Research On The Methods Of Change Detection And Image Annotation For Remote Sensing Images

The task of change detection and image annotation for remote sensing image is oneof the important issues of remote sensing image analysis and interpretation. Traditionalmethods of change detection consist of three steps, i.e., preprocessing, obtaining thechanged information, and postprocessing of the changed content. These methods aim tosolve three problems, i.e., judging and labeling the changed regions, and identifying theclassification of the changed regions. With the extension of the application fields ofremote sensing image, the demands for change detection and image annotation haveincreased. On one hand, people want to obtain the changed information convenientlyand rapidly, and at the same time, to maintain the spectral characteristics of the changedregions, and to reduce the effect of preprocessing error on change detection. On theother hand, people also want to know the change process and abundant and accurateproperty information of the changed regions, such as saliency and class information.However, the traditional change detection and image annotation methods can not solveall of these problems very well.In order to address these problems in a better way, this dissertation focuses onexploring the methods of change detection and image annotation in the followingaspects, i.e., maintaining the spectral characteristics of the changed regions, weakeningthe effect of image registration error on change detection, obtaining the information ofchange process, introducing the saliency information, and improving the accuracy ofclass annotation. These methods are as follows.Based on the investigation of hard change detection with image registration, thedissertation proposes two methods, i.e., hard change detection methods based onquaternion principal component analysis and random walker algorithm, respectively.The former can simultaneously process three channel images under the model ofquaternion while maintaining the spectral characteristics between these channel images.The latter can integrate image registration and change detection using the feature pointsof image registration to represent the initial seed nodes of random walker algorithm.In the section of hard change detection for the unregistered remote sensing image, in order to perform change detection on the unregistered remote sensing image directly,the dissertation proposes two methods, i.e., hard change detection methods based onclustering algorithm and global contrast, respectively. The former performs changedetection by combining two clustering algorithms, i.e., bottom-up unsupervisedclustering algorithm and top-down discriminative clustering algorithm. The latterperforms change detection by analyzing the color differences between the changed andunchanged regions.In order to describe the change process of the changed regions, the dissertationintroduces the concept of soft change detection to the research field of change detectionfor remote sensing image. It proposes a soft change detection method under the modelof Bayesian. The result of soft change detection can represent two kinds of information,i.e., whether the regions have changed or not, and the progress of the change. Therefore,it is useful for analyzing and interpreting the information of the changed regionsin-depth.In the research of image annotation of the changed regions, the dissertationproposes two kinds of methods, i.e., saliency annotation based on visual attention andclass annotation based on probabilistic graphical model. With the salient features of thechanged regions, the former can perform saliency annotation for the changed regionsrapidly. The latter uses a color and shape descriptor (CSD) for multispectral remotesensing images, and employs the author topic model (ATM) to perform class annotationof the changed regions. In addition, by employing the genre information of the author,the dissertation proposes an author-genre topic model (AGTM), and uses it to promotethe accuracy of class annotation for remote sensing image successfully.The dissertation uses the proposed change detection and image annotation methodsto process the multispectral remote sensing images, and has obtained expected results.