Sea ice types and concentrations are of great importance for ship navigation in or near the ice. The evaluation of ice types and properties using synthetic aperture radar (SAR) imagery has attracted much attention in recent years. SAR sea ice images usually have consistent textures that can be utilized for sea ice description and classification. Therefore, methods based on texture discrimination could be designed to identify ice types and evaluate ice properties by machine without human intervention.; This thesis contributes to the ice identification problem mainly by investigating the feature extraction phase in a texture classification process. A review is given of several different approaches including Gray Level Co-occurrence Matrices and Gabor filtering, while the emphasis is on those based on the wavelet transform techniques. Comparative studies have been conducted on both the selection of wavelet band signatures and of wavelet kernels.; A new wavelet band signature, named wavelet entropy , is proposed and applied to texture classification with encouraging results. This technique extracts features from wavelet band histograms. A promising aspect of this new technique is that it provides estimates of probability measures of the texture memberships. These membership probabilities have been used in a ship navigation application with interesting results presented in the thesis.; Texture orientation issues are also addressed in this thesis. Because of the oriented structures apparent in some SAR sea ice textures, it is desirable to extract rotation invariant features. Some new work is presented that has achieved this goal to some degree by DFT encoding on the features of different orientations, obtained via the complex wavelet transform instead of the traditional discrete wavelet transform to separate the mixed diagonal directions.
展开▼
