Semisupervised Multiview Feature Selection for VHR Remote Sensing Images With Label Learning and Automatic View Generation

Abstract

The features of very high resolution (VHR) images can be considered as multiview data. For better analysis of intrinsic data structure, a semisupervised multiview feature selection (SemiMFS) method is proposed to exploit the multiple views in this paper. In SemiMFS, feature views are automatically generated by decomposing features into multiple disjoint and meaningful groups. Each feature group represents a view, and each view describes a data characteristic. Then, features are evaluated and selected within each view. The contributions of SemiMFS are listed as follows: 1) A SemiMFS is proposed for VHR images. 2) ${\ell _{1,2}}$ -norm regularization and automatic view generalization are utilized in semisupervised feature selection for the intragroup sparsity, not the conventional intergroup sparsity, without any prior knowledge. Thus, SemiMFS reduces the redundancy within views by selecting features within each view, and simultaneously preserve as much information as possible by only shrinking the weight corresponding to different views. 3) An improved iterative method is developed in an ${\ell _{1,2}}$ -norm-based minimization problem together with label learning of unlabeled objects. The experiments on three VHR satellite images verify the effectiveness and practicability of the method, compared with traditional single-view algorithms. The experiments demonstrate that the views and the intraview features make sense, and they offer a new way to analyze data structure of VHR images.