China is the world’s largest tea producer, and the tea industry is an important specialty industry for increasing agricultural efficiency and farmers’ income. Frequent tea diseases severely affect yield and quality. Traditional manual diagnosis is inefficient and error-prone, while conventional plant protection operations rely mainly on large-scale pesticide application, resulting in high costs and serious pollution. Current deep learning-based disease detection technologies are easily affected by complex environments in real tea plantations, such as lighting, occlusion, and rainy weather, and their accuracy and stability are insufficient to meet real-time field detection needs, becoming a key bottleneck in the development of smart tea plantations.
Recently, the team led by Shuhe ZHENG and Wuxiong WENG at Fujian Agriculture and Forestry University published a research paper titled “GDE-YOLO: a robust and accurate method for real-time tea leaf disease detection in complex plantation environments” in Engineering Agriculture (DOI: 10.15302/J-FASE-2026669). Based on the lightweight YOLOv8n model, the study developed the GDE-YOLO detection model, achieving high-accuracy, high-robustness real-time identification of tea leaf diseases in complex tea plantation environments.
Real-time detection of tea leaf diseases in complex tea plantation environments faces multiple technical challenges. First, disease targets are small and symptoms are similar; under backlight, shadow, and overlapping leaf occlusion, features are weak, and general-purpose models are prone to missed and false detections. Second, lightweight models lack sufficient accuracy, while high-accuracy models have large computational demands and slow speed, making them difficult to deploy on edge devices. Third, existing methods have weak generalization ability, with a large gap between laboratory and field performance, and cannot provide reliable perception support for intelligent agricultural machinery, restricting the intelligent upgrading of tea plantations.
The research team constructed a field dataset covering multiple diseases and complex environments, and improved performance through three key modifications: introducing a global attention mechanism (GAM) into the neck network to strengthen disease features and suppress background noise; optimizing the C2f module with a diverse branch block (DBB) to enhance multi-scale feature representation without increasing inference cost; and replacing the complete intersection over union (CIoU) loss function with the efficient intersection over union (EIoU) loss function to improve bounding box regression accuracy and convergence speed. Experimental results showed that under complex tea plantation scenes and disease types, GDE-YOLO achieved an overall accuracy of 91.7%, 3.1 percentage points higher than the baseline model, with detection of tea white scab improving by 12.4 percentage points, while maintaining a speed of 80 FPS and meeting real-time detection requirements. The model was further deployed on the NVIDIA Jetson Orin Nano embedded platform, achieving a field inference speed of 18 FPS. Under complex conditions such as strong light, occlusion, and after rain, it still stably output detection results with confidence greater than 0.8, realizing the leap from a laboratory algorithm to a field-ready system.
This study breaks through the technical limitations of accurate crop disease perception in complex natural scenarios, provides an efficient solution for intelligent monitoring of tea throughout the entire growth period, and injects new quality productive forces into the digital transformation of the traditional tea industry. The model is lightweight, highly accurate, and strongly generalizable, and can be widely integrated into tea plantation inspection robots, UAVs, precision sprayers, and other equipment. It promotes disease diagnosis from “manual judgment” to “real-time online sensing,” and control methods from “large-scale blanket prevention” to “precise site-specific pesticide application,” helping reduce pesticide use, lower agricultural non-point source pollution, and improve tea quality and production efficiency. The study also verified the feasibility of deploying lightweight deep learning models on agricultural edge devices, providing a replicable technical path for intelligent monitoring of cash crops, smart plant protection, and precision management, with important practical value for promoting agricultural digital transformation and enhancing the intelligence level of agricultural equipment.
DOI: 10.15302/J-FASE-2026669