SLDATU: Short-Long Distance Attention Transformer-Based U-Net for Hyperspectral Image Classification

Transformer usually achieves good performance in hyperspectral image classification (HSIC), and existing transformer-based HSIC methods still face multiple challenges, such as insufficient multiscale adaptability, cross-level feature fusion strategy needs to be improved, and the ability to capture detailed information is limited. To alleviate the aforementioned challenges, this paper integrates short-long distance attention, Transformer, U-Net, and skip connection cross attention into a unified framework for the first time. A short-long distance attention transformer-based U-Net for HSIC, namely SLDATU, which contains two main components, a short-long distance attention transformer and a U-Net with skip connection cross attention (SCCA), is then proposed. The designed short-long distance attention transformer extracts both local details and global dependencies by alternating between long distance attention mechanism and short distance attention mechanism. The U-Net with SCCA is constructed to build different scale hierarchies by downsampling and upsampling, and the dynamic interactions of cross-scale features are realized by SCCA. To validate the effectiveness of proposed SLDATU, five real hyperspectral datasets are used for experiments, and the results are compared with several recent related state-of-the-art HSIC methods. The proposed SLDATU method can achieve optimal classification performance.