An interpretable machine learning technique to identify the key meteorological factors influencing the incidence of wheat Fusarium head blight in Korea

• Fusarium Head Blight (FHB) of wheat in Korea depends more on yearly meteorological variations than geographical features. • Machine learning was used to investigate the key meteorological factors influencing the FHB epidemic in Korea. • Interpretable machine learning technique identified not only important variables but also their critical thresholds. • 75 % relative humidity during heading and flowering and 60 mm precipitation during flowering are critical for FHB. Fusarium Head Blight (FHB), predominantly caused by Fusarium asiaticum , represents a major threat to wheat production in Korea, resulting in yield losses, severe economic consequences, and increased risks of mycotoxin-induced health disorders, including liver damage, immune dysfunction, carcinogenesis, and reproductive impairments, in both humans and animals. Nevertheless, there is a limited understanding of the growth stage-specific environmental conditions favoring FHB occurrence in wheat growing fields. In this study, we successfully applied an interpretable machine learning technique to identify the key meteorological variables influencing FHB in Korea. Nationwide FHB incidence data, collected from all wheat-growing regions between 2015 and 2021, were utilized for this analysis. Two machine learning models, Random Forest (RF) and Boosted Regression Trees (BRT), were employed because they generally exhibit lower sensitivity to correlations among variables than statistical models and require smaller datasets than deep learning methods. Using these models, we identified three key variables and their critical thresholds for FHB occurrence in Korea: a relative humidity (Rhum) of 75 % during the heading period, and an Rhum of 75 % combined with 60 mm of precipitation during the flowering period. Furthermore, exceeding two or all three of these thresholds significantly increased FHB incidence compared to exceeding only a single threshold. Overall, this study revealed the feasibility and potential applicability of interpretable machine learning techniques to better understand the relationship between disease incidence and environmental conditions not only for wheat FHB but also other plant diseases.