Context-Based Max-Margin for PolSAR Image Classification

Context-based method for classification has been successfully applied in image. However, most of these classifiers work in stages. This paper presents a novel discriminative model named context-based max-margin to perform the task of classification for polarimetric synthetic aperture radar (PolSAR) images. Based on the max-margin frame, support vector machine (SVM), and conditional random fields (CRF) are used to describe the spectral and spatial information of polarimetric synthetic aperture radar (PolSAR) image, respectively. First, the probabilistic result which is obtained from SVM can be applied as the spectral term of the discriminative classifier. Second, CRF is used to describe the spatial information of PolSAR image. The contextual information of both label and observation field are built as the spatial term, by which the smoother region is obtained and the spatial information is preserved. Finally, a discriminative classifier can be learned by means of integrating the spectral and spatial terms. Compared with other state-of-the-art classification methods, our method exhibits higher accuracy, which indicating the effectiveness of our scheme. Here, the total classification accuracy of the proposed model increases by about 10% and 3% compared with the other methods for two data sets.