Band selection framework for hyperspectral images based on discriminability-oriented optimization mechanism

Hyperspectral images suffer from high dimensionality, inter-band redundancy, and noise interference, increasing computational cost and degrading classification performance. Conventional band selection methods often fail to capture complementary and interactive band relationships, while deep learning-based approaches lack interpretability. Moreover, band selection remains a high-dimensional combinatorial optimization problem, requiring balanced discrimination, redundancy suppression, and search efficiency.

To address the limitations of existing swarm intelligence-based band selection methods and the premature convergence of the original ant colony optimization for continuous domains (ACOR) algorithm, an improved ACOR-based framework, termed IACOR-BS, is proposed. First, a spectral relationship representation based on normalized discriminative scores is constructed to evaluate the discriminative capability and correlation of candidate band subsets. Second, an adaptive preprocessing strategy based on normalized matched filter weights is introduced to suppress low-quality and noise-dominated bands before optimization. Finally, k-means-based initialization, adaptive band index correction, and adaptive parameter adjustment are incorporated into ACOR to improve convergence efficiency and search stability.

Experiments on four widely used hyperspectral datasets show that IACOR-BS achieves superior or competitive classification performance under both KNN and SVM classifiers. The proposed framework reduces the number of selected bands by approximately 80% compared with the original spectral dimensionality while maintaining classification accuracy comparable to or higher than that of the best-performing methods. Statistical analyses based on the Friedman and Wilcoxon tests further confirm its superiority and robustness across different datasets and experimental settings.

IACOR-BS identifies compact and discriminative band subsets, achieving superior or near-optimal classification performance while maintaining strong band compression capability across diverse hyperspectral datasets.