Rich Features Aggregation For Hyperspectral Image Spectrum Reduction

Classification of hyperspectral images (HIC) plays an important role in remote sensing image studies. To address the issue of redundant spectral bands in hyperspectral photographs, both conventional and CNN-based solutions are proposed. However, present dimension reduction approaches frequently depend solely on spectral information and disregard the spatial information, so impairing classification accuracy. In addition, these methods lower spectral dimension without taking physical significance into account. In this study, we offer a new dimension reduction technique based on multi-scale spatial module. A physical-based feature extraction module and a rich spectrum feature selection module are included in this method. Experiments on three benchmark datasets demonstrate that our strategy can outstandingly improve performance on various HIC model.