Ensemble of Winter’s belief based frameworks for Hyperspectral Endmember Extraction

Abstract

The goal of Hyper-Spectral Unmixing (HSU) is to decompose HSC imagery into a group of pure material spectral signatures also known as endmember signatures and fractional proportions weights that characterize the contribution of each endmember in forming a sample. In 1999, Winter's proposed an idea to address HSU that considers vertices of a simplex whose volume is maximum as pure pixel vectors. In hyperspectral remote sensing, this belief has much influence on HSU, especially on endmember extraction methods. Moreover, this belief has inspired much attention, resulting in various endmember extraction frameworks such as Simplex Growing Algorithm (SGA), N-point FINDeR (NFINDR), Alternating Volume Maximization (AVMAX), Successive Volume Maximization (SVMAX). In this paper, we propose an ensemble of these frameworks intending to utilize the best part of the result of each framework. The proposed ensemble framework uses a majority voting approach. Our experiments, applied on four hyperspectral datasets (Cuprite, Urban, Samson and Jasper), expose that the ensemble framework by majority voting can provide efficient and competitive performance compared to individual winter's belief-based endmember extraction frameworks.