ANSGA-III: A Multiobjective Endmember Extraction Algorithm for Hyperspectral Images

Abstract

Endmember extraction is the foremost step in hyperspectral unmixing and has always been one of the most challenging tasks in hyperspectral image processing due to the intrinsic complexity of hyperspectral images. Over the past few decades, a number of endmember extraction methods have been proposed, most of which may not be able to express all the characteristics of each endmember and cannot meet the assumptions of the simplex structure because of the complexity of remote sensing images. Recent progress has shown that endmember extraction methods can be transformed into a multiobjective optimization problem, with the aim of generating a set of Pareto-optimal solutions. However, the existing multiobjective endmember extraction methods cannot obtain complete nondominated solutions and have a high time complexity. To resolve this problem, this paper proposes an adaptive-reference-point-based nondominated sorting genetic algorithm (ANSGA-III) for endmember extraction, which adaptively updates and includes new reference points on the fly. In ANSGA-III, two objective functions, i.e., volume maximization and root-mean-square error minimization, are simultaneously optimized. Experimental results on three real hyperspectral remote sensing images show the superior performance of the proposed ANSGA-III approach.