Multi-temporal Satellite Images Change Detection Algorithm Based on NSCT

(1)

Procedia

Engineering

www.elsevier.com/locate/procedia

2011 International Conference on Advances in Engineering

Multi-temporal Satellite Images Change Detection Algorithm

Based on NSCT

Wei Cui

a

, Zhenhong Jia

a*

, Xizhong Qin

a

, Jie Yang

b

, Yingjie Hu

c

,a*

a_{College of Information Science and Engineering, Xinjiang University, Urumqi 830046, China;} b_{Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University, Shanghai 200240, China;} c_{Knowledge Engineering and Discovery Research Institute, Auckland University of Technology, Auckland 1020, New Zealand}

Abstract

In order to get the change detection image.An unsupervised change detection algorithm for multi-temporal satellite image based on NSCT（non-subsampling contourlet transform）and k-means clustering is proposed in this paper. For each pixel in the log-ratio image, multi-scale and multi-direction feature vector is extracted by NSCT and the reconstruction of the log-ratio image is obtained. The threshold is produced by using the k-means clustering algorithm and can distinguish between the unchanged and the change region. Finally, the change detection map is achieved. Some satellite images are used to verify the proposed method and the results shows that it has a higher stability and accuracy against Gaussian and speckle noise than traditional algorithms.

Selection and/or peer-review under responsibility of ICAE2011.

Keywords: NSCT；k-means clustering；multi-temporal satellite images；change detection

1. Introduction

Change detection technique is given with the same target multiple regions or single-phase or multi-band multi-band remote sensing images using image processing methods to detect whether changes in surface features of the region, and through a comprehensive multi-temporal remote sensing images analysis of change information can be extracted to achieve the remote monitoring of a comprehensive technology[1]. The technology can be divided into different levels according to the analysis level: pixel level, feature level and target level. It can also be divided into supervised classification algorithm and un-supervised classification algorithm according to classification process [2]. A change detection algorithm of un-supervised and pixel level is adopted in this paper.

Recently, changed detection technology of remote sensing image is developing rapidly, and variety of change detection algorithms are proposed by some researchers [3-8]. For example, sub-band decomposition NSCT(non-subsampling contourlet transform) coefficient selection algorithm [4],

* Corresponding author. Zhenhong Jia. Tel.: +086+0991+8582028. E-mail address: [email protected].

Open access under CC BY-NC-ND license.

(2)

NSCT shift invariance of the image texture extraction algorithm [6], PCA and k-means clustering combined change detection algorithm [7], dual-threshold wavelet value change detection algorithm [2], UDWT and k-means clustering combined change detection algorithm [8].

The NSCT and k-means clustering [7, 8] are combined in the proposed change detection algorithm. The algorithm can well inhibit transformation NSCT remote sensing images of the speckle noise [9], and can keep more edge and detail information [10]. And at the same time, the k-means clustering algorithm classifies the multi-scale and multi-direction texture points into change classes and un-change class. Finally, we get the change detection of the image. Experimental results show that the proposed algorithm has higher noise immunity and detection accuracy than the other algorithms mentioned in this paper.

2. Theoretical description and algorithms introductions 2.1 Theoretical description

NSCT is a redundant transform which consists of NSP (nonsubsampled pyramid) and NSDFB (nonsubsampled directional filter bank). NSCT inherits the contourlet’s features, for example, multi-scale, multi-direction and shift invariance [11]. Besides, it is a flexible methods MSD-based and owns more strength expression ability of edge information than wavelets or contourlet transform. Firstly, source image is decomposed into different sub-bands by NSP. Secondly, these sub-bands are filtered by NSDFB to get different scales and directions coefficients [12]. It avoids the frequency aliasing by removing up-sampling and down-up-sampling in the process of decomposition and reconstruction which emerges in contourlet or wavelets. Therefore, it satisfies perfect reconstruction conduction and shift invariance.

