SAR Image Change Detection Based On Feature Analysis And Bayesian Posterior Model

Because of the active imaging mechanism of the synthetic aperture radar(SAR)system,the SAR image can be acquired at any time and under any weather conditions.Therefore,the SAR image has become an indispensable information source in change detection.However,due to the usage of synthetic coherent imaging,multiplicative speckles always accompanied with SAR images.So,how to effectively design SAR change detection algorithms has attracted more and more attention of scholars.In this paper,we mainly focus on the two-temporal SAR image change detection problem,and the aspects of the construction of the difference map and the analysis of the difference map.The specific content is as follows.(1)In order to reduce the influence of speckle on the change detection,the algorithms based on feature extraction and analysis are studied for an effectively extraction of the change information,where the effective feature extraction algorithm is the key point.Here,with the embedding of the feature subspace,a two-stage clustering algorithm was designed for SAR change detection.In this algorithm,three classes(namely,changed class,unchanged class and confused class)were built,and the subspaces were specially constructed to represent the changed class and unchanged class.Then,by projecting on the built subspaces,the samples corresponding to the confused class are reclassified as the changed and unchanged classes.The experimental results show that the algorithm obtains the highest G-mean.Therefore,the algorithm is reasonable for the classification of changed and unchanged parts.(2)The construction of difference map is one of the most important steps in change detection.Based on this,an iterative Bayesian algorithm was designed to construct the difference map.In this algorithm,by constructing the prior probability and likelihood function of the change information,Bayesian posterior probability formula was adopted to realize continuous modification of difference map iteratively.Finally,OTSU thresholding algorithm was used to separate the changed and the unchanged information.The experimental results show that the proposed algorithm can obtain the highest G-mean and the discriminability of the difference map can be improved by iteration.(3)Given the robustness of patch-based similarity to speckle and the usage of neighborhood of markov random field(MRF),local spatial correlation(LSC)was computed for each pixel and a FCM-LSCMRF algorithm was designed for SAR change detection.In this algorithm,FCM algorithm was firstly used to get the membership of each pixel belongs to the changed and unchanged classes.Secondly,using the original two-temporal SAR images,patch-based similarity was used to compute LSC of each pixel.Then,LSC-based MRF algorithm was designed and Bayes maximum posterior criterion(MAP)was used to extract the change information of SAR image.The experimental results show that the algorithm can detect the details of the changed parts more completely.Moreover,the algorithm can obtain the highest G-mean basically.