SAR Image Change Detection Method Research Based On Feature Fusion

With the rapid development of aerospace engineering,sensor technology,computer technology and it's related disciplines,modern Satellite Remote Sensing Technology has been able to obtain multiresolution,multiangle and multisensor images whose geometric resolution from a few kilometers to few centimeters,spectral resolution from hundreds of nanometers to 10 nm,revisit cycle from one time a few days to several times a day.Abundant remote sensing data with different precision meet the needs of different levels in different applications which greatly expanded the application of remote sensing images.Remote sensing image change detection is one of the most important applications.Traditional change detection framework is the process to get difference image from two or more images with different time,and then do clustering analysis.With the appear of huge amounts of data,traditional change detection methods faced enormous challenges.Recent years,matrix decomposition has attracted more and more attention,thus provided a new train of thought for change detection.According to characteristic of remotes sensing image,we first get difference figures of temporal sequences that have different patterns in this thesis and then use the information of neighborhood as a character description.Next two matrix fusion methods are used for remote image change detection.One is Non-negative Matrix Factorization,the other one is Spectral Clustering method based affinity aggregation.In this thesis,the specific work summarized as follows:1.The remote sensing image change detection method based on Non-negative Matrix Factorization feature fusion is proposed.In view of the problem of performance degradation caused by high dimension feature of image data in traditional change detection method based on pixel,we proposed one feature fusion method based on non-negative matrix decomposition which take advantage of non-negative physical interpretability and dimension reduction characteristics of Non-negative Matrix Factorization method.First we got difference figures of temporal sequences that have different patterns and then we stitched matrices to realize the combination of multichannel,multifeature matrices on the basis of the feature description.Finally we realized the fusion of the characteristic matrix by using non-negative matrix decomposition iterative approach to obtain the final results ofremote sensing image change detection.This method have achieved effective compression of data and taken advantage of different feature matrices so as to improve the accuracy of change detection.2.The change detection method based on optimization of weights about affinities is proposed.According to the traditional spectral clustering algorithm's effect excessively depend on the construction of similar matrices and its running speed reduce and clustering performance get weak caused by high image data dimension,one change detection framework based on AANMSC(Affinity Aggregation for Spectral Clustering Using Nytr?m method)algorithm is proposed.First of all,we structured affinities' submatrices of sample data using difference figures and different similarity descriptors.And then we got matrix of similarity by linear superposition of affinities' submatrices.We used the Lagrangian function optimization method to study optimal weight ratio and achieved unequal weight features fusion method next.This method have improved the accuracy and efficiency of change detection and provided a new train of thought about application for change detection in real-time scene.