Research On Several Methods In Change Detection Of Multispectral/SAR Images

Change detection techniques have been used in environmental monitoring,urban studies,forest monitoring,agricultural surveys,disaster and battle damage assessment,and dynamic monitoring for national geographical state of China,etc.However,the existing methods for change detection in multispectral or Synthetic Aperture Radar(SAR)images have some defects,for instance,the use of multitemporal spectral vectors is insufficient;the lack of support of Human Visual System(HVS)theory;and the accuracy limit caused by speckle noise in SAR images.Therefore,the study of change detection methods on multispectral/SAR images is of great significance to promote the application of this technology in the aforementioned fields.Based on these above,the change detection research in this dissertation will focus on the combination of Change Vector Analysis(CVA)and Spectral Angle Mapper(SAM),the use of structural information based on HVS theory and spatial information of SAR images.The major achievements are as follows:(1)Three methods combining CVA and SAM are proposed for change detection in multispectral images.Theoretical analysis on the performance of CVA and SAM when they are used for change detection in multitemporal multispectral images indicates that the use of multitemporal spectral vectors is insufficient and they are complementary.To integrate the advantages of CVA and SAM in change detection of multispectral images and improve the robustness of change detection,Generalized Change Vector Analysis(GCVA),Autoadapted Fusion Atrategy(AFS)and Approach Based on Discrete Wavelet Transform(ABDWT)are reported in this dissertation.Experimental results indicate that: 1)three change categories may occur in a changed pixel and CVA or SAM alone can only detect two of the three change categories properly;2)the proposed methods perform better than the state-of-the-art methods in change detection of multispectral images.(2)To overcome the shortcoming that the commonly used methods in change detection of multispectral/SAR images are defined mathematically and lack the support of HVS theory,the methods based on structural information are proposed for change detection in multispectral/SAR images.The improved method for extracting structural information and two stratagies for extracting multiband structural information are first proposed,which are based on an assumption that the process of extracting structural information by HVS can be divided into two steps.Then,the difference image is generated by using structural information.Experimental results show that: compared with the state-of-the-art methods,the proposed methods can well describe the difference between multitemporal multispectral/SAR images and improve the accuracy of change detection.(3)Accurate performance of Ratio(R)and Log Ratio(LR)methods in change detection of SAR images is reported in this dissertation.Firstly,we theoretically proved that the monotonically increasing linear or non-linear transformation cannot change the Receiver Operating Characteristic(ROC)curve of difference image.Secondly,we confirmed the reason that LR performs better than R in many experiments on the basis of theoretical analysis and experiments.Then,we summarized the accurate performance of R and LR operators in the data set with one or two change types.Experimental results from three original and stitched data sets clearly proved that LR performs better than R in only a few special cases.(4)Improved Neighborhood-based Ratio(INR)is proposed for change detection in SAR images.Firstly,we analyse three drawbacks of the neighborhood-based ratio approach.Then,we reconstruct the neighborhood-based ratio operator to overcome its drawbacks.Experimental results show that INR can well balance between restraining speckle noise and preserving the detail information and improve the accuracy of change detection in SAR images effectively.(5)A filtering approach to binary change map of SAR images is proposed based on the voter model and spatial-contextual information.Given the fact that the false and missed alarm noise are discretely distributed in unchanged and changed areas in change map,the local continuity assumption of the changed and unchanged areas is first proposed.During the filtering process,this method simulates election behavior,which originated from the democratic life of Ancient Greece,and identifies the detection noise with the help of eight connected domains.Experimental results show that the proposed method can effectively decrease the detection noise and improve the detection accuracy of binary change map.This approach has a wide potential application in the postprocessing of the binary change map generated from SAR images.