GAEEII: An Optimised Genetic Algorithm Endmember Extractor for Hyperspectral Unmixing

Abstract

Endmember Extraction is a critical step in hyper-spectral unmixing and classification providing the basis to applications such as identification of minerals [1], vegetation analysis [2], geographical survey [3] and others [4] [5]. It determines the basic constituent materials contained in the image while providing the requirements to the abundance inversion process, used to obtain the percentage of several endmembers in each pixel. Nevertheless, low spatial resolution and computing time are two major difficulties, the first due to the spatial interactions of different fractions of mixed endmembers and the second due to the strict and extensive search utilized in state-of-the-art methods. In this paper, we propose a novel endmember extractor, so-called GAEEII, based on a multi epochs genetic algorithm with enhancements to the naive genetic algorithm endmember extractor (GAEE). We introduce the following additions to the GAEE: a two-dimensional gene initialization, a permutation crossover, a 2D step Gaussian mutation, and an epoch ensemble. To demonstrate the superiority of our proposed method, we conduct extensive experiments on several well-known real and synthetic datasets, as well as a possible relation to the spectral angle distance (SAD) and the volume of the simplex. The results confirm that the proposed method considerably improves the performance in accuracy and computing time compared to the state-of-the-art techniques in the literature including recent developments.