Forecasting severe grape downy mildew attacks using machine learning.

Area of Science:

• Agricultural Science
• Plant Pathology
• Computational Biology

Background:

• Grape downy mildew (GDM) significantly impacts grapevine yield and fruit quality, necessitating frequent fungicide applications.
• Current GDM management in Bordeaux involves an average of 10.1 fungicide treatments annually, raising environmental and public health concerns.
• Reducing fungicide use requires targeted interventions based on predicted disease risk.

Purpose of the Study:

• To develop and compare predictive models for grape downy mildew incidence and severity in Bordeaux.
• To identify key factors influencing GDM risk, including disease onset date and weather variables.
• To assess the potential of predictive modeling to reduce fungicide applications.

Main Methods:

• Utilized a 9-year dataset of GDM observations for model development.
• Compared generalized linear models with machine learning algorithms (LASSO, random forest, gradient boosting).
• Employed year-by-year cross-validation to assess model accuracy using disease onset, temperature, and precipitation data.

Main Results:

• Machine learning algorithms, particularly LASSO, random forest, and gradient boosting, outperformed generalized linear models.
• The date of GDM onset was a more influential predictor than weather variables (precipitation > temperature).
• Under projected climate scenarios, reduced April-May rainfall and increased temperatures correlate with lower GDM risk.

Conclusions:

• Predictive models incorporating disease onset date and weather data can accurately forecast GDM risk.
• Implementing fungicide treatment decision rules based on these models can reduce applications by at least 50% in Bordeaux.
• Targeted GDM management strategies are crucial for sustainable viticulture and reduced environmental impact.