Change Detection Of High Resolution Remote Sensing Imagery Combining Multi-scale Feature And Active Learning

Land-cover and land-use change information plays an important role in practical applications,including land reclamation,urban expansion,disaster monitoring and assessment,as well as land planning.Land use change information can be obtained accurately and timely through remote sensing image change detection,with its advantages of wide coverage and short cycle.Change detection of remote sensing imagery provides basic information for urban planning and sustainable development.However,while the resolution of remote sensing images is significantly improved,there are multiple challenges in change detection of high resolution imagery,which is different from traditional change detection techniques.Processing high resolution remote sensing imagery has limitations caused by abundant spacial information,spectral heterogeneity and high cost of obtaining a large number of labeling samples.In view of these problems,multi-scale multi-scale image segmentation was applied and multiple features were extracted,such as spectral,texture and shape features,based on object-based change detection.Active learning method,which combined with Support Vector Machine,was introduced to change region detection.And the change type detection was realized by multi-level sequential spectral analysis.The main research contents were as follows:(1)Considering the objects with different scales and the spatial characteristics of high resolution remote sensing imagery,the images were hierarchically segmented by Region Adjacency Graph based image segmentation method.After obtaining image objects with different scales,multiple features of each level were extracted,including spectral,texture and shape features.Then the optimal feature vector was determined by Random Forest feature selection,followed by computing the feature change vector for change detection.(2)In order to detect change regions,SVM based on active learning was applied in this thesis.Euclidean distance of change vector was calculated to obtain the histogram of change magnitude.The original training set for Support Vector Machine was automatically obtained according to the histogram.And then Margin Sampling and the Entropy Query-By-Bagging sampling strategy were applied to select informative unlabeled samples and add labels for SVM model training.Thus the separating hyperplane was optimized and thus the accuracy of change region detection was improved,with less sample labeling cost.(3)In order to determine change types,on the basis of change region,a multi-level change type analysis was performed by combining Sequential Spectral Change Vector Analysis and SVM.Firstly,by calculating the S~2CVA change angle of the changed pixels,the change angle and magnitude were presented in the polar coordinate system.The number of change types and training samples of each type were determined according to the clustering characteristics.SVM classification was used to distinguish the main types of change,followed by the iterative analysis until subtle change types were detected.It ensures the detection of various types of change,when the number of change types is unavailable.Finally a hot spot change map was generated to further analyze the changes.