(PDF) A New Artificial Intelligence Optimization Method for Pca Based Unsupervised Change Detection of Remote Sensing Image Data

A NEW ARTIFICIAL INTELLIGENCE OPTIMIZATION METHOD FOR PCA BASED UNSUPERVISED CHANGE DETECTION OF REMOTE SENSING IMAGE DATA U.H. Atasever, M.H. Kesikoglu, C. Ozkan∗ Abstract: In this study, a new artificial intelligence optimization algorithm, Differ- ential Search (DS), was proposed for Principal Component Analysis (PCA) based unsupervised change detection method for optic and SAR image data. The model firstly computes an eigenvector space using previously created k × k blocks. The change detection map is generated by clustering the feature vector as two clusters which are changed and unchanged using Differential Search Algorithm. For clus- tering, a cost function is used based on minimization of Euclidean distance between cluster centers and pixels. Experimental results of optic and SAR images proved that proposed approach is effective for unsupervised change detection of remote sensing image data. Key words: unsupervised change detection, differential search, PCA Received: May 26, 2014 DOI: 10.14311/NNW.2016.26.008 Revised and accepted: March 10, 2016 1. Introduction One of the most important subjects of remote sensing discipline is to examine the temporal variation of a specific zone. The existing differences in land cover or land use can be identified by using satellite images of different periods. Change detection approaches can effectively be used for solving important problems such as observing forest and agricultural areas, identifying the changes in land cover and identifying the spatial difference after a forest fire [7, 10, 12, 14, 20]. Change detection is separated into two groups as supervised and unsupervised. But in solv- ing real world problems, unsupervised approach is more preferable than supervised approach. The reason lying under preference is that to provide the training data to be used in supervised approach is difficult while unsupervised approach does not need a training data [17]. Although there are a lot of unsupervised change detection methods for both optic and SAR remote sensing images, pre-processing steps consisting of geometric and radiometric corrections are generally needed [10, 17]. ∗ Umit Haluk Atasever, Mustafa Hayri Kesikoglu – Corresponding author, Coskun Ozkan, Department of Geomatics Engineering, Erciyes University, 38039, Kayseri, Turkey,

,

,

c CTU FTS 2016 141 Neural Network World 2/2016, 141–154 Change vector analysis (CVA), image differencing, principal component analysis (PCA), image rationing are the basic approaches used in change detection process [4, 6, 18]. Among them, image differencing is the most commonly used approach. Many unsupervised change detection methods have been improved over the last years. Some of them are EM+MRF based [2, 24], PCA and K-Means based [5], Fuzzy Local Information C-Means Clustering Algorithm (FLICM) [11], Refor- mulated FLICM (RLICM) [11], Hard C-Means (HCM), Hard C-Means hybridized with Simulated Annealing (SA-HCM), Hard C-Means Hybridized with Genetic Algorithm (GA-HCM), Fuzzy C-Means (FCM) Fuzzy C-Means Hybridized with Genetic Algorithm (GA-FCM) [9]. HCM, SA-HCM, GA-HCM and GA-FCM are generally used for optic images, FLICM and RFLICM are used for SAR images and PCA K-Means, FCM and EM+MRF can be used for both optic and radar images [2, 5, 9-11, 24]. After pre-processing, some of the change detection methods need to generate the feature space from the images of different dates by the aid of image differenc- ing or image rationing. And then this feature space is clustered into two clusters as changed and unchanged [9, 16, 21]. One of the most important components increasing the success of the operation is the performance of the clustering algo- rithm. A number of algorithms have been used with this purpose in many studies [3, 5, 9, 11]. Over the last decade, artificial intelligence (AI) optimization algo- rithms have been tested for unsupervised