SAR Sea Ice Recognition Using Texture Methods

With the development of remote sensing techniques, a vast amount of SAR sea ice imagery is being provided by satellite platforms. As an important aspect of measurement, monitoring, and understanding of sea ice evolution during the seasons, the generation of ice type maps is a fundamental step in the interpretation of these data. The abundant texture information in SAR imagery is useful for segmentation of the pertinent ice types. Many texture analysis approaches have been proposed in the literature which can be identified into three categories: geometrical-based, statistical-based and model-based. Among these methods, to explore a coherent theoretical framework to support robust and powerful algorithms for sea ice segmentation is necessary. This thesis will focus on two methods: gray level co-occurrence probability (GLCP) method and Markov random field (MRF) method. GLCP method can extract texture features with different frequencies along different directions in the image space. Using MRF method, texture is analyzed as having preferred relations and interactions that can be articulated mathematically, and then a Bayesian framework can be employed to make inferences. Quantizations, displacement, orientation, window size and texture statistics are important parameters in GLCP method. Accurate estimation of the texture models is an important concern for a successful segmentation using MRF method. In this thesis, the potential of these two methods for texture analysis and unsupervised segmentation are investigated and tested using synthetic, Brodatz and SAR sea ice images, and their texture distinguishing abilities are also compared.