Predicting wheat powdery mildew epidemics in China using meteorological data and machine learning approaches

BACKGROUND: Prediction is vital for plant disease management. This study developed machine learning models that used meteorological data to predict wheat powdery mildew (WPM) occurrence severity degree and area in China. Six machine learning algorithms were trained and cross-validated to predict WPM severity degree with 411 pieces of meteorological data from 48 counties (1981–2021) across China. Areas of WPM occurrence were also derived from WPM severity degrees [which were predicted by the K-Nearest Neighbor (KNN) model] with spatial interpolation models. RESULTS: The best-performing machine learning severity prediction models were based on meteorological data during the coldest month (January) of the wheat overwintering period, and also the wheat jointing stage–heading stage. In each case the times were subdivided into 5-day periods. In particular, the prediction model showed that the best performance was based on the support vector machine algorithm. Climate variable importance ranked via random forest identified eight key predictors. Using these, KNN achieved high performance, demonstrating its suitability for predicting WPM severity degree. Nationwide severity distributions were produced using inverse distance weighted (IDW) and ordinary kriging methods, based on severity degrees predicted by the KNN model from 1990 to 2019. Validation via chi-squared and error reference methods confirmed that the IDW_4.0 model outperformed the others. CONCLUSIONS: Machine learning models effectively predict WPM severity degree and area of occurrence at a national scale using meteorological data. The disease severity distribution of WPM displays disease severity spatial patterns visually and can improve management strategies for WPM across China.