classification, i.e. clustering. The studies show that these algorithms have the potential to outperform the classical clustering techniques [1, 15, 25]. Although there are a number of AI optimization algorithms such as Partial Swarm Optimization (PSO), Genetic Algorithm (GA) and Artifi- cial Bee Colony (ABC) used for clustering, as a new AI optimization algorithm, Differential Search Algorithm (DS) has not been tested for clustering purpose in change detection so far. DS algorithm outperformed many conventional AI algo- rithms in many standard multimodal tests [8]. It is a very powerful algorithm with fast convergence ability. In this study, Differential Search Algorithm (DS) was used as a clustering method in an automatic PCA based unsupervised change detection approach de- veloped by Celik (2009) [5]. This change detection approach is automatic, fast and can be used in both optical and SAR remote sensing images. Its clustering phase is the most arguable and developable step because of the fact that resulting change and unchanged regions are categorized by clustering process and many dif- ferent methods exist. In this approach, after a difference image is obtained from multi-temporal images, the difference image is separated into k ×k non-overlapping blocks and transformed into a feature space by PCA. Then, a min-max normal- ization is applied to the feature space and the obtained data set is clustered into two groups with Differential Search Algorithm (DS) as changed and unchanged areas. According to this, the relationship between PCA and DS algorithm is that PCA constructs features and DS transforms these features into clusters. Finally, performance tests are done through ground truth maps of optic and SAR image data sets. 142 Atasever U.H., Kesikoglu M.H., Ozkan C.: A new artificial intelligence. . . 2. PCA Based Unsupervised Change Detection with DS (PCA-DS Approach) In the first step of the method, two multi-temporal images must be registered to each other. If it is necessary, the radiometric distortions coming from sensor, at- mosphere, etc. must be corrected at this step. Then, second step is the calculation of the difference image. Let Xtime1 and Xtime2 be the images belonging to same geographical areas in different dates and Xdiff be difference image. For optical images, Xdiff is calculated by subtracting the image density values of Xtime1 from Xtime2 and then absolute value of Xdiff is calculated [2, 5]. The equation for optical images is defined as Xdiff = |Xtime2 − Xtime1 | . (1) For SAR images, Xdiff is calculated by subtracting the logarithmic values of Xtime2 and Xtime1 then absolute value of Xdiff is calculated. The equation for SAR images is defined as [3] Xtime2 Xdiff = log = |log(Xtime2 ) − log(Xtime1 )| , (2) Xtime1 where log is natural logarithm. In the third step, the difference image Xdiff is separated into k × k blocks in a lexicographical order. Let this obtained matrix be Mpdiff . Eigenvectors of the Mpdiff matrix are generated with PCA. In Mpdiff , p is an index value with 1 ≤ p ≤ S b(R × C)/(k × k)c , (3) where S is the number of column vector sets in Mpdiff , R is the number of rows of Xdiff , C is the number of columns of Xdiff . Since it has shown that k = 3 generates more successful results [5], k value is set as 3 in this study. A mean vector (Φ) is calculated from Mpdiff along the columns as S 1X p Φ= M . (4) S p=1 diff After that, this mean vector is subtracted from each column vector of Mpdiff matrix ∆p = Mpdiff − Φ, (5) where ∆p difference vector set. Then, covariance matrix (CM) is calculated with matrix algebra as S 1X CM = ∆p (∆p )T , (6) S p=1 where ()T is transposing process. Eigenvalues (λi ) and corresponding eigenvectors (ei ) of the covariance matrix (CM) are computed. Then the related eigenvectors for reducing of effective dimension are selected by Cumulative Percent Variance 143 Neural Network World 2/2016, 141–154 (CVP) selection criteria which are subjective but highly effective selection criteria for real data sets [23] as follows: Pn ! i=1 λi CVP(n) = 100 Pk×k . (7) i=1 λi In CVP, the eigenvectors are sorted in descending order in terms of eigenvalues and the first bundle of eigenvectors that provides a reasonable percent (e.g. 90%, 95% or 99%) is selected which reduces the effective dimension. These selected eigenvectors are used to project pixel values in Mdiff (x, y) into feature vector space as   v1  v2    v(x, y) =  v3  , (8)    ..   .  vσ where 1 ≤ n ≤ k 2 , vσ = (ei )T (Mdiff (x, y) − Φ), 1 ≤ i ≤ n, 1 ≤ x ≤ R, 1 ≤ y ≤ C and n is the number of eigenvectors used for projecting Mdiff (x, y) onto eigenvector space. The later step is the normalization of the feature vector space v(x, y) into [0, 1] interval in order to boost the performance of DS optimization. This technique, called also min-max normalization, is performed according to the following equa- tion: ϕ − ϕmin ϕnew = . (9) ϕmax − ϕmin In Eq. (9), ϕnew is the normalized value; ϕ, ϕmin and ϕmax are the original value, minimum value and maximum value in the feature space, respectively. The tests of this study show that this type of normalization increases the effectiveness of DS in terms of convergence speed. In the last step, feature vector space which min-max normalization has been done is clustered into two classes by using DS algorithm. As in all of the optimiza- tion methods, DS algorithm also needs an objective function for minimization. The objective function used in this study is given as cj = xi ; min(kxi − cntj k2 ) , (10) arg min X kcj − cntj k2 . (11) cntj In Eqs. (10) and (11), kk2 is Euclidean distance, xi is i-th pixel, cntj is the center of the j-th class, cj represents all of the pixels in the j-th class. As can be seen in the objective function, DS optimizes the cluster centers so that these centers provide the least minimum total distance between the cluster centers and the pixels of each cluster in each iteration. After the clustering, change detection map (CDM) is ready. Since DS algorithm initializes the cluster centers randomly, the cluster labels of change and unchanged regions can vary from simulation to simulation, i.e. the change cluster doesn’t 144 Atasever U.H., Kesikoglu M.H., Ozkan C.: A new artificial intelligence. . . always obtain the same label such as 1 or 2. In order to solve this problem, a simple label correction process can be done as CDM m1 = 2 and m2 = 1 CDM = , (12) (CDM × (−1)) + 3, m1 = 1 and m2 = 2 where m1 and m2 are label values of pixels in changed and unchanged areas, re- spectively. One of the most important characteristics of the proposed approach is a PCA block based approach: Firstly, the difference image prepared and then pixel blocks with size of k×k are used instead of individual pixels. These blocks are transformed into a new feature space by PCA. By this way, local consistency can be improved upon classical approaches. Another prominent feature of the approach is that a new evolutionary optimization DS algorithm is used for unsupervised classification instead of Fuzzy C-means or K-Means. Thus, the potential of DS algorithm in the change detection applications of remotely sensed images can be revealed. 3. Differential Search (DS) DS is a population-based iterative meta-heuristic optimization algorithm. It was developed by Civicioglu to solve optimization problems having real values [8]. DS algorithm simulates Brownian-like random walk action used by a superorganism with the purpose of migrating. As the fertility of food areas changes depending on climate, a lot of species migrate to more fertile fields. A lot of species such as birds, fire ants and bees have seasonal migration cycle. Migrating species generate superorganisms involving a number of individuals. After that, superorganisms start to change the location towards more productive areas. The movement of the superorganisms can be explained by Brownian-like random walk [8]. Many of predatory species control the fertility of that habitat before migrating to a new habitat. If the potential of this habitat is enough for the needs of