclidean distance. In
regions are separated by boundaries, the boundary pixels do order to make full use of polarimetric characteristic, the
not belong to any regions. proposed method uses the Wishart distance [10] as the
The oversegmentation can be used as the first step of the similarity between two regions:
region-based classification algorithm for PolSAR data [11],
each region of oversegmentation consists of some pixels.
1
 
Dreg  v, w  ln Z v  ln Z w  tr Z v1Z w  Z w1Z v
2
(3)

The image can be represented as a region adjacency graph where Zv and Zw are the average covariance matrices of the
(RAG)   (V ,  ) [16] by regarding the regions as a set of regions v and w . The similarity D reg is obtained by
vertices V and the spatially adjacent regions are connected computing the Wishart distance of each pair of regions.
by the arcs e  . v V is a region of image, the set of After performing the region-based affinity propagation
image pixels belonging to region v can be represented as clustering with D reg , the regions are divided into different
S v . e   is an arc representing the boundary between two
clusters    {i | i  1, , n} , where n is number of clusters.
adjacent regions. Each region v V and the pixel s  Sv has Each region is then assigned with a new label based on the
clustering result. The region-based affinity propagation

5104
clustering can produce a stable initial classification map. TABLE I
CLASSIFICATION ACCURACY (%) OF THE
2.4. Adjacent Wishart classification DIFFERENT METHODS FOR DATASET1
Method Proposed H/α-
PolarIRGS
Class Method Wishart
The RAG  * and initial classification map are used to get
Rapeseed 98.4 94.4 68.8
*
the final result of PolSAR image. For an arbitrary region v Bare soil 95.9 97.2 98.2
which can be regarded as a vertex in RAG  * , if the region Wheat 98.3 93.5 96.2
* Barley 93.8 85.0 90.4
which have an arc connected with v is labeled by the Lucerne 99.0 93.7 66.6
region-based affinity propagation clustering with a new label Pea 98.0 60.6 74.4
r *
set L  {1,..., n} , the label xv* of region v can be Potato 99.6 84.8 67.4
estimated using the Wishart classifier proposed by J. S. Lee Beet 98.9 58.8 91.6
et al. [10] as follows, Grass 99.7 61.1 39.4

tL

xvr*  arg min W  v* ,t   (4) Overall
Kappa
96.9
0.95
85.3
0.67
76.0
0.63
W (v ,t )  ln Zv*  tr Z Z
*
 1
v* t  (5)
Dataset1 is the subimage of the L-band multilook fully
where W (v* ,t ) is the Wishart distance between region v*
polarimetric SAR image provided by AIRSAR of Flevoland
and the class  t , t is the label of adjacent region of v* , in 1989 and a dimension of 380×430 pixels. There are about
which is obtained by initial classification. Z v* and Zt are nine classes including rapeseed, baresoil, wheat, beat, grass,
pea, barley, lucerne, and potato in Fig. 1(a). The ground
the average covariance matrices of regions v* and class  t . truth is shown in Fig. 1(b). Figs. 1(c)-1(e) show the
As the boundary points between regions produced by the classification results obtained by the proposed, PolarIRGS
watershed algorithm are not labeled, our method also uses and H/α-Wishart methods, respectively. Table I shows that
the adjacent Wishart classifier to obtain the class of the the overall accuracy and kappa of our method are 96.9% and
unlabeled boundary points. In this paper, only an 8-neighbor 0.95, respectively, which are higher than that of other
set centered around the boundary point is considered. methods. Compared with Figs. 1(d) and 1(e), the proposed
method has good visual effects and improves the uniformity
3. EXPERIMENTS for each homogenous region, with better connectivity, as
shown in Fig. 1(c). Our proposed approach uses a new
In this paper, we use two real PolSAR datasets to validate formula to get the gradient of each pixel in order to obtain a
the effectiveness of our method. All the experimental data reasonable edge strength map. At the same time, both the
are preprocessed by refined Lee filter [17] with windows of polarimetric scattering information and spatial context
size 33. We also make comparisons between the proposed information between pixels are considered, so it outperforms
method and the other two commonly used methods: H/α- other methods.
Wishart [10] and PolarIRGS [11]. The maximum Dataset2 is the L-band, multilook PolSAR image in
number of iterations dmax of boundary-preserving region Oberpfaffenhofen, Germany of ESAR platform, whose size
growing is empirically set to be 15. is 1300 1200 pixels. There are three classes including
build-up areas, wood land and open areas in Fig. 2(a). The
ground truth is shown in Fig. 2(b). Figs. 2(c)-2(e) show the
classification results obtained by the proposed, PolarIRGS
and H/α-Wishart methods, respectively. Fig. 2 and Table II
indicate that the H/α-Wishart method yields poor results [see
(a) (b) (c) Fig. 2(e)]. For example, the proportion of built-up areas and
wood land are 26.8% and 38.7%, respectively. The kappa
coefficient is 0.44, and it is the lowest. Most pixels of the
two categories are misclassified into other categories.
Compared with the H/α-Wishart method, the result with the
PolarIRGS method significantly improves for wood land.
(d) (e) However, the proportion of built-up areas is still low. As
Fig.1. Classification results of Dataset1. (a) Pauli RGB shown in Table II, the accuracy of the proposed method is
composite. (b) Ground truth map. (c) Proposed method. (d) the highest at 91.4%, which outperforms the PolarIRGS
PolarIRGS. (e) H/α-Wishart. method by 70.9% and the H/α-Wishart method by 60.7%.

