Change Detection Of Remote Sensing Imagery Based On Active Ensemble Learning And Uncertainty Analysis

The change detection technology based on multi-temporal remote sensing images has the great significance in land use monitoring,urban construction and vegetation cover research etc..The spatial resolution is improved in high-resolution remote sensing images and the characteristics contain more abundant spatial feature structure information,followed by the reduction of spectral information,more complicated spatial structure and higher labelling cost.In addition,the characteristics of the spectral differences of similar ground objects become larger,the spectral domain separability become worse.Therefore,considering of pixel-based and object-based change detection approaches,an automatic change detection algorithm using multiple classifiers and multi-scale uncertainty analysis based on active learning in highresolution remote sensing images has been proposed in this paper.Moreover,a novel post-classification algorithm based on uncertainty analysis and Bayesian soft fusion is also introduced to obtain the reliable change types of the detection maps.The main research contents are as follows.(1)Texture features,morphological features,Gabor filtering features and spectral features are extracted to make full use of the rich spatial information of high-resolution remote sensing images.And then,the input data set for change detection has been built with the optimal feature vector.(2)In order to utilize the multi-scale information and the advantages of multiple classifiers to detect all kinds of change types effectively,a change detection algorithm based on multiple classifiers and multi-scale uncertainty analysis has been proposed.More specifically,according to the heterogeneity,K-nearest neighbor(KNN),Support Vector Machine(SVM)and Extra trees(ExT)are utilized to composed our multiple classifiers system to extract the change information.Confronting with the limitation of training samples in the monitoring change detection,spatial optimization and break ties(BT)algorithm are introduced to select unlabeled samples with abundant information.What' more,the uncertainty analysis is carried out through multi-scale layers with propagation relations to effectively take advantage of multi-scale information.The final results of change detection are formed by combined the “certain” objects on all scales.Not only can this algorithm reduce the influence of “salt-and pepper” noise in pixelbased algorithms,but also reduce the dependence of the change detection precision on the segmentation scale in the traditional object-oriented algorithms.(3)A novel post-classification algorithm based on uncertainty analysis and Bayesian soft fusion has been proposed in order to achieve the change types.Firstly,SVM and unsupervised change detection algorithm are utilized to obtain the change probability map,ExT is utilized to obtain the classification probability map.Then,the uncertainty analysis of classification probability maps in two-phase is carried out respectively.After that,the pixel-based detection result is determined by transformed the post probability map and the change probability map using Bayesian fusion.The final change type detection result is determined by object-oriented category determination.