Detecting Changes in Synthetic Aperture Radar Images using Image Fusion and Fuzzy Clustering

Detecting Changes in Synthetic Aperture Radar Images using Image Fusion and Fuzzy Clustering

Mayura M. Mane Mrs. R. M. Mandi

M.E. (VLSI and Embedded Systems) Asst. Prof. (SITS, Narhe, Pune)

Abstract-There are various change detection techniques that are presented previously but those were not much efficient. An unsupervised distribution-free change detection approach is presented in this paper for synthetic aperture radar (SAR) images based on an image fusion strategy and a novel fuzzy clustering algorithm. The image fusion technique will be used to generate a difference image by using a mean- ratio image and a log-ratio image. Wavelet fusion rules based on an average operator and minimum local area energy are chosen to fuse the wavelet coefficients for a low-frequency band and a high- frequency band respectively are used in image fusion. A reformulated fuzzy local-information C- means clustering algorithm is proposed for classifying changed and unchanged regions in the fused difference image. It enhances the changed information and of reduces the effect of speckle noise.

Keywords- SAR, Speckle noise, Image fusion, Change detection, Reformulated fuzzy local information C-means clustering.

1. INTRODUCTION

   Synthetic aperture radar is a imaging radar which operates in any weather conditions and provides high resolution. So it has various applications like Reconnaissance, Surveillance, and Targeting, Foliage and Ground Penetration etc. Image change detection [9] is a process that analyzes images of the same scene taken at different times in order to identify changes that may have occurred between the considered acquisition dates [1]. Change detection in synthetic aperture radar (SAR) images exhibits some difficulties than optical ones due to the fact that SAR images suffer from the presence of the speckle noise which degrades the quality of image. However, SAR sensors are independent of atmospheric and sunlight conditions, which make the change detection in SAR images still attractive [1],[4].

   An unsupervised change detection [8] in SAR images can be divided into three steps: 1) image preprocessing; 2) producing difference image between the images; and 3) analysis of the difference image. The tasks of the first step mainly include coregistration, geometric corrections, and noise reduction. In the second step, two images are compared pixel by pixel to generate the difference

   image. For the remote sensing images, differencing (subtraction operator) and rationing (ratio operator) are well-known techniques for producing a difference image. In differencing technique, changes are measured by subtracting the intensity values pixel by pixel between the considered couple of temporal images. And in rationing, changes are obtained by applying a pixel-by-pixel ratio operator to the considered couple of temporal images. However, in the case of SAR images, instead of differencing the ratio operator is typically used operator because the image differencing technique is not adapted to the statistics of SAR images and nonrobust to calibration errors [4]. In addition, because of the multiplicative nature of speckles, the ratio image is usually expressed in a logarithmic or a mean scale [2]. With the log-ratio operator, the multiplicative speckle noise can be transformed in an additive noise component. Furthermore, the range of variation of the ratio image will be compressed and thereby enhances the low-intensity pixels while weakening the pixels in the areas of high intensity; therefore, the distribution of two classes (changed and unchanged) could be made more symmetrical. However, the information of changed regions that is obtained by the log-ratio image may not be able to reflect the real changed trends in the maximum extent because of the weakening in the areas of high- intensity pixels. As for the ratio mean operator, the background (unchanged regions) of mean-ratio image is quite rough, for the ratio technique may emphasize the differences in the low intensities of the temporal images. In the third step, changes are usually detected by applying a decision threshold to the histogram of the difference image. Several thresholding methods have been proposed in order to determine the threshold in an unsupervised manner, such as Otsu, the Kittler and Illingworth minimum- error thresholding algorithm (K&I), and the expectation maximization (EM) algorithm [5]. It is apparent that this kind of methods requires an accurate estimation of the decision threshold. Moreover, they need to select a proper probability statistical model for distribution of change and unchanged classes in the difference image, which leads to significant restrictions on their application prospect. Various fuzzy c-means clustering algorithms [6],[7] were used in change detection like FCM_S, fast generalized FCM, FGFCM, and FLICM but all those were not much efficient. FCM was very sensitive to the speckle noise, in FCM_S the labelling of a pixel to be influenced by the labels

   in its immediate neighbourhood. However, compared with the original FCM, the computational complexity of FCM_S is significantly increased since it computes the neighbourhood term in each iteration step FGFCM significantly reduced the execution time by clustering on gray-level histogram rather than on pixels; meanwhile, It is less sensitive to noise to some extent because of the introduction of local spatial information. An artificial parameter is applied in their

   objective functions in order to balance between robustness to noise and effectiveness of preserving the details of the image. The selection of parameter is not easy to implement since there is no prior knowledge about the speckle noise level. FLICM introduced a term fuzzy factor which uses gray-level difference and spatial distance and use of spatial distance is not appropriate in some cases. Therefore a new algorithm is introduced.

