<p>Hyperspectral image classification focuses on accurately classifying each pixel in hyperspectral images using partially labeled data, which is crucial for the precise identification and differentiation of various land cover types. In recent years, Graph Neural Networks (GNNs)-based hyperspectral image classification have exhibited exceptional performance when a substantial number of labeled samples are available. However, the acquisition of annotation data is often a labor-intensive and time-consuming process in practical scenarios. When the number of available labeled samples is limited, the model’s performance can easily degrade. To tackle the above issue, a Graph Knowledge Distillation-guided Few-Shot Learning method (GKDFSL) is proposed for hyperspectral image classification. Specifically, for a comprehensive modeling of hyperspectral image features, a graph structure is first constructed from segmented superpixels. Subsequently, graph augmentation is performed to generate two separate graph perspectives. Teacher and student GNNs are then applied to each perspective to extract superpixel features. Knowledge distillation techniques are then utilized to transmit diverse knowledge learned by the teacher GNNs to the student GNNs, thereby enabling the student model to effectively capture the latent associations between unlabeled superpixels. To further enhance the model’s understanding and learning capabilities, GNNs are integrated with the mean teacher framework to improve generalization. Moreover, ensemble learning is introduced to bolster the robustness of the method. Extensive experiments conducted on four benchmark datasets have demonstrated that the proposed method surpasses other competitive approaches in both classification accuracy and generalization capabilities.</p>