Research On SAR Image Change Detection Method

The synthetic aperture radar(SAR)microwave imaging technology can acquire highresolution data in all day and under all weather conditions.The acquired SAR data is increasingly applied to civilian,military and other fields.Interpretation of SAR data has also become more and more important.As an important content of SAR image interpretation,SAR change detection is a key technology for environmental monitoring,agricultural surveys,natural disasters detection and other ground observation applications.SAR image change detection is a technology to qualitatively or quantitatively analyze and determine the characteristics and processes of surface changes from multi-temporal SAR images of the same geographic area acquired at different times,and is used to detect some changes in the same place over a period of time.The microwave imaging mechanism makes SAR images have the shortcomings of complicated background information and serious feature aliasing between different ground object areas,and there is coherent speckle noise in SAR images.At present,one of the most important target in the process of SAR image change detection is to eliminate the influence of coherent speckle noise,and thereby to improve the detection accuracy.This paper studies the SAR image change detection based on the traditional methods and the network learning methods,and embeds the studied algorithms into the software.The specific work done is as follows:1.The traditional SAR image change detection methods based on clustering and threshold segmentation are studied,which mainly include SAR image change detection clustering algorithm and threshold segmentation algorithm.First,based on the Fuzzy CMeans(FCM)clustering algorithm,the difference image generation algorithm is deeply analyzed,and the FCM clustering results of the difference images obtained by different difference image generation algorithms are compared through experiments.Then the KI threshold selection algorithm based on the generalized Gaussian distribution is studied.The threshold algorithm and the misclassified pixels are analyzed,and a SAR image change detection method based on threshold fusion and neighborhood voting is proposed.The results on the actual measured SAR data show that the proposed method can mitigate the influence of noise on SAR images,reduce the number of false alarm pixels and missed pixels for change detection,and effectively improve the accuracy of change detection.2.The change detection methods based on shallow network model and deep network model are studied.For the shallow network model,the change detection method based on Extreme Learning Machine(ELM)is studied.For the deep network model,Principal Component Analysis Network(PCANet)and Convolutional Neural Network(CNN)are studied.The performance of the three types of network model is analyzed through experiments,and then based on CNN,the sample imbalance problem and the typical sample selection problem of the network is studied.Aiming at the ubiquitous problem of imbalance between the changed and unchanged samples in the change detection,this paper studies the problem of the sample imbalance in the change detection by augmenting the data.Aiming at the problem that the samples in the network are randomly selected from the pre-classification results and are not representative,the method of extracting representative training samples from the pre-classification results through edge detection is studied,and then the samples selected by the studied method are verified through experiments.3.The SAR image change detection software is designed,and the studied algorithms are embedded into the human-computer interaction interface.First,the software and hardware conditions required for software development and the software development process are introduced.Then the function of each module of the software is introduced in detail.Finally,the instructions for using the software are introduced.