Hyperspectral Image Classification Based on Multi-Level Spectral-Spatial Transformer Network

Hao Yang, Haoyang Yu, Danfeng Hong, Zhen Xu, Yulei Wang, Meiping Song

Abstract: Deep learning methods have been widely used in hyperspectral image classification (HSIC). In recent years, Convolutional Neural Network (CNN) has become a mainstream model of deep learning for HSIC. Although the CNN-based method has made great progress, it still faces a series of challenges such as insufficient use of long-distance information, limited receiving domain, and high computational overhead. In order to overcome these issues, this paper proposes a multi-level spectral-spatial transformer network (MSTNet) for HSIC through the image-based classification framework. The proposed network learns feature representation through a transformer encoder, and integrates multi-level features through a decoder to generate classification results. Finally, the experimental results on two real hyperspectral data sets verified the superiority of the method.