Polarimetric scattering topic model for Pol-SAR image annotation

In this paper, we propose a novel multi-level max-margin discriminative analysis (M³DA) framework, where the maximum entropy discrimination latent Dirichlet Allocation (MedLDA) model is introduced, for semantic annotation task of high-resolution polarimetric synthetic aperture radar (PolSAR) image. In our framework, low-level intrinsic feature description is represented from polarimetric scattering signature by spherical local embedding (SLE); Based on polarimetric bag of words representation (PBOW), word-level and topic-level max-margin discrimination are achieved by SVM and MedLDA respectively. Moreover, the large margin nearest neighbor (LMNN) classifier takes another important part in the M³DA, since it can optimize a joint soft label posterior composed of word-level and topic-level discriminative probability. The experimental results demonstrate that the proposed polarimetric topic model in multilevel semantics scheme can capture a suitable mid-level representation for complex target polarimetric scattering and improve the annotation performance.