Machine Learning Approaches for Predicting Fusarium Head Blight Epidemic Levels from Climatological Time Series Features.

Abstract

Fusarium head blight (FHB), caused by the FHB species complex, is one of the most damaging diseases affecting wheat. Accurately predicting FHB occurrence prior to infection is crucial for preventing outbreaks, minimizing crop losses, and reducing the risks of mycotoxins entering the food chain. This study utilized 55 years of historical weather data and the level of the primary Fusarium inoculum in crop debris to predict FHB severity. Time series features were extracted from daily average temperature, relative humidity, precipitation, and sunshine hours recorded from 1 January to 31 March each year. Random forest (RF), statistical analysis, binary enumeration, and feature interpretability analysis were employed to identify features most strongly associated with FHB occurrence. Six machine learning (ML) models, including artificial neural networks (ANNs), logistic regression, K-nearest neighbors, support vector machine, RF, and extreme gradient boosting, were applied to the selected features to classify FHB epidemic levels. The results revealed that (i) the inclusion of inoculum strength did not significantly enhance the predictive accuracy of models based solely on climatological data; (ii) the nine complex time series features derived from the four types of climatological data were effective in classifying FHB epidemic levels through cross-regional validation by capturing critical conditions linked to the pathogen lifecycle; and (iii) ANNs outperformed the other five ML models in classifying observed FHB epidemic levels using the selected nine features. Further research should focus on applying this model to larger datasets to enhance its practical utility for FHB management.