Research On Object-based Decomposition And Classification Of High Resolution PolSAR

As the spatial resolution of Polarimetric Synthetic Aperture Radar(Pol SAR)increases,traditional pixel-based imagery interprating methods for low or medium resolution became no longer applicable.Aim at the new quality of high resolution Pol SAR images,this dissertation proposes a research line of “matrix estimation – target decomposition – object construction – object-based classification”,and adopted objectbased thinking to solve the unabundant information utilization and the vulnerability to speckle noise of traditional pixel-based methods.Some adaptive methods are proposed to fit the high resolution Pol SAR images.The major research contents and results are as follows:1)Aiming at the problem that the traditional Sample Covariance Matrix(SCM)estimation is no longer applicable in high-resolution polarimetric SAR images with the mixture of homogeneous and heterogeneous regions,this dissertation proposes a self-selected matrix estimation scheme based on the polarimetric heterogeneity index.The polarimetric heterogeneity index divides homogeneous scene from heterogeneous scene,and leverages SCM estimation in homogeneous scenes while utilizes Fixed Point(FP)estimation in heterogeneous scenes.Additionally,after comprehensively analysis the transition scenes between homogeneous and heterogeneous scenes,a totally adaptive matrix estimation is further proposed,which uses a linear combination of SCM and FP estimation as output in mixture regions.2)Aiming at the problem that traditional superpixel can't utilize all information including in the image,this dissertation proposes a Wishart energy based revised Superpixels Extracted via Energy-Driven Sampling(SEEDS)method to generate superpixel.All information contained in polarimetric coherency matrix are considered during the block level update,while in the pixel level update,this dissertation raises Wishart energy to improve the accuracy for evaluating the pixel movement.Finally,an elimination process is incroperated to reduce the potential fragmented superpixels of SEEDS.3)Aiming at the traditional target decomposition only serves coherent or incoherent scenes,a target coherence parameter is proposed to clarify the coherency of targets.Additionally,SDH decomposition and Yamaguchi decomposition are adopted for coherent and incoherent scenes,respectively.The unified decomposition result is further obtained by fusing two above decomposition components.To obtain object-based unified decomposition,merge the components within superpixel,and eliminate the error based on pauta criterion.4)Combining with previous research contents,an object-based classification method with probability Latent Semantic Analysis(p LSA)is proposed to solve the vulnerability of Support Vector Machine(SVM).Firstly,the adaptive coherency matrix estimation method is adopted to get accurate polarimetric coherency matrixs.Then,features are extracted based on pixels,including textures and unified decomposition components,to generate object-based feature vectors based on superpixels.Finally,utilize p LSA to convert low level features to high level semantic features,and get classification results accordingly.The innovations of this dissertation are as follows: 1)The heterogeneity coefficient is raised which can characterize the image heterogeneity,on this basis,the self selective matrix estimation is proposed,and the adaptive coherency matrix estimation is further improved after comprehensively analyzing transiting regions.2)The revised SEEDS algorithm with Wishart energy is proposed to utilizing all information contained in Pol SAR images,especially the polarimetric information to enhance the performance of superpixel segmentation.3)The target coherency parameter is proposed to obtain the object-based fusion of coherent-incoherent unified decomposition,an object-based SVM classification is further raised by combining the adaptive coherency matrix estimation,the superpixel generation,the unified decomposition,and the p LSA.