Image Classification And Dielectric Information Extraction Based On Fully Polarimatric SAR Data

Compared with traditional visible, near-infrared sensor, SAR has an unparalleled advantage, which works all-time, all-weather and worldwide over the ground, and it could avoid clouds, rain and other adverse weather. The mathematical basis of polarimetric SAR is Maxwell’s equations, and the basic principles are radar equation, polarized electromagnetic wave theory, polarimetric target decomposition theory, microwave scattering theory. Moreover, the polarimetric SAR provides new methods for image classification, object recognition and surface parameter inversion by detecting richer target scattering information, in a word, the study of polarimetric SAR applications have important theoretical and practical significance.The Paper Select Xingcheng, in Liaoning Province as a study area with certain representation, the experimental area covers most of the feature types, such as coastal, floodplains, towns, farmland, woodland and other types of regions. Two aspects of the thesis were performed by using ALOS PALSAR fully polarimetric SAR data in study area, which are image classification and dielectric constant inversion, and obtained the following innovative results:1. Study on unsupervised image classification method based on polarization decomposition theory.Thesis build upon the decomposition polarization theory, and furthermore, we get the parameters named entropy(H) and anti-entropy(A), which describe objects scattered randomly degree, as well as another parameter named average scattering angle(α) stand for the differences of scattering mechanisms of scatterers. Authors determine the type of feature by studying the position of scatterer in H/α scattering plane, and took Wishart classifier to cluster analysis. Wishart distance estimation which established on complex Wishart probability density function is very steady and easy to use in polarimetric SAR classification application, because the polarization scattering matrix obey multivariate complex Wishart distribution and do not susceptible to be influenced by polarization scaling. After taking anti-entropy into H/α/Wishart classification, polarization information is so richer as to improve the situation that similar natural features clustering may leap decision plane boundaries so that they may randomly assign to different categories, as a result for this, the classification results generate less noise, and category meticulous accuracy is improved.2. Discussion on the applicability of microwave scattering models in study area.Different microwave scattering model is suitable for different surface roughness scales, the small perturbation model is suitable for a relatively smooth surface, the quantitative description is kh0<0.3,k L<3,s<3; Oh model is widely used because of a broader roughness, and the quantitative description is 0.1 <kh0 <6, 2.5 <k L <20. Taking the actual situation into account and combining with results of previous studies, thesis inverses roughness parameters in the study area, root-mean-square height(h0), root-mean-square slope(s) and the correlation length(L), the inversion proved that the selected models are appropriate in study area in terms of roughness.3. Thesis simplifies the small perturbation model and inverses relative dielectric constant by the simplified model.The existing small perturbation model involves various parameters, such as polarization backscattering coefficient, permittivity, permeability, rms height, spectral density, Fresnel reflectivity, shadow function, radar incidence angle, wave numbers, and many other parameters, so that it isn’t easy to calculate the unknown quantities. During the study on the model, we find that it will be able to simplify the process if we make ratio operation by horizontal polarization and vertical polarization backscatter coefficient. By studying the magnetic permeability and combining with previous research results, after make the permeability constant value equal one, the model only remain the relationship among the polarization, wave numbers, the incident angle and the relative dielectric constant. The inversion result of simplified small perturbation model is consistent of actual dielectric situation in the study area.4. Try to use Oh model in light snow-covered area, and make a good dielectric constant inversion result.Compared with the regional lithological geological map, the inversion result of Oh model with a high dielectric constant value could distinguish the red granite well, and the shape of dielectric constant is consistent with the boundary of red granite. However, the dielectric constant distinction of other lithologic is not clear. It may be influenced by the actual surface(such as vegetation and snow cover, etc.). Compared with the same-phase Landsat7 742 band composition image, we find the study area is covered with snow when ALOS transited, so the authors tried to use Oh model in light snow cover area. The study results show that the dielectric constant is small where is obviously snow-covered, maybe because dielectric constant of dry snow is relatively small, on the contrary, the high dielectric constant value is in the snow melt zone because the soil moisture increasing may led to the dielectric constant increase after snow melting. Absolute error of inversion dielectric constant verifying by rock samples collected in the field is about 10%, and initially, we prove that Oh model can be used in inversing relative dielectric constant in shallow snow-covered area.