Combination of contextual information and optimal texture features for improving the accuracy of SAR image classification

This paper employs the full advantages of contextual information and optimal texture features for improving the accuracy of pixel-based classification. In the proposed novel classification method, first, optimal texture features are selected based on the genetic algorithm (GA) and Jeffries-Matusita (JM) distance criterion. Second, the selected texture features are combined with backscattering SAR data, and a support vector machine (SVM) pixel-based classification is done. Finally, integration of Gaussian Markov random field (MRF) model with SVM classifier obtains final classification map. Comparison of the proposed method with pixel-based classification shows a 13.77% improvement in overall classification accuracy of TerraSAR-X images.

Keywords:

synthetic aperture radar (SAR) , Markov random field (MRF) , Texture features , support vector machine (SVM)