Coastline Detection And Terrain Classification In SAR Images

SAR(Synthetic Aperture Radar) image processing, which has been widely applied in many fields, is one of the core ways to gain the information. In nowadays, the SAR imaging technology has been mature, and the speed of gaining the SAR data, as well as the quality of the SAR data, have increased greatly, which will increase the demands for the SAR image processing technology. More and more high-resolution SAR data measured make the fast and automatic SAR image processing, become an important research field. This thesis, mainly focuses on the introduction of coastline detection and image classification methods. When introducing the coastline detection algorithms, we mainly study two basic detection algorithms. When introducing the image classification algorithms, we study the algorithms based on texture features. We first extract the texture features based on the superpixel segmentation, and then an appropriate classifier is adopted for classification. By further combining with the contextual information of the SAR image based on the markov random field model, the classified result can be improved. The main work of this thesis is summarized as follows:1.Two kinds of basic coastline detection methods are introduced. The first one is the edge tracing, which was proposed by J.Lee and I.Jurkevich in 1990. This is a method that is fit for SAR image coastline detection, and many techniques of the method are still used today. This method includes two parts: the first one is image preprocessing, and sigma filtering is introduced mainly; another part is the boundary tracing, which means tracing the continuous coastline in the detection result. The second one is the threshold method, which is one of the simplest segmentation method, but it is sensitive to speckle. A filtering method using variance is mainly studied in this paper. The Otsu method is used to gain the optimal global threshold value which is used to segment the sea and the land to gain the coastline.2. Texture features are studied in this part. The thesis studies a pixel-level feature extracting method, which extracts the multilevel local pattern histogram(MLPH) using the sliding window. However, the speed of performing the sliding window is slow, and many parameters need to be considered. The superpixel can solve these problems. Therefore, a superpixel-level MLPH feature extracting method is studied. In order to acquire the superpixel, a likelihood ratio based superpixel segmentaion algorithm is introduced. Then the sliding window using in extracting the MLPH is replaced by the superpixel. Finally, the LIBSVM classifier is selected to classify the test samples. In order to improve the stability of the classification algorithm, the thesis also studies combing the classification results using superpixels of various sizes.3.Further improvement of the classification result using texture features is studied. Neighbourhood and cliques are defined based on the superpixel, and by combining with the basic theory of the markov random field model, the priori information of the sample is taken into consideration in the classification. Likelihood energy is defined based on the distance of each sample from the SVM seperating plane. Based on the defined energy function and the MAP-MRF framework, the energy minimization algorithm is used to accomplish the scene classification.