ten super organisms, it settles in this new area temporarily and goes on migration with goal of finding more productive regions. The population consisting of candidate solutions in DS algorithm means a migrating superorganism. This superorganism selects the good locations as stop-over site in habitat. Then the superorganism clans that find a good accommodation area migrate there. The number of artificial organisms (clans) (i, e., Xi , i = {1, 2, 3, . . . , N }) form- ing superorganism (i.e., Superorganism g, g = 1, 2, 3, . . . , maxgen) is equal to the dimension related optimization problem. In DS algorithm a member of superor- ganism’s initial position is calculated by using following equation: xi,j = rand(up lim −low lim) + low lim, (13) j j j where N, maxgen, uplimj and lowlimj values are population number, generation number and habitat boundaries defined for j-th elements of clans, respectively [8]. In DS algorithm, stopover site search process can be explained with a sim- ple Brownian-like random walk model [22]. Randomly selected individuals from the clan in limited number move on from important parameters to their donor = 145 Neural Network World 2/2016, 141–154 XRand−Shuffling(i) targets for the success of migration process so as to find stop- over sites. Rand-Shuffling function is used for changing the placement of the com- ponents in Xi set. The changes in the locations of artificial organisms are identified by using scale factor. The scale factor is obtained by gamma random number gen- erator which is controlled by uniform number generator. The operating range of this number generator is [0,1]. Besides that, artificial superorganism can move in many different directions with this kind of tip number generator. A new stop-over site in DS is calculated as [8]: StopoverSite = (SOrganism + SFactor) × (Donor − SOrganism), (14) where SOrganism is a superorganism and SFactor is a scale factor. In Differential Search Algorithm, three different stop-over site producing mech- anisms are defined as Bijective DS, Surjective DS and Elitist DS. In Bijective DS, every clan searches its habitat by going to a different clan. Bijective DS is suitable for using multi-modal problems. In Surjective DS, a clan has the tendency of going another clan existing in a more productive location. It can be preferred for partly multimodal and mostly unimodal problems. Elitist DS makes all the clans go to the clan presenting in the most productive area. It is a quite fast mechanism and can be used especially for unimodal problems. In addition to this, there are dif- ferent defined random number generators which can be used for generating Scale Factor. In this study, Standard normal distributed random number generator was preferred. Which individuals in artificial organism generating superorganism will join the searching is determined at the end of a random process. The parameters of DS algorithm are p1 , p2 , stop over site production mechanism and scale factor. These parameters effect the convergence speed of DS for calculating the class cen- ters. p1 and p2 control the mutation variable: p1 is used to determine the type of mutation strategy randomly and p2 defines the variation for the mutation strategy. The experimental tests confirm that the optimum value for p1 and p2 was 0.3×rand [8]. Stop over site production mechanism determines the search direction of DS algorithm: Elitist DS focuses on the best solution in each generation, Bijective DS searches randomly all over the solution space; Surjective DS partially focuses on the best solution. As given in Experimental study-II section, Bijective DS gives the best results for all datasets. Scale factor is a parameter which is updated according to the type of the random number generator in each generator and aims to accelerate the convergence speed by searching solution space by multiplying with mutation variable. Although there are 6 different scale factor production mechanisms, the tests show that function of R = 1/normrnd(0, 5) gives the best