Research On Change Detection Of High Resolution Remote Sensing Images Based On Mid-level Feature

With the development of remote sensing technology and digital image processing technology,the use of remote sensing imagery for change detection can provide more accurate and larger-scale land surface change information for mankind with low-cost,which is important for coordinating population,environment,resources and achieving sustainable development.Recently remote sensing change detection is widely used in agricultural monitoring,urban planning,disaster assessment and many other fields.The rise and prevalence of high resolution remote sensing images provide the possibility for mankind to obtain land surface change information from smaller scales,but it also puts forward higher requirements for existing change detection technologies.Although the change detection methods for high resolution image are constantly evolving,these algorithms only use the low-level feature in the images to detect changes.How to effectively and accurately extract change information for high resolution images is still a problem that needs to be solved.This paper proposes a high resolution image change detection method based on mid-level semantic feature,aiming at solving the existing problems and the key issues in the field of change detection for high resolution remote sensing image,and explore the new application of change detection.The main research work of this paper is as follows:Firstly,this paper summarized the representative achievements of middle-level semantic features and high-resolution image change detection in the existing literature,and systematically analyzed the difficulties and problems existing in the detection of high resolution remote sensing images.Secondly,starting from the object level,a change detection method based on the semantic features in the middle layer of objects is proposed,and the process of change detection based on the semantic features in the middle layer of objects is designed-aiming at the problem of insufficient performance of the low-level feature,this paper introduce the Bag of Word model into the feature extraction of high resolution remote sensing image objects,and deeply mining the mid-level semantic information of objects.And similarity measurement algorithm was used to effectively complete change detection research for high resolution image.Compared with the change detection method based low-level features and OCVA method,the proposed method based on mid-level semantic feature of objects eliminates the problem of pepper-and-salt appearing in traditional pixel-based change detection methods,effectively improves the problem of insufficient low-level feature expression abilities,and improve the accuracy of change detection.Finally,from the perspective of scenario level,four analysis strategies based on middle-level semantic features were proposed:PS-LDA,PT-LDA,FS-LDA and FT-LDA.This paper considers the change detection of high resolution remote sensing images from the perspective of scene.At the same time,four different strategies of change analysis are proposed for the problem of"semantic gap" between low-level features and high-level semantics in scenario level change analysis.Through comparison and analysis of four kinds of analysis strategies,it is found that time characteristics play a very important role in the study of scene-level change detection for high resolution images.When using LDA as a feature extraction tool,all samples share the same topic space and avoid two classification processes,which is conducive to improving the accuracy of scene-level change detection.