Light Self-Gaussian-Attention Vision Transformer for Hyperspectral Image Classification

In recent years, convolutional neural networks (CNNs) have been widely used in hyperspectral image (HSI) classification because of their exceptional performance in local feature extraction. However, due to the local join and weight sharing properties of the convolution kernel, CNNs have limitations in long-distance modeling, and deeper networks tend to increase computational costs. To address these issues, this article proposes a vision Transformer (VIT) based on the light self-Gaussian-attention (LSGA) mechanism, which extracts global deep semantic features. First, the hybrid spatial-spectral tokenizer module extracts shallow spatial-spectral features and expands image patches to generate tokens. Next, the light self-attention uses Q (query), X (origin input), and X instead of Q, K (key), and V (value) to reduce the computation and parameters. Furthermore, to avoid the lack of location information resulting in the aliasing of central and neighborhood features, we devise Gaussian absolute position bias to simulate HSI data distribution and make the attention weight closer to the central query block. Several experiments verify the effectiveness of the proposed method, which outperforms state-of-the-art methods on four datasets. Specifically, we observed a 0.62% accuracy improvement over A2S2K and a 0.11% improvement over SSFTT. In conclusion, the proposed LSGA-VIT method demonstrates promising results in the HSI classification and shows potential in addressing the issues of location-aware long-distance modeling and computational cost. Our codes are available at https://github.com/machao132/LSGA-VIT.