Gradient boosting machine learning model to predict aflatoxins in Iowa corn

Area of Science:

• Agricultural Science
• Environmental Science
• Data Science

Background:

• Aflatoxin (AFL) contamination in corn poses significant health risks due to its toxic and carcinogenic properties.
• Ensuring food and feed safety in the US necessitates effective AFL mitigation strategies for this vital commodity.

Purpose of the Study:

• To develop and evaluate an Iowa-centric predictive model for AFL contamination in corn.
• To utilize historical data, meteorological, satellite, and soil properties for AFL risk assessment.

Main Methods:

• Gradient Boosting Machine (GBM) learning was employed for AFL prediction.
• Two AFL risk thresholds (20-ppb and 5-ppb) were evaluated using a 90%-10% training-to-testing ratio and independent validation.
• Feature engineering was applied to identify key predictive variables.

Main Results:

• The GBM model achieved high overall accuracy (96.77% for 20-ppb, 90.32% for 5-ppb) but showed low sensitivity for detecting high contamination events.
• Satellite-derived vegetation index in August significantly improved end-of-season contamination prediction.
• Aflatoxin Risk Indices (ARI) in May and July, latitude, and soil-saturated hydraulic conductivity (Ksat) were identified as influential factors.

Conclusions:

• Predictive AFL models are practical for grain handling, enabling preventative rather than reactive mitigation.
• Identifying annual AFL risk predictors is crucial for cost-effective hazard management and optimal corn crop utilization.