Study On Segmentation And Classification Algorithm Of Polarimetric SAR Images

Synthetic aperture radar (SAR) instruments have been widely used in the past years for remote sensing applications such as agriculture, geology and military surveillance. Precise segmentation and classification of different types of targets in SAR images is a crucial step for SAR image understanding and interpretation. Particularly, segmentation and classification of polarimetric SAR (PolSAR) image is a hot topic in SAR applications since its polarimetric scattering matrix consists of more ground target information.Due to the complexity of land feature, however, there are still many problems such as lack of statistical prior knowledge and insufficient description of physical property of target by the present eigenvalues. The problems are blocking the application of PolSAR processing methodology and how to improve the accuracy and robustness of segmentation or classification algorithms is generally acknowledged as a difficult problem.Recently, more and more study has been concentrating on the partial differential equations (PDE) based image processing approaches and their applications. Based on the analysis of the existing SAR image segmentation and classification methods, this dissertation proposes to study the SAR image segmentation and classification by using curve evolution theory and level set method which are both under the framework of PDE. The primary contents and the academic contributions are as follows:1) Following the detailed analysis of the region as well as boundary properties presented in SAR images, both parametric and geometric active contour models are presented. A more appropriate energy functional is derived by sufficiently using the image information. A level set SAR image segmentation method with joint region-boundary information is presented in this dissertation. The level set based approach has better segmentation performance and it has the ability to deal with the topology variation of active contours and to partition the multiple regions simultaneously.2) A variational model for SAR image classification is presented. The model conserves the edge information and is suitable for image restoration since it integrates with regularization. The variational classification method which is easy to implement and has little time cost classifies the ground objects in PolSAR images with high accuracy and restrains the influence of the speckle noise.3) A multi-region SAR image segmentation approach based on partial differential equation is proposed. The method sufficiently exploits the edge information and avoids the drawback when the segmentation merely depends on image gradient since it well integrates the gradient information and the statistical property of different regions. It needs no additional parameters and can estimate the region numbers. Hierarchical splits energy functional is used to get better segmentation results.4) A PDE model adapted for PolSAR segmentation is presented. The segmentation of PolSAR is implemented by using curve evolution and level set method. The polarimetric information is included in the model and it is employed as a criterion of curve evolution to control the movement of active boundary The criterion guarantees the correct segmentation and avoids the over segmentation problem commonly occurs in level set method.5) Detailed experiments are designed and different types of data, such as synthetic images, real SAR images and PolSAR images, are used to verify the performance of the segmentation and classification approaches.In general, study on the PDE based SAR image processing method provides a new way for SAR image interpretation and target recognition and in turn is a notable promotion for the development of image processing methodology under PDE framework.