Research On Polarimetric Synthetic Aperture Radar Interpretation

Synthetic Aperture Radar(SAR)is a type of active microwave imaging system,with advantages of working with all-weather and all-time,which has been extensively applied in military and civilian areas,and shows remarkable ability and irreplaceable in investigation and battlefield situation assessment.Compared with the optical remote sensing image,SAR images contain an abundant of polarimetric information which effectively explore physical and structure features of target.However,since SAR images are formed by unique imaging mechanism,the readability of SAR images is often poor,and lots of polarimetric information embedded in SAR image is not perceived by human eyes.So,SAR image interpretation is the key to the SAR image application.To facilitate human interpret SAR image,three key and widely applied interpretation techniques—SAR images despeckling,polarimetric information visualization and polarimetric SAR images classification—are researched and a useful SAR image interpretation system is built in this thesis.Since the speckling noise is very strong,suppressing strong speckling meanwhile preserving weak details is still an opening question.To achieve a better balance between suppressing strong noise and preserving weak details,based on the sparse representation which has the advantage of preserving image details,for the images with strong noise,we introduce sparse domain subspace decomposition over the overcomplete dictionary,which use classical sparse representation to obtain an over complete dictionary and identifies the important atoms from the dictionary to form the main sparse domain.The details can be preserved by the overcomplete dictionary,especially weak details,and the strong noise are rejected by the main sparse domain,so the aims of preserve weak details while rejecting strong noise can be achieved.Considering the signals embedded in the group of similar image patches grouped by the nonlocal similarity are almost homogeneous,which can enhance the sparsity level of signals in sparse domain to benefit sparse domain subspace decomposition method to separate the signals and noise in sparse domain.Thus,we combine the nonlocal similarity and the sparse domain subspace decomposition into a unified framework to despeckling,first,the structure similar image patches are stacked into a group according to the nonlocal similarity,and then the sparse domain subspace decomposition method is applied to each group.The experiments show that the proposed method can achieve high performances in terms of both structure details preservation and speckle noise reduction.Polarimetric SAR images information visualization as one of the most methods for human interpretation,aims to display the rich of polarimetric information contained in SAR images with an intuitive manner.For certain military domain,interpreters tend to pay more attention to spatial information of target,and wish to obtain distribution information of all targets by an image.However,some limitations inhere in the existing methods,such as traditional SAR images(an color coded image constructed by polarimetric channel intensity or amplitude images)benefit to show the target which have a large different in polarimetric channel intensity,but are difficult to show the other targets with the similar intensity.In addition,the visualization methods based on the polarimetric targets decomposition can display the polarimetric information by their decomposition components.However these decomposition components are not independent,so the information of all targets cannot be displayed by one color coded image synthesized by three decomposition components.In this thesis,we propose a new polarimetric SAR image information visualization method,which content adaptively selects the main information components from the quasi-overcomplete feature space spanned by lots of polarimetric decomposition components to form an SAR "information image".The selected main information components can almost cover the whole feature space,so the distribution information of all targets can be shown in such an "information image".It is benefit for interpreters to make a quick and reliable assessment of the battlefield situation and striking effect.Sparse representations have shown great potential for polarimetric SAR images classification.However,since the targets belonged to different class often share the similar features,and the accurate training sets are difficult to obtain,the learned dictionaries for each class have to represent these shared features,which make them unable to focus on obtaining discriminative features.To enhance the discriminative capability of dictionary classifiers,this thesis proposes a new discriminative structure dictionary learning model.The proposed model learns an overcomplete structure dictionary consisted of two kinds of sub-dictionaries by imposing the sparse and low-rank regularity,that is,the common sub-dictionary and class-specific sub-dictionaries.So the inner-class features and inter-class features can be separated,and the class-specific intrinsic features can be learned from the inner-class features for classification.More specifically,the representation coefficients corresponding to common dictionary are imposed on the sparse regularity to make common dictionary obtain the inter-class features,thus the inner-class discriminative features are concentrated on the class-specific dictionaries.And the coefficients corresponding to class-specific dictionaries are imposed sparse and low-rank regularity to make class-specific dictionaries capture the discriminative intrinsic features.So such overcomplete structure dictionary has both good ability of representation and strong ability of discrimination.Based on the proposed dictionary learning model,we present a polarimetric SAR image classification method based on the discriminative structure dictionary learning model,considering the common dictionary without ability of discrimination,the classifications of test samples are performed according to the reconstruction error based on the class-specific dictionaries,with the discriminative instrinc featurs captured by the class-specific dictionaries,the proposed method can achieve a good classification performance.Polarimetric SAR image interpretation system extracts the information from the SAR images to visualize them for helping the manual interpretation,according to the application requirements.It aims to visualize the rich of information contained in SAR images by a scientific and intuitive manner,so as to the ordinary people can also interpret the SAR images.It has following features,compatible with various SAR systems,the richness of SAR image features,the diversity of visualization and real-time visualization.