Local Prototype Space-based Band Selection for Hyperspectral Subpixel Analysis

In this paper, two unsupervised local band selection (BS) methods in the prototype space (PS) for improving subpixel analysis performance are proposed. Here, the PS is a two dimensional space which is constituted of the target spectrum and that of the local background. The proposed methods benefit from local background characterization through image clustering. These BS methods select the discriminative bands in two ways: 1) selecting the bands which form a convex hull in the PS, and 2) using a cluster-based approach in the PS to select bands. An experiment applied to real-world hyperspectral data showed that the proposed BS methods improve the performance of constrained energy minimization (CEM) and adaptive matched filter (AMF) subpixel detection methods in terms of the number of false alarms.