What does this research mean for the field?

Machine learning models can effectively predict aflatoxin contamination risk in Texas maize, achieving nearly 99% accuracy with the best-performing model. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The aim is to develop machine learning models that predict aflatoxin contamination risk in maize based on historical data and environmental factors.

March 1, 2026

Application of machine learning for predicting aflatoxin contamination risk in maize from Texas.

Key Points

The aim is to develop machine learning models that predict aflatoxin contamination risk in maize based on historical data and environmental factors.
Developed models using machine learning algorithms based on historical aflatoxin records.
Utilized weather, AFRI, NDVI, and soil property data as input variables for model development.
Evaluated model performance using training/validation splits and one-year validation datasets.
Assessed contamination at two threshold levels: 5 μg/kg and 20 μg/kg.
GBDT models achieved nearly 99% accuracy in predicting aflatoxin presence/absence.
Other models showed accuracy between 0.73 and 0.87 at the 20 μg/kg threshold.
Aflatoxin contamination linked to higher precipitation and NDVI pre-flowering and lower during flowering.
Significant soil properties include moisture content, clay percentage, pH, and calcium carbonate.

Abstract

In this study, we developed models for predicting aflatoxin contamination in Texas maize samples. These models were trained on individual or median historical aflatoxin contamination records using machine learning algorithms. The weather, engineering feature (AFRI), remote sensing (NDVI), and soil property data available in the public domain were used as the main input variables for model development. The study utilised 9,694 maize samples collected from 973 Texas counties spanning the period from 2010 to 2023. The performance of all models was evaluated using either a 70/30 training/validation split (30% for validation) or a one-year validation dataset (year 2017). The evaluation was conducted at two threshold levels (5 μg/kg and 20 μg/kg) for high contamination events. The gradient boosting decision tree (GBDT) models developed on median data achieved the highest correct classification rate and balanced accuracy of nearly 99%, while the hierarchical logistic regression (HLR) models for individual data points showed the lowest overall accuracy. The GBDT models were highly effective at predicting the presence or absence of aflatoxin contamination, displaying exceptional sensitivity and specificity. The overall accuracy of other models was also acceptable, particularly at the 20 μg/kg threshold, with values ranging from 0.73 to 0.87. The developed models revealed that aflatoxin contamination was significantly associated with higher precipitation/AFRI/NDVI during the pre-flowering period, but was highly linked to lower precipitation/AFRI/NDVI during the flowering and post-flowering periods. Of the 25 soil properties investigated, moisture content, clay percentage, pH, and calcium carbonate were also significantly associated with aflatoxin contamination in maize. The developed models, along with the key influential predictors identified in this study, should be applicable and useful for growers and other relevant parties to take preventive action against aflatoxin contamination in maize produced in Texas.

Bookmark

Application of machine learning for predicting aflatoxin contamination risk in maize from Texas.

Key Points

Abstract

Cite This Study