IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 4, DECEMBER 2011

949

Unsupervised Classification of PolInSAR Data Based on Shannon Entropy Characterization With Iterative Optimization Wei Yan, Wen Yang, Member, IEEE, Hong Sun, Member, IEEE, and Mingsheng Liao Abstract—In this paper, we propose a modified unsupervised classification method for the analysis of polarimetric and interferometric synthetic aperture radar (PolInSAR) images using the intensity, polarimetric and interferometric contributions to the Shannon entropy characterization. In order to improve the classification accuracy where the polarimetric information is similar, the method gives intensity, polarimetric and interferometric information equal weighting to more effectively use the full range of information contained in PolInSAR data. In addition, this method uses an iterative clustering scheme which combines the expectation maximization (EM) and fast primal-dual (FastPD) optimization techniques to improve the classification quality. The first step of the method is to extract the Shannon entropy characterization from the PolInSAR data. Then, the image is initially classified respectively by the spans of the intensity, polarimetric and interferometric contributions to Shannon entropy. Finally, classification results of different contributions are merged and reduced to a specified number of clusters. An iterative clustering scheme is applied to further improve the classification results. The effectiveness of this method is demonstrated with DLR (German Aerospace Center) E-SAR PolInSAR data and CETC (China Electronics Technology Group Corporation) 38 Institute PolInSAR data. Index Terms—Expectation maximization, iterative optimization, polarimetric SAR interferometry, shannon entropy, unsupervised classification.

I. INTRODUCTION

P

OLARIMETRIC and interferometric synthetic aperture radar (PolInSAR) data contain a wealth of information provided by intensity, polarization and interference measurements [1]. Intensity information, which reflects the backscattering strength of the radar wave, makes it possible to classify

Manuscript received November 16, 2010; revised April 21, 2011 and revised April 21, 2011; accepted July 21, 2011. Date of publication October 13, 2011; date of current version December 14, 2011. This work was supported in part by the National Key Basic Research and Development Program of China under Contract 2007CB714405 and in part by the National Natural Science Foundation of China (No. 40801183, 41021061, 60872131, 60890074) and LIESMARS Special Research Funding. W. Yan and H. Sun are with the Signal Processing Laboratory, School of Electronic Information, Wuhan University, Wuhan 430079, China (e-mail: [email protected]; [email protected]). W. Yang is with the Signal Processing Laboratory, School of Electronic Information, Wuhan University, Wuhan 430079, China (corresponding author, e-mail: [email protected]). M. Liao is with the State Key Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University., Wuhan University, Wuhan 430079, China (e-mail: [email protected]). 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/JSTARS.2011.2164393

scatters with different intensity values [2], [3]. Polarimetric information, which relates to geometric and material properties, makes it possible to separate different scattering mechanisms within a given resolution cell [4]–[6]. Interferometric information, which reflects the topographical characteristics, makes it possible to distinguish variations in height [7]–[11]. PolInSAR data can provide better performance for a diverse range of classification applications by analyzing the large amount of information it contains. Lee et al. [12] proposed a method for forest classification based on PolInSAR data. The method used Freeman decomposition to obtain the initial forest clusters and introduced optimal coherency of interferometric information to distinguish different types of forest with a high accuracy. The contribution of interferometric information was limited as it was only used to subdivide the forest cluster. Ferro-Famil et al. [13] derived interferometric entropy and anisotropy on the basis of optimal coherency to increase accuracy in the scene description. The introduction of the coherence information permitted retrieval of some important features. The advantage of interferometric information in PolInSAR data was not fully utilized as the method reclassified the scene based on the PolSAR classification result. Guillaso et and which indicated al. [14] introduced two parameters the distributions of the optimized coherence in different channels to improve the accuracy of building detection by analyzing the optimized coherence phase obtained with the ESPRIT algorithm [15]. The advantage of a wealth of information in PolInSAR data was not effectively utilized as the intensity information and other available information were ignored. As a whole, the general framework for unsupervised PolInSAR classification is built on the unsupervised PolSAR classification. Interferometric information is imported into the PolSAR classification to improve the classification accuracy of particular categories or areas. As a consequence, the classification result more responses the differentiation of polarimetric information and the advantage of containing a wealth of information in PolInSAR data is not used fully and effectively. In this paper, we propose a modified unsupervised classification method that adopts an improved framework to utilize the wealth of information available in PolInSAR data. The first step of the method is to extract the Shannon entropy characterization from the PolInSAR data [16]. The Shannon entropy characterization of PolInSAR data has recently been proposed by Morio et al. [17], [18], and employs the intensity, polarimetric and interferometric contributions expressed in the same “unit”. Next, the image will be initially classified, respectively, by the span of intensity, polarimetric and interferometric contributions

