Hyperspectral Image Classification Using Attention-Only Spatial-Spectral Network Based on Transformer

Abstract

Hyperspectral image (HSI) classification has drawn increasing attention in the last decade. HSIs accurately classify terrestrial objects by capturing approximately contiguous spectral information. Owing to their excellent performance in image classification and semantic segmentation, many of the latest deep learning approaches, which can extract complex spatial and spectral characteristics compared to traditional machine learning methods, have been applied in HSI classification. The paper proposes a new HSI classification network based on pure multihead attention mechanisms based on a vision transformer. Due to the unique spatial and spectral attention modules, the network can derive long-range spatial and spectral contextual relations between pixels in images. The spatial and spectral features are effectively fused and interacted through the cross-field gating module. The paper evaluates the classification performance of the proposed network on three HSI datasets by conducting extensive experiments, showing its superiority over standard convolutional neural networks and achieving a significant improvement in comparison with other networks. In addition, due to the complete abandonment of the convolution layer and the application of multihead attention mechanisms, the number of parameters of the network is greatly reduced.

Keywords

• Hyperspectral image classification
• remote sensing
• multihead attention mechanism
• vision transformer