Using Combined Difference Image and k-Means Clustering for SAR ...

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 11, NO. 3, MARCH 2014

691

Using Combined Difference Image and k -Means Clustering for SAR Image Change Detection Yaoguo Zheng, Xiangrong Zhang, Member, IEEE, Biao Hou, Member, IEEE, and Ganchao Liu

Abstract—In this letter, a simple and effective unsupervised approach based on the combined difference image and k-means clustering is proposed for the synthetic aperture radar (SAR) image change detection task. First, we use one of the most popular denoising methods, the probabilistic-patch-based algorithm, for speckle noise reduction of the two multitemporal SAR images, and the subtraction operator and the log ratio operator are applied to generate two kinds of simple change maps. Then, the mean filter and the median filter are used to the two change maps, respectively, where the mean filter focuses on making the change map smooth and the local area consistent, and the median filter is used to preserve the edge information. Second, a simple combination framework which uses the maps obtained by the mean filter and the median filter is proposed to generate a better change map. Finally, the k-means clustering algorithm with k = 2 is used to cluster it into two classes, changed area and unchanged area. Local consistency and edge information of the difference image are considered in this method. Experimental results obtained on four real SAR image data sets confirm the effectiveness of the proposed approach. Index Terms—Change detection, difference image, k-means clustering, synthetic aperture radar (SAR) images.

I. I NTRODUCTION

A

S ONE of the most crucial applications in interpretation and understanding of synthetic aperture radar (SAR) images, change detection technique has been studied for many years. Compared with traditional optical remote sensing images, a SAR image is actively obtained and can be captured under all illuminations and weather conditions. These specialties make it widely used in many applications, such as agricultural survey [1], forest monitoring [2], natural disaster monitoring [3], and urban change analysis [4]. Change detection can be defined as a clustering process which clusters the pixels of the two SAR images acquired

Manuscript received December 14, 2012; revised April 18, 2013 and July 2, 2013; accepted July 25, 2013. This work was supported in part by the National Basic Research Program (973 Program) of China under Grant 2013CB329402, by the National Natural Science Foundation of China under Grants 61272282, 61072106, 61173092, 61271302, 61001206, 61202176, and 61271298, by the Fundamental Research Funds for the Central Universities under Grant K50511020011, by the National Science Basic Research Plan in Shaanxi Province of China under Grant 2011JQ8020, by the National Research Foundation for the Doctoral Program of Higher Education of China under Grant 20100203120005, by the Fund for Foreign Scholars in University Research and Teaching Programs (the 111 Project) under Grant B07048, and by the Program for Cheung Kong Scholars and Innovative Research Team in University under Grant IRT1170. The authors are with the Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University, Xi’an 710071, China (e-mail: [email protected]; [email protected]; [email protected]; [email protected]). Digital Object Identifier 10.1109/LGRS.2013.2275738

on the same geographical area but at different times into two classes: changed area and unchanged area. Traditional change detection methods can be classified into two categories: supervised approaches and unsupervised approaches. The supervised methods need some labeled samples for the training of a sequential classifier, while the unsupervised methods generate the difference image directly depending on the two SAR images. Lacking of the ground truth in real applications makes the unsupervised approaches much more popular than the supervised ones. In this letter, we also focus on the unsupervised change detection. As mentioned in the literature [5], unsupervised change detection algorithms of SAR images can be mainly summarized into three steps: 1) image preprocessing; 2) obtaining the difference image; and 3) analyzing the difference image and postprocessing. The multiplicative speckle noise in SAR images makes the change detection much more difficult than that of optical images. Therefore, speckle noise reduction in the first step is very important for the entire process of change detection. For the second step, the difference image can be obtained through various methods. The subtraction operator and the log ratio operator are often used in obtaining the difference image. Although the subtraction operator can find the details of weakly changed area easily, it is not suitable for the statistical nature of the SAR image. The ratio operator can suppress the noise to a certain extent; however, it will enlarge the change information when the values in the two SAR images of the same position are all very small (e.g., six to three), which obtains the same ratio value to that of 200 and 100 in the change map, and also, it will miss the weakly changed information (e.g., 120 to 100, and 240 to 200 also obtain the same ratio value). All these difference images obtained via a single difference operator make the sequential analysis much more difficult. The third step is mainly to partition the difference image into changed area and unchanged area. Traditional clustering methods are fuzzy c-means [6], k-means [7], normalized cut [8], etc. In this letter, in order to solve the aforementioned problems of the traditional difference image generation operators, we propose a simple and effective method which concentrates on the processes of speckle noise reduction and difference image generation. First, in order to suppress the effect of speckle noise, the probabilistic-patch-based (PPB) algorithm [9], [14] is used in the first step, and the subtraction operator and the log ratio operator are used to generate two kinds of different difference images on the two multitemporal SAR images. Then, the mean filter is applied on the difference image obtained by the subtraction operator to smooth the image, and the median filter is applied on the difference image obtained by the log ratio

