An Information Entropy Feature Preservation-Based Band Selection and Noise Reduction Method for Hyperspectral Image Classification

Effectively reducing spectral redundancy while preserving the distinguishability of land cover is key to hyperspectral image classification (HSIC). Moreover, the presence of noise also interferes with the discriminability of spectral features. To address these issues, this paper introduces a novel HSIC method that integrates band selection and discriminative feature extraction to achieve high performance of land cover classification. Firstly, by incorporating the concept of information entropy, the information entropy feature (IEF) is designed to describe the relationship between spectral bands, then the representative band group is selected based on the designed IEF. Secondly, a novel tensor subspace representation model is proposed to separate the land cover of interest and noise component for the selected band group. Next, in light of the different shapes and sizes exhibited by different land cover, multi-scale spatial-spectral features are used as supplementary information. Finally, the extracted class feature is used to guide and refine the extraction of land cover, further optimizing the classification performance. The results on multiple datasets show that the proposed method outperforms existing methods.