Sar Image Noise Reduction And Polarization Sar Image Classification Research

Classification of polarimetric synthetic aperture radar (SAR) images is an important research in SAR image interpretation, which has great research significance and appliance value. In this dissertation, image denoising and supervised classification technologies are researched for polarimetric SAR images. Firstly, improved Non-Local Means filter (NLM) is used in SAR images denoising, as a preprocessing for following research. Then spatial pyramid matching technology is introduced for polarimetric SAR images. In this classification method, polarimetric information and texture information around pixels are extracted, and sparse coding is introduced for feature coding. The main work and contributions of this dissertation include:The combination algorithm of the non-local means filter (NLM) and two-dimensional principal component (2DPCA) for SAR image is researched in this dissertation. To overcome the disadvantages of traditional denoising algorithms for SAR images, NLM used in optical images denoising is introduced to SAR images. Meanwhile, it is improved by using pre-selection and2DPCA. Experiments are carried out on the simulated SAR images and the real SAR images, the influence of parameters and the complexity of this algorithm are analyzed at last, which prove the efficiency of this algorithm.In this dissertation, spatial pyramid matching algorithm is researched for polarimetric SAR images. The basic theory of bag of features algorithm and pyramid matching kernel are summarized, as well as the general work framework of this classification method are given in the dissertation. For the low efficiency of vector quantization in feature coding, sparse coding is used as a replacement to reduce the training time. In order to confirm the validity of this algorithm, corresponding experiments are done in a polarimetric SAR data of San Francisco, using MOD dictionary and K-SVD respectively. The results show that training method using K-SVD is better for this study area.On the basis of the spatial pyramid matching using sparse coding algorithm presented above, Locality-constrained Linear Coding (LLC) method is applied to replace Orthogonal Matching Pursuit (OMP) coding. In the experiments of a polarimetric SAR data of Flevoland, Netherlands, twenty polarimetric features are extracted, and then LLC and OMP optimization algorithm are used to coding these features respectively. The results of these two encoding methods show the superiority of LLC in the classification for polarimetric SAR images. And multi-features SVM classification method is applied to the same study data as a comparison of the algorithm in this session, which prove the effectiveness of this algorithm.