Automated detection of downy mildew and powdery mildew symptoms for vineyard disease management

Abstract

This work focuses on developing an automated system for detecting downy mildew and powdery mildew symptoms in grapevines, with particular attention to the role of data partitioning and dataset diversity in ensuring reliable model performance. Leveraging deep learning techniques, specifically the YOLO (You Only Look Once) object detection model, we aimed to provide a robust tool for disease detection, which is crucial for optimizing vineyard management, increasing crop yield, and promoting sustainable agricultural practices. Over two years, we collected and expertly annotated a large dataset of images depicting downy and powdery mildew symptoms in field conditions. The YOLO model was trained and validated on this dataset, achieving a mean Average Precision (mAP) of 0.730, demonstrating good detection accuracy. A key contribution of this study is the emphasis on the importance of proper data partitioning strategies, showing that random image partitioning can lead to an overestimation of model performance. Our findings underscore that true improvements in detection accuracy are driven not merely by increasing the number of images but by enhancing the diversity of the dataset, particularly for the areas, seasons, growth stages, and conditions in which the images are captured. This approach ensures a more realistic assessment of the system's performance, critical for deploying such systems in practical, real-world agricultural scenarios. The results highlight the potential of deep learning models to enhance vineyard management through a reliable and efficient detection of diseases in real-world conditions.