A Novel Hyperspectral Image Band Selection Method: Self-Attention based Reconstruction Network

Hyperspectral image (HSI) consists of hundreds of bands, carrying abundant but redundant information. Band selection (BS) has been proven to be an effective way to reduce the redundancy of HSI. Previous studies have shown that there are correlations between bands. The greater the correlations between one band and others, the more important this band is, which should be prioritized for selection. The self-attention mechanism can automatically discover potential correlations between bands. Therefore, this paper proposes a Self-attention-based Reconstruction Network, named SRN, for band selection. SRN discovers band correlations and generates a band correlation matrix, and then reconstructs HSI based on this. To optimize SRN, the reconstruction error between the original and the reconstructed HSI is used as the loss function. After converges, the band correlation matrix is summed along each row as band weight vector. Then, a special peak-first strategy is implemented to select bands. Experimental results on real HSI datasets demonstrate the effectiveness of SRN.