Image analysis with deep learning for early detection of downy mildew in grapevine

Abstract

Downy mildew is a major disease of the grapevine that can severely reduce crop quality and yield. Its assessment in the laboratory is time-consuming, usually carried out by experts, and can require expensive and complex tools. For this reason, there is an opportunity to apply sensor technologies and artificial intelligence to plant disease detection. In this study, deep learning applied to RGB images was investigated to early detect downy mildew and the infection stage in grapevine leaf discs under laboratory conditions. Leaf discs of Tempranillo grapevine variety from 3 to 9 days post-inoculation located in Petri dishes were imaged using controlled conditions. Leaf disc images were extracted using computer vision techniques. Convolutional Neural Networks were used to classify the infected and healthy discs and to identify the disease infection. 10-fold cross-validation was used to evaluate the network results and Grad-CAM was used to interpret model prediction. An accuracy around 99% and a f1-score of 0.99 was achieved in downy mildew detection after DPI 3. An accuracy of 81% and a f1-score of 0.77 was obtained in infection stage identification. The developed method offered objective, rapid and accurate results, giving the possibility of early detecting downy mildew in grapevine leaf discs using low-cost techniques.