DSFormer: An Enhanced Dual-Selective Fusion Transformer Framework for Hyperspectral Image Classification++

Abstract

Hyperspectral image classification requires joint modeling of spectral signatures and spatial structure. In patch-based classification, each input patch is assigned the label of its center pixel, although neighboring pixels may belong to different land-cover classes. This setting motivates coordinating shallow spatial features, contextual refinement within the patch, and supervision of the resulting representation. We propose DSFormer++, an enhanced framework based on the original Dual Selective Fusion Transformer Network (DSFormer). Shallow Feature Compensation (SFC) combines shallow and patch-embedded features through gated residual fusion before the backbone. The Collaborative Global–Local (CGL) module applies an additional refinement operation with shared parameters before token selection at two backbone stages. Pixel Contrastive Loss (PCL) uses center-pixel labels to supervise projected patch embeddings during training. The proposed method achieves overall accuracies of 96.50%, 97.60%, 95.35%, and 94.83% on the Pavia University, Houston13, Indian Pines, and WHU-HH datasets, respectively. Component comparisons and approximately parameter-matched controls characterize configuration-dependent effects. Numerical checks assess consistency between the stated operations and their implementation, and an additional Botswana experiment extends the evaluation to a satellite wetland and woodland scene.