Spatial Contextual Modeling And Application Of SAR Image Based On Nonstationary Random Field Model

Synthetic aperture radar(SAR)is a kind of active aerospace and aviation microwave imaging system,and it has the all-time,all-weather,multiband,and multiview abilities to acquire ground observation data,which can be widely used in various military and civil applications.For the large number of image data acquired from SAR systems,the SAR image interpretation technology has become the research hot spot for foreign and domestic researchers.Due to the existed speckle noise and the limit of feature extraction on complex ground targets in SAR images,the interpretation labeling process based on local feature extraction can be easily affected by these negative factors to produce inaccurate labeling result that does not possess spatial consistency.Based on the random field model that effectively incorporates the spatial contextual information within images,this dissertation depends on the local image feature extraction and analysis to develop nonstationary SAR image spatial contextual modeling methods through nonstationary random field theory.The nonstationary random field model in this dissertation can be applied to describe the nonstationary contextual information within a single SAR image for segmentation,and it can also be used to describe the the nonstationary contextual information between two SAR images for change detection.The main content and the acquired research results in this dissertation can be summarized as the following four parts:1.Considering the limitation of the current triplet Markov fields(TMF)model under low-order neighborhood for complex structures description in SAR images,a higher-order neighborhood based TMF(HN-TMF)model has been proposed for describing the nonstationary spatial contextual information within higher-order neighborhood to improve the SAR image segmentation accuracy.In the HN-TMF,with the autocovariance analysis applied to reveal the local fluctuation at each site,the auxiliary field is redefined to denote homogeneity or heterogeneity.Based on the auxiliary field,the local energy function in HN-TMF is constructed either in a homogeneous or heterogeneous way,and hence,the local structure can be embedded in the energy function to improve the prior modeling ability.Along with the newly constructed energy function,new initialization method of HN-TMF parameters is given to fulfill the physical interpretation of the prior energy function.The experimental results on real SAR images indicate that HN-TMF is able to properly characterize the SAR image nonstationary contextual information,and it can further incorporate the contextual information into the segmentation procedure to get satisfactory segmentation results in both homogeneous regions and heterogeneous regions containing complex structured features.2.To solve the observation information uncertainties caused by speckle noise in SAR images,an ambiguity label information fusion based TMF(ALF-TMF)has been proposed for SAR image segmentation.The ALF-TMF takes into account the nonstationary and uncertain properties associated with SAR images simultaneously to improve the segmentation accuracy.The auxiliary field in the ALF-TMF is redefined to specify a dominant direction for each site according to local image gradient analysis.The label field is extended by adaptively generating ambiguity class at each site according to pixel observation and contextual information.According to the generalized label field with adaptive ambiguity class in the ALF-TMF,likelihood and nonstationary prior energy are constructed.Therefore,uncertainties of SAR image pixels are considered in the statistical segmentation procedure.The experimental results on real SAR images indicate that ALF-TMF model can properly characterize the SAR image nonstationary contextual information under the influence of speckle noise.Therefore,the proposed model can obtain segmentation results with better robustness against speckle noise and more accurate preservation of detailed features.3.In order to solve the problem that regional CRF heavily depends on the superpixel generation accuracy,an adaptive hybrid nonstationary CRF(AHNCRF)has been proposed for SAR image segmentation.Based on the generation of superpixels and their boundaries features analysis,the proposed method adaptively divides SAR image into different parts,namely homogeneous regions,heterogeneous regions and edges.In homogeneous regions,the regional-level CRF is defined on superpixels,and the pixels within each superpixel force to have the same segmentation label.Oppositely,the pixel-level CRF is defined on pixels within heterogeneous regions or edges,and local autocovariance features are extracted for constructing the unary and pairwise potentials to incorporate effective local contextual information.Experimental results on real SAR images indicate that AHNCRF model can effectively divide the SAR image into homogeneous and heterogeneous regions.Moreover,it can properly characterize the nonstationary contextual information for the detailed features in heterogeneous regions,and improves the heterogeneous regions segmentation accuracy accordingly.4.Taking the nonstationary correlation between multitemporal SAR images into account,a TMF model has been proposed for SAR image change detection.The auxiliary field for change detection is redefined to describe the nonstationary textural similarity in two images,and the initialization of the auxiliary field is executed through spatially nonstationary anisotropic texture analysis.Then,the corresponding prior energy function is reconstructed.To distinguish the emphasis between prior energy and likelihood locally,a weight parameter adaptively varied under the auxiliary field is introduced,and that helps to overcome the compromise between accurate detection rate and false alarm rate in traditional MRF methods.Experimental results on real SAR images indicate that the proposed TMF model can properly characterize the nonstationary contextual information between the two SAR images through conjoint analysis of the features extracted from the same location.The proposed model can combine the nonstationary contextual information with the adaptive weight to obtain change detection results with satisfactory accurate detection rate and false alarm rate.