Multiview Feature Selection for Very High Resolution Remote Sensing Images

Object based image analysis on very high resolution (VHR) remote sensing imagery often ignores the heterogeneous constitution of feature spaces. In this paper, a supervised multiview feature selection (SMFS) method is proposed. In this method, features are decomposed into multiple disjoint and meaningful feature subsets by employing affinity propagation, where each feature subset represents a view, and each view describes a data characteristic. Features are evaluated and selected within each view. The experimental results on two VHR satellite images, including Quickbird-2 and Worldview-2 images attest to the effectiveness and practicability of the method in compared with traditional single-view algorithms. The results also demonstrate the utility of multiview information in processing VHR datasets.