ETD

• CatBoost
• decision support system
• downy mildew
• hybrid model
• machine learning
• mechanistic model
• Plasmopara viticola
• precision viticulture

Downy mildew, caused by the oomycete Plasmopara viticola, cut Italian wine production by roughly 23% in the 2023 outbreak, yet most small producers still rely on calendar-based fungicide schedules or closed proprietary tools that sit poorly with the European Farm-to-Fork strategy. This thesis, developed within the publicly funded wAIne project, presents an open-source machine-learning decision support system that predicts daily downy mildew risk in Tuscan vineyards. Its core is a Rossi (2008) mechanistic epidemiological engine, deployed on Google Cloud Platform with a bud-burst gate and ERA5 weather gap-filling. A hybrid grey-box model then feeds the engine's internal states as engineered features into a CatBoost regressor. Under station-grouped cross-validation on the 2025 Tuscan season, the hybrid improves on a weather-only baseline only by a small, not statistically significant margin, and the hyperspectral channel does not help at the sparse coverage available. The strongest empirical result is a timing analysis showing that the Rossi thresholds, calibrated for the Po Valley, are crossed far too early under Tuscan weather and require regional recalibration. The contributions are the production-deployed engine and its refinements, this recalibration finding, a reproducible multi-modal data corpus and quality audit, the grey-box and treatment-decay designs, and a comparative evaluation of where each information channel helps and where it does not.