SSBFNet: a spectral-spatial fusion with BiFormer network for hyperspectral image classification

Recently, convolutional neural network (CNN) methods have gained widespread popularity in hyperspectral image (HSI) classification, due to their remarkable ability to capture local spatial features. However, HSI data imply rich spatial and spectral information, making it still a challenging task to efficiently model local and global features with full exploitation of its spectral and spatial information. In this letter, a spatial-spectral fusion with BiFormer network (SSBF) is proposed. Specifically, after dimensionality reduction and shallow feature extraction of the HSI, a spatial-spectral attention module (SSAM) is introduced to extract and fuse spatial and spectral features. Then, a textual feature weighting (TFW) module is leveraged for adaptive semantic information conversion. Finally, a dynamically sparse attention-based BiFormer encoder module is proposed to efficiently capture global features of the image with reduced computing overhead. The experimental results on three widely used HSI datasets prove the advantage of the SSBF method. The source codes are available at: https://github.com/DrWuHonglin/SSBFNet.