Researches On Multispectral Remote Sensing Image Change Detection Method Based On Multiscale Subspaces Analysis

Change detection is an important way to obtain useful information from massive high-resolution remote sensing images.It utilizes multi-temporal remote sensing images in same geographical area at different time to identify changes of the state of ground objects or the difference of natural phenomena by collaborative processing between people and computers.Change detection technology is widely used in various practical applications,e.g.,land resource management,ecological environment monitoring,natural disaster assessment and urban planning,which is of significant for promoting the coordinated and sustainable development of human society.In this thesis,the imaging mechanism and spectral feature distribution of ground objects are researched in multispectral images.Multiscale analysis tools are introduced to address the problem that structure,shape,size,spectral and texture feature are mixed in high-resolution remote sensing images by mapping different distribution of features to matched subspaces to obtain the consistency of features.The better change detection result is obtained through the multiscale subspaces joint optimization.The specific work is summarized as follows:1.A multiscale model of structure is introduced,and an effective approach is proposed based on optimized fusion of multiscale subspaces for multispectral images change detection.Firstly,structure elements that match ground objects structure are selected to expand the phase images to construct multiscale subspaces aiming at obtaining the joint distribution of space-spectrum information in multispectral images.Subsequently,in order to reduce difference interference of the near infrared channel,the scale change vector analysis and scale spectral angle mapper model are presented to extract change feature.Finally,a feature subspaces fusion algorithm is proposed named stochastic sampling affinity aggregation based on spectral clustering,and the final change detection results are achieved by optimizing the weight allocation of the subspaces to reconstruct the affinity matrix of spectral clustering.2.A multiscale segmentation model is introduced,and the proposed multispectral images change detection method incorporates multiscale segmentation and waterfall Gaussian mixture model is presented.First of all,combining change vector analysis and spectral angle mapper are employed to build multichannel difference images to improve the constraint ability of segmentation model.Subsequently,dynamic sorting statistical region merging model is adopted to separate the ground objects with different sizes and geometric structure.Finally,waterfall Gaussian mixture model is proposed to fuse multiscale segmentation result to complete change detection.Hierarchical iteration and parameter transmission mechanism is designed to obtain the mixed Gaussian parameters optimized in waterfall Gaussian mixture model,and then the change decision happens according to the consistency decision strategy in label vector space and probability vector space.For experiment,we select two groups of real multispectral datasets from bi-temporal remote sensing images of Xi'an Chanba Ecological District to validate the effectiveness of the proposed methods.The experimental results show that the proposed change detection methods based on multi-scale analysis can not only detect weak change objects and maintain the geometric structure of the change objects completely,but also achieve good detection performance.