The threshold value is determined by k-means clustering algorithm. In this algorithm, the parameter k is 2 and class n is 2. Firstly, two objects are selected as initial cluster centers. Secondly, the distance is calculated from the object to .these centers. Lastly, according to the minimum distance re-classification of the corresponding objects updates the parameters in turn to each iteration cluster until no changing. Ultimately, changing region and are an un-changing region are got namely the detected image.

The advantages of NSCT in extracting feature information and k-means clustering algorithm are combined in this paper to get good effect.

2.2 Algorithm introductions

Figure 1 is the flow chart of the change detection algorithm which combines NSCT with k-means clustering algorithm. The concrete realization steps of the algorithm are as follows:

(3)

Fig. 1.Proposed unsupervised change detection algorithm.

 Get the remote sensing images of different time phases. The remote sensing images obtained in different time t1 and t2 are X1 and X2, whose size are I×J.

 The different time phase remote sensing images are preprocessed. The pre-processing consists of the radiation, the geometric correction and the registration of the remote sensing images at different time phases.

 Compared the image pixel of the remote sensing images named X1 and X2 one by one for operations,

then take the logarithm of both sides, to be log-ratio images, namely:

XR L( , )i j  lo g₁₀



X₂( , ) /i j X i j₁( , )



（1）  The multi-scale and multi-orientation coefficients are gotten by the NSCT.





D S R L

X  N S C T X （2） Then get a set of images with the same size as the original image.



0 _, 1 _, 2 _, 3 _, 4 _, 5 _, 6 _, 7 _, 8



D S R L R L R L R L R L R L R L R L R L

X  X X X X X X X X X （3）

Reconstruct set of images of XDS, then get the specific value image XF.





F D S

X  resh ap e X （4） Using the K-MEANS algorithm to cluster the NO.2 class for the specific value logarithm image XF, to

determine the threshold.



F



T  K M E A N S X （5）  Get change image.

c F 1 F 2 u 2 F 1 X (i,j) T o r X (i,j) T ( , ) T X (i,j) T F i j         _ _  ，，，（6） Produce a different image. In which, L is the height of the specific value logarithm image, W is the width of specific value logarithm image,_i

_

_1,_L

_

, j



1,W



The specific value logarithm image

formed by two types of XRL 

 

i j, , one is the class changed, the other is the class not.

3. Experimental results and analysis

In order to verify rightness of the algorithm, two groups of remote sensing images are adopted to compare the algorithm which the PCA-KMEANS algorithm proposed by dissertation [7]. Figure (a) and (b) in figure 2 have four small map in all. Figure 1 is a reference, figure 2 is a detection map, figure 3 is the change detection algorithm result map of PCA-KMEANS algorithm proposed by dissertation [7]. Figure 4 is the proposed change detection algorithm result in this paper. Using algorithm (a4, b4) and comparison algorithm (a3, b3) to detect the changes of images (a1, a2, b1, b2) at the same area, but different time phases, from the visual point of view, this algorithm is better. Thus, the detection accuracy of the algorithm in this paper is better than the comparison algorithm.

a1 a2 a3 a4 (a)

(4)

b1 b2 b3 b4 (b)

Fig.2 Comparison of change detection results for different approaches.

In order to verify the detection performance of the algorithm, this paper chooses the anti-noise performance as an indicator. For a size L×W image I, add Gaussian noise and the speckle noise ,then get the image

I

, then the peak signal to noise ratio of two images is defined as:

1 0 2 i 1 1 2 5 5 2 5 5 P S N R = 1 0 l o g ( ) ( ( , ) ( , ) ) L W j L W I i j I i j          （7） The anti-noise performance of the detection algorithm can be measured by the difference between the change detection results F1 and F2. For the two input images I1and I2, add noise to I1,get

I3. Input I1and I2, get F1; input I3and I2, get F2.The anti-noise performance of the detection algorithm can

be determined by the following formula [7]:

1 2 i 1 1 ( , ) ( , ) 1 L W j F i j F i j L W  _ _       │ │_（8） 20 25 30 35 40 45 50 0.85 0.9 0.95 1 PSNR(dB)



PCAKMeans-Based NSCTKMeans-Based 20 25 30 35 40 45 50 0.993 0.994 0.995 0.996 0.997 0.998 0.999 1 1.001 PSNR(dB)



PCAKMeans-Based NSCTKMeans-Based

Fig.3 Comparisons against Gaussian noise in different levels. Fig.4 Comparisons against Speckle noise in different levels.

