Comparison and fusion of co-occurrence, Gabor, and MRF texture features for classification of SAR sea ice imagery

Image texture interpretation is an important aspect of the computer-assisted discrimination of Synthetic Aperture Radar (SAR) sea-ice imagery. Co-occurrence probabilities are the most common approach used to solve this problem. However, other texture feature extraction methods exist that have not been fully studied for their ability to interpret SAR sea-ice imagery. Gabor filters and Markov random fields (MRF) are two such methods considered here. Classification and significance level testing shows that co-occurrence probabilities classify the data with the highest accuracy, with Gabor filters a close second. MRF results significantly lag Gabor and co-occurrence results. However, the MRF features are uncorrelated with respect to co-occurrence and Gabor features. The fused co-occurrence/MRF feature set achieves higher performance. In addition, it is demonstrated that uniform quantization is a preferred quantization method compared to histogram equalization.