Object-based Change Detection For High Resolution Remote Sensing Images Based On Multi-scale Image Segmentation

Change detection is one of the research hotspots of remote sensing applications.With the rapid development of remote sensing platforms,the image resolution has reached sub-meter level.Using traditional pixel-based image analysis methods to detect the changes of highresolution remote sensing images always lead to serious speckle phenomenon in the detection results.The emergence of object-based image analysis technology provides an opportunity to solve the problem,but there are still many problems that restrict the accuracy of change detection.The problems can be enumerated as follows: the image segmentation scale is difficult to determine;the measurement methods of image object diversity are too simple,resulting in poor detection;the use of a single feature for change detection cannot take into account the characteristic of different geo-objects with spectral overlap;and single scale detection results are difficult to consider the size difference of different objects.Accounting for the issues above,by exploring both object-oriented image analysis methods and change detection methods,this thesis carry out the optimal scale research of multi-scale segmentation and introduced a similarity measure to analyze objects differences in the framework of object-oriented image analysis.On this basis,a spectral-texture feature adaptive fusion method is proposed,and a multi-scale voting mechanism combined with multi-valued logic is constructed to implement object-oriented change detection with multiscale and multi-feature fusion.The specific research contents are as follows:(1)An optimal scale selection method based on Grey Level Co-occurrence Matrix(GLCM)texture feature is proposed.The gray level co-occurrence matrix is used to extract the texture average statistics of images,then the optimal segmentation scale is determined according to the relationship between the texture average and the segmentation scale,and the optimal scale extraction of joint texture features and spectral features is achieved.Experiment results of different image scenes indicated that this method is more applicable to images of different scenes than the existing methods,and the accuracy of the optimal scale extraction result is higher.(2)The similarity measurement method is introduced into object-oriented change detection to measure the difference between image objects in different periods,so as to realize object-oriented change detection based on similarity measure.Using the traditional difference method for comparative experiments,the results show that the similarity measurement method has higher detection accuracy in different scenes,and mades a greater improvement for complex image scenes than simple scenes.(3)A feature adaptive fusion method is proposed to realize the change detection method of fusion spectral feature and texture feature.On this basis,a multi-scale fusion mechanism,which combining multi-valued logic rules and majority voting principles,is constructed to achieve the vote fusion of optimal scale sets.Experimental results show that the detection accuracy of feature fusion is higher than that of single feature change detection in different scenes,and the accuracy of multi-scale fusion is also higher than that of single-scale detection,furthermore,these methods are more effective in complex scenes.