Change Detections In Multi-temporal Satellite Images

The process of obtaining the changed information of the earth by making use of multi-temporal satellite images is called change detection. According to the level of analyzing image, the change detection algorithms can be divided into pixel level class, characteristic level one and target level one. According to the mechanism of processing data, they can be divided into supervised class and unsupervised one. The kind of the supervised change detection algorithms are based on method of supervised classifying and require training to get the parameters of network. While the kind of the unsupervised change detection algorithms generate the change map by making a comparison of bi-temporal satellite images automatically without manual operation. The proposed algorithms belong to the kind of unsupervised change detection algorithms in pixel level.An unsupervised change detection algorithm in multi-temporal satellite images based on principal component analysis and up-down-set fuzzy Kohonen clustering network is proposed. This method makes a combination of both PCA and UDSFKCN initially, and applies it to change detection. This method generates eigenvector corresponding to every pixel combining itself with its neighbors using principal component analysis. At the same time, solving the detection of the changed pixel in a region is to divide the pixel into two groups, changed class and unchanged class. Since every pixel is described as a eigenvector, therefore to obtain a changed map of the changed region in pixel level, up-down-set fuzzy Kohonen clustering network is applied to divide all the eigenvectors into changed ones and unchanged ones.An unsupervised change detection algorithm in multi-temporal satellite images based on non-sub-sampled Contourlet transform and pulse coupled neural network is proposed. This method makes a combination of both non-sub-sampled Contourlet transform and pulse coupled neural network, and applies it to change detection initially. An unsupervised multi-scale change detection algorithm in multi-temporal satellite images is also proposed. This method makes a combination of both non-sub-sampled Contourlet transform and up-down-set fuzzy Kohonen clustering network, and applies it to change detection initially. For each pixel in the log-ratio image, multi-scale and multi-direction feature vector is extracted using non-sub-sampled Contourlet transform. The final change detection map is achieved by clustering the multi-scale and multi-direction feature vectors using up-down-set fuzzy Kohonen clustering network into two classes: changed and unchanged.Through three specific change detection algorithms, summarized the change detection algorithm for general research ideas.