Spectral Spatial Window Attention Transformer for Hyperspectral Image Classification

Hyperspectral image (HSI) data have a wide range of spectral information that is valuable for numerous tasks. HSI data encounter some challenges, like insufficient representation of spectral spatial data, data scarcity, and redundant information. Researchers introduce various research works to address these challenges, with convolutional neural networks (CNNs) finding widespread use in HSI classification because of their capacity to extract features from HSI data. Due to a limited receptive field, CNN cannot handle long-range dependencies. To overcome the above-mentioned issue, we proposed a cross-window spectral spatial transformer model for HSI classification. First, the fused-MBConv module is used to extract low-level spectral spatial features from HSI data. Second, we propose cross cross-window attention module to model the long-range spatial dependencies and obtain refined spectral spatial features. Finally, we proposed an interactive feature-enhanced (IFE) module to obtain global contextual features, leading to more discriminative representations. We tested the proposed model on four publicly available HSI datasets: Xuzhou, Salinas Pavia University, and WHU-Hi-HC. Using only 0.5% and 0.05% of the training samples from the four datasets, we achieved the best classification in terms of overall accuracy (OA), average accuracy (AA), and Kappa coefficient.