1545-598X © 2013 IEEE

692

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 11, NO. 3, MARCH 2014

operator to preserve the edge information. Second, a simple combination framework is proposed to integrate the difference images obtained by the mean filter and median filter to a new one, and then, k-means clustering algorithm is used to partition it into two classes: changed area and unchanged area. Such a framework takes the advantages of the smoothness and local consistency of the difference image obtained by the mean filter and the preservation of edge information of the difference image obtained by the median filter. The main contributions of this work include the following: 1) We propose a new difference image generation framework which combines the local consistency and edge information of two kinds of classical difference images and can be implemented easily, and 2) the number of missed alarms and false alarms can be tuned in the combined difference image and kmeans clustering (CDI-K) via a regularization parameter, which makes the proposed approach suitable for specific applications. The remainder of this letter is organized as follows. Section II provides the proposed method. Experimental results on four real SAR image data sets are presented in Section III, where we compared CDI-K with the methods presented in [10]–[12]. Section IV concludes this letter. II. CDI-K The PPB filter [9] has better performance in image smoothness and shape preservation, so we use it for noise reduction. After using the PPB filter for the two SAR images I1 and I2 obtained in the same geographical area with different times, we obtain the denoised images X1 and X2 corresponding to I1 and I2 . Then, we use the traditional difference image generation operators to generate two kinds of difference images. The subtraction operator and the log ratio operator are used in this letter (1) Ds = |X1 − X2 |    X2 + 1  = |log(X2 + 1) − log(X1 + 1)| . (2) Dl = log X1 + 1  In the process of using the log ratio operator, we use Xi + 1 (i = 1, 2) instead of Xi (i = 1, 2) to avoid the case that the pixel values in Xi (i = 1, 2) be zeros which make (2) nonsense. After using the subtraction operator and the log ratio operator, we normalize Ds and Dl to the range [0, 255] for the sequential processing. Then, the mean filter is used for the change map Ds obtained by the subtraction operator. We use the mean filter because there are a lot of isolated pixels in Ds , and a filter with a large window size can make the region much more complete and the local area consistent. The median filter is then used for the change map Dl obtained by the log ratio operator. The median filter has better performance in edge information preservation. Thus, we use the median filter for Dl to suppress the isolated pixels and preserve the edge pixels to some extent. The window sizes for the mean filter and the median filter are empirically set as 11 × 11 and 3 × 3 in all our experiments, respectively. To get the final difference image, we propose a simple combination framework to combine Ds and Dl using the expression

Fig. 1.

Framework of CDI-K proposed in this letter.

Fig. 2. Multitemporal images relating to the city of Bern used in the experiments. (a) Image acquired in April 1999, (b) image acquired in May 1999, and (c) reference.

in (3), where Ds and Dl are the results after filtering of Ds and Dl , respectively. The combination can take advantage of the smoothness and enhanced local consistency provided by Ds and the edge preservation by Dl . In order to make pixel values in the combined difference image D between 0 and 255, we use one regularization parameter α > 0 to balance the effects of the smoothness term and edge preservation term. The weight of Ds is set as α, and the weight of Dl is set as 1 − α. The mean filter obtains many locally smooth regions which have strong local consistency. The median filter has a good nature in edge information preservation, so it preserves the shape of the changed area. In theory, the weight should be small for the smoothness term and be large for the edge preservation term. Too large weight for the smoothness term will lead to a large number of false alarms, and too small weight for the edge preservation term will lead to a large number of missed alarms. The quantitative analysis of the effect of the regularization parameter α will be given in Section III-C. Finally, k-means clustering algorithm is used to partition the difference image D into two clusters. For a good visual effect of the clustering result in R, according to the clustering labels, we set the gray values in the changed areas to 255 and the unchanged areas to zero. The entire framework of our CDI-K is shown in Fig. 1 D = αDs + (1 − α)Dl .

(3)

III. E XPERIMENTAL R ESULTS A. Description of Data Sets We have tested our method on four real SAR image data sets. The first data set represented a section (301 × 301 pixels) of two SAR images. Fig. 2(a) and (b) shows the images acquired by the European Remote Sensing 2 satellite SAR sensor over an area near the city of Bern, Switzerland, in April and May 1999, respectively. A reference map was manually defined as shown in Fig. 2(c). The second data set represented a section (290 × 350 pixels) of two SAR images. Fig. 3(a) and (b) shows the

ZHENG et al.: USING CDI-K FOR SAR IMAGE CHANGE DETECTION

693

Fig. 3. Multitemporal images relating to Ottawa used in the experiments. (a) Image acquired in July 1997, (b) image acquired in August 1997, and (c) reference.

