Abstract: Conventional deep learning-based algorithms require substantial annotated samples as input. However, high-quality remote sensing image samples are constrained by acquisition challenges and high cost, limiting both performance and application effectiveness of these algorithms. Given the heavy reliance of conventional algorithms on substantial annotated samples, this paper proposed a vegetation classification method for hyperspectral images based on the model-agnostic meta-learning (MAML) algorithm. This method aims to achieve efficient classification using a small number of samples under conditions of sample scarcity. Based on the GF-5 hyperspectral remote sensing images of the Weifang region captured on April 7, 2023, this study constructed two meta-learning models within the MAML framework using spectral features and spatial-spectral features, respectively. Subsequently, it conducted vegetation classification experiments using the two models. Experimental results show that, with only 20 samples per class, the spatial-spectral model achieved an overall accuracy of 84%, with a maximum single-class accuracy of 89%. Compared with conventional classification methods, the MAML-based method presented significant enhancements in overall accuracy and the classification accuracy across all vegetation types. This verified the effectiveness and accuracy of this method under few-shot conditions.