Abstract: Aming at the problems  of diversity and poor image quality in agricultural disease image datasets, we  proposed a multisource agricultural disease image dataset construction method based on hierarchical annotation and adaptive preprocessing. Firstly, images were collected from different regions, crop types, and growth stages by using devices such as smartphones, professional cameras, and drones to ensure data diversity. Secondly, we constructed a hierarchical annotation system that  covered three levels of agricultural disease type, severity, and location, we used tools such as LabelImg and LabelMe for annotation, and requested expert review. Finally, we applied adaptive preprocessing methods, including automatic cropping, normalization, denoising and enhancement, to adjust parameters based on image features to improve quality. The experiment used a convolutional neural network (CNN) model based on the ResNet-50 architecture for validation, and the results show that hierarchical annotation and adaptive preprocessing methods significantly improve the quality of the dataset and model performance, the model achieves accuracy, recall, and F₁ score of 92.5%, 91.8%, and 92.1%, respectively, which are better than the training results of other datasets.

Key words: agricultural disease image, dataset construction, hierarchical annotation, adaptive preprocessing, multisource data