Abstract: In the process of wheat pest identification, the traditional deep learning model is unstable, the recognition accuracy is low,and the generalization ability is limited. A new dual-flow network model is proposed, which combines ResNet and ViT to improve the recognition accuracy. This method integrates the convolutional neural network to process the local structure of the image, while Transformer is used to capture the long-distance dependency, which improves the recognition performance. Through the training verification of 2 070 images of wheat pests and diseases, the parameters of ResNet50 and ViT pre-training models were adjusted. The results showed that the double-flow model achieved 96.5% accuracy in the training set and 0.94 F₁ score in the verification set, which was significantly better than other mainstream single models. The results confirm that the new model has excellent performance in wheat pest and disease identification, providing potential for its wide application in intelligent agricultural systems.