Research On High-resolution SAR Image Segmentation And Classification Methods

In recent years, because of the ability to acquire images under all-time and all-weather conditions, synthetic aperture radar(SAR) has become an important tool for remote sensing based earth observation. SAR image segmentation and classification are fundamental problems for SAR image processing and interpretation, playing a key role in putting SAR techniques into practical applications. Therefore, it is worthy for studying the theories and methods of SAR image segmentation and classification.One major trend in the development of SAR imaging techniques is the constantly improvement of the spatial resolution. High-resolution(HR) SAR images are able to provide rich ground information. Meanwhile, new problems are also put forward for SAR image segmentation and classification. Compared to low- and middle-resolution SAR images, HR SAR images have unique characteristics: the statistical properties change, texture patterns present in the image, the scenes become very complex, and the data amount is huge. Segmentation and classification methods that are applicable for low and middle resolution SAR images are not able to cope with HR SAR images due to aforementioned characteristics. In view of the above problems, after surveying and analyzing existing methods, in this dissertation several segmentation and classification methods are proposed for HR SAR images in the consideration of accurate modeling of HR SAR image properties and fully exploiting SAR image information.The main work and contributions are summarized as follows:1. A survey of existing theories and methods for SAR image segmentation and classification is carried out. With the characteristics of HR SAR images in mind, the problems of applying existing segmentation and classification methods on HR SAR images are analyzed.2. Based on the study of high-resolution SAR image statistical modeling theories and parameter estimation techniques, the energy functional model for segmenting high-resolution SAR images is built based on the G0 statistical model, and a G0 model and level set method based SAR image segmentation approach is proposed. Moreover, by further studying the convex relaxing and convex optimization based energy minimization schemes, a G0 model and split-Bregman method based SAR image segmentation algorithm is presented, which can improve the efficiency of SAR image segmentation.3. Considering that HR SAR images contain rich structural information of ground objects and the scene is often very complex, a novel variational SAR image segmentation method which incorporates discriminative learning is presented. A novel variational model is proposed, which is able to combine feature information learned by discriminative methods and a prior regularization information. Based on the proposed variational model, a morphological texture feature based segmentation approach is proposed for HR SAR images. Experimental results reveal that the proposed approach can effectively utilize discriminative information contained in high-dimensional SAR image features, leading to enhanced segmentation accuracy for HR SAR images with complex scenes.4. For classification of SAR images, pixel based methods are often affected by noise and clutters. To tackle this problem, superpixel based HR SAR image classification method is investigated. Firstly, a SAR image superpixel generation approach is proposed based on statistical model and mean intensity ratio. Experimental results demonstrate that this approach is well adapted to SAR images. With generated superpixels, the problem of superpixel feature extraction is studied and a SAR image classification method using sparse coding based superpixel feature extraction is presented. The merits of the proposed classification method are validated with real HR SAR images based classification experiments.5. For classifying HR PolSAR images, novel methods that combine polarimetric information and contextual information are proposed. In the proposed methods, comprehensive polarimetric information is captured by collecting multiple polarimetric features, and the classification is performed based on the theory of sparse representation based classification. To take advantage of contextual information, two novel classification schemes are proposed by unifying the theory of sparse representation based classification with the concept of superpixel. With the proposed approaches, not only is the classification accuracy for PolSAR images improved, the theories and methods of sparse representation based classification are also extended.In summary, a series of studies are carried out centered on the problem of HR SAR image segmentation and classification. This study enriches the theories of SAR image segmentation and classification. It also provides effective methods for information extraction and interpretation of HR SAR images.