Two-Stage Evolutionary Algorithm Based on Subspace Specified Searching for Hyperspectral Endmember Extraction

Abstract

In recent years, the introduction of multiobjective evolutionary algorithms (MOEAs) into the field of endmember extraction (EE) in hyperspectral unmixing has demonstrated a breadth of results that surpass those derived from single-objective-based methodologies. Despite these advancements, the adaptation of MOEAs to EE and the attainment of globally optimal solutions represent unresolved challenges meriting continued exploration. This study addresses two principal obstacles in MOEA-based EE: the notorious “curse of dimensionality” in high-dimensional optimization, and the difficulty in striking a balance between convergence and population diversity. We propose a two-stage, evolutionary-based EE algorithm, referred to as TSEA, designed to confront these issues. A novel solution space splitting strategy is incorporated into TSEA that efficiently mitigates the curse of dimensionality by strategically contracting the search space. This advantage is largely attributed to the significant reduction of invalid solutions achieved through the simple application of a clustering procedure. Furthermore, a two-stage optimization approach is employed to meticulously uphold the convergence and diversity of the population, aiming to attain the optimal solution within the realm of high-dimensional optimization. Empirical evidence from four real hyperspectral images demonstrates that the proposed TSEA outperforms other comparison multiobjective optimization algorithms. Thus, this study contributes to the ongoing discourse on the optimization and applicability of MOEAs in the context of EE.

Keywords

• Endmember extraction (EE)
• hyperspectral remote sensing
• multiobjective optimization