Semiautomatic Extraction Of Typical Linear Target From High-Resolution Polarimetric Synthetic Aperture Radar Imagery

Synthetic Aperture Radar (SAR) is an active microwave sensor, and has the advantage of working all-weather and all-time. Recently, SAR has been widely applied in both military and civil. With the development of radar system, SAR remote sensing is with multi-frequency, multi-polarization, high-resolution. However, compared with the data acquisition, image interpretation was less developed. Typical linear targets are important to military, politics, and economics, such as roads and transmission lines. Therefore they become significant parts of topographic features to be interpreted. However, due to the disturbance by speckle and the complxity of linear features, it is very challenging to interpret these targets with high confidence. Furthermore, the existing methods were mainly focused on the target extraction from the low-and median-resolution images. The high-resolution polarimetric SAR images provide another way to this problem for their rich spatial details and full backscattering information from the terrain targets. Currently, to process and to analysis such information is one of key issues for linear target extraction. Based on the current theory and technologies, it would be hard to realize fully automatic linear feature recognition. Therefore, the semiautomatic method is more prefered, because it can make use of both the accurate and fast computing ability of computers and the high reliable pattern recognition ability of humankind, and has good potentials in applications.According to the published papers, there are kinds of linear features in high-resolution polarimetric SAR images, and different methods are required to deal with different targets for their individual geometric and scattering characteristics. For example, roads, appearing as long and dark areas, varies with shapes and sizes, which are depressed by speckle seriously and disturbed by the objects besides them, over-line bridge, river bridge, cars or trucks, etc. transmission line as one of other typical linear targets, is of low signal-to-clutter ratio, and difficult to detect by traditional methods using intensity images. To solve these problems, aiming at better utilizing the scattering and geometric characteristics of interesting targets, we focused our research on key problems of the semiautomatic extraction of typical linear features from high-resolution polarimetric SAR images. After the research of the edge detector, speckle filter, road seeds extractor using one class classifier of polarimetric SAR images, semiautomatic road tracer, and transmission line detector using polarimetric coherence, and we finally constructed a semiautomatic working framework of "edge detecting—line preservative preprocessing—seed extraction—target tracing—special target detecting". The validity of the proposed method is demonstrated by a set of experiments. Overall, the research work in the paper contains some contributions as following:Firstly, aiming at to preserve linear features, a filter fusing a spatial detail-preserving filter and polarimetric scattering preserving filter was proposed. To avoid enhancing false edges and smearing true edges, two more sub-windows--the ribbon-like and the square-shaped masks--are introduced compared with the refined Lee filter. To preserve the scattering mechanism in the homogeneous areas, the filter is carried out only for the pixels with the same scattering class. Finally, both filters are fused based on the spatial structure type. The speckle reduction, spatial preservation, and scattering preservation are evaluated by the experiments qualitatively and quantitatively.Secondly, to make full use of polarimetric data, we proposed AOPCE-ROA for edge detection by combing adaptive polarimetry optimal contrast enhancement and ratio of averaging, and derived its fast version. The speckle is suppressed, and a maximum response of edges is obtained, because more scattering information is used by polarimetry synthesis.Thirdly, for the tracer initialization is very important for the efficiency and accuracy, a seed extractor was proposed based on the road seed index and polarimetric Support Vector Data Description (SVDD) one-class classifier. After discussing the one-class classification, the SVDD based on the structure risk minimization and kernel tricks was first introduced to the supervised polarimetric classification, and the optimal feature vector and optimal classifier parameters were discussed. The validity of the method for small training samples was tested by experiments. Using the classification results and the line detector response, a road seed index was proposed and used to extract road seeds.Fourthly, to reduce the disturbance of the obstacles on the road in SAR images, a semiautomatic road tracer by the "predict-measure-correct" under the framework of adaptive Bayesian filter using the profile and rectangular matching methods to obtain the road center measurement was proposed. Specifically, we firstly discussed the road tracer initialization using the seeds extracted previously. Then the observation is obtained by the weighted least square errors methods using the linear geometric and radiometric distortion model. Finally, based on the discrete road tracing system model, the road is traced using the adaptive Bayesian filter, utilizing all the observations current and before, with few supervision from users, which automatically switches the two form of Bayesian filter, the Kalman filter and Particle filter, and automatically adjust the tracing steps. It is demonstrated by the experiment using the developed prototype software that, the improved tracer makes good use of the both filter, can avoid most of the disturbance from obstacles on the road. The roads can be extracted reliably and efficiently, supervised by the user, with the observations measured by the profile and rectangular matching for the narrow and wide roads, respectively.Finally, the scattering from transmission lines is modeled by dipoles, and the transmission lines arranged with different azimuthal angles are detected using a CFAR method of the coherence of the co-and cross-polarizations estimated by the Hough domain, based on the difference of azimuth symmetry of the transmission line the background clutter. The airborne P-band polarimetric SAR data were used to the test the validity of the method.