1939-1404/$26.00 © 2011 IEEE

950

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 4, DECEMBER 2011

of the Shannon entropy. The classification results of different contributions are then combined. Finally, the fusion results are merged to a specified number of clusters on the basis of the between-cluster Wishart distance [19]. An iterative clustering optimization algorithm is applied to improve the classification result. In general, an iterative clustering technique is applied to improve the final classification result [20]. Since traditional clustering is based on the assumption that the pixels are independent, the classification results are sensitive to noise. Wu et al. [21] improved the accuracy and clarity of the classification results by exploiting the statistical a priori knowledge and the spatial relation between adjacent pixels with the Wishart Markov random field (MRF) method. Jäger et al. [22] used the expectation maximization and graph cut optimization techniques to utilize the spatial dependencies for PolSAR classification. A significant increase in classification quality was obtained. In this paper, we use the expectation maximization and fast primal-dual (FastPD) [23], [24] optimization techniques in the iterative clustering. By exploiting smoothed prior information and local spatial information, noise disturbance is decreased and the clarity of classification results is improved. We validate the effectiveness of this method using PolInSAR data of German Aerospace Center’s (DLR) E-SAR and the CETC 38 Institute, which is the first set of PolInSAR data collected in China. In order to justify the effectiveness of the method, we compare it with the traditional PolSAR classification [20] and the PolInSAR classification [13]. The paper is structured as follows. In Sections II and III, the Shannon entropy characterization and the EM and FastPD optimization techniques are briefly introduced as the basic theory underpinning this paper. The classification procedure is described in Section IV. In Section V, the experimental results are presented and analyzed. A conclusion is provided in Section VI.

where , are the directions in which the waves are emitted and the operator denotes the conjugate transpose. Generally, we consider the covariance matrix of the PolInSAR data which is a 6 6 Hermitian matrix and follows a complex Wishart distribution. The definition of the covariance matrix has the following structure [9] (2) stands for ensemble averaging and matrices and where are 3 3 Hermitian polarimetric covariance matrices for each of the two antennae. Each matrix contains polarimetric information for one antenna. represents the correlation between the two antennae containing the interferometric information. B. Shannon Entropy Characterization The Shannon entropy of a six dimensional complex-valued random vector with probability density function is defined as: (3) where

denotes complex six-dimensional integration and follows a six-dimensional Gaussian distribution. From (3), we can obtain the Shannon entropy characterization written as (4) Furthermore, the expression of the Shannon entropy in one homogeneous region of PolInSAR data can be decomposed into a sum of three terms, respectively, representing the intensity, polarimetric and interferometric contributions. (5)

II. SHANNON ENTROPY OF POLINSAR DATA

where the terms are

In this section, we briefly review the Shannon entropy characterization of PolInSAR data. The Shannon entropy is a way of quantifying the “disorder” of a random variable and represents a new approach for evaluating the contributions of the different channels of PolInSAR data proposed in [17], [18]. It can be interesting to determine how the quantified disorder is respectively partitioned into the intensity fluctuations, depolarization and interference [25], [26]. A. Basic Model of PolInSAR Data In the basic model of PolInSAR data, the backscattered signal can be described by a complex six-dimensional vector corresponding to the polarimetric performances at the two antennae

(6) The first term represents the intensity contribution of the Shannon entropy which is only a function of the intensities and which are received by the two antennae. The second term represents the polarimetric contribution of the Shannon entropy which only depends on the degrees of polarization and of the polarimetric covariance matrices. The last term represents the interferometric contribution of Shannon entropy which only depends on the optimized degrees of coherence , and . III. EM AND FASTPD OPTIMIZATION

(1)

Unsupervised classification methods of PolInSAR data generally need an iterative clustering approach to stabilize the clas-

YAN et al.: UNSUPERVISED CLASSIFICATION OF PolInSAR DATA BASED ON SHANNON ENTROPY CHARACTERIZATION WITH ITERATIVE OPTIMIZATION

sification results. However, most classification methods use iterative clustering schemes which assume that the pixels are mutually independent. As a consequence, the results are sensitive to noise and lack of spatial coherence. In this paper, we extend the work of Jäger et al. [22] to PolInSAR data by combining the EM and FastPD optimization techniques to solve the problem by introducing smoothed prior information and local spatial information into the iteration [25]. It can be proved that FastPD is as powerful as alpha-expansion, in the sense that it computes exactly the same solution, but with a substantial speedup-of a magnitude ten-over existing techniques.