5105
Most pixels of all categories are correctly classified, and the IEEE Geosci. Remote Sens. Lett., vol. 5, no. 4, pp. 668–672, Oct.
proposed method also achieves the best kappa of 0.90. As 2008.
shown in Fig. 2(c), the proposed method can better [5] Y. H. Wang, C. Z. Han, and T. Florence, “PolSAR data
segmentation by combining tensor space cluster analysis and
distinguish the edge between different categories compared
markovian framework,” IEEE Geosci. Remote Sens. Lett, vol.7,
with Figs. 2(d) and 2(e). no.1, pp.210–214, Jan. 2010.
[6] T. Ince, S. Kiranyaz, and M. Gabbouj, “Evolutionary RBF
classifier for polarimetric SAR images,” Expert Syst. Appl., vol. 39,
no.5, pp. 4710–4717, Apr. 2012.
[7] Y. Maghsoudi, M. J. Collins, and D. G. Leckie, “Radarsat-2
polarimetric SAR data for boreal forest classification using SVM
and a wrapper feature selector,” IEEE J. Sel. Topics Appl. Earth
Observ. Remote Sens., vol. 6, no. 3, pp. 1531–1538, Jun. 2013.
(a) (b) (c) [8] B. Banerjee, A. Bhattacharya, and K. M. Buddhiraju, “A
generic land-cover classification framework for polarimetric SAR
images using the optimum Touzi decomposition parameter subset-
an insight on mutual information-based feature selection,” IEEE J.
Sel. Top. Appl. Earth Obs. Remote Sens., vol. 7, no. 4, pp. 1167–
1176, Apr. 2014.
[9] B. Liu, S. Member, H. Hu, H. Wang, K. Wang, X. Liu, and
W. Yu, “Superpixel-based classification with an adaptive number
(d) (e) of classes for polarimetric SAR images,” IEEE Trans. Geosci.
Fig.2. Classification results of Dataset2. (a) Pauli RGB Remote Sens., vol. 51, no. 2, pp. 907–924, Feb. 2013.
composite. (b) Ground truth map. (c) Proposed method. (d) [10] J. S. Lee, M. R. Grunes, T. L. Ainsworth, L. Du, S. Member,
PolarIRGS. (e) H/α-Wishart. D. L. Schuler, and S. R. Cloude, “Unsupervised classification
TABLE II using polarimetric decomposition and the complex Wishart
CLASSIFICATION ACCURACY (%) OF THE classifier,” IEEE Trans. Geosci. Remote Sens., vol. 37, no. 5, pp.
DIFFERENT METHODS FOR DATASET2 2249–2258, Sept. 1999.
[11] P. Yu, A. K. Qin, D. A. Clausi, and S. Member,
Method Proposed H/α-
PolarIRGS “Unsupervised polarimetric SAR image segmentation and
Class Method Wishart classification using region growing with edge penalty,” IEEE
Build-up Areas 90.1 54.5 26.8 Trans. Geosci. Remote Sens. vol. 50, no. 4, pp. 1302–1317, Apr.
Wood Land 92.3 93.3 38.7 2012.
Open Areas 91.2 70.4 83.3 [12] V. Akbari, A. P. Doulgeris, G. Moser, T. Eltoft, S. N.
Overall 91.4 70.9 60.7 Anfinsen, S. B. Serpico, “A textural–contextual model for
Kappa 0.90 0.66 0.44 unsupervised segmentation of multipolarization synthetic aperture
radar images,” IEEE Trans. Geosci. Remote Sens., vol. 51, no. 4,
pp. 2442–2453, Apr. 2012.
ACKNOWLEDGEMENT [13] A. P. Doulgeris, S. N. Anfinsen, and T. Eltoft, “Automated
non-Gaussian clustering of polarimetric synthetic aperture radar
This work was supported in part by the National Natural images,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 10, pp.
Science Foundation of China under Grant 61072106 and 3665–3676, Oct. 2011.
61271302. [14] F. Cao, W. Hong, Y. Wu, and E. Pottier, “An unsupervised
segmentation with an adaptive number of clusters using the
REFERENCES 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.
[1] P. Mishra, S. Member, D. Singh, and S. Member, “A
[15] B. J. Frey and D. Dueck, “Clustering by passing messages
statistical-measure-based adaptive land cover classification
between data points,” Science, vol. 315, no. 5814, pp. 972–976,
algorithm by efficient utilization of polarimetric SAR
Feb. 2007.
observables,” IEEE Trans. Geosci. Remote Sens., vol. 52, no. 5, pp.
[16] M. Sonka, V. Hlavac, and R. Boyle, Image Processing,
2889–2900, Mar. 2014.
Analysis and Machine Vision, 3rd ed. Belmont, CA: Thomson
[2] H. Skriver, “Crop classification by multitemporal C- and L-
Learning, 2008.
band single- and dual-polarization and fully polarimetric SAR,”
[17] J. S. Lee, M. R. Grunes, and G. de Grandi, “Polarimetric
IEEE Trans. Geosci. Remote Sens., vol. 50, no. 6, pp. 2138–2149,
SAR speckle filtering and its implication for classification,” IEEE
May. 2012.
Trans. Geosci. Remote Sens., vol. 37, no. 5, pp. 2363–2373, Sept.
[3] J. Lee, M. R. Grunes, E. Pottier, and L. Ferro-famil,
1999.
“Unsupervised terrain classification preserving polarimetric
scattering characteristics,” IEEE Trans. Geosci. Remote Sens., vol.
42, no. 4, pp. 722–731, Apr.2004.
[4] Y. H. Wu, K. F. Ji, W. X. Yu, and Y. Su, “Region-based
classification of polarimetric SAR images using wishart MRF,”

5106