Wheat Fusarium Head Blight Severity Prediction in Southern Henan: K- Means-SMOTE and Ensemble Learning Comparison

Abstract

Fusarium Head Blight (FHB) severely affects wheat yields and food safety, especially in Southern Henan, China. This study proposes a hybrid predictive framework combining K-means-SMOTE for class imbalance correction and ensemble learning algorithms LightGBM, CatBoost and Decision Tree —to enhance FHB severity classification. Using meteorological and disease data from 2016 to 2023, the model was trained on both original and oversampled datasets. Feature selection via Kendall’s tau identified key meteorological indicators critical for prediction. Experimental results show that LightGBM with SMOTE achieved the highest accuracy (0.9242) and F1 score (0.9242), followed closely by CatBoost with K-means-SMOTE. The study also finds that default hyperparameters often outperform tuned configurations, and ADASYN is less effective than SMOTE or K-means- SMOTE. Despite strong performance, the model’s regional specificity limits its generalisability, requiring adaptation for broader climatic and crop contexts. This work demonstrates the potential of integrating oversampling and ensemble techniques for early, accurate FHB severity prediction, supporting targeted disease management in agriculture.