951

TABLE I ITERATIVE PROCESS OF THE COMBINATION TECHNIQUE

A. Expectation Maximization For PolInSAR data, the EM algorithm iteratively maximizes the probability of for a complex Wishart mixture distribution [27]. Here, denotes a set of model parameters which are the centers of each classes expressed by 6 6 covariance matrices. As the PolInSAR data follow a complex Wishart distribution, the probability of the pixels is (7) and

where

is the observed variable and

is the latent variable.

is the covariance matrix of PolInSAR is the trace operation, denotes the gamma funcdata, tion, is the number of looks, and is the expected covariance matrix in class ( i.e. the weighted average covariance matrix of all pixels which are belong to class ). In the maximization step, the model parameters are updated:

(8)

where is the posterior density over labels and is the covariance matrix with observation . The EM algorithm begins with the E step, in which the probability is computed. These probability is used to obtain an updated set of model parameters in the M step. The estimate of is improved by the EM iteration. Then, an optimal Y label set is easily derived from a given . B. Energy Minimization by FastPD Optimization Aiming at introducing the smoothed prior information and local spatial information, a MRF model is embedded which yields an energy minimization problem. In our iterative clustering, the FastPD method [23], [24], comprising a new primaldual MRF framework optimization, is used to minimize the energy. Choosing a labeling scheme that minimizes MRF energy is equivalent to solving the following optimization problem:

(9) where is a discrete set of independent pixels, is the set of interacting pairs of pixels, represents a singleton potential assigning label to observation data, is a pairwise potential representing the interact of pixels pairs, and is a weight per edge, which can be used for scaling the pairwise potential function associated with that edge. FastPD algorithm solves the optimization problem under a primal-dual schema which keeps generating pairs of integralprimal and dual solutions until the elements and of the last pair are both feasible and satisfy the relax primal complementary slackness condition. The algorithm assumes that is merely a nonmetric distance and then tries to find feasible , satisfying complementary conditions. It is equal to restrict the active label should have the lowest height at each vertex, require that any two active labels should be raised proportionally to their separation costs and there is an upper bound on how much we can raise a label. Base on it, the algorithm updates the primal and dual variables at the outer iteration. At the dual variables updating process, the current active labels are given. Any nonactive labels are raised until either reach the active labels or attains the maximum raise allowed. At the primal variables updating process, the new height is given. The nonactive labels which below but have already reached the maximum raise and their relative active labels are updated. C. The Combination of EM and FastPD The EM and FastPD optimization techniques can be combined to utilize smoothed prior information and spatial information in a MRF model. The iterative process is similar to the EM algorithm iteration. In the iterative process, initially give a set of parameters and obtain improved estimate from as the E and M steps of (7) and (8) are modified to accommodate FastPD optimization. The parameters are successively refined until the label set is convergent to a stable configuration. The overall iterative process is described in Table I. In summary, the modified iterative clustering inserts the FastPD optimization into the iteration process of the EM algorithm. As the MRF model is embedded, the smoothed prior information and local spatial information are introduced into the iteration process. Consequently, the noise disturbance is

952

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 4, DECEMBER 2011

V. EXPERIMENTAL RESULT AND ANALYSIS

Fig. 1. Flowchart of PolInSAR Classification with Shannon entropy characterization.

decreased and the clarity of the classification result is improved. We will discuss the effectiveness of the modified technique in detail in Section V.

In this section, we will validate the effectiveness of the above proposed method using PolInSAR data from the German Aerospace Center’s (DLR) E-SAR and the CETC 38 Institute. The E-SAR data are L-band and have dimensions of 1300 1200 pixels collected in the Oberpfaffenhofen area of Germany. The resolution of the data is 3.00 2.20 m and the look angle varies from 25 to 60 . The average baseline across the scene is approximately 26 m and the flight height of the platform is about 3 km. The true terrain classes for each region are labeled according to optical remote sensing and related geographic information. The airborne data from the CETC 38 Institute are the first set of PolInSAR data acquired in China. These data are X-band and have dimensions of 400 1200 pixels acquired in Lingshui City, Hainan Province, China. Its aquisition date is January 31, 2010. The resolution of the data is 0.50 0.50 m and the look angle is 57 . The distance between platform and the closest scene edge is about 7 km. The original ground truth map is according to on-the-spot field investigation.

