Classifying PolSAR Images Based on Multiview Unsupervised Feature Selection via View-Specific and Cross-View Representations

Multiview features can be acquired from multiple feature extractors or frequency bands to describe polarimetric synthetic aperture radar (PolSAR) images. These multiview features usually have high dimensionality and contain some redundant and irrelevant information leading to classification performance deterioration. Additionally, acquiring label information on PolSAR data is challenging. To resolve these difficulties, we put forward a multiview unsupervised feature selection method via view-specific and cross-view representation (VSCVR) to classify PolSAR images, resulting in retaining the global and local structure of the data and exploring higher order relationships between different views. Specifically, we learn view-specific self-representation with the tensor low-rank constraint for the global structure preservation and exploration of high-order consensus and complementarity between different views. We adopt tensor nuclear norm by tensor singular value decomposition (t-TNN) to achieve the low-rank constraint for multiview self-representation tensor. In addition, we incorporate cross-view common representation with manifold regularization to sustain the local structure information and ensure consistency across multiple views. To accurately describe the local structure, we construct a similarity graph for manifold regularization via the Wishart distance reflecting the statistical property of PolSAR data. Furthermore, a group sparsity constraint is imposed on view-specific projection matrices for feature selection. We propose the corresponding algorithm to handle the VSCVR model and conduct complexity and convergence analysis. Numerical and visual results on some real PolSAR datasets certify that VSCVR can acquire a few important features and improve classification performance effectively.