A NEW UNSUPERVISED CHANGE DETECTION APPROACH BASED ON DWT IMAGE FUSION AND BACKTRACKING SEARCH OPTIMIZATION ALGORITHM FOR OPTICAL REMOTE SENSING DATA

Change detection is one of the most important subjects of remote sensing discipline. In this paper, a new unsupervised change detection approach is proposed for multi-temporal remotely sensed optic imagery. This approach does not require any prior information about changed and unchanged pixels. The approach is based on Discrete Wavelet Transform (DWT) based image fusion and Backtracking Search Optimization Algorithm (BSA). In the first step of the approach, absolute-valued difference image and absolute-valued log-ratio image is calculated from co-registered and radiometrically corrected multi-temporal images. Then, these difference images are fused using DWT. The fused image is filtered by median filter for edge information preservation and by wiener filter for image smoothing. Then, a min-max normalization is applied to the filtered data. The normalized data is clustered into two groups with BSA as changed and unchanged pixels by minimizing an objective function, unlike classical methods using CVA, PCA, FCM or K-means techniques. To show effectiveness of proposed approach, two remote sensing data sets, Sardinia and Mexico, are used. False Alarm, Missed Alarm, Total Alarm and Total Error Rate are selected as performance criteria to evaluate the effectiveness of new approach using ground truth images. Experimental results show that proposed approach is effective for unsupervised change detection of optical remote sensing data. * Corresponding author, uhatasever@erciyes.edu.tr, +90 352 2076666-32678


INTRODUCTION
The detection of changes on earth surface is one of the most important subjects of remote sensing discipline.In many applications (monitoring urban growth, disaster management, identifying the spatial difference after a forest fire, etc.), change detection based on remote sensing is vital to extract meaningful information (Cihlar et al 1992, Gopal andWoodcock 1996).Change detection is separated into two main topics as supervised and unsupervised.But in many applications, unsupervised change detection techniques are preferred due to not need any training data.Basic approaches used in change detection are change vector analysis (CVA), image differencing, principal component analysis (PCA) and image rationing.Among them, image differencing is the most common used approach in real world problems (Mishra et al. 2012).Over the last years, many unsupervised change detection methods based on differencing image have been improved such as PCA-Kmeans (Celik 2009), PCA-FCM, HCM Based (Ghosh et al. 2011), G-HCM Based (Ghosh et al. 2011), EM+MRF Based (Bruzzone and Prieto, 2000).
In some of the change detection methods, firstly, feature space is calculated from the multi-temporal images with the help of image differencing and then this feature space is divided into two clusters as changed and unchanged (Ghosh et al. 2011).In this step, performance of the clustering algorithm is very important.Over the last decade, artificial intelligence (AI) optimization algorithms have been tested for clustering.The studies show that these algorithms are more successful than the classical clustering techniques such as K-means, Fuzzy Cmeans (Zou et al. 2010, Karaboga andOzturk 2011).In this study, a new artificial intelligence optimization algorithm, BSA, has been preferred for clustering in this approach.
In this study, a new automatic change detection approach based on DWT image fusion and BSA clustering for optic images is proposed.Experimental results prove that proposed approach is effective for unsupervised change detection.

BACKTRACKING SEARCH OPTIMIZATION ALGORITHM (BSA)
Optimization is one of the important research area in engineering problems and applied mathematics (Civicioglu, 2013).This contribution has been peer-reviewed.doi:10.5194/isprsarchives-XL-7-15-2014effectiveness.BSA outperformed many algorithms such as ABC, DE in many benchmark tests and real world problems (Civicioglu, 2013).
BSA is a population-based iterative optimization algorithm.BSA has five basic steps: Selection-1, Mutation, Crossover and Selection-2.Unlike many AI algorithms, BSA has single control parameter.Also, BSA has a very simple and strong structure which can be adapted to different numerical optimization problems (Civicioglu, 2013).BSA's strategies for generating trial populations have powerful exploration capabilities.
Table .1General Structure of BSA

