Weather-based logistic regression models for predicting wheat head blast epidemics

Abstract

Wheat head blast is a major disease of wheat in the Brazilian Cerrado. Empirical models for predicting epidemics were developed using data from field trials conducted in Patos de Minas (2013 to 2019) and trials conducted across 10 other sites (2012 to 2020) in Brazil, resulting in 143 epidemics. Each epidemic was classified as either outbreak (≥ 20% head blast incidence) or non-outbreak. Daily weather variables were collected from the NASA Power website and summarized for each epidemic. Wheat heading date (WHD) served to define four time-windows, each comprising two seven-day intervals (before and after WHD), combined with weather-based variables, resulting in 36 predictors (9 weather variables × 4 windows). Logistic regression models were fitted to binary data, with variable selection using LASSO and sequentially best subset analyses. The models were validated using LOOCV, and their statistical performance was compared. One model was selected, implemented in a 24-year series, and assessed by an expert and literature. The logistic models, with 2 to 5 predictors, showed accuracies between 0.80 and 0.85, sensitivities from 0.80 to 0.91, specificities from 0.72 to 0.86, and AUC from 0.89 to 0.91. The accuracy of LOOCV ranged 0.76 to 0.81. The model applied to a historical series included temperature and relative humidity in pre-heading date, as well as post-heading precipitation. The model accurately predicted the occurrence of epidemic outbreaks, aligning closely with observations, specifically tailored for locations with tropical and subtropical climates.