A deep learning-based two-stage method for severity grading of corn leaf diseases

The deep learning-based detection of corn diseases has proven effective for images with simple backgrounds. However, for images of complex field environments, diverse backgrounds lead to decreased model accuracy in lesion segmentation, further affecting the accurate grading of disease severity. Hence, based on the corn leaf images acquired by unmanned aerial vehicles (UAVs), this study proposed a deep learning-based two-stage model for grading the severity of corn leaf diseases: the YOLOv5-VGG-Seg model. This model first separates the corn leaf areas from complex field backgrounds and then performs lesion segmentation on the background-removed leaf images. Ultimately, it completes the automatic quantitative grading of disease severity according to the proportions of lesion-covered leaf areas. The experimental results show that in the leaf extraction stage, the YOLOv5-VGG-Seg model achieved an F1 score of 96.0% on the test set. In the lesion segmentation stage, the background removal strategy based on leaf segmentation increased the F1 score for the extraction of the northern corn leaf blight (NCLB) lesions from 66.6% to 72.1%, improving the overall accuracy by 5.5%. Finally, based on the segmentation results of corn leaf lesions, the model performed a statistical analysis of disease severity, achieving an F1 score of 72.0%. Therefore, this two-stage method enables accurate segmentation and reliable grading of NCLB by effectively eliminating the interference from complex field environments. It provides a feasible technical solution for automatic crop disease monitoring under UAV conditions.