Texture representation of SAR sea ice imagery using multi-displacement co-occurrence matrices

In this paper, we describe multi-displacement co-occurrence matrices for representing sea ice textures of SAR imagery. Our design of co-occurrence matrices captures local relationships among neighboring pixels and global links among distant pixels, an advantage over other existing versions of co-occurrence matrices. As a result, it can adequately represent micro textures, such as grainy details, and macro textures, such as patchy blocks. We have conducted experiments to compare our multi-displacement co-occurrence matrices with other existing versions using Bayesian linear discrimination. We have found that our design is the most texturally representative in terms of classification accuracies in both training and test datasets. In addition, we have applied this design to sea ice texture analysis which includes detection and localization, and subsequent image-texture mapping.