Aiming at the problems of the high number of crop disease recognition parameters and gradient disappearance of the network in the existing few-shot learning, a multi-branch grouping convolution feature cascade crop disease recognition model is proposed. Firstly, the model applies grouped convolution to the residual block, which is used to alleviate the gradient disappearance problem of the network. Secondly, the input of the network is divided into isomorphic multi-branch embeddings, and more abundant features are obtained and the number of network parameters is reduced through the feature supplementation between different branches. Finally, a multi-layer feature cascade backbone network is designed to mine different levels of information in the image, and more location and detail information in the shallow features and stronger semantic information in the deep features were cascaded out to improve the generalization ability of the network. On the PlantVillage dataset, the accuracy of the model on the 5-way 1-shot and 5-way 5-shot tasks reached (79.68±0.35)% and (93.25±0.21)%, respectively, which were 3.74% and 4.32% higher than the prototype network. The experimental results show that the proposed model can better identify crop diseases, and reduce the number of network parameters while deepening the depth of the network, alleviating the gradient disappearance and mining different levels of semantic information, which effectively solves the problem of few-shot learning of crop diseases.