Terrain Classification For PolSAR Image With Scattering Mechanism And Target Decomposition

Polarimetric synthetic aperture radar(polarimetric SAR) is a multi-parameter, multi-channel imaging radar system, which is to obtain polarimetric information about the target by measuring the scattering echo of each resolution cell on the ground, such as scattering matrix, coherency matrix and so on. Compared with conventional SAR images, polarimetric SAR images can provide more information and features for classification. Due to the complexity of polarimetric information, classification for polarimetric SAR image is not only popular but also difficult research. On developing the polarimetric scattering characteristics of targets and improving the interpretation of polarimetric SAR images, the research of polarimetric SAR images classification has important theoretical and practical value. Domestic research of polarimetric SAR image processing is still in the early stage of data interpretation, and the accuracy and stability of classification methods are needed to improve.In this thesis, we focus on target decomposition and terrain classification for polarimetric SAR images, and propose a series of practical and effective target decompositions and terrain classification methods. The main achievements are:1. Terrain classification method with new scattering coefficients for polarimetric SAR image is proposed. Since the eigenvalues and eigenvectors of the coherency matrix is too complex to describle the polarization characteristics of the targets, according to the physical meaning of the eigenvalues and eigenvectors, three coefficients associated with the scattering mechanism are proposed. They are named scattering coefficients, which are single-target scattering coefficient, double-target scattering coefficient and three-target scattering coefficient. The scattering coefficients imply the number of scattering mechanisms on each pixel. The availability and efficacy are proved by real polarimetric SAR data. For describing the inherent physical characteristics of polarimetric SAR data more accurately, we improve the proposed scattering coefficients. According to the entropy of targets, the different scattering vectors are set. Moreover, the priori probability of the scattering coefficients is considered. By comparing with the current scattering coefficients, the improved scattering coefficients are more realistic for the common targets. Via the actual polarimetric SAR data, the validity of the improved scattering coefficients on classification of polarimetric SAR image is proved.2. A novel version of Freeman/Eigenvalue decomposition with non-reflective symmetry scattering models is proposed. Existing Freeman/Eigenvalue decompositions requires the assumptions of reflection symmetry for the polarimetric SAR data, but the actual data are often unable to meet the reflection symmetry. Aiming to this problem, surface scattering model and double-bounce scattering model with non-reflective symmetry are proposed to improve Freeman/Eigenvalue decomposition. The proposed Freeman/Eigenvalue decomposition has three advantages: First, it donot need reflection symmetry, which is more suitable for the characteristics of the actual data, especially in complex artificial areas; Secondly, the scattering powers are linear combinations of the eigenvalues of the coherency matrix, therefore the scattering powers are all rotation invariant parameters; Thirdly, there is no power restriction, and the surface scattering power and double-bounce scattering power are nonnegative values. The actual polarimetric SAR data are used to prove the effectiveness of the proposed Freeman/Eigenvalue decomposition.3. Two novel versions of Freeman/Eigenvalue decomposition with modifed volume scattering models are proposed. The first modifed version is based on the extended volume scattering model. Because random orientation angle of polarimetric SAR data may cause various targets with the same scattering characteristics, two unit transformation matrices are used on polarimetric SAR data for deorientation, which lead to minimize the cross-polarized term and eliminate the negative surface scattering powers and double-bounce scattering powers. Four volume scattering models are used for various types of the targets. Different from the current Freeman/Eigenvalue decompositions, a new discriminant is used to determine the volume scattering from vegetation or artificial areas. If the volume scattering is derived from the vegetation areas, the co-polarized ratio is used to select the appropriate volume scattering model. The second improved Freeman/Eigenvalue decomposition is used the volume scattering models derived from the eigenvectors of the coherency matrix, in addition, we use the differen volume scattering model for vegetation or artificial areas. Via the actual polarimetric SAR data, availability and efficacy of the two modified Freeman/Eigenvalue decompositions are verified.4. The proposed scattering coefficients and Wishart MRF classifier are combined for polarimetric SAR terrain classification. Firstly, according to scattering coefficients derived from the coherency matrix of the polarimetric SAR data, the data are categoried into three rough classes; Then, the scattering powers of Freeman-Durden decomposition are used to refine the data into ten classes; Finally, an adaptive Wishart MRF classifier is iteratively used to obtain a final classification result. In the final step, a limited label switching mechanism is set, the scattering characteristics in the rough and refined classifications are limitedly preserved, and thus we can take the advantages of statistical features, and reduce the error of Wishart classfier. The effectiveness of the proposed classification is proved on actual polarimetric SAR data.