Objective In order to improve the grade classification accuracy of damage degree by wheat stripe rust, the automatic, accurate and rapid identification method of damage degree by wheat stripe rust was studied.

Method Under complex field conditions, images were taken by mobile phones, and data sets of wheat leaves with different grades of stripe rust were constructed. The combination of GrabCut and YOLOv5s was used to automatically segment wheat leaves from complex background. The Inception module was added to enhance the ability of ResNet50 in extracting phenotypic features. The disease grades of wheat stripe rust were identified according to the classified disease grade standards of wheat stripe rust. The performance of the improved ResNet50 model (B-ResNet50) on the data set was analyzed using evaluation indexes such as accuracy, recall and precision.

Result Wheat leaf images were segmented automatically, accurately and quickly by the combination of GrabCut and YOLOv5s under complex background in the field. The recognition rate of B-ResNet50 on the data set of wheat stripe rust leaves was 97.3%, which was obviously higher than that of InceptionV3 (87.8%), DenseNet121 (87.6%) and ResNet50 (88.3%). The accuracy rate was greatly improved, and nine percentage points more than that of the original model(ResNet50).

Conclusion Using deep learning to identify the disease grade of wheat stripe rust is of great significance to applying accurate pesticide for its control, and provides technical support for the control of wheat stripe rust under complex field conditions.