MTEA-BS: a multi-objective multi-task evolutionary algorithm for unsupervised band selection in hyperspectral images

Hyperspectral images (HSIs) suffer from high dimensionality, significant redundancy, and noise interference, which make direct processing inefficient. As a result, band selection (BS) is a crucial preprocessing step. Evolutionary algorithms (EAs) have been widely applied to BS problems due to their flexible search mechanisms and strong global optimization capabilities. However, single-objective EAs, while converging quickly, often fail to balance multiple conflicting objectives. Multi-objective EAs, on the other hand, can theoretically balance multiple objectives but incur substantial computational costs in the process of approaching optimal solutions. To address these challenges, this study introduces a multi-task evolutionary algorithm-based band selection method, MTEA-BS. Specifically, MTEA-BS incorporates a new encoding strategy, neighborhood band grouping with integer encoding (NBG-IE), by combining NBG grouping and integer encoding to represent bands. Two single-objective tasks are constructed to assist the multi-objective task, accelerating convergence while maintaining balance across multiple objectives. Additionally, a band-quality-based Pbest operator is proposed for population updates in both tasks. Finally, the effectiveness of MTEA-BS is validated through experiments using two widely used classifiers, KNN and SVM, on four standard hyperspectral datasets. Experimental results show that MTEA-BS outperforms comparable state-of-the-art methods.