Quantile interval prediction strategy enhances quantitative characterization of mycotoxin content in wheat using Near-Infrared spectroscopy

Abstract

Rapid and reliable detection of mycotoxins in wheat is of great significance for ensuring food safety and risk warning. This study combines portable near-infrared spectroscopy technology with quantile regression deep learning (QRDL) to construct an efficient and information-rich method for predicting the range of mycotoxin content. This study compared classification models and point prediction regression models using convolutional neural networks (CNN) and long short-term memory (LSTM), as well as interval prediction models based on quantile regression convolutional neural networks (QR-CNN) and quantile regression long short-term memory (QR-LSTM). The results show that the interval prediction model demonstrates excellent performance in terms of interval width control, median fitting accuracy, and control of the number of samples with inter-class bias. It effectively enhances the ability to express uncertainties, and the interpretability of predictions. This study provides more robust information support for the identification of mycotoxin risk levels and on-site rapid detection in wheat.