Predicting Mycotoxin Contamination in Irish Oats Using Deep and Transfer Learning

TL;DR

This study tackles the challenge of predicting multi-toxin contamination in Irish oats by framing it as a multitask learning problem that jointly predicts toxin concentrations and presence/absence. It compares five neural-network architectures, including a pre-trained MLP and several transfer-learning models, with TabPFN delivering the strongest overall performance (RMSE $=0.81$, $R^2=0.56$, AUC $=0.96$, F1 $=0.77$), while the Baseline NN remains competitive and unfrozen TL yields the best results among transfer approaches. A key finding is that weather history over the 90 days before harvest, especially humidity, rainfall, and temperature, alongside seed moisture content, are the most influential predictors across both tasks. The results demonstrate that transfer-learning-enabled tabular models can achieve robust mycotoxin risk predictions with limited data, offering a pathway to early warning and targeted sampling in cereal supply chains, though geographic specificity and model choice should be considered for broader generalization. The study also validates a masked-loss framework that enables learning from incomplete toxin measurements, which is crucial for real-world agricultural data.

Abstract

Mycotoxin contamination poses a significant risk to cereal crop quality, food safety, and agricultural productivity. Accurate prediction of mycotoxin levels can support early intervention strategies and reduce economic losses. This study investigates the use of neural networks and transfer learning models to predict mycotoxin contamination in Irish oat crops as a multi-response prediction task. Our dataset comprises oat samples collected in Ireland, containing a mix of environmental, agronomic, and geographical predictors. Five modelling approaches were evaluated: a baseline multilayer perceptron (MLP), an MLP with pre-training, and three transfer learning models; TabPFN, TabNet, and FT-Transformer. Model performance was evaluated using regression (RMSE, $R^2$) and classification (AUC, F1) metrics, with results reported per toxin and on average. Additionally, permutation-based variable importance analysis was conducted to identify the most influential predictors across both prediction tasks. The transfer learning approach TabPFN provided the overall best performance, followed by the baseline MLP. Our variable importance analysis revealed that weather history patterns in the 90-day pre-harvest period were the most important predictors, alongside seed moisture content.

Figures (11)

• Figure 1: Sampling site locations.
• Figure 2: Total annual rainfall (mm) and annual mean temperature ($^{\circ}$C) for 2022.
• Figure 3: Spatial distribution of contamination concentrations by individual mycotoxin.
• Figure 4: Detection rates for all mycotoxins measured, showing the percentage of samples in which each compound was detected above the limit of quantification.
• Figure 5: Concentration distributions for the nine most frequently detected mycotoxins. Histograms show log$(x+1)$ transformed concentrations for samples where each mycotoxin was detected above the limit of quantification.