Deoxynivalenol prediction and spatial mapping in wheat based on online hyperspectral imagery scanning

Abstract

Deoxynivalenol (DON), a harmful mycotoxin produced by several Fusarium species, poses critical challenges to wheat production and food safety. However, reducing risks of human toxicity requires pre-harvest detection of DON concentration across different zones of a field. This study investigates the potential of integrating hyper spectral imaging (HSI) in the 400–1000 nm range with machine learning (ML) models for online field detection and mapping of DON contamination in wheat (Triticum aestivum). Using a tractor-mounted push-broom hyper spectral camera, spectral data were collected across four commercial fields in Lithuania and Belgium. A total of 76 wheat samples collected during crop scanning were analyzed for DON levels using liquid chromatography- mass spectrometry (LC-MS). Initial analysis of spectral data alone revealed relatively low classification accu racy, with light gradient boosting machine (LGBM) achieving 55.92 % and decision tree classifier (DTC) achieving 51.97 %. However, the inclusion of fusarium head blight (FHB) severity as an additional feature significantly improved performance, boosting accuracy to 90.79 % for LGBM (a 62.4 % increase) and 86.18 % for DTC (a 65.8 % increase). Moreover, the use of mutual information (MI) for feature selection enhanced model accuracy, achieving 93.42 % for LGBM and 90.13 % for DTC. Spatial mapping of DON contamination demon strated fair to substantial agreement with ground truth maps, providing valuable tools for farmers to understand DON distribution and implement targeted harvesting strategy. This study highlights the potential of integrating online HSI, ML, and feature selection techniques, for pre-harvest DON detection and mapping, providing valu able information for reducing risks of human toxicity and improving the economic value of wheat grain