   In general, it appears clearly from the literature that the whole performance of SAR-image change detection is mainly relied on the quality of the difference image and the accuracy of the classification method. In order to address the two issues, in this paper, we propose an unsupervised distribution-free

   SAR-image change detection approach. It is unique in the following two aspects: 1) producing difference images by fusing a mean-ratio image and a log-ratio image, and 2) improving the fuzzy local-information c-means (FLICM) clustering algorithm [18], which is insensitive to noise, to identify the change areas in the difference image, without any distribution assumption.

   This paper is organized into five sections.

2. PROPOSED METHOD

   Let us consider the two coregistered intensity

   time and space, allowing detail information to be easily extracted from images.

   First the mean-ratio operator and the log-ratio operator are applied to the two source images used for

   Original SAR image 1

   Original SAR image 2

   Original SAR image 1

   Original SAR image 2

   Log Ratio Image

   Log Ratio Image

   Mean Ratio Image

   Image Fusion

   Fuzzy Clustering

   Fuzzy Clustering

   Fig. 1. Flowchart of the proposed change detection approach.

   fusion respectively, which are commonly given by

   Xm= 1-min (µ1/µ2 , µ2/µ1) (1) Xl= |log(X2/X1)| (2)

   where µ1 and µ2 represent the local mean values of multitemporal SAR images X1 and X2, respectively.

   In the image fusion scheme based on the wavelet transform wavelet coefficient map is obtained by taking DWT of log ratio and mean ratio images. After that fusion rules are applied. Here, two main fusion rules are applied: the rule of selecting the average value of corresponding coefficients for the low-frequency band, and the rule of selecting the minimum local area energy coefficient for the high- frequency band. The fusion rules can be described as

   SAR images X1= {x1(i,j), 1<i<H, 1<j<W} and X2=

   follows:

   F m l

   {x2(i,j), 1<i<H, 1<j<W} of size , i.e., acquired over

   D LL=(D LL+D LL)/2 (3)

   the same geographical area at two different times

   DF = Dm (i,j) , Em (i,j) < El (i,j) or

   and, respectively. Our objective is aiming at

   producing a difference image that represents the change information between the two times; then, a binary classification is applied to produce a binary image corresponding to the two classes: change and unchanged. As shown in Fig. 1, the proposed unsupervised distribution- free change detection approach is made up of two main phases: 1) generate the difference image using the wavelet fusion based on the mean-ratio image and the log-ratio image; and

   1. automatic analysis of the fused image by using an improved fuzzy clustering algorithm.

      1. Generate the Difference Image Using Image Fusion

         A discrete wavelet transform (DWT) is used for image fusion. The DWT isolates frequencies in both

         Dl (i,j) , El (i,j) >= Em (i,j) (4)

         Where m and l represent the mean-ratio image and the log-ratio image, respectively. F denotes the new fused image. DLL stands for low-frequency

         coefficients. Dl (i,j) represents three high-frequency

         coefficients at point (i,j) in the corresponding sub- images. The local area energy coefficient E(i,j) can be computed as follows:

         E(i,j)=kNi,j |D(k)| (5)

         Where E(i,j) represents the local area energy of the wavelet coefficient at point (i,j) in the corresponding sub-image, and Ni,j represents the local window centred on (i,j). D(k) denotes the value of the kth wavelet coefficient that is around the local window.

         After that inverse DWT is taken to get the fused image.

      2. Detect Changed Areas in the Fused Image Using the Improved FCM

   In reformulated fuzzy local information c-means clustering algorithm the local coefficient of variation is adopted to replace the spatial distance in FLICM which is defined as:

   Cu=var(x)/( x)2 (6)

   where var(x) and are the intensity variance and the mean in a local window of the image, respectively.

   The value of Cu reflects the gray-value homogeneity degree of the local window.