Figure3 and Figure 4 show that the anti-noise performance of the proposed algorithm is better than PCA-KMEANS algorithm in Gaussian noise and speckle noise.

4. Summarize

This is the first time to combine the NSCT (non-sampling contourlet transform) with k-means clustering algorithm, and proposes an unsupervised multi-temporal remote sensing image change detection algorithm. The NSCT is used to extract the texture of each pixel in the specific value logarithm image in multi-scale and multi-direction and reconstruct the image by multi- direction and multi-scale coefficients. The k-means clustering algorithm is used to generate the threshold and classify the reconstructed images. At the end, we get two classes, namely, changed area and unchanged area. The change detection image is got by the proposed method. The experimental result shows that the proposed algorithm has a higher noise immunity and detection accuracy compared with the other dissertation [7] in this paper.

(5)

Acknowledgements

We gratefully thank the financial support by International Cooperative Research Project of the Ministry of Science and Technology of the P. R. China (Grant number: 2009DFA12870) and the Ministry of Education of the People's Republic of China (Grant number: 2010-1595).

References

[1] Zhen Mei. Journal of Computer Applications.Journal of Computer Applications.vol.29, 2009, pp. 2402-2405.

[2] Shiqi Huang. Daizhi Lu, Mingxing Hu, Shicheng Wang. Multi-temporal SAR Image Change Detection Technique Based on Wavelet Transform, Acta Geodaeticaet Cartographica Sinica. vol .39, 2010, pp.180-186.

[3] Xiaohua Zhang, Le Wang, LC Jiao. An Unsupervised Change Detection based on Clustering combined with Multiscale and Region Growing. IEEE Geoscience and Remote Sensing Letters.vol.11, 2011, pp.78-82.

[4] Qiang Zhang, Baolong Guo. Multifocus image fusion using the nonsubsampled contourlet transform .Signal Processing. vol.89, 2009, pp.1334–1346.

[5] Xiaojun Wang. Multisensor Image Edge Detection Based on Nonsubsampled Contourlet Transform. Modern Electronics Technique. vol.34 ,2011, pp.197-199.

[6] G.Y.Chen, B.Kegl. Invariant pattern recognition using contourlets and AdaBoost. Pattern Recognition. vol.43,2010, pp.579 -583.

[7] Turgay Celik. Unsupervised Change Detection in Satellite Images Using Principal Component Analysis and k-Means Clustering. IEEE Geoscience and Remote Sensing Letters. vol.6, 2009, pp.772-776.

[8] Turgay Celik. Multiscale Change Detection in Multitemporal Satellite Images. IEEE Geoscience and Remote Sensing Letters. vol.6,2009, pp.820-824.

[9] F.Argenti, T.Bianchi, G.M. Scarfizzi, L.Alparone. LMMSE and MAP estimators for reduction of multiplicative noise in the nonsubsampled contourlet domain. Signal Processing. vol.89, 2009, pp.1891-1901.

[10] Arthur L. da Cunha, Jianping Zhou, and Minh N. Do. The Nonsubsampled Contourlet Transform: Theory, Design, and Applications. IEEE Transactions on Image Processing. vol.15,2006, pp.3089-3101.

[11] Xingmiao Liu, Shicheng Wang, Jing Zhao, Zhiguo Liu, Tai-yang Liu. Small Infrared Target Detection Based on Nonsubsampled Contourlet Transform. Infrared (monthly). vol.32,2011, pp.35-40.

[12] Turgay Celik and KaiKuang Ma. Unsupervised Change Detection for Satellite Images Using Dual-Tree Complex Wavelet Transform. IEEE Transactions on Geo Science and Remote Science. vol.48,2010, pp.1199-1210.