Spectral-Spatial Convolutional Hybrid Transformer for Hyperspectral Image Classification

The combination of Convolutional neural networks (CNNs) and Transformers has achieved excellent performance in hyperspectral image classification due to the characteristics of local features extraction and long-range dependencies capture. However, how to integrate the spectral and spatial features effectively after dimensionality reduction while retaining key information remains to be optimized further. To address this issue, the Spectral-spatial convolutional hybrid Transformer (SSCHFormer) hyperspectral classification model is proposed in this article. First, the spectral pyramid 3D convolution and 2D convolution are combined to extract joint and detailed spectral-spatial features. Then, the hybrid convolutional transposed attention mechanism is introduced to enhance the correlation between the local and global features further and perceive the complex spatial and spectral information while reducing computational complexity. Furthermore, the gated feedforward network is designed to enhance the ability to distinguish subtle differences between different classes. Extensive experiments are implemented on four public datasets and demonstrate the superior classification performances of the proposed SSCHFormer to those of other advanced networks.