Data driven predictive models based on Artificial Intelligence to anticipate the presence of Plasmopara viticola and Uncinula necator in southern European winegrowing regions.

Proceedings of the International Conference of the Catalan Association for Artificial Intelligence (CCIA), Sitges, Barcelona (Spain), 19 – 21 October 2022

by Marta Otero, Luisa Fernanda Velasquez, Jordi Onrubia, Alex Pujol, Jordi Pijuan (Eurecat); Boris Basile (Department of Agricultural Sciences, University of Naples Federico II)


Downy and powdery mildews are two of the main diseases threatening grapevine cultivation worldwide caused by the phytopathogens Plasmopara vitícola and Uncinula necator, respectively. These diseases may cause severe damage to grapevines by inducing wilting of plant organs, including bunches, especially when vines are untreated. This fact, together with the widespread of these pathogens due to the large extensions of land dedicated to grapevine monoculture, makes necessary to develop new predictive modeling tools that allow anticipating disease appearance in the vineyard, minimizing the losses in fruit yield and quality, and helping farmers in defining appropriate and more sustainable disease management strategies (fungicides applied at the right time and dose). For this purpose, farms located in three countries (Portugal, Spain, and Italy) were selected to study the relationship between the microclimatic characteristics of the plots, the phenological stage of the plants throughout the annual cycle, and the presence of both pathogens using different Machine and Deep Learning classification algorithms: Logistic Regression, Decision Trees, Random Forest, Gradient Boosting, K-Nearest Neighbors, Naïve Bayes, Support Vector Machines, and Deep Neural Networks. The results showed that, after an entire annual grapevine cycle, the best performing models were Support Vector Machines for downy mildew and Random Forest for powdery mildew, providing a prediction accuracy of more than 90% for the infection risk and more than 80% for the treatment recommendation. These models will be fine-tuned during two additional vegetative seasons to ensure their robustness and will receive short- and medium-term climatological and phenological forecasts to make recommendations. The preliminary results obtained show that these models are a promising tool in the field of plant disease prevention and resource saving.