Fig. 5. Change detection results by using different methods on the Ottawa data set. (a) By LN, (b) by PCA-K, (c) by NR, (d) the difference image generated by CDI-K, and (e) by CDI-K. TABLE I C OMPARISON OF D ETECTION R ESULTS ON THE B ERN AND OTTAWA DATA S ETS

Fig. 4. Change detection results by using different methods on the Bern data set. (a) By LN, (b) by PCA-K, (c) by NR, (d) the difference image generated by CDI-K, and (e) by CDI-K.

images taken in July and August 1997, respectively. A reference map was manually defined as shown in Fig. 3(c). The last two SAR image data sets were parts of SAR images of Yellow River Estuary, which were taken over Dongying in the Shandong Province of China. The spatial resolution is 8 m × 8 m. It should be noted that the last two SAR data sets have different levels of noise. The images in Figs. 6(a) and 7(a) were four looks and acquired in June 2008. The images in Figs. 6(b) and 7(b) were single look and acquired in June 2009. Due to lack of reference maps, we only give out the visually experimental results obtained by different methods on the last two data sets. B. Results and Analysis In order to demonstrate the effectiveness of our CDI-K, we select three related methods for comparison: log-normal (LN) method [10], principal component analysis and k-means clustering (PCA-K) [11], and neighborhood-based ratio approach (NR) [12]. The performance is evaluated by false alarms, missed alarms, overall error, and the kappa index which is a statistical measurement of accuracy or agreement [13]. The visual results of different algorithms on the Bern data set are shown in Fig. 4. Fewer isolated pixels can be found in the result obtained by CDI-K compared with those of the other methods. The difference image generated by CDI-K is shown in Fig. 4(d). Compared with PCA-K, CDI-K has a better performance in the preservation of edge information. Our CDI-K

obtains the least overall error of 289 pixels, while the other methods have 314, 304, and 309 pixels, respectively. Using the mean filter results in fewer isolated pixels, and with the median filter, the edge information can be well preserved. In this experiment, the regularization parameter α in CDI-K is set to 0.2. The visual results of different methods on the Ottawa data set are shown in Fig. 5. Compared with LN and NR, CDI-K results in fewer isolated pixels and large region of smoothed areas. The difference image generated by CDI-K is shown in Fig. 5(d). Our CDI-K obtains the overall error of 1704 pixels, much smaller than those of LN, PCA-K, and NR of 2257, 2470, and 2126, respectively. The kappa index of CDI-K is 0.9353, which is superior to 0.9187, 0.9049, and 0.9224 of the other three methods. Compared with Fig. 5(a), (c), and (f) shows the superiority in removing isolated pixels of CDI-K, and compared with Fig. 5(b), it shows the better nature in the preservation of edge information of CDI-K. The regularization parameter α in CDI-K is set to 0.1 in this experiment. The Bern and Ottawa data sets are quantitative analysis of the experimental results, as shown in Table I. As we can see from it, our method has better performance in terms of overall error and kappa index.

694

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 11, NO. 3, MARCH 2014

Fig. 6. Experimental results on Yellow River data set A. (a) Image taken in 2008, (b) image taken in 2009, (c) result by LN, (d) result by PCA-K, (e) result by NR, and (f) result by CDI-K.

Fig. 8. Process of difference image generation using filters. (a) Difference image Ds , (b) image after mean filter with a window size of 11 × 11 on Ds , (c) log ratio image Dl , (d) image after median filter with a window size of 3 × 3 on Dl , and (e) final change detection result.

pixels and relatively well preserved edges can be found in Figs. 6(f) and 7(f). A detailed representation of CDI-K is shown in Fig. 8. We can see from Fig. 8(a) that there are many isolated and small regions in the difference map Ds and fewer isolated pixels in Dl . In order to reduce the effect of isolated pixels, a filter with a large window size is needed. After filtering it using the mean filter with a large window size, the local consistency is enhanced. However, it results in an inexact edge in the difference map. Another important aspect to be considered in difference image generation is edge information. Here, we use the median filter with a small window size to preserve the edge information. Combining the two parts, we can obtain a relatively satisfactory clustering result via k-means. C. Influence of the Regularization Parameter

Fig. 7. Experimental results on Yellow River data set B. (a) Image taken in 2008, (b) image taken in 2009, (c) result by LN, (d) result by PCA-K, (e) result by NR, and (f) result by CDI-K.