results for all data sets. 4. Datasets, Experimental Results and Discussions In this paper, both SAR (Bern and Ottawa dataset) and optic (Sardinia and Mex- ico dataset) images are used. All datasets were downloaded from Pudn website (Accessed 12 September 2013). In order to remove the speckle noise, SAR images were filtered by the Enhanced Lee Filter (5 × 5 windows). The fourth band of 146 Atasever U.H., Kesikoglu M.H., Ozkan C.: A new artificial intelligence. . . the multispectral images was selected for change detection process because of the reflectance characteristics of water, land and vegetation objects [13]. Moreover, a ground truth map was employed to calculate the performance [3-5, 10, 21]. The measures below were used for quantitative analysis of developed approach: – False Alarm (FA): Number of unchanged pixels identified as changed pixels. – Missed Alarm (MA): Number of changed pixels identified as unchanged pixels. – Total Error (TE): Total number of incorrect categorized, which is sum of FA and MA. – Total Error Rate (TER): Ratio of sum of FA and MA to number of image pixels (PN). (TER = (FA+MA)/PN×100) 5. Experimental Study-I In this section, Sardinia (Landsat 5 TM) dataset and Bern dataset were used. Sardinia image set belongs to Mulargia Lake in Sardinia Island from 1995 and 1996 years. In Sardinia data, the water level of Mulargia Lake increased between two dates and its nearby sites were flooded. The image sizes are 300 × 412 pixels. Because the histograms of multi-temporal Landsat 5 images are very similar to each other, any radiometric correction was not applied. Sardinia dataset and ground truth image are shown in Fig. 1. The SAR image set is a C-band ERS-2 dataset of Bern, Switzerland obtained by European Remote Sensing (ERS)-2 satellite in April and May in 1999. In Bern data, The Aare River overflowed and submerged the partial part of Bern city. The image sizes are 301 × 301 pixels. The filtered Bern dataset and ground truth change detection map is shown in Fig. 2. In the application of Sardinia Landsat 5 image data, PCA-DS approach was compared with HCM, EM+MRF, PCA-KMeans, FCM, SA-HCM, G-HCM and G- FCM methods. Except PCA-K Means, results of the other methods were taken from the original papers cited in Tab. I. Thus, the efficiency of the PCA-DS ap- proach was analyzed for optic images. The visual results and error values obtained from ground truth map data are given in Fig. 3 and Tab. I, respectively. As seen from Tab. I, PCA-DS approach yielded the best change detection accuracy in all Fig. 1 Multitemporal Landsat 5 TM images of Mulargia (Band 4): (a) September 1995, (b) September 1996, (c) Ground truth. 147 Neural Network World 2/2016, 141–154 Fig. 2 Multitemporal SAR Images of Bern: (a) April 1999, (b) May 1999, (c) Ground truth. Fig. 3 Change detection result of Sardinia data set, (a) PCA-DS approach, (b)Error Map (False and Missed Alarms) of PCA-DS, (c) Ground truth. False Missed Total Total Change Detection Method Alarm Alarm Error Error Rate HCM [9] 275 4133 4408 3.56 EM+MRF(β = 1.7) [19] 3856 289 4145 3.35 SA-HCM [9] 425 2727 3152 2.55 G-HCM [9] 3006 132 3138 2.53 FCM [9] 494 2246 2740 2.21 G-FCM [9] 627 1983 2610 2.11 PCA-Kmeans Based [5] 1603 832 2435 1.97 Proposed Approach 744 1405 2159 1.73 Tab. I The quantitative results of Sardinia Landsat 5 TM image dataset. methods. These results were obtained from DS with Bijective stop-over site pro- duction mechanism. In this mechanism, the scale factor is an inverse random num- ber coming from a normal distribution with 0 mean and standard deviation of 5. 