CSFRT: Cross‐Spectral Feature‐Refined Transformer for Hyperspectral Image Classification

ABSTRACT Hyperspectral images (HSIs) contain rich spectral information, which enables accurate target classification. However, HSIs also contain a significant amount of redundant spectral information, and the spectra of different objects often overlap. This overlap in spectral features increases the similarity between different objects' spectra, posing a challenge for classification. Thus, suppressing redundant spectral information and refining key features are critical tasks. To address these challenges, we propose a cross‐spectral feature‐refined transformer (CSFRT) based on the vision transformer (ViT). In the proposed CSFRT, a two‐branch gated‐refined feed‐forward network (TBGFN) module is introduced to suppress redundant information and enhance key spectral features by utilizing branches with and without gated mechanisms. Additionally, a cross‐layer spectral feature‐fusion (CLSF) module is proposed to integrate feature information and facilitate information complementarity across different encoder blocks. Extensive experiments are conducted on five different HSI datasets to verify the classification performance of the proposed CSFRT, demonstrating the effectiveness of the architecture.