A Two-Stage Feature Selection Method for Hyperspectral Image Classification: Hybrid Filter Preprocessing and Improved Dwarf Mongoose Optimization

Hyperspectral image classification is the foundation of numerous hyperspectral image applications, but its inherent data characteristics such as high-dimensional feature space and significant spectral redundancy often impede both classification accuracy and computational efficiency simultaneously. To address these challenges, this paper proposes a two-stage feature selection method for hyperspectral images. The method integrates a hybrid Filter approach with an improved Dwarf Mongoose Optimization (DMO) algorithm to extract highly discriminative features, while simultaneously enhancing classification accuracy and computational efficiency. First, a hybrid Filter method synergistically integrating Fisher and ReliefF algorithms is used to quickly screen the original feature set, enabling efficient initial pre-selection of informative features and reducing subsequent computational overhead. Second, the DMO algorithm is improved by integrating symbiotic organisms search mechanism, adaptive mutation operator, adversarial learning strategy, and feature weight guidance strategy. The improved DMO algorithm is applied for secondary screening of the pre-selected feature set to further eliminate redundant features that are unfavorable for classification. Finally, experimental evaluations are conducted on three benchmark hyperspectral datasets (Indian Pines, Salinas, and Pavia University) and one self-developed GF-5 satellite dataset (HU ZHOU-WuXing). Results demonstrate that compared with state of-the-art techniques, the proposed method achieves significantly higher classification accuracy with smaller feature subsets and substantially improves feature screening efficiency, fully verifying its excellent comprehensive performance with both accuracy and efficiency advantages in hyperspectral image classification tasks.