In recent years, the graph transformer architectures have gained widespread attention due to its local and global capabilities. However, these architectures ignore the dynamic updating of adjacency structure and lack the ability to extract representative features in limited labeled data. In this paper, we integrate the connection between graph convolution and Transformer, creating a novel Graph Sampling Transformer framework (GST). Specifically, we enhance key HSI pixel features based on the inductive summary of the graph structure. Next, we perform sampling on the graph structure, filtering out noise features, while restricting the global range of Transformer attention. Building upon this, we further strengthen the consistency between target pixels and global context features, effectively updating the sampling neighbors by fully exploiting relevant features. On real HSI classification datasets, the proposed method outperforms state-of-the-art classifiers, showcasing its strong feature extraction capability and broad application prospects.