A. Analysis of the Shannon Entropy IV. PROCEDURE FOR THE PROPOSED METHOD In this section, we present the detailed procedure for the modified method. The modified method uses the Shannon entropy characterization of PolInSAR data to initially classify the image and applies a clustering scheme which combines the EM and FastPD optimization techniques to improve the classification quality. Fig. 1 shows the classification scheme. Firstly, we extract the Shannon entropy characterization from the PolInSAR data and use a filter with 3 3 independent sample averages. According to the definition of Shannon entropy characterization, we can obtain three terms which are , and , respectively representing the intensity, polarimetric and interferometric contributions. Secondly, we initially classify the image according to the spans of the intensity, polarimetric and interferometric contributions. As the intensity, polarimetric and interferometric information are independent and response the different properties, we then segment the image according to each span, respectively. In this work, we segment the image to three clusters according to each span. As the three contributions are expressed in the same “units”, we utilize the intensity, polarimetric and interferometric information equally and effectively. Next, we fuse the classification results in each of the three contributions. As we segment the image into three clusters in each of the three spans, the image is segmented into 27 clusters. But human beings can only identify a limited number of clusters. So we merge the clusters to a certain number of clusters based on the between-cluster Wishart distance. In this work, we merged the 27 clusters into eight clusters. Finally, an iterative clustering scheme which combines the EM algorithm and FastPD optimization technique is applied to stabilize the classification results.

The Shannon entropy characterization is extracted from the PolInSAR data. The intensity, polarimetric and interferometric contributions of the Shannon entropy in the Oberpfaffenhofen area of Germany are represented in Fig. 2. In addition, we compare the polarimetric characterization parameter proposed by Cloude and Pottier [28] and the interferometric characterization parameter proposed by Ferro-Famil [13] with the different contributions to the Shannon entropy. It can be observed in Fig. 2(b) that the intensity contribution is clearly distinct in different areas, especially in the grassland and the farmland. We can effectively utilize the differences in intensity contributions in the two areas to separate them. Also, the intensity contributions of the airport runway and the bare ground are clearly different. These differences can be used to accurately segment the two areas. Further, it can be seen in Fig. 2(c) that the polarimetric contributions of the airport runway and the bare ground are similar. Compared with other polarimetric characterizations, we find in Fig. 2(e) that the parameter is also similar in the two area. As a result, the performance of the classification carried out only using the polarimetric information in the two area is poor. Meanwhile, we notice that the polarimetric contribution has a low value in the bare ground, an intermediate value in farmland and a high value in forest. Therefore, the classification method using the polarimetric information can easily separate these different areas into different clusters. We can see in Fig. 2(d) that the interferometric contribution of the airport runway has a high value. However, the interferometric contribution of the bare ground is low. Good separation between these two areas is possible when using the interferometric information. Compared with other interferometric characterizations, we can see in Fig. 2(f) that the parameter is also well discriminated in these two areas.

YAN et al.: UNSUPERVISED CLASSIFICATION OF PolInSAR DATA BASED ON SHANNON ENTROPY CHARACTERIZATION WITH ITERATIVE OPTIMIZATION

Fig. 2. (a) E-SAR data of berpfaffenhofen area; (b) the intensity contribution; (c) the polarimetric contribution; (d) the interferometric contribution; (e) parameter. eter; (f)

Overall, the three contributions of the Shannon entropy reflect the different properties of each category. The three contributions yield good discrimination in different areas. For ex-

953

param-

ample, the intensity contributions of grassland and farmland are different, the polarimetric contributions of the bare ground and forest are clearly distinct, and the interferometric contributions

954

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 4, DECEMBER 2011

Fig. 3. (a) CETC 38 data of Lingshui City area; (b) the intensity contribution; (c) the polarimetric contribution; (d) the interferometric contribution; (e) parameter. eter; (f)

of the airport runway and the bare ground are obviously different. Therefore, we can achieve a good segmentation of different areas by effectively taking advantage of different contributions of the Shannon entropy. In order to further illustrate the features of the three contributions, we show the three contributions of the CETC 38 Institute PolInSAR data for the area of Lingshui City in Fig. 3. The polarimetric characterization parameter and the interferometric characterization parameter are also shown and compared. As shown in Fig. 3, we compare the paddy land I and paddy land II areas, which are rice crops at different stages of growth in the marked regions. We can see in Fig. 3(c) that the polarimetric contributions of the two areas are similar. It is hard to segment the two areas accurately. However, we can see in Fig. 3(b) and Fig. 3(d) that the intensity contributions and the interferometric contributions are obviously different between the two areas yielding an improvement in the classification accuracy of the two areas. Next, we compare the three contributions of Shannon entropy in Fig. 3 as a whole. We can see in Fig. 3(c) that the discrimination of the polarimetric contribution in different areas is not obvious, since the polarimetric scattering mechanism is similar in each area. Other polarimetric characterizations, such as the parameter , also show little variation, as shown in Fig. 3(e).

