An Object-based Unsupervised Change Detection Method Based On Hybrid Spectral Difference

Spectral variation is the most commonly used feature in unsupervised change detection.Traditional methods assume that the degree of change is in proportion to the difference of spectral values or spectral shapes.Normally,the distance between two spectral vectors are used to describe the difference of spectral values,which however will not be appropriate if spectral values are dark,so these change types are hard to be detected by method based on spectral values.Spectral Angle Mapper(SCM)and Spectral Gradient Difference(SGD)are method based on the difference between spectral shapes.SCM is insensitive to the affine transformation of the spectra of original images,but it indirectly describes the difference between spectral shapes by the Pearson's correlation without representing the spectral shape of each spectral curve.SGD used the spectral gradient to represent the spectral shape directly,so it can describe the difference between spectral shapes better.However,SGD is sensitive to multiplicative gains of spectra which cannot be avoided in change detection.Besides,some object types have similar spectral shapes especially when their spectral curves are close to flat,which makes it difficult to detect the change among these objects by the method based on the difference between spectral shapes.Consequently,this paper proposed a new object-based unsupervised change detection method based on hybrid spectral difference(OBHSD)which is totally automated,and the main ideas are as follows:1 To normalize all bands to ensure that each band has the same impact on the spectral values and spectral shapes.Then,segment images to get objects.2 Use the Euclidean distance between spectral change vectors to characterize the difference between spectral values(CDSV).Then,calculate the spectral magnitudes of two objects and use the larger one to represent the credibility since the credibility of CDSV is low when both two objects are dark.3 This paper proposes a new method to describe the difference between spectral shapes(CDSS)to fuse the results of SCM and SGD.Then,calculate the spectral gradients of two objects and use the larger one to represent the credibility since the credibility of CDSS is low when the spectra of both two objects are close to flat.4 OBHSD fuses the difference image of CDSV and CDSS with the weights which are calculated by their credibility and get the final result.5 Two experiments were carried out on WorldView-2/3 and Landsat-7 Enhanced Thematic Mapper Plus(ETM+)datasets.We found that hybrid spectral difference method had better adaptability and superior capabilities of object-based change detection compared with standard change vector analysis(CVA),spectral correlation mapper(SCM),spectral gradient difference(SGD)and multivariate alteration detection(MAD),yielding Kappa coefficients of 0.798 and 0.835 for two experiments respectively.