CHANGE DETECTION BASED ON DWT IMAGE FUSION AND BSA CLUSTERING
Firstly, two multi-temporal images must be registered to each other and if it is necessary the radiometric distortions must be corrected.Secondly, absolute-valued difference image and absolute-valued log-ratio image are calculated with equation (1) and equation (2), respectively: In equation ( 1) and ( 2), X D is absolute-valued difference image; X LR is absolute-valued log-ratio image; X T1 and X T2 are multitemporal images (Celik, 2010).
In the third step, absolute-valued difference image and absolutevalued log-ratio image are fused with Discrete Wavelet Transformation (DWT).In DWT image fusion, Discrete Wavelet Transformation is applied to X D and X LR images and wavelet coefficients of each image are calculated.After that, combined wavelet coefficients are calculated by combination rules such as min, max or rand (Yingchun et al. 2011).Lastly, fused image (X Fus ) is obtained by applying Inverse Discrete Wavelet Transformation (IDWT) to combined wavelet coefficients.In this step, db8 for decomposition, mean for approximations and min for details have been used.
In the fourth step, 2-D median filter with 3x3 window sizes and 2-D wiener filter with 11x11 window sizes are applied to X Fus , respectively .Median filter has been used to preserve edge information and to suppress the isolated pixels in X Fus .Wiener filter has been used for smoothing of X Fus without losing edge information.
After that, filtered data is normalized into [0, 1] interval in order to boost the performance of BSA.This technique, called also min-max normalization, is performed according to the equation (3): min max min (3) In equation ( 3), P new is the normalized value; P, P min and P max are the original value, minimum value and maximum value in the filtered data, respectively.
In the last step, normalized data is clustered into two classes by using BSA algorithm.Also, BSA needs an objective function for clustering.The objective function is given in equation ( 4) and ( 5): arg min (5) In equation ( 4) and ( 5), 2 is Euclidean distance, i x is i th pixel; j cent is the centre of the j th class; j d represents all of the pixels in the j th class.As can be seen in the objective function, BSA optimizes the cluster centres so that these centres provide the least minimum total distance between the cluster centres and the pixels of each cluster in each iteration.General scheme of proposed approach is shown in figure 1.In the application of Sardinia dataset and Mexico dataset, proposed approach is compared with HCM, G-HCM, PCA-FCM and PCA-K Means methods.Except PCA-K Means and PCA-FCM, results of the other methods are taken from the original papers cited in table 2 and table 3. Thus, the efficiency of the proposed approach is analyzed for optical remote sensing images.The visual results and error values obtained from ground truth map data are given in figure 3, figure 4 and table 2, table 3, respectively.As seen from table 2, proposed approach gets the best change detection accuracy in all methods for Sardinia dataset.Preferred scale factor is an inverse random number coming from a Normal distribution with 0 mean and standard deviation of 5. Population and generation number values of BSA are 10 and 100, respectively.These parameters are also have been used for Mexico dataset.
As seen from figure 4 and table 3, proposed approach can generate quite successful results, i.e.Except G-HCM method, the proposed approach outperforms other change detection methods for Mexico dataset.

CONCLUSIONS
In this paper, a new unsupervised change detection approach is proposed using DWT image fusion and BSA clustering for multi-temporal optical data.The performance of proposed approach is compared with a number of change detection techniques.The applications with Sardinia dataset (Landsat 5) and Mexico dataset (Landsat 7) show that the blending with DWT based image fusion and BSA based clustering make the approach more effective and successful.Considering the calculation time, the proposed approach is slower than other methods because it combines DWT and BSA.

Figure 1 .
Figure 1.General scheme of proposed approach

Figure 3 .
Figure 3. Change detection result of Sardinia data set, (a) Proposed approach, (b) Ground truth (a) (b)

Figure 4 .
Figure 4. Change detection result of Mexico data set, (a) Proposed approach, (b) Ground truth

Table 2 .
The quantitative results of Mexico Landsat 7 image data set.

Table 3 .
The quantitative results of Sardinia Landsat 5 image data set.