基于双波段高光谱与机器学习的花生中黄曲霉毒素B1污染程度快速检测

Rapid Detection of Aflatoxin B1 Contamination Levels in Peanuts Based on Dual-band Hyperspectral Imaging and Machine Learning

摘要:

【目的】通过双波段高光谱与机器学习技术实现花生中黄曲霉毒素B₁（AFB₁）污染程度的快速检测。【方法】采用400~1 000 nm可见近红外与900~1 700 nm短波近红外双波段高光谱，结合机器学习构建花生AFB₁污染程度预警模型。通过采集4个污染梯度的花生样本，基于超高效液相色谱-串联质谱测定AFB₁的质量分数；获取双波段光谱数据后，比较4种预处理方法[一阶导数、二阶导数、多元散射校正、标准正态变换（standard normal variate，SNV）]、2种特征波长选择方法[连续投影算法、非信息变量剔除技术（uninformative variables elimination，UVE）]和2种机器学习方法[支持向量机（support vector machine，SVM）、反向传播神经网络]对分类效果的影响。【结果】SNV-UVE-SVM的组合模型在双波段均能较好地判断AFB₁污染程度，在900~1 700 nm波段该模型性能最优，其中特征波长主要集中在1 126~1 308 nm和1 588~1 720 nm，其训练集和测试集准确率均为100.00%。【结论】首次采用双波段高光谱与机器学习算法构建花生AFB₁污染程度智能预警模型，为农产品真菌毒素污染的无损检测提供了新型技术范式，对实现加工环节的实时在线监测具有重要应用价值。

Abstract:

[Objective] This study aims to achieve rapid detection of AFB₁ contamination levels in peanuts through dual-band hyperspectral imaging and machine learning.[Method] A peanut AFB₁ contamination level warning model was constructed via a dual-band hyperspectral imaging system consisting of visible near-infrared spectroscopy at 400~1 000 nm and short-wave near-infrared spectroscopy at 900~1 700 nm, combined with machine learning. Peanut samples with four contamination levels were collected, in which the mass fraction of AFB₁ was determined by ultra-high performance liquid chromatography-tandem mass spectrometry. The effects of four pretreatment methods [first derivative, second derivative, multiplicative scatter correction, and standard normal variate (SNV)], two characteristic wavelength selection methods [successive projections algorithm and uninformative variables elimination (UVE)], and two machine learning methods [support vector machine (SVM) and back-propagation neural network] on classification performance were compared.[Result] The combined model SNV-UVE-SVM effectively determined AFB₁ contamination levels in both bands, with optimal performance in the 900~1 700 nm band. The characteristic wavelengths were primarily concentrated in the ranges of 1 126~1 308 nm and 1 588~1 720 nm, achieving 100.00% accuracy in both the training and test sets.[Conclusion] This paper pioneered the construction of an intelligent early-warning model for AFB₁ contamination levels in peanuts by using dual-band hyperspectral imaging and machine learning algorithms, providing a novel technical paradigm for non-destructive detection of mycotoxin contamination in agricultural products. This approach holds significant engineering application value for real-time online monitoring in processing stages.