Nonlinear Compressed Sensing-Based LDA Topic Model for Polarimetric SAR Image Classification

In this paper, a nonlinear compressed sensing-based LDA Topic (NCSLT) model is proposed for the classification of polarimetric synthetic aperture radar (PolSAR) images. The CS theory shows that when a signal is sparsely rendered on some basis, it can be recovered exactly by a relatively small set of random measurements of the original signal. In this paper, such notion is applied to a more general case to analyze nonlinear PolSAR data. Therefore, the NCSLT model is presented with the following two objectives: 1) to capture the nonlinear structure of PolSAR data on a manifold surface using the CS theory and 2) to provide a generative explanation for the relationship between the image pixels and high-level complex scenes for image classification by establishing a Texture-CS-Topic model. Compared with the other traditional SAR image-classification methods, the proposed method displayed potential achievements when applied to two sets of PolSAR image data.