Research On The Superpixel-based Change Detection Methods Based On Remote Sensing Images

In recent years,the rapid development of satellite technology has provided highresolution remote images with more abundant spatial information for earth observation,which has greatly promoted the development of change detection technology.It makes change detection technology widely used in disaster assessment,urban planning,land use type monitoring and many other fields.However,uncertain factors such as lighting conditions,sensor platform limitations,and the complexity of the surface environment will affect the accuracy of the final change detection results.It's not only a hot topic but also a crux to effectively use the spatial and spectral information to improve the accuracy of change detection.So,this paper takes superpixel as the basic processing unit of change detection,and improves the accuracy of change detection results by constructing a change detection model,which takes into account the spectral and texture features of superpixel at the same time.The main research contents of this paper are as follows:(1)The basic process of change detection is thoroughly sorted out,the principle of change detection and several commonly used methods are explained.It focuses on analyzing the problems of superpixel segmentation and neighborhood definition in superpixel change detection,and proposes corresponding solutions to these problems.(2)Aiming at the problem of over-utilizing spatial information by Markov random field(MRF)method,the initial change membership information of superpixel is used to construct a new spatial labeling field model.And a new method based on superpixels is proposed.In order to improve the accuracy,change probability information on the basis of the original MRF model is added.Three remote sensing datasets of change detection experiments show that the proposed method can obtain higher-precision change detection results,compared with traditional MRF,support vector machine(SVM)and EM methods.(3)The accuracy of the graph cut change detection result is highly rely on the accuracy of initial threshold.To solve this problem,this paper uses the classification probability obtained by the FCM algorithm to improve the graph cut algorithm.An improved graph cut change detection algorithm based on superpixels is proposed by constructing the change feature energy function.This method extracts the change area based on the superpixel classification probability and change intensity information,and selects four features of the image gray level co-occurrence matrix to calculate the spatial feature energy.Experimental results show that the method is robust and usable,and the accuracy of the change detection results is higher than that of FCM and FLGICM.