The experimental results of the last two data sets are shown in Figs. 6 and 7. The parameter α is set to 0.3 in CDI-K in the two experiments. As can be seen from Figs. 6(c) and 7(c), the LN gets many isolated pixels in the final change detection results, which show that it is not very robust to strong level of noises. The PCA-K can reduce the effect of speckle noise to some extent; however, it loses some edge information of the changed areas. For the NR, due to its neighborhood operations applied in the difference image generation, although the phenomenon of isolated pixels and the effect of noise can be reduced, it also isolates the connected regions to some extent. Fewer isolated

As our CDI-K is based on the combination of Ds and Dl with a regularization parameter α, we analyze the influence of α on the performance in this section. All the quantitative analyses with the increase of α from zero to one for the Bern and Ottawa data sets are shown in Fig. 9. In order to preserve the properties of local smoothness and edge information simultaneously, the weight should be neither too large nor too small. Because there also has the property of smoothness in Dl , we set α which controls the contribution of smoothness with a relatively small value. As we can see, a relatively small value of α can get good results on the two data sets. A finely tuned parameter will improve the performance of CDI-K. IV. C ONCLUSION In this letter, an unsupervised change detection technique has been proposed by using k-means clustering on the combined difference image. The motivation of our method is to use the local consistency of the difference image after using the mean filter and the edge information preservation of the difference image after using the median filter for better difference image representation. Experiments on four real SAR image data sets

ZHENG et al.: USING CDI-K FOR SAR IMAGE CHANGE DETECTION

695

Fig. 9. Performances of CDI-K on the Bern and Ottawa data sets against the parameter α. (a) Result on the Bern data set. (b) Result on the Ottawa data set. (c) Kappa index.

have demonstrated the effectiveness of the proposed method. Our work is based on the PPB filter, so a better denoised image which considered the sequential applications (e.g., classification, recognition, and change detection) in the process of denoising will improve the performance. Experiments show that our approach has better performance in the preservation of changed area, but how to make the edge pixels classified more exactly is still a research direction. Thus, an exact edge detection algorithm in SAR images should be considered to improve the change detection performance in our future works. The change detection of point targets will also be our future work.

R EFERENCES [1] L. Bruzzone and S. B. Serpico, “An iterative technique for the detection of land cover transitions in multitemporal remote sensing images,” IEEE Trans. Geosci. Remote Sens., vol. 35, no. 4, pp. 858–867, Jul. 1997. [2] T. Hame, I. Heiler, and J. San Miguel-Ayanz, “An unsupervised change detection and recognition system for forestry,” Int. J. Remote Sens., vol. 19, no. 6, pp. 1079–1099, 1998. [3] G. Di Martino, A. Iodice, D. Riccio, and G. Ruello, “A novel approach for disaster monitoring: Fractal models and tools,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 6, pp. 1559–1570, Jun. 2007. [4] M. K. Ridd and J. Liu, “A comparison of four algorithms for change detection in an urban environment,” Remote Sens. Environ., vol. 63, no. 2, pp. 95–100, Feb. 1998.

[5] L. Bruzzone and D. F. Prieto, “An adaptive semiparametric and contextbased approach to unsupervised change detection in multi-temporal remote sensing images,” IEEE Trans. Image Process., vol. 11, no. 4, pp. 452–466, Apr. 2002. [6] A. Ghosh, N. S. Mishra, and S. Ghosh, “Fuzzy clustering algorithms for unsupervised change detection in remote sensing images,” Inform. Sci., vol. 181, no. 4, pp. 699–715, Feb. 2011. [7] T. Celik, “Unsupervised change detection of satellite images using local gradual descent,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 5, pp. 1919–1929, May 2012. [8] X. Zhang, Z. Li, B. Hou, and L. Jiao, “Spectral clustering based unsupervised change detection in SAR images,” in Proc. IEEE IGARSS, 2011, pp. 712–715. [9] C. Deledalle, L. Denis, and F. Tupin, “Iterative weighted maximum likelihood denoising with probabilistic patch-based weights,” IEEE Trans. Image Process., vol. 18, no. 12, pp. 2661–2672, Dec. 2009. [10] G. Moser and S. B. Serpico, “Generalized minimum-error thresholding for unsupervised change detection from SAR amplitude imagery,” IEEE Trans. Geosci. Remote Sens., vol. 44, no. 10, pp. 2972–2982, Oct. 2006. [11] T. Celik, “Unsupervised change detection in satellite images using principal component analysis and k-means clustering,” IEEE Geosci. Remote Sens. Lett., vol. 6, no. 4, pp. 772–776, Oct. 2009. [12] M. Gong, Y. Cao, and Q. Wu, “A neighborhood-based ratio approach for change detection in SAR images,” IEEE Geosci. Remote Sens. Lett., vol. 9, no. 2, pp. 307–311, Mar. 2012. [13] G. H. Rosenfield and A. Fitzpatrick-Lins, “A coefficient of agreement as a measure of thematic classification accuracy,” Photogramm. Eng. Remote Sens., vol. 52, no. 2, pp. 223–227, 1986. [14] X. Zhang, Y. Zheng, J. Feng, and S. Gou, “SAR image change detection based on low rank matrix decomposition,” Proc. IEEE IGARSS, pp. 6271– 6274, Jul. 2012.