Research On Edge And Region Information Extraction For Polarimetric SAR Images

Polarimetric synthetic aperture radar(Pol SAR) is an advanced form of imaging radar and focuses on transmitting and receiving polarized radar waves to characterize the observed land covers and targets. In the last two decades, research has shown that Pol SAR is capable of providing more useful information than single-polarization cases in target detection, terrain classification, parameter inversion, topography extraction,and so on. Several operating spaceborne platforms are providing an enormous quantity of Pol SAR data; most of the spaceborne SAR systems, which will be launched in the future or are being researched and developed, are capable of providing multipolarization data. To extract the information in these highly complicated images is the key to breaking through the bottleneck of their applications. Therefore, developing automatic or semi-automatic systems for Pol SAR image interpretation and information mining is urgently required and widely studied.Researchers have developed numerous analysis methods for Pol SAR images. Statistical modeling and scattering analysis are two main ways. This paper is dedicated to spatial analysis. Spatial information is the information which is reflected and provided by some patterns generated by pixels in the spatial domain, such as edges, lines, regions, and textures. We consider this for three main reasons: a) it exploits the information of Pol SAR data in a deeper way; b) it takes the fact that radar images are sensitive to structure into consideration; c) it supports the upcoming interpretation tasks, such as high resolution Pol SAR image understanding, multi-source data fusion, and multi-temporal data analysis.This paper researches on the extraction of typical spatial information, edges and regions, for Pol SAR images in a comprehensive and systematic way, and provides methods specially designed for Pol SAR image objectification and description.It also exploits the spatial information to improve the performance of Pol SAR image interpretation, and develops Pol SAR image processing chains integrated with spatial information, such as land cover classification and change detection; this provides evidences for the availability and advantage of the proposed designs and algorithms.The theoretical bases of this paper are Pol SAR image statistical modeling and similarity measures. The core principle of spatial information extraction is to numerically measure the similarity between one pixel and its neighbor in the spatial domain.Since the Pol SAR data are contaminated by inherent speckle noise, the similarity measure reflects the similarity in the statistical sense between two entities in the image.As the core bases of Pol SAR image spatial information analysis, first of all, this paper introduces statistical modeling for Pol SAR images and its scope of application; based on the statistical modeling, similarity measures between two pixels in Pol SAR data are presented; with the research focus of this paper taken into consideration, the probabilistic similarity measures used for Pol SAR image spatial information analysis are studied in a deeper way.Based on the statistical modeling and similarity measures, according to the needs of applications and the state of the art, this paper focuses on developing extraction methods of edge and region information for Pol SAR images. The main contributions are listed as follows:Resolution-kept and precisely-localized edge extraction – This paper presents an edge extraction method for Pol SAR images based on the degenerate filter and weighted maximum likelihood estimation. The probability distribution parameters needed in the degenerate filter are estimated by the principle of the weighted maximum likelihood estimation. The proposed method addresses the problems which limit the performance of the traditional method. The experimental results show that, compared to the traditional method, the proposed method provides better performance, finer extraction and more precise localization.Superpixel generation – This paper presents a superpixel generation method for single- and multi-temporal Pol SAR images. The edge information is used as the information source, and used as the bridge to link the local statistical information and global similarity measure. The superpixels are generated by a process of optimal oversegmentation. On the basis of the superpixels, this paper presents the superpixel-based number-of-classes-adaptive classification method and the multi-temporal superpixel-based change detection method for Pol SAR images. The experimental results demonstrate the availability of the proposed methods.Spatially adaptive region generation – This paper presents a spatially adaptive representation and region generation method based on the wedgelet approximation and analysis specially designed for Pol SAR data. The data fidelity measure derived from the statistical information is used for the multiscale wedgelet decomposition, the bottom-up quadtree pruning is used for the multiscale wedgelet representation, which provides the initial spatially adaptive region generation, and the Wishart Markov random field model is used for the region generation refinement. The experimental results demonstrate the advantage of the proposed method for region generation for Pol SAR images with heterogeneous scenes.Globally dominant land cover boundary delineation – This paper presents a delineation method of globally dominant land cover boundaries for Pol SAR images based on image representation. This paper presents the regularized L0 gradient minimization representation specially designed for Pol SAR data, proposes to analyze the image in a global and whole way, and combines the statistical characteristics and spatial information. The experimental results demonstrate that the proposed method gets effective contours of land covers for Pol SAR image representation, namely, smooth contours in homogeneous areas, as well as detailed contours in heterogeneous areas.In this paper, the performance of each proposed method is presented and analyzed on simulated and/or real experimental data sets, with visual presentation as well as numerical evaluation. The experiments and results confirm that it achieves the goal of design, and demonstrate its availability and advantage. The experimental results demonstrate that the proposed methods provide means for Pol SAR image objectification and description, and they can improve the performance of Pol SAR image interpretation. It is worth mentioning that, at this stage, the methods are tested on limited,often-used experimental data sets. In the future, we need to apply the proposed methods to a larger Pol SAR image database to develop fully operational procedures and to allow end users to benefit from them.