In the field of agricultural pest image classification, limited pest image data or the requirement to quickly identify new categories of pest images in dynamic environments pose challenges to current deep learning methods. This article constructs a rice pest training and testing set based on the IP102 dataset. On the basis the prototype network, VGG16 is used as a feature extractor to extract pest features, and a metric space based prototype network algorithm is used as a metric learner to construct a VGG-ML rice pest classifier. Train VGG-ML through meta learning. The experimental comparison showed that the accuracy of VGG-ML rice pest classification recognition was 69.5% and 81.5% under 5-way, one shot, and 5-way, 5-shot conditions, respectively. Compared with the original prototype network, the accuracy has been improved by 3.53% and 4.4%, and compared with the transfer learning network, the accuracy of pest classification results has been improved by 12.78% and 25.94%, respectively. Using the VGG-MG rice pest classifier in this article to learn agricultural pest classification tasks and achieve rice pest classification experiments with very few samples.