Research On Statistic Segmentation Method Of High Resolution Remote Sensing Image Based On Curvelet Feature Weighted

With the rapid development of spatial resolution,the differences between homogeneous regions in the remote sensing image exist not only in spectral feature,but also in boundary feature,texture feature and so on.Therefore,the pixel-based traditional spectral segmentation method can not reach the segmentation accuracy of high resolution remote sensing image.To this end,this paper proposes a statistic-and curvelet feature weighted-based segmentation method for high resolution remote sensing image.The proposed method can not only adaptively determine the importance of each feature in the segmentation processing,but also effectively improve the segmentation accuracy of high resolution remote sensing image.First of all,curvelet transform is used to extract multiscale spectral features,texture feature and boundary feature of pixel,thus two different feature vectors are obtained,and the corresponding feature sets are formed from feature vectors of all pixels.In order to adaptively determine the importance of every feature in the segmentation processing,every feature component in the feature vector is given a contribution weight.Then Bayesian paradigm and energy function are respectively used to combine the contribution weight set and the feature set to build two statistic segmentation model based on feature weighted,including the Bayesian segmentation model based on feature weighted and the energy segmentation model based on feature weighted.In the simulation processing of the Bayesian segmentation model based on feature weighted,Reversible Jump Markov Chain Monte Carlo（RJMCMC）and Expection Maximization（EM）algorithms are used to segment regions and estimate the values of contribute weight.In the RJMCMC algorithm,three move types,including:sampling parameter vector,samping label field and sampling the number of regular blocks,are designed by the segmentation model.In the simulation processing of the energy segmentation model based on feature weighted,the RJMCMC algorithm is used to segment regions and determine the values of contribute weight.In the RJMCMC algorithm,samping label field,sampling the number of regular blocks,and sampling contribution weight,are designed by the segmentation model.Experiments are tesed on high resolution remote sensing images（including Synthetic Aperture Radar（SAR）images,panchromatic remote sensing images and color remote sensing images）.The results of quantitative evaluation and qualitative evaluation demonstrate the feasibility and effectiveness of the proposed methods.There are 110 figures,30 tables and 126 references.