In order to address the problem of excessive reliance on annotated datasets in existing apple leaf disease detection methods, a few-shot early apple leaf disease detection algorithm based on weakly supervised learning was proposed. Firstly, a set of backbone networks with shared weights was used to map the diseased leaves into a high-dimensional feature space. Secondly, a dual-branch feature semantic correlation module was established by using a multi-layer attention mechanism, and prototype sets for classifying new disease types in query images were generated on the correlation semantic feature map. Thirdly, the similarity between the prototype sets and the features of diseased leaf in the query image was calculated by using a nonparametric matching method, and the disease regions were located and recognized based on the similarity values. Finally, a weakly supervised learning mechanism was established by using dashed rectangle annotations, and the model was optimized end-to-end with the help of label smooth cross-entropy loss. Experimental results on the open-source Plant Village dataset and a self-built early apple leaf disease dataset demonstrated that the proposed method achieved precision rates of 96.39%、9481%, recall rates of 96.71%、94.67%, and F1 scores of 97.24%、95.20%.