In order to reduce the influence of leguminivora glycinivorella on soybean production and quality, and to realize the rapid detection of leguminivora glycinivorella, this paper proposed a leguminivora glycinivorella detection model based on 3D-Realtion Network (3D-RN) model. Firstly, collect the hyperspectral images of 20 soybeans that are attached to eggs, larvae, gnawed and normal soybeans, respectively, and extract the region of interest (ROI) to establish a 3D-RN model based on hyperspectral images. The accuracy of the final model reached 82%±2.50%. Compared to the Model-Agnostic Meta-Learning (MAML) and Matching Network (MN) models, the 3D-RN model can fully measure the distance between sample features, and the recognition effect is greatly improved. Thus, this research shows that the 3D-RN model based on the hyperspectral image can detect leguminivora glycinivorella in a small number of samples. The method of combining few-shot meta-learning with hyperspectral provides a new idea for pest detection.