148 Atasever U.H., Kesikoglu M.H., Ozkan C.: A new artificial intelligence. . . Population and generation number values were 10 and 500, respectively. These pa- rameters were also used for SAR application. However, PCA-DS approach yielded better results than other methods for Sardinia dataset. In the application of ERS-2 SAR image data of Bern, PCA-DS approach was compared with PCA-KMeans, FCM, FLICM and RFLICM methods. Except PCA- KMeans, the results of other the methods were taken from the original papers cited in Tab. II. Quantitative change detection results are listed in Tab. II. As seen from Fig. 4 and Tab. II, PCA-DS approach could generate quite successful results, i.e. except RFLICM method, PCA-DS outperforms other change detection methods for SAR data. Fig. 4 Change detection result of Bern data set. (a) Proposed approach, b) Error Map (False and Missed Alarms) of PCA-DS, (c) Ground truth. False Missed Total Total Change Detection Method Alarm Alarm Error Error Rate FCM [11] 507 61 568 0.626 FLICM [11] 137 169 306 0.337 PCA-Kmeans Based [5] 146 158 304 0.335 RFLICM [11] 133 159 292 0.322 Proposed Approach 160 133 293 0.323 Tab. II The quantitative change detection results of Bern dataset. 6. Experimental Study-II In this section, Mexico and Ottawa image set was used for experimental studies. Also proposed approach was examined in point of DS parameters in this exper- imental study. The Ottawa image set is two SAR images (290x350 pixels) over 149 Neural Network World 2/2016, 141–154 Ottawa acquired by the Radarsat SAR sensor at Ottawa, Canada in July 1997 and August 1997. Ottawa data consists of before and after images of flood occurred in 1997 by Red River. The Mexico data set belongs to an area in Mexico acquired by Landsat 7 ETM+ from April 2000 and 2002. In Mexico data, the fire occurred between two dates destroyed the vegetation in a wide region and its environmental effect was imaged in 2002. In the applications of Ottawa and Mexico image sets, PCA-DS approach was compared with PCA-KMeans which can generate good results for both optic and SAR images. Quantitative change detection results and computation times are listed in Tab. III. As seen from Tab. III and Fig. 5 and 6, proposed approach yielded better results for Ottawa and Mexico image sets. From a computational burden point of view, PCA-DS approach is more complex and certainly slower PCA-Kmeans method because of clustering phase takes longer for all applications since DS is not a built-in function as Kmeans in Matlab as seen in Tab. III. Change Detection False Missed Total Total Comp. Method Alarm Alarm Error Error Rate Time [s] Ottawa PCA-Kmeans [5] 583 1901 2484 0.024 0.7 Image set PCA-DS 856 1574 2430 0.023 26.68 Mexico PCA-Kmeans [5] 1024 3662 4686 0.018 1.38 Image set PCA-DS 1589 2428 4017 0.015 74.76 Tab. III The quantitative results of Ottawa and Mexico dataset. Because of Differential Search Algorithm’s and dataset structure; accuracy can vary depending on stop-over site mechanism and scale factor. To examine ef- fectiveness of stop-over site producing mechanism and scale factor, PCA-DS ap- proach was run with Ottawa image set with each mechanism and scale factor and best results are shown in Tab. IV and Tab. V (Population and generation number values are 10 and 200, respectively.). In this study, the best accuracy for all datasets was obtained by Bijective DS for stop-over site mechanism and R = 1/normrand(0, 5) for scale factor. Tab. V illustrates effect of different scale factors: R = lognrnd(rand, 5 × rand) returns an array of random numbers gener- ated from the lognormal distribution with rand mean and standard deviation of Stop-over Site Total Total Producing Mechanism Error Error Rate Surjective 3339 0.033 Elitist 3274 0.032 Bijective 2430 0.023 Tab. IV The quantitative results of Ottawa dataset in point of stop-over site pro- ducing mechanism. 150 Atasever U.H., Kesikoglu M.H., Ozkan C.: A new artificial intelligence. . . Fig. 5 Multitemporal Radarsat images of Mexico: (a) April 2000, (b)May 2002, (c) Result of PCA-DS Approach, (d) Error Map (False and Missed Alarms) of PCA-DS (e) Result of PCA-Kmeans Approach, (f ) Ground Truth. Fig. 6 Multitemporal Radarsat images of Ottawa: (a) July 1997, (b) August 1997, (c) Result of PCA-DS Approach,(d) Error Map(False and Missed Alarms) of PCA- DS (e) Result of PCA-Kmeans Approach, (f ) Ground Truth. 