Research On Change Detection For Remote Sensing Data Based On Fuzzying Classification

Change detection for remote sensing image is the technology of identifying land cover change using remotely sensed imagery of the same scene acquired at different times.Recently,with the maturity of change detection method,change detection for remote sensing image has become an indispensable technology of monitoring the surface of the earth,and it has successful been used in urban development,forest cover change,forest fire,wetland change and other fields.Supervised classification method has better classification result because of the information of labeled samples.But,it is difficult to acquire the information of labeled samples,because it requires the help of domain experts and it is exhausting and time-consuming.Unsupervised classification method has no prior knowledge,so,it has is no ideal classification accuracy.Semi-supervised classification method combined with the advantages of supervised classification method and unsupervised classification method and it improved classification accuracy by labeling a small number of sample information.Based on the above analysis,this dissertation adopted semi-supervised classification method in the change detection.A novel labelling method is suggested in this dissertation to ovcome the difficulty of labeling sample information.Based on existing change detection methods,this dissertation proposes a novel semi-supervised classification method in which space information is considered.The main works of this dissertation dissertation are summarized as follows:(1)This dissertation introduces a novel labelling method.Traditional semi-supervised clustering requires real information about the ground,generally,it is difficult to acquire the information of labeled samples.In traditional semi-supervised clustering,the information of labeled samples is usually obtained by random labeling.According to the problem of change detection,this dissertation given a novel labelling method in the case of the sample has no class label,to mark the change point and no-change point in the difference image.This method sorted the pixels according to the gray value in the difference image,and marked the small gray value points(it is the gray value is 0 in the experiment)as unchange points,the bigger gray value points as change points.(2)Based on existing change detection methods and the novel labelling method,this dissertation proposes a novel semi-supervised classification method in which space information is considered.After getting the difference image,firstly,using the novel labellingmethod suggested in this dissertation to mark the change point and no-change point in the difference image,then,the improved semi-supervised fuzzing c-means classification algorithm is applied to partition the data,lastly,the Markov random field model is employed to reduce the noise points by considering the space information,and getting change detection results.To verify the and usefulness of the method,this dissertation selected the Landsat7 data do experiment in the area of Borneo,Brazil and DaLian,respectively.Through the experimental result,it is proved that the novel semi-supervised classification method proposed in this dissertation is effective,the Kappa coefficient is more than 85%.