param-

However, we note that the discrimination in the intensity contribution is much more apparent, so that intensity contribution can be used to segment the different areas efficiently. Moreover, the interferometric contribution shows a sharp distinction between different areas. We can use the interferometric information to improve the accuracy of classification. From the analysis of the different contributions of Shannon entropy, we find that the three contributions of the Shannon entropy are linked to the scattering mechanism which can be exploited in the classification of PolInSAR images. The intensity, polarimetric and interferometric contributions can be equally used to improve the accuracy of the classification. B. Analysis of Classification Results In this section, we will analyze the classification results. From the analysis of different contributions in Section V-A, we find that the three contributions of the Shannon entropy can be used to classify the PolInSAR images as the three contributions show clear discrimination between different areas. We will also analyze the effectiveness of the modified clustering algorithm, which combines the EM and FastPD optimization technique. Fig. 4 and Fig. 5 show the results obtained for the two sets of PolInSAR data. The images Fig. 4(b) and Fig. 5(b) are the results of PolSAR classification [20]. The images Fig. 4(c) and Fig. 5(c) are the results of PolInSAR

YAN et al.: UNSUPERVISED CLASSIFICATION OF PolInSAR DATA BASED ON SHANNON ENTROPY CHARACTERIZATION WITH ITERATIVE OPTIMIZATION

955

Fig. 4. (a) E-SAR data of Oberpfaffenhofen area; (b) the classification; (c) the classification; (d) the Shannon entropy classification using Kmeans clustering; (e) the Shannon entropy classification using EM clustering without FastPD optimization; (f) the Shannon entropy classification using EM clustering with FastPD optimization.

956

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 4, DECEMBER 2011

Fig. 5. (a) CETC 38 data of Lingshui City area; (b) the classification; (c) the classification; (d) the Shannon entropy classification using Kmeans clustering; (e) the Shannon entropy classification using EM clustering without FastPD optimization; (f) the Shannon entropy classification using EM clustering with FastPD optimization.

classification [13]. The images (d)–(f) of Fig. 4 and Fig. 5 are the results of Shannon entropy classification respectively using the Kmeans clustering, EM clustering and the modified clustering. Table II and Table III show the classification accuracies of typical categories areas which are the marked regions in the images. The classification results of the E-SAR data for the Oberpfaffenhofen area are shown in Fig. 4. We can compare the areas in Fig. 4 where there is an airport runway. In the results of the classification, we cannot distinguish the airport runway from the surrounding bare ground accurately, since the polarimetric information of the two areas is similar which was seen in Section V-A. The classification accuracy of the

runway and the bareland are 80.4% and 80.9%, respectively. However, in the results of the PolInSAR classification, the airport runway is separated as the PolInSAR data contains much more available information. Fig. 4(c) shows that the classification separates the airport runway from the surrounding bare ground because the values of the interferometric characterization parameter of the two areas are different. The classification accuracy of the runway reaches 87.4% and the classification accuracy of the bareland reaches 85.6%. As shown in Fig. 4(d), Fig. 4(e) and Fig. 4(f), the results of the Shannon entropy classification method can segment the two areas more accurately. From the analysis in Section V-A, we can conclude that the intensity

YAN et al.: UNSUPERVISED CLASSIFICATION OF PolInSAR DATA BASED ON SHANNON ENTROPY CHARACTERIZATION WITH ITERATIVE OPTIMIZATION

957

TABLE II CLASSIFICATION ACCURACIES USING E-SAR POLINSAR DATA (%)

contributions and the interferometric contributions in the two areas are clearly discriminated, but not the polarimetric contribution. The proposed method utilizes the three contributions to initially classify the image and gives intensity, polarimetric and interferometric information equal weight. Consequently, better classification results are obtained using the proposed method. The classification accuracy of the runway is 95.3% and the classification accuracy of the bareland is 97.7%. In addition, we compare the different classification results for the whole image and find that the proposed method is more effective for the classification of forest, farmland and other categories because of its more effective utilization of the intensity, polarimetric and interferometric information contained in the PolInSAR data. However, the classification accuracy of the buildings is only 51.1% since the Shannon entropy cannot express the difference of buildings and other categories. It can be improved if other characterizations are used to remedy the shortage of Shannon entropy expression. Nevertheless, Shannon entropy characterization expresses the intensity, polarimetric and interferometric information in a simple way and could have a good utilization in classification. Moreover, the discrimination of the buildings and other categories will be better if the results are merged to more classes. In the proposed method, we apply a modified iterative clustering which combines the EM and FastPD optimization techniques to replace the traditional iterative clustering. In Fig. 4(d) to Fig. 4(f), we show the results of the Shannon entropy classification obtained, respectively, using the Kmeans clustering, EM clustering and modified clustering. Fig. 4(f) shows that the effect of the modified clustering is obvious as it utilizes smoothed prior information and spatial coherence.

