Remote Sensing Change Detection Based On Two-level Classification Considering Uncertainty Analysis

Remote sensing change detection is a process of extracting land surface change information by analyzing remote sensing images obtained at different times in the same geographical location,which is widely used in disaster assessment,natural resources monitoring and urban research.In recent years,the change detection technology based on two-level classification has been proposed and satisfactory detection results were obtained.However,existing change detection methods based on two-level classification do not fully consider the uncertainty of change detection,which leads to the unsatisfactory results of change detection sometimes.This paper conducts an in-depth study on this problem.With the help of fuzzy clustering,uncertainty analysis theory,K-nearest neighbor(KNN)algorithm,Extreme learning machine,ELM),Support vector machine(SVM),Majority voting(MV),etc.,proposed three two-level classification remote sensing change detection methods considering uncertainty analysis,and the main work is as follows:(1)To reduce the influence of uncertainty on change detection,a two-level classification change detection method based on Fuzzy C-means(FCM)clustering and KNN is proposed.To begin with,FCM is utilized to analyze the difference image and produce a fuzzy membership matrix for both changed and unchanged pixels.Then,the uncertainty of each pixel was calculated by Gini coefficient based on fuzzy membership degree,and the difference images were divided into three categories: change,unchanged and uncertainty according to the uncertainty.Finally,KNN is used to reclassify the uncertain class pixels with the help of the pixels that have been divided into changed and unchanged classes,and the change detection graph is generated.The proposed method identifies uncertain pixels by FCM clustering and Gini coefficient calculation of pixel uncertainty in first-level classification,and reclassifies them by KNN in second-level classification,which can effectively improve the classification effect of uncertain pixels.The experimental results of three sets of real remote sensing data show that this method can obtain higher precision change detection results.For example,for farmland dataset,the Kappa coefficient of this method is 87.39%,1.19%～16.98% higher than that of other methods.(2)In order to reduce the influence of uncertainty and integrate the advantages of different classifiers in the second-level classification,a two-level classification change detection method based on FCM clustering and multi-classifier fusion is proposed.This method is an improvement of the previous method.The main innovations and contributions focus on the secondary classification steps.Firstly,the same steps as the previous method are used for first-level classification,and the difference image elements are divided into three categories: changing,unchanged and uncertain.Then an effective integrated classification system based on KNN,SVM and ELM classifiers is proposed to reclassify the uncertain class elements.By integrating the advantages of different classifiers,the second-level classification step of this method can obtain better change detection results for uncertain class pixels,thus,enhancing the overall precision of change detection.Experimental results of three sets of real remote sensing data verify the effectiveness of the proposed method.For example,for Madeirinha dataset,the Kappa coefficient of this method is 90.20%,2.73% ～ 54.95% higher than that of other methods.(3)In order to reduce the influence of uncertainty and utilize the complementarity of different difference images,a two-level classification change detection method based on FCM clustering and multi-difference image fusion is proposed.Firstly,three typical difference image generation algorithms are selected to generate three complementary difference images.Secondly,FCM clustering is used to calculate the fuzzy membership matrix of each group of difference images with respect to the variable class and the unchanged class.Then,three fuzzy membership matrices were fused based on MV and information entropy to calculate the uncertainty of the pixel,and the pixel was divided into three categories: changing,unchanged and uncertain.Finally,the uncertainty class pixel is reclassified by the spatial correlation of the image.The experimental results of three sets of real remote sensing data show that this method is capable of efficiently combining the strengths of disparate difference images to address uncertainties during the fusion process,and the method achieve better change detection results.For example,for farmland dataset,KC of this method is 91.55%,which is5.35%～21.14% higher than other methods.