Weather‐Driven Ensemble Logistic Models for Predicting Fusarium Head Blight of Wheat in Brazil

ABSTRACT Fusarium head blight (FHB), caused by members of the Fusarium graminearum species complex, is a major constraint to wheat production in southern Brazil. In this study, we used a functional data analysis (FDA) framework to identify critical weather variables around flowering and help guide the development of models predicting an epidemic (FHB > 10% severity; n = 125 cases). This was evaluated in terms of predictive performance and economic benefit. FDA revealed that daily records of mean relative humidity, mean minimum temperature, precipitation and mean dew point temperature (summarised as average or totals, particularly within 2–10 days after the beginning of flowering), were strongly associated with the epidemics. These insights informed logistic regression models, either assuming linear effects or incorporating restricted cubic splines to account for potential non‐linearities. All individual models showed good predictive performance (ROC‐AUC 0.79–0.81). Ensemble approaches further improved discrimination (ROC‐AUC up to 0.84), with the stacked ensemble providing the best balance between sensitivity, specificity and probability calibration, thus representing the most reliable option. This model also demonstrated good temporal and spatial generalisation when evaluated under strict leave‐one‐year‐out and leave‐one‐location‐out validation schemes (ROC‐AUC 0.75–0.83). When evaluated within a Net Monetary Benefit (NMB) framework, the stacked ensemble showed an 80% probability of positive economic returns at risk thresholds ≥ 0.45, indicating that fungicide application is economically justified under moderate to high epidemic risk. The simplified models developed in this study offer practical decision‐support tools for managing FHB.