A two-stage method for wheat stripe rust grading based on YOLOv11-leaf and convnextv2-rust

Wheat stripe rust is a major fungal disease severely impacting global wheat production. However, in complex field environments, the disease’s early lesions are small and its severity grading is highly subjective, leading to low accuracy and deployment challenges in automated recognition. To address this challenge, this study aims to develop an efficient and precise automated recognition framework for wheat stripe rust. We propose a two-stage recognition framework integrating lightweight detection with fine-grained classification. First, based on a constructed field image dataset covering the international standard 0–9 disease severity scale, we designed a lightweight object detection model, YOLOv11-leaf, to precisely localize and extract leaf regions from complex backgrounds. Subsequently, for the fine-grained task of disease severity grading, we modified the ConvNeXtV2 network to propose the ConvNeXtV2-rust classification model, enabling automatic determination of disease severity levels on extracted leaves. Experimental results demonstrate that YOLOv11-leaf achieves 96.71 % accuracy in detection tasks, representing a 7.06 % improvement over YOLOv11n. For classification tasks, ConvNeXtV2-rust achieves 75.94 % accuracy, surpassing the original ConvNeXtV2 by 1.75 %. This study provides a high-precision, easily deployable practical solution for crop disease-resistant breeding screening and intelligent field forecasting systems. It advances the precise prevention and control of crop pests and diseases while accelerating the development of precision agriculture.