Noise disturbance in the classification results is reduced significantly and the clarity of the classification is greatly improved. However, the performance of point-like target detection becomes poorer as the smoothed prior information makes the point-like target is homogenized by the area surrounding. As we used FastPD optimization to replace the graph cut optimization, more than seven times speed-up is gained. It only costs about 2.7 second, while graph cut costs 19.1 second, to complete one iteration process on a 2.6 GHz PC with 4 GB memory. We also applied the proposed method to the CETC 38 Institute PolInSAR data in the area of Lingshui City to demonstrate its effectiveness. The classification results are shown in Fig. 5 which reveals that the two kinds of rice crops at different stages of growth in the marked regions are wrongly classified into a same cluster in the classification result, since their polarimetric properties are similar. In Section V-A, we concluded that the intensity contributions and interferometric contributions show strong contrast between the two areas which aids their distinction. This result demonstrates how the modified method significantly improves separation of the two areas by taking advantage of the intensity, polarimetric and interferometric contributions of the Shannon entropy in the PolInSAR data. The classification accuracy of the two area is about 95.0%. The modified clustering technique improves the classification quality significantly. Fig. 5(f) shows that the classification result obtained using the modified clustering is much smoother than the results obtained with the Kmeans or EM iterative clustering. Noise is evident in results based on the Kmeans or EM clustering since the noise amplitude of the original PolInSAR data is large. However, noise is greatly reduced in results obtained using the modified iterative clustering since the clustering ef-

958

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 4, DECEMBER 2011

TABLE III CLASSIFICATION ACCURACIES USING CETC38 POLINSAR DATA (%)

fectively introduces smoothed prior information and local spatial information into the iteration process. Also, we note that the clarity of the classification results from the modified clustering is improved greatly. As the FastPD optimization is used, it only costs about 1.8 second, while the computation time of graph cut optimization is 14.6 second, to complete one iteration process on a 2.6 GHz PC with 4 GB memory. VI. CONCLUSION The PolInSAR data contain intensity, polarimetric and interferometric information which reflect the different properties of the various surface categories. Shannon entropy characterization utilizes the intensity, polarimetric and interferometric contributions of PolInSAR data and expresses them in the same “units”. In Section V-A, we analyze the data from the discrimination of the three contributions and other polarimetric or interferometric characterizations. From this analysis, we find that the discrimination of the polarimetric information is unsatisfactory in terms of identifying different categories, and in some areas, such as the airport runway and the bare ground, even the polarimetric properties are similar. However, discrimination using intensity information is much clearer between different categories. Meanwhile, the interferometric information provides good discrimination in areas where the topographical properties are different. The main problem in PolInSAR classification is how to effectively utilize the different information. In PolSAR classification, only the polarimetric information is utilized, leading to incorrect classification in areas where the polarimetric information is similar. In the traditional PolInSAR classification, the interferometric information is used to improve classification accuracy based on the PolSAR classification. As the polarimetric information in a major position, the classification results do not perform well in where the polarimetric information is similar even though the interferometric information is imported into the classification process. Therefore, a more reasonable framework which effectively utilizes the polarimetric

information, interferometric information and other available information can make classification accuracy further improved. In this paper, we proposed a modified method of unsupervised classification using the three different contributions of the Shannon entropy characterization. As the Shannon entropy expresses the three contributions in the same “units” and because the method gives equal weight to the three contributions when initially classify the image. In this way, the intensity, polarimetric and interferometric information in PolInSAR data is used equally. Therefore, the rich information contained in the PolInSAR data is utilized more effectively, yielding better results, as shown in Section V-B. In addition, we introduced a modified iterative clustering to the classification. In a combination of EM and FastPD optimization technique, the smoothed prior information and spatial coherence were introduced into the iterative process. From the analysis and comparison with the classification results, we can conclude that this approach is an effective technique to suppress the degree to which speckle produces isolated pixels and improve the clarity of classification results. ACKNOWLEDGMENT The authors would like to thank the CETC 38 Institute for their sustained effort in providing PolInSAR data. REFERENCES [1] J. S. Lee and E. Pottier, Polarimetric Radar Imaging: From Basics to Applications. Boca Raton, FL: CRC Press, Taylor & Francis Group, 2009. [2] C. J. Oliver and S. Quegan, Understanding Synthetic Aperture Radar Images. New York: Artech House, 1998. [3] D. H. Hoekman, M. A. M. Vissers, and T. N. Tran, “Unsupervised full-polarimetric SAR data segmentation as a tool for classification of agricultural areas,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens. (JSTARS), vol. 4, no. 2, pp. 402–411, 2011. [4] S. R. Cloude and K. P. Papathanassiou, “Polarimetric SAR interferometry,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 5, pp. 1551–1565, Sep. 1998.