151 Neural Network World 2/2016, 141–154 Scale factor Total Error Total Error Rate R = lognrnd(rand, 5 × rand) 4424 0.043 R = 1/gamrnd(1, 0.5) 3893 0.038 R = 1/normrnd(0.5, 0.5) 3675 0.036 R = 4 × randg 3225 0.031 R = 4 × randn 2737 0.027 R = 1/normrnd(0, 5) 2430 0.023 Tab. V The quantitative results of Ottawa dataset in point of scale factor. 5 × rand; R = gamrnd(1, 0.5) generates random numbers from the gamma distribu- tion with shape parameters in 1 and scale parameters in 0.5; randg returns a scalar random value chosen from a gamma distribution with unit scale and shape; randn returns a pseudorandom scalar drawn from the standard normal distribution; rand returns a pseudorandom scalar drawn from the standard uniform distribution on the open interval (0,1) [8]. 7. Conclusions The clustering, i.e. unsupervised classification, has an important role in change detection analysis of remotely sensed images as well as computer vision applica- tions. Since any priori thematic information does not exist, different approaches in clustering produce different clusters. So, in this study a new artificial intelligence optimization method, Differential Search, was proposed for clustering in change de- tection analysis of both optic and SAR remote sensing images. DS algorithm was blended with an automatic PCA based unsupervised change detection method, i.e. PCA-DS approach. The performance of PCA-DS approach was compared with a number of change detection techniques. The applications with Sardinia, Bern, Ottawa and Mexico data sets showed that the blending with DS based clustering made the approach more effective and successful. Also proposed approach was examined in point of stop-over site production mechanism and scale factors. The optimum values of DS parameters for both optical and Radar data sets were de- termined and reported in the study. DS relatively needs less parameters and it is a robust algorithm. With the help of quantitative results, it can be concluded that DS algorithm had a very promising potential as a clustering method in PCA based change detection. Acknowledgement This work was supported by Scientific Research Projects Department of Erciyes University [FBA-2013-4643]. 152 Atasever U.H., Kesikoglu M.H., Ozkan C.: A new artificial intelligence. . . References [1] ATASEVER U.H., CIVICIOGLU P., BESDOK E., OZKAN C. A New Unsupervised Change Detection Approach Based On DWT Image Fusion And Backtracking Search Optimization Algorithm For Optical Remote Sensing Data. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. 2014, XL-7, pp. 15–18. doi: 10.5194/isprsarchives-XL-7-15-2014. [2] BRUZZONE L., PRIETO D.F. Automatic analysis of the difference image for unsupervised change detection. IEEE Transactions on Geoscience and Remote Sensing. 2000, 38, pp. 1171–1182, doi: 10.1109/36.843009. [3] CELIK T. Change Detection in Satellite Images Using a Genetic Algorithm Approach. IEEE Geoscience and Remote Sensing Letters. 2010, 7, pp. 386–390, doi: 10.1109/LGRS.2009. 2037024. [4] CELIK T. Multiscale Change Detection in Multitemporal Satellite Images. IEEE Geoscience and Remote Sensing Letters. 2009, 6, pp. 820–824, doi: 10.1109/LGRS.2009.2026188. [5] CELIK T. Unsupervised Change Detection in Satellite Images Using Principal Component Analysis and k-Means Clustering. IEEE Geoscience and Remote Sensing Letters. 2009, 6, pp. 772–776, doi: 10.1109/LGRS.2009.2025059. [6] CHEN J., CHEN X., CUI X., CHEN J. Change Vector Analysis in Posterior Probability Space: A New Method for Land Cover Change Detection. IEEE Geoscience and Remote Sensing Letters. 2011, 8, pp. 317–321, doi: 10.1109/LGRS.2010.2068537. [7] CIHLAR J., PULTZ T.J., GRAY A.L. Change detection with synthetic aperture radar. International Journal of Remote Sensing. 1992, 13, pp. 401–414, doi: 10.1080/ 01431169208904045. [8] CIVICIOGLU P. Transforming geocentric cartesian coordinates to geodetic coordinates by using differential search algorithm. Computers & Geosciences. 2012, 46, pp. 229–247, doi: 10. 1016/j.cageo.2011.12.011. [9] GHOSH A., MISHRA N.S., GHOSH S. Fuzzy clustering algorithms for unsupervised change detection in remote sensing images. Information Sciences. 