Abstract: A certain correlation exists between the changes of an object (phenomenon) at different time,which can be used as one of the important features to detect changes. Many traditional change detection methods like layer arithmetic,change vector analysis,and multi-data direct comparison take no consideration of context information and spectral and spatial correlation. Neighborhood correlation images (NCIs) and cross correlation analysis (CCA) change detection methods that use spectral and spatial correlation information to extract changes in land-use both have been verified to have a good accuracy. In this paper,these two methods are applied and evaluated in order to verify their applicability. Dual phase images from GF-1 satellite of yanse lake disaster areas are chosen to conduct the experiments,where the supervised NCIs and the supervised post-classification comparison (PCC) based on support vector machine (SVM) and maximum likelihood,and the unsupervised CCA and the unsupervised PCC based on ISODATA have been compared. The results indicate that the supervised NCIs method shows the highest overall accuracy with a 3-pixel radius of neighborhood,93.05%,and it is a little higher than the supervised PCC method based on SVM whose overall accuracy is 90.52%. The demerit of the NCIs method is the constraint of neighborhood size. Moreover,the unsupervised CCA method performs best with a threshold of 20 pixels,where the overall accuracy is 93.95%. It shows more advantages than the unsupervised PCC method based on ISODATA,the overall accuracy of which is 90.52%.However,it is not easy for the CCA method to choose the proper threshold and obtain the ground-truth of a single phase image.

Key words: remote sensing imagery, change detection, neighborhood correlation images, cross correlation analysis