Over the past few years, the application of few-shot learning (FSL) in hyperspectral image (HSI) classification has emerged as a popular research area. Therefore, this paper proposes a lightweight network named UNet-PRFN (U-Net with principal component analysis and rotate-flip-noise augmentation) based on 3D U-Net. Firstly, we reduce redundant data and enhance sample diversity through dimensionality reduction by using principal component analysis (PCA) and data augmentation by using rotate-flip-noise (RFN). Then, we design a 3D-U-Net-like structure aimed at generating more discriminative features by employing segmentation and recombination strategies. Finally, we apply small-sample loss function and comparative loss function to minimize the sample distribution entropy. Our experiments demonstrate that the proposed approach surpasses multiple current leading models across four publicly accessible datasets.