2011, 181, pp. 699–715, doi: 10. 1016/j.ins.2010.10.016. [10] GHOSH S., BRUZZONE L., PATRA S., BOVOLO F., GHOSH A. A Context-Sensitive Technique for Unsupervised Change Detection Based on Hopfield-Type Neural Networks. IEEE Transactions on Geoscience and Remote Sensing. 2007, 45, pp. 778–789, doi: 10. 1109/TGRS.2006.888861. [11] GONG M., ZHOU Z., MA J. Change Detection in Synthetic Aperture Radar Images based on Image Fusion and Fuzzy Clustering. IEEE Transactions on Image Processing. 2012, 21, pp. 2141–2151, doi: 10.1109/TIP.2011.2170702. [12] GOPAL S., WOODCOCK C. Remote sensing of forest change using artificial neural net- works. IEEE Transactions on Geoscience and Remote Sensing. 1996, 34, pp. 398–404, doi: 10.1109/36.485117. [13] HAO M., ZHANG H., SHI W., DENG K. Unsupervised change detection using fuzzy c-means and MRF from remotely sensed images. Remote Sensing Letters. 2013, 4, pp. 1185–1194, doi: 10.1080/2150704X.2013.858841. [14] HUANG X., FRIEDL M. A. Distance metric-based forest cover change detection using MODIS time series. International Journal of Applied Earth Observation and Geoinforma- tion. 2014, 29, pp. 78–92, doi: 10.1016/j.jag.2014.01.004. [15] KARABOGA D., OZTURK C. A novel clustering approach: Artificial Bee Colony (ABC) algorithm. Applied Soft Computing. 2011, 11, pp. 652–657, doi: 10.1016/j.asoc.2009.12. 025. [16] MA W., JIAO L., GONG M., LI C. Image change detection based on an improved rough fuzzy c-means clustering algorithm. International Journal of Machine Learning and Cybernetics. 2013, 5, pp. 369–377, doi: 10.1007/s13042-013-0174-4. [17] MISHRA N.S., GHOSH S., GHOSH A. Fuzzy clustering algorithms incorporating local in- formation for change detection in remotely sensed images. Applied Soft Computing. 2012, 12, pp. 2683–2692, doi: http://dx.doi.org/10.1016/j.asoc.2012.03.060. 153 Neural Network World 2/2016, 141–154 [18] PACIFICI F., FRATE F.D., SOLIMINI C., EMERY W.J. An Innovative Neural-Net Method to Detect Temporal Changes in High-Resolution Optical Satellite Imagery. Ieee Transac- tions on Geoscience and Remote Sensing. 2007, 45, pp. 2940–2952, doi: 10.1109/TGRS. 2007.902824. [19] PATRA S., GHOSH S., GHOSH A. Unsupervised Change Detection in Remote-Sensing Images Using Modified Self-Organizing Feature Map Neural Network. In: Proceedings of the International Conference on Computing: Theory and Applications, (ICCTA ’07), Kolkata, India. IEEE, 2007, pp. 716–720, doi: 10.1109/ICCTA.2007.128. [20] SCHROEDER T.A., WULDER M.A., HEALEY S.P., MOISEN G.G. Detecting post-fire salvage logging from Landsat change maps and national fire survey data. Remote Sensing of Environment. 2012, 122, pp. 166–174, doi: 10.1016/j.rse.2011.10.031. [21] SUBUDHI B.N., BOVOLO F., GHOSH A., BRUZZONE L. Spatio-contextual fuzzy cluster- ing with Markov random field model for change detection in remotely sensed images. Optics & Laser Technology. 2014, 57, pp. 284–292, doi: 10.1016/j.optlastec.2013.10.003. [22] TRIANNI V., TUCI E., PASSINO K.M., MARSHALL J.A.R. Swarm Cognition: an interdis- ciplinary approach to the study of self-organising biological collectives. Swarm Intelligence. 2011, 5, pp. 3–18, doi: 10.1007/s11721-010-0050-8. [23] VALLE S., LI W., QIN S.J. Selection of the Number of Principal Components: The Variance of the Reconstruction Error Criterion with a Comparison to Other Methods. Industrial & Engineering Chemistry Research. 1999, 38, pp. 4389-4401, doi: 10.1021/ie990110i. [24] YA-QIU J. Change detection of enhanced, no-changed and reduced scattering in multi- temporal ERS-2 SARimages using the two-thresholds EM and MRF algorithms. In: Proceed- ings of the 2005 IEEE International Geoscience and Remote Sensing Symposium (IGARSS ’05). IEEE, 2005, 6, pp. 3994–3997, doi: 10.1109/IGARSS.2005.1525789. [25] ZOU W., ZHU Y., CHEN H., SUI X. A Clustering Approach Using Cooperative Artificial Bee Colony Algorithm. Discrete Dynamics in Nature and Society. 2010, 2010, pp. 16, doi: 10. 1155/2010/459796. 154