Research On Change Detection Methods For Multitemporal SAR Images

With the increasingly rapid development of techniques and tools of acquiring Synthetic Aperture Radar(SAR)images and mass accumulation of SAR image data,change detection technique is more and more widely used in many areas and is playing a more important role.Therefore,it is significant to do further research on the change detection and discrimination algorithms.The main research work in the dissertation is as follows:(1)Due to the inconsistency of multitemporal images' boundaries and spatial correspondence in the tasks of object-based SAR image change detection,a superpixel cosegmentation algorithm for SAR image change detection is proposed.Firstly,the pixel intensity similarities between the two pixels of the multitemporal SAR images are calculated respectively,which are then combined using a weight factor to form a new similarity measurement.Additionally,the edge magnitudes of the two multitemporal SAR images as well as their log-ratio image are detected,the maximum value among which is chosen to form an edge map image.Finally,the weighted similarity based on pixel intensity,location distance and edge information is used to replace the CIELAB space similarity for local clustering in simple linear iterative clustering(SLIC).The multitemporal SAR images are then cosegmented with accurate boundaries and spatial correspondence.Experiments on simulated and real-world datasets show that the proposed method can obtain higher boundary recall and lower undersegmentation than other compared methods.(2)To solve the problem that pixel-wise change detection methods are vulnerable to speckle noise and lack of robustness,a superpixel-based change detection method is proposed.Firstly,the superpixel cosegmentation algorithm is performed to simultaneously segment the SAR image pairs to take superpixels as basic processing units for the subsequent change detection task.Then,the superpixels are modeled by generalized gamma distribution which is widely used to precisely model the statistics and Scale Independent Shape Estimation(SISE)formula is used to estimate the parameters of each superpixel.Finally,the superpixels are classified into being changed or unchanged by Kullback-leibler(KL)distance which is able to measure similarity between two distributions.By using a thresholding method,the sueperpixels are classified into the changed and unchanged categories respectively.Morphology filtering is also used to reduct isolated regions in order to obtain the final change map.Experiments on simulated and real-world datasets validate the the performance of the proposed method.(3)To solve the limitations of poor generalization and discrimination in traditional change detection methods,a SAR image change detection method is realized based on PCANet.Firstly,interested pixels which are highly possible of being changed or unchanged are selected in the preprocessing procedure.Then image patches centered at interested pixels are obtained to train the model.After that,the rest pixels are classified by the trained PCANet.Finally,the final change map is a combination of the preclassified classification and the classification result.Experiments on simulated and real-world datasets show the effectiveness of the proposed method.