YAN et al.: UNSUPERVISED CLASSIFICATION OF PolInSAR DATA BASED ON SHANNON ENTROPY CHARACTERIZATION WITH ITERATIVE OPTIMIZATION

[5] J. S. Lee, S. R. Cloude, K. P. Papathanassiou, M. R. Grunes, and I. H. Woodhouse, “Speckle filtering and coherence estimation of polarimetric SAR interferometry data for forest application,” IEEE Trans. Geosci. Remote Sens., vol. 41, no. 10, pp. 2254–2263, Oct. 2003. [6] M. Jäger, M. Neumann, S. Guillaso, and A. Reigber, “A self-initializing PolInSAR classifier using interferometric phase differences,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 11, pp. 3503–3518, Nov. 2007. [7] K. Papathanassiou and S. R. Cloude, “Single-baseline polarimetric SAR interferometry,” IEEE Trans. Geosci. Remote Sens., vol. 39, no. 11, pp. 2352–2363, Nov. 2001. [8] V. Alberga, “Volume decorrelation effects in polarimetric SAR interferometry,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 11, pp. 2467–2478, Nov. 2004. [9] L. Ferro-Famil, F. Kugler, E. Pottier, and J. S. Lee, “Forest mapping and classification at L band using Pol-inSAR optimal coherence set statistics,” in Proc. EUSAR, Dresden, Germany, May 16–18, 2006. [10] S. Sauer, L. Ferro-Famil, A. Reigber, and E. Pottier, “Physical parameter extraction over urban areas using L-band PolSAR data and interferometric baseline diversity,” in Proc. IGARSS, Barcelona, Spain, Jul. 23–28, 2007, pp. 5045–5048. [11] J. M. Lopez-Sanchez, J. D. Ballester-Berman, and I. Hajnsek, “First results of rice monitoring practices in spain by means of time series of TerraSAR-X dual-Pol images,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens. (JSTARS), vol. 4, no. 2, pp. 412–422, 2011. [12] J. S. Lee, M. R. Grunes, T. Ainsworth, I. Hajnsek, T. Mette, and K. P. Papathanassiou, “Forest classification based on L-band polarimetric and interferometric SAR Data,” in PolInSAR 2005 Workshop, Frascati, Italy, Jan. 17–21, 2005. [13] L. Ferro-Famil, E. Pottier, and J. S. Lee, “Unsupervised classification and analysis of natural scenes from polarimetric interferometric SAR data,” in Proc. IGARSS, Sydney, Australia, Jul. 9–13, 2001, pp. 2715–2717. [14] S. Guillaso, L. Ferro-Famil, A. Reigber, and E. Pottier, “Buliding characterization using L-band polarimetric interferometric SAR Dat,” IEEE Geosci. Remote Sens. Lett., vol. 2, no. 3, pp. 347–351, Jul. 2005. [15] H. Yamada, Y. Yamagushi, Y. Kim, E. Rodriguez, and W. M. Boerner, “Polarimetric SAR interferometry for forest analysis based on the ESPRIT algorithm,” IEICE Trans. Electronics, vol. E84-C, no. 12, pp. 1917–1924, 2001. [16] S. Maity, C. Patnaik, J. S. Parihar, S. Panigrahy, and K. A. Reddy, “Study of physical phenomena of vegetation using polarimetric scattering indices and entropy,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens. (JSTARS), vol. 4, no. 2, pp. 432–438, 2011. [17] J. Morio, P. Réfrégier, F. Goudail, P. Dubois-Fernadez, and X. Dupuis, “Information theory-based approach for contrast analysis in polarimetric and/or interferometric SAR images,” IEEE Trans. Geosci. Remote Sens., vol. 46, no. 8, Aug. 2008. [18] J. Morio, P. Réfrégier, F. Goudail, P. C. Dubois-Fernandez, and X. Dupuis, “A characterization of Shannon entropy and Bhattacharyya measure of contrast in polarimetric and interferometric SAR image,” Proc. IEEE, vol. 97, no. 6, pp. 1097–1108, Jun. 2009. [19] 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. [20] J. S. Lee, M. R. Grunes, T. L. Ainsworth, L. J. Du, D. L. Schuler, and S. R. Cloude, “Unsupervised classification using polarimetric decomposition and the complex Wishart classifier,” IEEE Trans. Geosci. Remote Sens., vol. 37, no. 5, pp. 2249–2258, Sep. 1999. [21] Y. H. Wu, K. Ji, W. Yu, and Y. Su, “Region based classification of polarimetric SAR images using Wishart MRF,” IEEE Geosci. Remote Sensing Lett., vol. 5, no. 4, pp. 668–672, Oct. 2008. [22] M. Jäger, A. Reigber, and O. Hellwich, “Unsupervised classification of polarimetric SAR data using graph cut optimization,” in Proc. IGARSS, Barcelona, Spain, Jul. 23–28, 2007, pp. 2232–2235. [23] N. Komodakis and G. Tziritas, “Approximate labeling via graph-cuts based on linear programming,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 29, no. 8, pp. 1436–1453, Aug. 2007. [24] N. Komodakis, G. Tziritas, and N. Paragios, “Performance vs computational efficiency for optimizing single and dynamic MRFs: Setting the state of the art with primal dual strategies,” Comput. Vis. Image Understand., vol. 112, no. 1, pp. 14–29, Oct. 2008. [25] R. Fallourd and O. Harant et al., “Monitoring temperate glacier displacement by multi-temporal TerraSAR-X images and continuous GPS measurements,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens. (JSTARS), vol. 4, no. 2, pp. 372–386, 2011.

