Several Studies In Polarimetric Information And Image Processing For Polarimetric SAR Images

Full polarimetric synthetic aperture radar(SAR)usually alternately transmits two orthogonal polarizations and then both received polarizations,such as the horizontal polarization and vertical polarization.With higher dimensional information,Polarimetric SAR images provides richer structure features about ground targets than single polarimetric SAR.In the polarimetric SAR image,the scattering matrix expresses polarimetric information,and is used to extract target features and classify ground types.Therefore it is meaningful to study the polarimetric SAR image.The process of polarimetric SAR images can be divided into two categories:the polarimetric information extraction and remote image processing.Thus this dissertation focuses on the polarimetric information extraction according to the new SAR system,and improves the polarimetric parameters estimation with the structural feature in the image.Specifically three aspects of the contribution are as follows:In chapter two,The polarimetric orientation angle and the three-component decomposition for compact polarimetric SAR image are strictly derived.The compact polarimetric mode makes the system lower burden than the fully polarimetric mode,while there is more information loss.Much work have focused on the reconstruction of the compact polarimetric information.However,different compact polarimetric modes have not the same loss.In the order to reconstruct the fully polarimetric information from compact modes,the detailed analysis on the compact modes is done.Firstly due to the affect of the terrain slope,it is difficult to identify the ground object type and extract the scattering characteristics from the polarimetric information with orientation angles.Therefore the polarimetric orientation angle estimation and its impact on the scattering characteristics is still a hot research topic.Since the orientation angle estimation with the circular polarization state is accurate and efficient,which is commonly used in fully polarimetric mode,it is used to analyse the changes under different compact polarimetric modes.In addition,a three-component model-based decomposition algorithm is evaluated in the compact polarimetric mode.After that,three scattering mechanisms(plane scattering,double scattering,volume scattering)for terrain targets are available.According to the distinguishable information of gound targets,It studies the unsupervised classification for polarimetric SAR images in chapter three.Based on find of density peak method,it studies the classification from two perspectives.The first strategy is to construct a initial database with the k-mean method,where the metric is measured by the Chernoff distance,which is information entropy.So the next step is to find density peaks of the database,and pixels in the image are searching for their own category according to the density around the center.Another strategy is to apply the find of density peak method directly in the H/a/A/SPAN SPAN space.However,the following density is unstable,because of the unstable scattering of the boundary and the strong points in the polarimetric SAR image.The saliancy image which is based on information entropy is proposed to remove these points before classification.The feature in H/a/A/SPAN space of the remaining pixel is weighted by the saliency value.And the find of density peak method is used to realize classification.The main works in chapter four are speckle filtering for the polarimetric SAR image and network training.In the traditional polarimetric SAR denoising process,the improved Nonlocal method is proposed which is based on the polarimetric entropy.It uses the local structure information to measure the similarity of pixels.And it needs to design a nonlinear function which is a time-consuming work.With increasing resolution in SAR images,the structure information will become more important.Deep learning has already proved its effective in image processing.During the multi-layers denoising encoder network training,the network captures the structure information of the image,and it is also useful for SAR images.Due to the limited SAR image resourses,we use optical images for training the multi-layers network.