Event-based early warning of vineyard disease risk from environmental time series

Abstract

Accurate early warning of vineyard disease risk can support timely intervention and more targeted crop-protection decisions. However, conventional daily-status classification does not directly measure whether a model can anticipate the beginning of a new risk episode. This study investigates an event-based formulation for short-horizon warning of vineyard mildew risk from agro-meteorological time series. The available downy- and powdery-mildew annotations are derived from fungicide-treatment records and are therefore interpreted as management-oriented risk proxies rather than direct pathological observations. Because the two labels agree on 98.9% of observed days and are identical in the validation and test years, they are combined into an operational any-mildew target without implying epidemiological equivalence. New events are defined after a minimum disease-free gap, and models estimate whether an event will begin within the following 3–7 days. Using a chronological 2020–2021/2022/2023 train–validation–test split, we compare calendar-only, environmental-only, and combined representations with logistic regression, XGBoost, long short-term memory (LSTM) networks, and temporal convolutional networks (TCNs). Calendar-only models were highly competitive, with calendar-based LSTM and TCN models detecting 7 of 8 test events, while a monthly climatological baseline detected 6 events. Environmental-only models were substantially weaker, and adding environmental variables to calendar information produced inconsistent gains across model families. A comparison with conventional daily-status classification showed that the two formulations capture different objectives and exhibit different trade-offs in point-wise performance, event coverage, and alert burden, without uniform superiority of either formulation. Sensitivity analyses further showed that results depend on the historical window, prediction horizon, and event-separation gap. SHapley Additive exPlanations (SHAP) analysis indicated that the fitted model used both seasonal and meteorological information, without revealing a single dominant or consistently interpretable environmental precursor. These findings demonstrate the value of event-oriented evaluation for assessing anticipatory warning and emphasize the need for explicit seasonal baselines and cautious biological interpretation of management-derived proxy targets.