959

[26] A. Balenzano, F. Mattia, G. Satalino, and M. W. J. Davidson, “Dense temporal series of C- and L-band SAR data for soil moisture retrieval over agricultural crops,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens. (JSTARS), vol. 4, no. 2, pp. 439–450, 2011. [27] C. M. Bishop, Pattern Recognition and Machine Learning, 2nd ed. New York: Springer, 2007, pp. 423–455. [28] S. R. Cloude and E. Pottier, “An entropy based classification scheme for land application of polarimetric SAR,” IEEE Trans. Geosci. Remote Sens., vol. 35, no. 1, pp. 68–78, Jan. 1997.

Wei Yan received the Bachelor degree from Wuhan University, China, in 2010. He is currently a Master degree student at the School of Electronic Information, Wuhan University. His research interests include synthetic aperture radar image segmentation and classification, target detection and recognition.

Wen Yang (M’09) received the Ph.D. degree from Wuhan University, China, in 2004. From September 2008 to September 2009, he was a visiting scholar in the AI (Apprentissage et Interfaces) team in the Laboratoire Jean Kuntzmann (LJK) in Grenoble, France. He is currently an Associate Professor at the School of Electronic Information, Wuhan University. His research interests include image segmentation and classification, target detection and recognition, machine learning and data mining with applications to remote sensing. Hong Sun (M’01) received the Bachelor, Master, and Ph.D. degrees in communications and electrical systems from Huazhong University of Science and Technology (HUST), China, in 1981, 1984, and 1998, respectively. She held teaching and research positions, a professor at HUST from 1984 to 2000. She was a visiting scholar at Conservatoire National des Arts et Mtiers (CNAM), Paris, France, in 1997 and a visiting Professor at Ecole Nationale Suprieure des Tlcommunications (ENST), Paris, France, in 1998, 1999, 2000, 2001, and 2007. Since 2001, she has been with the School of Electronic Information at Wuhan University, where she is currently a Professor and Head of the Signal Processing Laboratory (http://dsp.whu.edu.cn). Her research covers digital signal processing theory and its applications, including works on SAR image interpretation, communication signal processing and speech signal processing. Mingsheng Liao received the M.Eng. degree in electronic engineering and the Ph.D. degree in photogrammetry and remote sensing from Wuhan Technical University of Surveying and Mapping(WTUSM), Wuhan, China, in 1985 and 2000, respectively. From 1985 to 1992, he was with the Department of Electronic Engineering, WTUSM, which is one of the universities that formed Wuhan University in 2000. From 1992 to 2000, he was with the State Key Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing(LIESMARS), WTUSM, and obtained a Professorship in 1997. Between 1995 and 1996, he was with the Department of Surveying and Photogrammetry, Technical University of Denmark, as a a visiting Ph.D. candidate. From 2000 to 2002, he was with the Joint Laboratory for Geo-information Science, Chinese University of Hong Kong, as a Research Associate. Since 2003, he has been at the LIESMARS, Wuhan University, where he is leading the SAR data analysis and processing group. His main research interests are in interferometric and polarimetric data processing, differential SAR interferometry, and multi-temporal SAR image analysis. He has published over 60 research articles in referred scientific journals and books.