   The fuzzy factor for RFLICM can be defined as

   transformation. Hence, complementary information from the mean-ratio image and the log-ratio image is utilized to fuse a new difference image. Here, the RFLICM algorithm that incorporates both local spatial and gray information is proposed, which is relatively insensitive to probability statistics model. The RFLICM algorithm introduces the reformulated factor as a local similarity measure to make a trade- off between image detail and noise. So, the proposed wavelet fusion strategy can integrate the advantages of the log-ratio operator and the mean-ratio operator and gain a better performance. The change detection results obtained by the RFLICM exhibited less spots than its pre-existence (i.e., FLICM) since it is able to incorporate the local information more exactly.

   G = [½+min((Cj /C )2,(C /Cj )2)]

   ki jNi

   u u u u

   REFERENCES

   u

   u

   [X(1-ukj)m||xj-vk||2 if Cj >= Cu

   Or

   1. Maoguo Gong, Member, IEEE, Zhiqiang Zhou, and Jingjing Ma, Change Detection in Synthetic Aperture Radar Images based on Image Fusion and Fuzzy Clustering, IEEE

      G =

      G =

      ki jNi

      [½-min((Cj /C )2,(C /Cj )2)]

      Trans. Image Process., VOL. 21, NO. 4, APRIL 2012.

      u

      u

      u u u u

      u u u u

      [X(1-ukj)m||xj-vk||2 if Cj < Cu (7)

      u

      u

      Where, vk is prototype value of kth cluster, ukj is fuzzy membership of ith pixel with respect to cluster k, Cj

      u

      u

      is local coefficient of variation of neighbouring pixels, Cu is local coefficient of variation of central pixel and Cu is the mean value of Cj that is located in a local window.

      The calculation of the membership partition matrix

      c

      c

      is performed as follows:

      u =1/{ {[(||x -v ||2+G )/( ||x -v ||2+G )]1/(m-

   2. R. J. Radke, S. Andra, O. Al-Kofahi, and B. Roysam, Image change detection algorithms: A systematic survey, IEEE Trans. Image Process., vol. 14, no. 3, pp. 294307, Mar. 2005.

   3. M. Sezgin and B. Sankur, A survey over image thresholding techniques and quantitative performance evaluation, J. Electron. Imag., vol. 13, no. 1, pp. 146165, Jan. 2004.

   4. E. J. M. Rignot and J. J. Van Zyl, Change detection techniques for ERS-1 SAR data, IEEE Trans. Geosci. Remote Sens., vol. 31, no. 4, pp. 896906, Jul. 1993.

   5. James C. Bezdek, Robert Ehrlich, William Full,

      ki =1

      i k ki

      i j ji

      Fcm: The Fuzzy C-means Clustering

      1)}} (8)

      uki represents the fuzzy membership of the ith pixel with respect to cluster k, and c is the number of clusters, ||xi-vk||2 is the Euclidean distance between object and the cluster centre.

      Here, the reformulated factor balances the membership value of the central pixel taking into account the local coefficient of variation, as well as the gray level of the neighbouring pixels. If there is a distinct difference between the results of the local coefficient of variation that are obtained by the neighbouring pixel and the central pixel, the weightings added of the neighbouring pixel in will be increased to suppress the influence of outlier.

3. CONCLUSION

In proposed method, the information of changed regions reflected by the mean-ratio image is relative in accordance with the real changed trends in multitemporal SAR images. On the other hand, the information of background obtained by the log-ratio image is relatively flat on account of the logarithmic

Algorithm , Computers & Geosciences Vol. 10, No. 2-3, pp. 191-203, 1984.

1. S. Krinidis and V. Chatzis, A robust fuzzy local information C-means clustering algorithm, IEEE Trans. Image Process., vol. 19, no. 5, pp. 13281337, May 2010.

2. W. Cai, S. Chen, and D. Zhang, Fast and robust fuzzy C-means clustering algorithms incorporating local information for image segmentation, Pattern Recognit., vol. 40, no. 3, pp. 825838, Mar. 2007.

3. L. Bruzzone and D. F. Prieto, An adaptive semiparametric and context- based approach to unsupervised change detection in multi-temporal remote-sensing images, IEEE Trans. Image Process., vol. 11, no. 4, pp. 452466, Apr. 2002.

4. A. Singh, Digital change detection techniques using remotely sensed data, Int. J. Remote Sens., vol. 10, no. 6, pp. 9891003, 1989.