Regional prediction of multi-mycotoxin contamination of wheat in Europe using machine learning

Abstract

Wheat is susceptible to fungal infection and mycotoxin contamination during its cultivation period, under opportune environmental conditions. The presence of mycotoxins, such as deoxynivalenol and zearalenone, in wheat destined for feed and food can affect animal and human health. This study aimed to develop a machine learning algorithm to predict the risk levels of one or more mycotoxins in wheat on a regional basis in Europe using crop phenological data, weather data, and satellite images as input. A dataset was obtained with 11 years of mycotoxin monitoring data (2010-2020), including historical records of the concentration of deoxynivalenol, zearalenone, T-2 toxin and HT-2 toxin, fumonisins, aflatoxins, and/or ochratoxin A in wheat in Europe. This dataset was linked, based on year and grid (25 × 25 km), to wheat phenology data, weather data, and satellite image data. The complete dataset for the years 2010-2018 was split into a model training (80%) and an internal model validation set (20%). Data for the years 2019 and 2020 were used for external validation. The random forest (RF) algorithm was applied to train the model using the model training data. The model predicts the probability (low, medium, high) of wheat grown in a certain grid in Europe to be contaminated with at least one of the six mycotoxins under study. Results showed high prediction performance of the model: internal and external validation resulted in 0.90-0.99 prediction accuracy. Based on the data for the Netherlands (case study), satellite images showed to improve the overall model performance. The current model with its mycotoxin predictions can be used by stakeholders to increase logistics in the wheat supply chain and for risk-based monitoring, in this way contributing to improving the safety of wheat derived products, and improving food security. For future development and improvement of prediction models, more mycotoxin data with detailed locations of the cultivated crop are needed.

Keywords: Chemical hazard; Climate change; Co-contamination; Contaminant; Food safety; Forecast; Map; Random forest algorithm; Satellite image.