PS2former: Parallel Spatial-Spectral Transformer Network for Hyperspectral Image Classification

Recent advances in hyperspectral image (HSI) classification demonstrate the potential of hybrid architectures that combine convolutional neural networks (CNNs) with Transformer modules. Yet, existing approaches fail to fully exploit the joint spatial–spectral characteristics of HSIs, as they often focus on either local or global features extracted separately by CNNs or Transformers. To address this limitation, we propose PS2former, a parallel spatial–spectral Transformer network. Specifically, we design a Parallel Extraction Module (PEM) that simultaneously captures primary spatial and spectral information and integrates them through feature fusion. Furthermore, we introduce a Parallel Hybrid Transformer (PHT) to jointly model local details and global context. The PHT incorporates a hollow CNN for parallel extraction of local features and a convolution Transformer for long-range global dependency modeling. Extensive experiments on two benchmark datasets demonstrate that PS2former consistently outperforms several recent state-of-the-art methods in HSI classification.