Global and Local Change Detection Using K-Means Clustering Improved by Particle Swarm Optimization

Change Detection (CD) considered as an important issue among researchers due to its applications in different aspect like urban management, environmental monitoring, and damage assessment and so on. Different methods and techniques have been proposed for CD process. One of the most common categories presented in the field of CD is supervised and unsupervised techniques. Unsupervised CD techniques are based on image information and do not require any additional information including training samples. Several studies have been done for unsupervised change detection methods. Some of the proposed algorithm are based on clustering technique and considered two cluster centers for entire image. It can be critical because changed and unchanged pixels might not be present consistent behavior at entire image so it can be conducted misclassification of changed and unchanged pixels. Another method considered clustering process at block levels of image. So changed or unchanged pixel might not be present at all block level of image simultaneously, it can be leaded misclassification of above mentioned pixels. In this paper, a novel unsupervised CD method is proposed based on K-Means clustering algorithm improved by particle swarm optimization method (PSO) to solve above mentioned problem. The proposed method comprises five main steps including: 1-preprocessing (radiometric and geometric correction), 2-generation of difference image and feature extraction (neighborhoods pixels information), 3- split of difference image into non-overlapping block (block analysis), 4-the proposed PSO-K-Means clustering and create binary change map, and 5-accuracy assessment (i.e. absolute and relative accuracy assessment). The main goal of the proposed PSO-K-Means method is an automatic detection of change area which occurred between bi-temporal remote sensing images. In the most region, there is different spectrum of changes, so the aim of proposed method is detecting the spectrum of change area at block level and also preserving global image information. To achieve the mentioned aim a novel cost function which considered K-Mean clustering at block level and at entire image simultaneously presented in this paper. To find optimum clustering centers with the minimum cost or in other word finding optimum feature vector corresponded to optimum cluster centers (changed and unchanged clusters), the PSO method was employed. Three cost function comparisons were implemented in order to verify the necessity of the proposed cost function. Moreover, maximum voting method applied in order to combine different band change maps to improve CD result. Finally, a sensitivity analysis was employed in order to confirm the validation of the proposed PSO-K-Means method. The sensitivity analysis employed against different block size, different initial population and iteration in the optimization process. The results show the stability of proposed method against initial population and iteration parameters. The proposed algorithm showed sensitivity against changing block size of image. Experiments applied on two data sets (i.e., Alaska and Uremia Lake). Both Data sets acquired by the Landsat satellite with seven spectral bands and 30 meters pixel size with the same image size (400 × 400 pixels). The ground truth image was created manually by an expert in visual analysis of the input images. The proposed method improved change accuracy 8%-12% rather than common methods, (i.e. FCM, Otsu thresholding, K-Means, K-Medoids) in both Alaska and Uremia. The optimum block size determined using experimental result.