A Kernel-based Change Detection Method For Remote Sensing Imagery

Remote sensing technology for earth observation technology provides more convenient and efficient service for the analysis of land use status,land resource management,disaster prevention and emergency response as well as urbanization projects with its advantages of wide coverage,strong macroscopic,multi-temporal,short cycle and abundant information.In the case that it is difficult to obtain the labeled samples and the change information may be nonlinear distribution in research areas,the kernel theory was introduced.The change detection based on kernel methods was designed for two purposes of change range and change type detection for remote sensing images in this thesis.The main research contents were as follows:(1)The basic theory of kernel method,the properties of kernel function and the necessary conditions of multi-kernel structure were summarized.And then several classical methods of change detection for remote sensing imagry were compared.The classics CVA method was realized,and the advantages and disadvantages among classics CVA,polarization CVA and C2 VA were analyzed while its change range was confirmed by using OTSU.Then,the traditional coding method was used to obtain the change types for the comparation with the experiments in this thesis.(2)In order to obtain change range,the adaptive fusion strategy was used to fuse the results of CVA and the spectral angle as well as spectral correlation match,that means that the fusion of multi-similarity measure were taken as the final change magnitude map.The kernel-based K-means clustering algorithm with multi objective optimization replaced the thresholding process in this thesis,so that it can increase the separability of data and improve the level of automation to a certain extent under the premise of the unknown data distribution.(3)In order to obtain change types,training samples were selected randomly within a small area of high confidence after getting the direction of change information in the polar coordinate system.And then change types were confirmed by multi-kernels SVM in different ways.Compared with the single kernel SVM,The design scheme save the process of artificial kernel function selection while having a certain improvement in accuracy compared with the traditional coding method.(4)A simple change detection system for remote sensing images based on kernel method was designed at the platform of Matlab 2013 a.Focus on the development of three modules: change magnitude fusion module,change range detection module and change types detection module,the system provided an interactive interface for easy operation and it can detect the change of two-phase image according to the specific methods under the selection by user.At the same time,two groups of simulation data and two sets of real images were used to verify the feasibility and effectiveness of the proposed system,and the change detection of different kinds of images with different distribution was realized.The experimental results showed that the scheme based on kernel methods showed better performances when comparing with the traditional detection methods.